Alibaba plans to introduce revenue-sharing terms for some commercial users of its next Qwen open-weight AI model, Reuters reported, citing two people familiar with the company’s plans.
The arrangement would require larger companies that generate revenue from offering the model as a service to reach a commercial agreement with Alibaba. The exact revenue-sharing rate has not been finalised, the sources said.
Alibaba is expected to introduce the measure with its next open-source model. The company has previously charged developers to access models hosted through its cloud platform, while generally allowing customers to deploy its open-source models in their own data centres without paying licensing fees.
The proposed arrangement would differ from the licensing used for the current Qwen3 open-weight models, which Alibaba says are released under the Apache 2.0 licence. The licence permits commercial use, modification, and redistribution subject to its conditions.
Open-source versus open-weight
Open-weight models make their trained parameters available for download, but that does not necessarily mean every part of the AI system is open or that all forms of commercial use are unrestricted.
Under the Open Source Initiative’s Open Source AI Definition, an open-source AI system should allow users to use, study, modify, and share it for any purpose without seeking permission. OSI’s definition also requires access to information about training data, relevant code, and model parameters.
Alibaba and other ******** AI developers have released large models with downloadable weights. OpenAI, Anthropic, and Google, by comparison, primarily distribute their main commercial models through closed systems and hosted services.
Alibaba’s planned terms resemble the licensing model adopted by ******** AI developer Moonshot for Kimi K3, an open-weight model released last month. Its licence includes separate conditions for companies operating Model-as-a-Service businesses above certain revenue thresholds.
Under Kimi K3’s published licence, a company operating such a service must reach a separate agreement with Moonshot when the combined revenue of the company and its affiliates exceeds $20 million during any consecutive 12-month *******. The provision applies to commercial use of Kimi K3 and derivative models.
The licence contains another requirement for large consumer-facing deployments. Commercial products exceeding either 100 million monthly active users or $20 million in monthly revenue must prominently display the Kimi K3 name, with exemptions covering internal use and services offered through Moonshot or certified inference partners.
According to two people familiar with Moonshot’s commercial arrangements, those agreements can include revenue sharing. One source said Moonshot can require partners to share up to 30% of the revenue involved.
******** IT services company Chinasoft International disclosed a revenue-sharing agreement with Moonshot in a regulatory filing last month. It did not disclose the percentage involved.
DigitalOcean Holdings is also among the companies offering Kimi K3 and other ******** models. Chief Executive Paddy Srinivasan confirmed that DigitalOcean has a commercial agreement with Moonshot but declined to provide details.
Srinivasan described the approach as an open-source “freemium” model, where companies can access software at little or no initial cost before paying for larger-scale commercial use, technical services, or earlier access to future releases.
The cost of running open models at scale
Companies can download an open-weight model without paying for access to an API, but large models still require substantial computing infrastructure when deployed at scale.
Kimi K3 contains 2.8 trillion total parameters and 104 billion activated parameters, according to Moonshot. Its mixture-of-experts architecture includes 896 experts, with 16 selected for each token.
The model’s size still places substantial hardware requirements on operators. Moonshot temporarily stopped accepting new Kimi K3 subscriptions in July after saying usage had placed pressure on its available GPUs, while Reuters reported that relatively few users were expected to self-host a model of that scale because of the infrastructure required.
Alibaba is using a similar architectural approach with Qwen3.8-Max. The model contains about 2.4 trillion parameters but activates around 95 billion parameters for each request, according to Reuters.
Moonshot says its mixture-of-experts design improves scaling efficiency by activating only a subset of the model’s experts for each token, rather than the full model.
Cloud providers can charge for hosting and inference, while AI infrastructure companies can generate revenue from deployment and optimisation services.
Dan Fu, vice president of kernels at Together AI, said companies providing AI services can differentiate their offerings through areas such as more efficient token use and deployment optimisation.
“At the application layer, there’s value out there for how you use it, how you actually get the models and the tokens to do something useful,” Fu said.
Model development presents a separate cost challenge. Research involving Epoch AI and Stanford researchers estimated that the cost of the most compute-intensive training runs had risen by about 2.4 times a year since 2016, while Stanford’s 2025 AI Index found that the price of accessing models at a given capability level had fallen sharply.
At published API prices at the time of release, Kimi K3 was priced at about one-third of Anthropic’s Fable model based on listed input and output token rates. Pricing is only one part of the deployment cost, particularly for companies running models on dedicated infrastructure or handling high volumes of requests.
These costs sit alongside the licensing arrangements being tested by model developers. Alibaba already charges developers that access Qwen through Alibaba Cloud. The proposed arrangement would also allow it to collect revenue from some companies deploying Qwen independently on their own infrastructure or through third-party services.
Moonshot has already attached commercial conditions to Kimi K3 while keeping its model weights available for download. DigitalOcean and Chinasoft International have both disclosed commercial arrangements with Moonshot, although the financial terms have not been made public.
The commercial arrangements are developing alongside wider tensions between China and the US over AI technology. The White House has accused Moonshot of using technology taken from Anthropic while developing its models, an allegation ******** officials have rejected.
Interest in releasing models with downloadable weights is not limited to ******** developers. Thinking Machines Lab, the San Francisco AI company founded by former OpenAI Chief Technology Officer Mira Murati, released its first open-source model last month.
Lin Qiao, chief executive and co-founder of Fireworks AI, said there was no fundamental technical barrier preventing US developers from releasing more capable open-source models. Fireworks AI works with models from developers including Moonshot, although Qiao declined to discuss its commercial arrangements.
Alibaba has not publicly announced the final licence for its next Qwen model or the revenue-sharing percentage it plans to seek from large commercial users.
(Photo by: Alibaba)
See also: Alibaba, DeepSeek push China’s AI model race towards lower costs
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Alibaba tests new business model for Qwen open-source AI appeared first on AI News.
View the full article
MIT researchers and collaborators found that AI explainability tools in the health sector can produce sharply different results depending on who uses them.
When applied to skin disease diagnosis, non-experts improved their accuracy with AI assistance, although the improvement largely came from deferring to the model. Primary care providers showed a different pattern: they performed best when they received an AI prediction without an explanation.
The study – which appears in Nature Medicine – examined dermatological diagnosis, where AI tools already support some clinicians and increasingly reach patients through AI-powered search products.
Marzyeh Ghassemi, an associate professor in MIT’s Department of Electrical Engineering and Computer Science, said the findings require care in the design of health AI interfaces.
“Good AI systems can improve performance in some health settings, but this has to be balanced carefully with algorithmic deference that can lead to more error,” she said. “We know that both AI and explainability methods can engage automation bias in humans, and this anchoring effect is something that must be accounted for when we design AI systems.”
The interface changes the diagnosis
Explainable AI aims to give users grounds to assess a model’s output. A system may highlight areas of a medical image that influenced its diagnosis. Another approach can show similar images that support a prediction.
Large language models offer a different route. They can produce a plain-language account of a model’s reasoning, presenting a diagnosis in terms intended for a general audience.
The MIT-led research tested several of these approaches. Participants saw medical images alongside an AI prediction of skin disease. One interface supplied a prediction and confidence level without any explanation. Another returned similar images, and a separate system used heat maps to identify regions of interest. Researchers also tested LLM-generated explanations.
Non-experts assessed whether images of skin ****** showed *******. Clinicians faced a broader task: they had to provide a differential diagnosis for dermatological disease.
Non-experts deferred most to language explanations
Every explainability approach improved non-expert accuracy in the study. The tools mainly helped participants identify non-cancerous ******.
Researchers also tested a fairness-constrained model intended to address bias against darker skin tones. That model improved accuracy and reduced diagnostic disparities based on skin tone. The performance gain came with a risk. Non-experts relied heavily on the model’s recommendation, and incorrect model output damaged their performance more than correct output improved it.
“The reason non-expert users are better is because they are more reliant on the models. When the model is wrong, it hurts performance more than it helps performance when the model is right. We were just able to train very good AI models for this setting,” Ghassemi said.
LLM explanations produced the strongest deference effect. Participants trusted those explanations whether the model output was correct or incorrect. They also found vague or generic explanations more convincing, according to the researchers.
Users who received LLM assistance reported greater confidence in wrong answers. That result puts pressure on interface design for consumer-facing diagnostic systems, where a plausible textual explanation can look authoritative even when the model has made an error.
Roxana Daneshjou, an assistant professor of biomedical data science and dermatology at Stanford University, said patients with limited medical knowledge face the greatest exposure to incorrect explainable AI output.
“These findings are important as patients increasingly turn to AI to help with their health care,” she said. “Our findings show that those with the least medical knowledge are most likely to be led astray when explainable AI models give an erroneous output.”
Primary care providers used AI differently
Clinicians did not follow incorrect AI explanations in the same way. They remained resilient when the system produced an erroneous recommendation or explanation. Their strongest performance came from a more limited interface where the system gave clinicians the model’s prediction without an accompanying explanation.
LLM explanations produced the smallest accuracy improvement among the tested explainability methods for clinicians. The result does not show that explanations have no role in clinical practice. It shows that an explanation format suited to a patient or novice may not fit a trained user performing differential diagnosis.
Lead author Orson Xu, an assistant professor in Columbia University’s Department of Biomedical Informatics, said: “It really comes down to how each group uses the explanation. A clinician already has a diagnosis in mind and checks the AI against their own training, so a bad explanation gets caught.
“Meanwhile, a non-expert can use that exact same explanation to form an opinion in the first place, so a plausible, confident-sounding rationale can pull them toward the wrong answer. The same tool ends up being an asset for one user and a liability for another.”
The study argues against treating explainability as a standard interface component that works identically for every role. The user’s baseline expertise affects whether an explanation acts as a check on the model or becomes a substitute for independent judgement.
Timing affects automation bias
The researchers also examined when users saw AI assistance. People became more deferential when the system showed an explanation before they had the opportunity to make their own diagnosis. That finding points to a practical design choice: an interface could ask the user for an initial diagnostic hypothesis, and then provide an AI recommendation that surfaces alternative conditions for consideration.
The study found that users who deferred most to AI were also the weakest performers when they completed the task without AI support. These participants may stand to gain from model assistance, although they also face the greatest risk when the model produces incorrect output.
The research compared human and AI performance across different presentations of disease. AI systems outperformed people when symptoms appeared subtly. Humans performed much better when an image contained atypical symptoms or unrelated features.
Clinician tools may need a direct model output that supports review against professional judgement. Patient-facing tools require particular care around LLM explanations, especially where the system presents a confident narrative for an incorrect recommendation.
See also: PRISM2 model uses clinical dialogue to interpret pathology slides
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Why health AI interfaces must adapt to user expertise appeared first on AI News.
View the full article
Built by Paige and Microsoft, PRISM2 reads whole-slide images through a perceiver-based encoder trained jointly on tissue tiles and clinical dialogue drawn from pathology reports. The model aggregates thousands of tile embeddings per slide into one representation, then generates text that answers diagnostic questions rather than simply classifying pixels.
Training data spans 2.3 million whole-slide images. The dialogue supervision comes from 685,507 pathology reports that Memorial Sloan Kettering ******* Center collected during routine care, converted into question-and-answer pairs by GPT-4o.
Architecture and embedding design
The architecture runs in two distinct phases. Stage one trains the slide encoder itself, teaching it to aggregate tile-level features into a single slide-level vector that correlates with report language.
Stage two freezes that encoder entirely and shifts the work onto the language model, fine-tuning it on dialogue so it learns pathology reporting conventions rather than encoder mechanics.
Single-turn dialogue supplies the supervision at this second stage. No multi-turn conversation history enters the training signal, which caps the kind of interactive back-and-forth a deployed system could support without further engineering.
A perceiver-based slide encoder sits at the centre of stage one, aggregating Virchow2 tile embeddings into the slide-level representation. Two loss functions train it simultaneously.
BioGPT text embeddings drive a contrastive objective, pulling slide representations toward matching report language and pushing them from mismatched pairs. Phi-3 Mini runs an autoregressive objective alongside it, forcing the encoder’s output to support direct text generation rather than just similarity scoring.
Contrastive training alone tends to produce embeddings good at retrieval but weak at generation, and autoregressive training alone can overfit to surface text patterns without learning transferable visual features. Running both against the same encoder output is the architectural bet PRISM2 makes.
PRISM2 exposes two separate embeddings rather than one. Base embeddings come straight from the slide encoder and transfer to biomarker prediction tasks. Diagnostic embeddings are pulled from the hidden state of the 4-billion-parameter language model instead, which is extracted only after that model has processed both the slide latents and the prompt text together.
That distinction shapes where each embedding performs best. Training tunes the diagnostic embedding specifically for ******* detection and subtyping tasks. Paper authors recommend the base embedding for anything sitting outside that diagnosis-focused distribution. Survival tasks get a third, separately fine-tuned embedding on top of the other two.
Performance and benchmark results
PRISM2 matches or exceeds the balanced accuracy of clinical-grade products calibrated for prostate and breast ******* detection, tested on those products’ own evaluation datasets. Breast lymph node classification joins the comparison, and PRISM2 outperforms Paige BLN there without any further training on that specific task.
The earlier foundation models in the comparison don’t hold up as well under contrastive classification: PRISM and TITAN both fall short of product-level performance, and the gap widens on breast lymph node testing specifically.
Pan-******* detection extends the benchmark further. Diagnostic embeddings reached 0.967 AUC, against 0.956 for the base embedding on the same task. PRISM scored 0.947 in that comparison while TITAN came in at 0.931.
Rare ******* detection dropped the diagnostic embedding’s score from 0.967 down to 0.957 AUC. The paper attributes that drop to sparse training examples for those tissue types.
Linear probing gives a cleaner read on representation quality than end-to-end fine-tuning does, since it holds the encoder fixed and tests only whether a simple classifier can extract the relevant signal.
Under that method, PRISM2 embeddings never statistically underperform prior foundation models across the diagnostic benchmarks tested, at P < 0.05. Results improve most where the pan-******* training set holds few samples from the tissue types under comparison.
Survival and biomarker tasks follow a related pattern under the same probing method. Researchers gathered over 225,000 cases tracking overall survival across nearly 100,000 patients, then compared a fine-tuned PRISM2 slide encoder directly against a survival specialist model trained from scratch on the identical dataset. PRISM2 embeddings won that comparison.
MSK colorectal ******* recurrence-free survival produced the widest gap: a 0.809 concordance index against 0.773 for the specialist model.
Base embeddings hold up here too, without any survival-specific fine-tuning at all. On biomarker tasks – which sit outside the report-dialogue training distribution – base embeddings actually beat diagnostic ones, averaging 0.854 AUC on MSK tasks and 0.784 on TCGA tasks. COBRA and TITAN track close behind on those same figures.
The ablation study isolates what the dialogue supervision contributes on its own. Adding dialogue templates to the original PRISM starting point lifted prompt-based inference from roughly 0.498 balanced accuracy to 0.653.
The question-answering dataset behind that jump runs 3.5 times larger than the PRISM subset it builds on, and the paper attributes about half of PRISM2’s diagnostic improvement to that scale increase alone rather than architectural changes.
Training records span 200,692 patients in total. GPT-4o generated every dialogue example from the underlying 685,507 reports, a synthetic-data pipeline that itself needed independent quality checking.
Data quality, error rates, and architectural limitations
A pathologist reviewed 50 held-out specimens across 10 tissue types to check both the generated training text and PRISM2’s own outputs.
Ground-truth question errors landed at 3 percent for open-ended and multiple-choice formats combined. Diagnostic summaries ran higher, at an 8 percent error rate. Complementary yes/no questions performed worst of all three categories: 18 percent were irrelevant or inaccurate.
PRISM2’s own question-answering errors sat between 7 percent and 11 percent in that same review. Hallucination and omission were the dominant failure types, not factual contradiction of the source slide.
The model carries no position encoding across tiles, meaning it has no built-in mechanism for reasoning about where structures sit relative to one another on a slide. Every scan in training and testing ran at a single fixed resolution: 0.5 microns per pixel.
Neither gap is trivial for an ML team evaluating the architecture. Spatial reasoning tasks and any need to work across variable magnification remain unaddressed problems rather than solved-but-unoptimised ones.
MSK scanned every slide used in training and evaluation. The authors call for external validation before deployment beyond that single-institution scanning pipeline.
Model weights are published on Hugging Face. Training and inference pipelines still depend on proprietary Paige and Microsoft infrastructure, which limits how far an outside team can reproduce the full pipeline rather than just running inference against the released weights.
Teams building on top of PRISM2 should test embedding transfer against their own scanner output before assuming parity with the MSK-trained baseline.
See also: Why biological data matters more in AI drug discovery
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post PRISM2 model uses clinical dialogue to interpret pathology slides appeared first on AI News.
View the full article
Alibaba has launched Qwen3.8-Max, its largest AI model to date, as DeepSeek’s latest V4-Flash model draws attention for inference pricing that is lower than several competing systems.
Qwen3.8-Max has 2.4 trillion parameters and uses a mixture-of-experts architecture, which activates only part of the model for each request. Alibaba said around 95 billion parameters are active at a time, reducing costs and response delays compared with activating the full model.
DeepSeek uses a similar sparse architecture at a smaller scale. Artificial Analysis lists V4-Flash at 284 billion total parameters, with 13 billion active during inference, while Moonshot AI’s Kimi K3 has 2.8 trillion total parameters and about 104 billion active.
Qwen3.8-Max can process text, images, and video and supports up to one million tokens of context. Alibaba also said the model completed a software engineering project over 16 days.
Its size places it close to Kimi K3, which Moonshot AI released in July. The two companies are also competing on price, with Qwen3.8-Max costing $2 per million input tokens and $6 per million output tokens, compared with $3 and $15, respectively, for Kimi K3.
Model size alone does not determine inference cost. Architecture, active parameter count, token consumption, and the number of calls required to complete a task also affect how much a model costs to run.
Qwen3.8-Max moved to the top position among ******** text models on crowdsourced comparison platform Arena.AI following its release, although it remained behind several Anthropic models in the overall rankings. It also ranked second on Arena.AI’s leaderboard for models that analyse images and other visual material, behind an Anthropic Claude Fable 5 variant.
DeepSeek pushes down inference pricing
DeepSeek has taken a different approach with V4-Flash. Rather than matching the overall scale of Alibaba’s and Moonshot AI’s latest models, it has priced the model below several widely used AI systems.
V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens, according to Artificial Analysis. The research firm lists the model with a one-million-token context window and 284 billion total parameters, of which 13 billion are active during inference.
Artificial Analysis lists cache-hit pricing of $0.003 per million tokens for the Max Effort version of V4-Flash, 98% below its standard input rate. Cached input covers previously processed context that can be reused across subsequent requests.
DeepSeek’s lower token rates also carried through to Artificial Analysis’ benchmark testing. Reuters reported that the research firm estimated V4-Flash’s average cost at three cents per test, compared with 86 cents for Kimi K3, $1.86 for OpenAI’s GPT-5.6 Sol, and $3.15 for Anthropic’s Claude Fable 5.
The comparison accounts for the amount of input and output each model uses to complete the benchmark. A lower per-token rate does not necessarily result in a lower task cost if a model generates more output or requires additional interactions.
Artificial Analysis gave the Max Effort reasoning version of DeepSeek V4-Flash a score of 40 on its Intelligence Index. The research firm also recorded an output rate of about 118 tokens per second during testing.
Token prices tell only part of the cost story
Moonshot AI’s Kimi K3 provides another example of how advertised API prices can differ from the cost of completing longer workloads. Artificial Analysis lists the model at $3 per million input tokens and $15 per million output tokens, with cached input priced at $0.30 per million tokens.
On Artificial Analysis’ AA-Briefcase benchmark for agentic knowledge work, Kimi K3 averaged $10.57 per task. It generated around 120,000 output tokens and used an average of 83 turns per task.
Artificial Analysis said the cost reflected Kimi K3’s token pricing, output volume, and number of model interactions. Repeated model calls and larger outputs can therefore raise the total cost of completing a workload beyond what the headline API rate suggests.
Kimi K3 recorded the second-highest overall score on the AA-Briefcase evaluation at the time of testing, behind Claude Fable 5. It also scored 57 on Artificial Analysis’ broader Intelligence Index.
The comparison with DeepSeek shows why cost-per-task measurements add useful context to standard API pricing. Models with different architectures and usage patterns can consume substantially different amounts of compute and tokens while working through the same type of task.
Open weights add another deployment option
Cost is also being shaped by how ******** developers distribute their models. Alibaba, DeepSeek, and Moonshot AI have continued to support open-weight releases alongside hosted API access, giving developers more options for how the models are deployed.
Artificial Analysis lists DeepSeek V4-Flash as an open-weight model under an MIT licence, with weights available through Hugging Face. Kimi K3 is also available as an open-weight model under Moonshot AI’s own licence.
Open weights allow developers to run models on their own infrastructure or through third-party providers instead of relying solely on a developer-hosted inference service. Deployment costs still depend on the hardware and infrastructure used, but access to the model is not tied to a single hosted API.
The approach differs from the main models offered by OpenAI, Anthropic, and Google, which generally keep their model weights closed.
Lian Jye Su, chief analyst at Omdia, said model selection for many business workloads does not depend solely on having access to the highest-performing system.
“Many business workflows do not need the industry’s very best model,” Su said. “They need models that are good enough, affordable, transparent and accessible, and open-weight models help meet that demand.”
(Photo by Solen Feyissa)
See also: Alibaba is designing AI chips around agents, and that changes what the race is actually about
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Alibaba, DeepSeek push China’s AI model race towards lower costs appeared first on AI News.
View the full article
Red Hat has launched asago, an open-source community project that aims to turn AI governance policy into production-ready deployment code.
The project describes itself as an automated, auditable workflow that connects the “fragmented steps, tools, and requirements” of engineering and compliance teams. With regulation such as the EU AI Act now taking effect, Red Hat frames the choice facing organisations as: either grind AI innovation down through manual review, or let ungoverned agents run in production without anyone checking their behaviour against policy.
asago builds on Red Hat and NVIDIA’s work inside the Open Secure AI Alliance. It is being released under the Apache License 2.0, and the project is currently in its formation phase, with a repository open on GitHub for developers, academic researchers, and enterprise early adopters to review and contribute to governance.
Four stages from policy text to running controls
The workflow Red Hat describes runs across four stages. Risk mapping comes first: the framework reads an organisation’s uploaded governance policy and maps its specific requirements against established frameworks, including the NIST AI RMF, the OWASP LLM Top 10, and the EU AI Act as catalogued via IBM’s AI Risk Atlas. Policy language becomes a risk profile automatically, rather than through a compliance team’s manual cross-referencing.
From there, asago moves into risk assessment. The project generates and runs scenarios tailored to the specific use case, probing for the harmful behaviours that its risk mapping flagged rather than testing against a standard checklist. Risk mitigation follows: the system recommends guardrails based on what the testing surfaced, and builds a rationale trail meant to survive a reviewer’s scrutiny.
asago orchestrates the recommended controls into deployment-ready configurations for hybrid cloud and Kubernetes environments, according to Red Hat, cutting out the manual infrastructure coding that would otherwise sit between a mitigation recommendation and a running control. Red Hat’s stated aim is to cut deployment timelines from months to days.
Audit-trail-as-a-product
Every stage is meant to feed a single, continuous audit trail. Each policy clause ties to a specific test, and each test ties to a runtime control. A reviewer, in principle, can trace any active control in a live deployment straight back to the policy line that justified it.
That traceability is the actual selling point. Red Hat’s own framing treats AI safety less as a one-off certification exercise and more as an ongoing enterprise utility (i.e. something that stays checkable as agents keep running, not just at the point they’re first approved.)
Steven Huels, Red Hat’s VP of AI Engineering, says: “As organisations transition from experimental AI pilots to long-running, autonomous agents, establishing clear operational guardrails becomes a critical infrastructure requirement.”
Huels connects asago to Red Hat’s separate Lightwell initiative, which focuses on securing the open-source supply chain from AI-driven vulnerabilities, calling asago “the next logical step for enterprise AI by automating the link between corporate policy definitions and live production agents.”
Stuart Battersby, Red Hat’s AI safety and model evaluation architect, is more direct about the project’s intended shape: “The asago project is a true collaborative, open-source endeavour bringing together stakeholders from the technology industry, academia, and government.
“We encourage more collaborators to join this community-driven effort, particularly from global jurisdictions, to ensure maximum coverage of AI safety viewpoints.”
A roster of major contributors, not a single vendor
The founding list runs far wider than Red Hat and NVIDIA. Brave Software, IBM Research, Microsoft, MIT Lincoln Laboratory, North Carolina State University, and The Alan Turing Institute all appear as contributors, alongside the EvalEval coalition and Austria’s Interdisciplinary Transformation University (IT:U). Alquimia AI, a partner rather than a founding research institution, is also named.
Sarah Bird, Chief Product Officer for Responsible AI at Microsoft, comments: “Many of the hardest AI safety and security challenges are still unsolved, and no single organisation can tackle them all alone.”
Academic voices push a similar line from a different angle. NC State’s Veena Misra, Interim Dean of the College of Engineering, calls AI safety “an engineering problem as much as a policy problem.”
asago’s outputs are meant to be infrastructure-agnostic: declarative configurations for Kubernetes, Terraform, and Ansible, according to Red Hat, so a safety posture set in one cloud doesn’t need re-engineering in another.
Nothing about the project is production-tested yet. There’s no deployed customer case study in Red Hat’s announcement, no benchmark showing the “days, not months” claim holding up under a live regulatory audit, and no indication of how disputes between contributing organisations over risk-mapping standards get resolved once the code moves past formation.
For now, the project exists as a repository and a governance structure on GitHub, open to developers, researchers, and enterprise teams willing to build alongside a list of contributors rather than adopt a finished product.
See also: OpenAI aligns safety practices with EU AI Act’s GPAI Code
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Red Hat, NVIDIA, IBM back project turning AI policy into code appeared first on AI News.
View the full article
Article 50 of the EU AI Act has entered into force, setting transparency obligations for AI providers and deployers operating across the bloc.
Enterprises running generative AI tools now have to comply with Article 50, which requires providers and deployers of certain AI systems to tell people when they’re interacting with a machine, and to mark AI-generated content so it can be flagged as such.
Advancements in generative systems have made it harder to tell AI interaction from human conversation. Distinguishing AI-generated images from authentic ones is getting harder too. People are now also being exposed to emotion recognition and biometric categorisation tools without knowing it.
The Commission links all of this to manipulation at scale and fraud, with impersonation and consumer deception following close behind on its list of concerns. Article 50 is the EU’s attempt to tackle the issue and ensure the responsible and safe rollout of AI across the bloc.
What providers have to build
Article 50 requires providers to design systems so that anyone interacting directly with an AI system knows it. The exception: cases obvious to a reasonably well-informed, observant, and circumspect person given the context.
Law enforcement systems used to detect, prevent, investigate, or prosecute criminal offences sit outside the rule too, provided safeguards protect third-party rights, unless the public can use the system to report a crime.
Providers of systems generating synthetic audio, image, video, or text face a separate duty under Article 50. Marking is the mechanism. The output needs a machine-readable mark, detectable as artificially-generated or manipulated.
The Act asks for the aforementioned marking to be effective and interoperable “as far as this is technically feasible,” weighing implementation cost against the state of the art. Assistive editing that leaves deployer-supplied input essentially untouched falls outside the requirement; a routine photo touch-up doesn’t trigger it, but a wholesale AI-generated replacement does.
What deployers must tell people
Anyone running an emotion recognition or biometric categorisation system must inform the people exposed to it. Personal data gathered through that system still falls under existing data protection law: the GDPR governs the general case, the EU institutions data protection regulation applies where an EU body is running the system, and the Law Enforcement Directive covers policing contexts.
Deepfakes get their own disclosure duty. Image, audio, or video content that’s been artificially-generated or manipulated has to carry a disclosure saying so. Artistic, satirical, or fictional work gets a lighter touch: the disclosure only needs to flag the content’s existence, worded so it doesn’t get in the way of enjoying the work.
Text published to inform the public on matters of public interest carries its own rule. Deployers must disclose AI generation or manipulation of that text unless a human has reviewed it and someone holds editorial responsibility for the publication. Standard newsroom review clears the bar. Unedited AI output published straight to a public interest story does not.
All disclosures need to land no later than the first interaction or exposure, in a manner that’s plain, distinguishable, and accessible under existing accessibility rules. No grace ******* covers informing someone after the fact.
The compliance path Brussels favours
Three bodies split enforcement. National market surveillance authorities handle most cases. The AI Office takes systems that fall under its own supervision. The European Data Protection Supervisor steps in when an EU institution itself acts as provider or deployer.
The guidelines set out how providers and deployers can show they’ve met the marking obligation in Article 50. Signing on to the Code of Practice on Transparency of AI-generated Content is one path.
Organisations that skip the Code have to demonstrate compliance through alternative means the Commission considers adequate. What those alternatives look like in practice isn’t spelled out in detail; that judgement falls to the market surveillance authorities doing the enforcing.
The other transparency duties don’t have an equivalent code. No code, no shortcut. Telling people they’re talking to an AI is one duty. Disclosing deepfakes and flagging AI-generated public interest text round out the rest, and providers and deployers work out their own adequate measures, with the guidelines serving as a reference point rather than a checklist.
Much of the document is definitional. It sets out what counts as a directly interactive AI system, what qualifies as synthetic content, and where the line sits between a deepfake and ordinary edited media. Standard editing sits outside scope by name, alongside assistive functions that leave deployer-supplied input intact.
The guidance also works through the value chain question of who counts as a provider and who counts as a deployer, and what happens when both roles sit with the same organisation. Which of the four Article 50 obligations apply, and to whom, comes down to that provider-deployer distinction.
Organisations weighing up the Code of Practice on Transparency of AI-generated Content against building their own labelling approach now have somewhere to start. The guidelines give them a Commission-endorsed reference point that goes beyond the bare text of the regulation.
See also: OpenAI aligns safety practices with EU AI Act’s GPAI Code
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post EU AI Act Article 50 transparency rules enter force appeared first on AI News.
View the full article
GSK has entered into a research collaboration with British biotechnology company Relation Therapeutics worth up to $110 million, expanding the companies’ existing work in AI-assisted drug discovery.
Under the agreement, Relation will generate large-scale datasets measuring how human cells respond to genetic changes and drug interventions. The data will be used to train AI models designed to identify potential drug targets, including models within Relation’s MORGAN platform.
The agreement places biological data generation alongside AI model development. Relation’s research approach links computational analysis with experiments that generate new information on human cells.
The collaboration builds on earlier agreements between GSK and Relation focused on fibrotic diseases and osteoarthritis. Those projects involved observational studies designed to create two functional disease datasets for analysis using Relation’s Lab-in-the-Loop platform.
The earlier work combined human genetics, single-cell multi-omics generated from human tissue, functional assays, and machine learning to identify and validate potential disease targets.
How Relation generates biological data
Relation describes its Lab-in-the-Loop approach as a combination of laboratory experimentation and computational analysis. Its work includes tissue profiling, single-cell and spatial transcriptomics, sequencing, and target validation, while machine learning is used for target identification, prioritisation, validation, and experimental design.
The company also conducts perturbation experiments that measure how genetic changes affect cellular characteristics associated with disease. Those results can then be analysed alongside genetic and patient-derived biological data.
Public repositories remain an important source of training material for biological foundation models, although combining information produced across different studies can introduce technical challenges.
A 2025 review in Experimental & Molecular Medicine noted that repositories including CZ CELLxGENE, the Human Cell Atlas, and NCBI Gene Expression Omnibus give researchers access to large volumes of single-cell data. CZ CELLxGENE alone provides access to more than 100 million standardised cells, according to the review.
Sampling methods, sequencing protocols, experimental procedures, and processing pipelines can differ between studies. Single-cell data can also contain technical noise and other artefacts, requiring careful dataset selection, filtering, composition balancing, and quality control during foundation-model training.
Dataset overlap presents another issue. The review noted that the same or similar cells can appear across multiple public resources, potentially giving them disproportionate influence during training and creating data-leakage risks when training and test datasets overlap.
The review found that assembling a high-quality, non-redundant dataset is as important as model architecture when building robust single-cell foundation models.
******* biological datasets do not guarantee better models
Research published in Nature Methods in June this year examined how the size and diversity of pretraining data affected single-cell foundation models using a corpus of 22.2 million cells. Researchers trained 400 models and evaluated them across 6,400 experiments.
The study found that current single-cell foundation models tended to reach performance plateaus after training on only a fraction of the available corpus. Unlike large language models, the systems assessed did not display clear data-scaling laws in which continually increasing training data consistently produced better results.
The researchers found that model capacity, dataset size, and computational resources need to be balanced rather than simply increased together. The study did not establish that smaller or proprietary datasets are inherently better, but it found that adding more biological training data did not consistently lead to further performance gains.
A separate study published in Genome Biology in 2025 assessed two single-cell foundation models, Geneformer and scGPT, across several zero-shot evaluation tasks. The models did not consistently outperform simpler approaches, while the researchers also identified challenges involving batch effects and cautioned against assuming that larger pretrained models automatically produce better biological representations.
Pharma companies pursue specialised datasets
Relation has already applied its data-generation approach to Osteomics, which it describes as a proprietary functional single-cell bone atlas. The project uses patient-derived samples and combines single-cell and spatial omics with imaging, genomics, proteomics, and clinical phenotype data.
According to the company, Osteomics is being used to investigate disease biology, therapeutic targets, biomarkers, and patient subgroups in osteoporosis. Hospitals and research partners in the *** and Australia are involved in the observational study.
Research published in Nature Genetics last month also examined the cellular and genetic determinants of skeletal disease using single-cell analysis, genetic data, and functional validation. Several Relation researchers were among the study’s authors.
A 2025 Nature Biotechnology analysis of AI-focused biopharma deals identified specialised dataset providers as one of several trends emerging from recent partnerships. Other trends included larger upfront payments, new therapeutic modalities, and greater participation from larger biotechnology companies.
The analysis said high-quality, disease-specific datasets are becoming an important input for causal and generative machine-learning models. It cited GSK’s separate agreement with Ochre Bio, worth $37.5 million for data licensing involving human liver single-cell and perfused-organ data.
Another example involved AstraZeneca and Pathos AI entering a $200 million agreement with Tempus in 2025. Under the arrangement, Pathos was to develop oncology foundation models using de-identified clinical, genomic, and imaging data covering more than 150,000 patients.
Access to sufficient high-quality data remains a constraint in AI drug discovery. A Nature research highlight on federated learning in pharmaceutical research identified limited access to suitable training data as a major bottleneck for AI applications, while noting that companies can also face restrictions on sharing proprietary information.
AI-biopharma agreements therefore vary in how companies obtain data and computational capabilities. Some centre on access to AI platforms, while others cover joint development, data licensing, or the creation of new biological datasets.
The GSK–Relation agreement includes both data generation and model development. Relation will produce human cellular datasets as part of the collaboration and use them to train AI models for identifying potential drug targets.
(Photo by CDC)
See also: How AI is shortening drug discovery timelines in China
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Why biological data matters more in AI drug discovery appeared first on AI News.
View the full article
OpenAI has outlined how it aligns safety, security, and transparency work with the EU AI Act’s GPAI Code as enforcement approaches.
The company has contributed to and endorsed the EU’s General-Purpose AI (GPAI) Code of Practice and the Code of Practice on Transparency of AI-Generated Content. Both emerged from multi-stakeholder processes.
The GPAI Code sets a shared bar for transparency, safety, and security across general-purpose models sold or deployed in the EU. OpenAI points to a stack of existing practices as evidence it already operates near that bar: pre-release testing of models, published system cards accompanying major launches, and outside red-teaming through what it calls its Red Teaming Network. The company also maintains a public Model Spec document describing how it shapes model behaviour.
Two internal frameworks sit underneath that work. The Preparedness Framework has been in place since 2023 and was updated in 2025; it sets out how OpenAI identifies, evaluates and manages serious risks from advanced systems. A separate Frontier Governance Framework builds on it, explaining how the company’s safety and security practices map onto legal requirements including the GPAI Code specifically.
Together, OpenAI says, those two documents govern risk assessment, safeguards, model reporting, security posture, incident response, and how external experts get pulled into the process.
OpenAI cites its participation in the Frontier Model Forum alongside collaborations with the US Center for AI Standards and Innovation and the *** AI Security Institute, plus contributions to third-party evaluation standards more broadly. The stated goal is shared safety research and clearer testing benchmarks across the industry, not just within one company’s walls.
Provenance gets harder as modalities multiply
The Transparency Code commitments centre on a different problem: helping people tell when content was made or altered by AI.
OpenAI’s approach rests on two mechanisms that are meant to reinforce each other. Content Credentials, built on the C2PA standard, attach context directly to a file. SynthID watermarking provides a fallback signal for cases where that metadata gets stripped out somewhere along the way.
Coverage is expanding from images into audio outputs, and OpenAI says it’s working toward extending provenance measures across further modalities, including text, as the underlying standards and tooling mature. The company is also building signals and guidance aimed at developers who need to meet their own transparency obligations when building on top of its models.
None of this solves provenance outright. Metadata gets lost and labels don’t always survive a transfer between platforms. No single signal, whether cryptographic or watermark-based, catches everything on its own. OpenAI’s response is a layered approach paired with continued work across the wider standards community rather than a claim that any one mechanism closes the gap.
Cybersecurity as the test case for adaptive governance
Capabilities that help defenders spot and patch vulnerabilities are the same capabilities that could help an attacker find them first. OpenAI believes the answer is its Trusted Access for Cyber programme, designed to give vetted defenders access to more advanced cyber capabilities while limiting exposure for misuse.
That programme now has a European deployment arm. OpenAI states it launched its EU Cyber Action Plan in early May 2026, working with EU and national cyber agencies, private sector partners, and infrastructure operators to give them access to its more advanced cyber models.
The stated aim of the plan is to strengthen cyber resilience across the continent. Whether “most advanced” translates into measurable defensive gains inside these agencies is a claim from OpenAI itself; the source material offers no independent verification of outcomes from the programme.
The company positions this work as consistent with the European Commission’s Action Plan on Cybersecurity and Artificial Intelligence, which calls for coordinated handling of AI’s risks alongside its use in strengthening defensive capability, including secure access arrangements for cybersecurity purposes specifically.
OpenAI says it will keep adjusting its compliance approach as EU AI Act implementation continues, and that it expects to keep learning from regulators and the wider community involved in shaping the rules. The company argues that rules need enough flexibility to adapt as the technology moves, so that businesses and organisations can keep benefiting from it.
The GPAI Code and the Transparency Code are still relatively new instruments, and OpenAI’s compliance documentation is a moving target rather than a finished product. Teams building on OpenAI’s models in regulated European markets should treat the current system cards and Frontier Governance Framework as a starting point for their own due diligence, not a substitute for it.
See also: Zuckerberg details Meta’s personal AI superintelligence strategy
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post OpenAI aligns safety practices with EU AI Act’s GPAI Code appeared first on AI News.
View the full article
Mark Zuckerberg has published a WSJ op-ed that argues superintelligence must reach individuals, not just a handful of institutions.
The Meta chief published the piece as a statement of company philosophy rather than a product announcement. It contains no release dates, no benchmark figures, no named models. What it does contain is an argument that the central question facing the industry is not whether superintelligence arrives, but who gets to use it once it does.
Zuckerberg believes the choice is binary: systems concentrated inside a small number of institutions, or systems distributed as tools that individuals control directly. He calls his preferred version “personal superintelligence” and commits Meta to building toward it on three stated principles: individual empowerment as the source of prosperity, invention as the primary purpose of superintelligence, and balance of power as the foundation of safety.
Meta takes aim at rival AI superintelligence labs
The op-ed spends more time criticising the tone of the industry than describing Meta’s own technical work.
Zuckerberg writes that it is “surprising that the discourse from many of those who are developing artificial intelligence is so filled with doom,” and questions why anyone convinced that AI will “eliminate most jobs and much of humanity’s relevance” would then rush to build it.
He goes further, calling the idea that AI danger justifies concentrating power in a few hands “dangerous” in itself, and adds that “hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened hasn’t led to safe or positive outcomes.”
None of the labs he’s describing are named in the piece, but I’m sure that readers can fill in that detail for themselves. What the op-ed does establish is a positioning claim: Meta wants to be read as the company arguing for broad distribution against a field it characterises as leaning toward centralisation.
To illustrate his point Zuckerberg asks readers to imagine a single person with access to a superintelligent lawyer, gaining “an unfair advantage in court—even if his position was wrong on the merits,” which he says “would lead to a worse society.”
However, Zuckerberg then flips the scenario: everyone has a superintelligent lawyer, and “justice would be carried out much more fairly and efficiently than it is today.”
It’s a thought experiment, not a deployment. There’s no pilot programme, court system, or legal-services partner referenced anywhere in the source material, and the article should be read as an analogy supporting Zuckerberg’s broader claim that widespread access checks concentrated power rather than a description of anything Meta has built or tested.
The same pattern holds for his historical references. He cites “the brothers in a bicycle shop who believed people could fly,” “the bookbinder’s apprentice with no schooling who figured out how to generate electricity,” and “the kid in a garage who thought personal computers could be for everyone” as evidence that progress tends to come from individuals rather than institutions.
The op-ed does make a claim with direct relevance to enterprise planning: the balance between AI as an automation tool versus an empowerment tool. Zuckerberg writes that “if the balance leans toward automation, the impact on jobs and the economy may be negative.”
Zuckerberg’s counter-position is that wide distribution of superintelligence produces more jobs, not fewer, largely because starting a business becomes possible “without raising large amounts of capital.” He expects the economy to tilt toward “a greater number of people working at small businesses rather than larger companies.”
AI risk categories get uneven treatment
Zuckerberg does distinguish between types of risk, and the distinction is useful for anyone trying to work out what Meta considers manageable versus what it considers to require outside coordination.
On cybersecurity, he argues that “the history of open-source software has shown that giving everyone full access to powerful systems will be the best way to protect safety and security over time.” On biological risks, his position is different: he calls for “more coordination between governments and other institutions on responsibly deploying capable models.”
That’s a meaningful split. One category gets an open-access argument grounded in software history; the other gets an admission that government coordination is necessary. The op-ed doesn’t reconcile how a company committed to broad distribution handles the second category in practice, and it doesn’t need to for the purposes of an opinion column.
Nothing in the op-ed specifies how “personal superintelligence” would be packaged, priced, or governed inside an organisation. There’s no mention of enterprise controls, data handling, deployment architecture, or model access tiers.
Zuckerberg closes by stating that “Meta is committed to building with the principles of individual empowerment, invention and balance of power,” which is a statement of intent rather than a roadmap. The op-ed sets out where Zuckerberg wants the argument to sit, but it doesn’t say what Meta will ship, when, or under what governance terms.
See also: Meta, Microsoft, Nvidia, IBM, and others back open-weight AI
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Zuckerberg details Meta’s personal AI superintelligence strategy appeared first on AI News.
View the full article
Cybersecurity wasn’t really something small and mid-sized businesses worried about too much a few years back. That’s changed fast. Once your customer data, your apps, your internal tools all end up online, protecting the server behind them stops being optional. And attacks aren’t getting any simpler either, which is part of the problem with sticking to old-school security methods on their own.
If your business runs on a secure Linux VPS, whether that’s hosting a website, a business app, or just a dev environment, pairing that server with some AI-based tooling actually makes a real difference. It’s not about replacing the basics. It’s more that AI catches suspicious behaviour faster, and honestly, faster than any human sitting there watching logs ever could.
Why AI Even Fits Here in the First Place
Old security setups run on fixed rules mostly. Block this IP after enough failed logins. Flag a file if it matches something already known. That works fine against threats people have already seen before. It falls apart against anything new, which, unfortunately, is most of what’s actually out there now.
AI comes at this differently. Instead of checking one login attempt or one file at a time, it’s looking at a pile of signals together, logins, traffic, resource usage, general behavior, trying to figure out if any of it looks off compared to what’s normal for that particular server. That’s really the whole trick behind it, and it’s why AI tends to catch stuff a lot earlier than someone manually scrolling through logs would.
The Risks a Linux VPS Actually Faces
Linux has a solid reputation for security, no argument there. But nothing’s bulletproof. A business running a VPS typically deals with a handful of recurring problems:
Brute-force login attempts that show up more often than most people expect
Malware sneaking in through software nobody got around to patching
DDoS attacks aimed at simply knocking things offline
Permissions that got misconfigured months ago and nobody noticed since
Unauthorised access that started small and just sat there quietly until someone finally found it
That last one is honestly the scary part. A lot of these problems don’t start as big dramatic breaches. They start as one tiny weakness that sits there for weeks, sometimes months, before anyone even realises it’s a problem.
Where AI Actually Helps With Detection
Say a server usually only sees admin logins during business hours, from one country. Then one day a login shows up from somewhere else entirely, at 3am, and whoever it is starts pulling sensitive files right away. A normal system checks the password, sees it’s correct, and moves on. That’s it. An AI-based system looks at the whole picture instead, notices this doesn’t match how that account usually behaves, and flags it before anything worse happens.
That gap, between “the password was right” and “but this doesn’t look like something this account would normally do,” is basically the whole value of AI in this context. In its context a rule-based system was never built to notice in the first place.
Keeping Software Patched Without Losing Your Mind
Updates are still one of the simplest ways to keep a server secure. The problem is a real server usually has dozens, sometimes way more, of installed packages running, and going through every single update by hand just isn’t realistic for most people.
AI tools help here by pointing out which software is outdated, which patches actually matter most, which vulnerabilities are being actively exploited somewhere else right now, and generally cutting down how much manual digging an admin has to do just to figure out what to fix first.
Watching Network Activity That Would Otherwise Get Buried
Attackers don’t always go straight for the kill either. A lot of the time they poke around quietly first, exploring a system before doing anything obvious. That’s exactly the kind of activity that’s easy to miss in a giant log file, but AI tools watching traffic, bandwidth, running processes, and file changes continuously tend to notice when something shifts, even subtly, and that gives a business a real shot at catching trouble before it turns into actual damage.
Access Control Still Matters More Than Anything
Weak logins are still one of the biggest reasons servers get compromised, full stop. None of the basics have changed here. SSH keys instead of passwords, root login turned off, multi-factor authentication wherever it’s supported, giving accounts only the access they actually need, checking who has access every so often instead of never. AI doesn’t replace any of that. What it adds is noticing when a login looks wrong even though the credentials technically check out, which a password check alone was never going to catch.
Responding Fast Matters Just as Much as Catching the Problem
How quickly a business reacts is often the actual difference between a small annoying incident and a genuinely bad one. AI-based systems can flag something suspicious, ping an administrator right away, block the offending IP, cut off a compromised service, and put together a report on what happened, all pretty much instantly. That doesn’t mean the human part goes away. It just means someone gets to spend their time figuring out why it happened instead of scrambling to contain it first.
None of This Replaces Actual Admin Work
AI helps a lot, but it’s not a substitute for someone who actually knows what they’re doing. Updates still need installing. Firewalls still need proper configuration. Old software still needs removing. Backups still need to exist, and actually get tested once in a while, not just sit there untouched. Security audits still need to happen periodically. AI works best as backup for an experienced admin, not as a stand-in for one.
Where This Is Probably Going
AI’s role in security is only going to grow as attacks keep getting more sophisticated. The models behind this keep getting better at spotting unusual behavior, predicting where the next vulnerability is likely to show up, and handling routine security tasks that used to eat up hours of someone’s week.
For a business running anything on a Linux VPS, pairing decent AI-based monitoring with the fundamentals still gets you the most resilient setup. Nothing eliminates every risk completely. But staying on top of updates, watching things proactively, and running your server responsibly, through something like BlueVPS or whichever provider fits your setup, is still what all of this is actually built on.
Treat AI as a genuinely useful assistant here, not a replacement for someone paying attention. That’s really the whole point. Businesses end up better protected, and they still keep the flexibility and performance that made a Linux VPS worth using in the first place.
The post How AI is Changing Linux VPS Security for Businesses appeared first on AI News.
View the full article
OpenAI has released a new field report tracking eight scientific computing projects where coding agents cut runtimes.
The report documents projects that used Codex on its own in five cases and a combination of Codex and Anthropic’s Claude Code in three others. Worth flagging upfront: this is a vendor publishing a survey of its own product’s application in research settings, built from case studies written by the contributors involved.
That doesn’t make the underlying pattern less worth examining. Research software has a documented maintenance problem. Tools built to accompany a single paper, coded by small academic teams without dedicated engineering support, tend to accumulate technical debt that nobody has the budget or mandate to pay down.
OpenAI’s report argues agents can address that debt, and the eight projects it cites span genomics, immunology, statistics, and RNA sequencing.
What tasks the agents undertook
The tasks split roughly into three categories: packaging and build-system cleanup, performance optimisation on existing code, and full language or backend ports.
cyvcf2, a Python library for reading genomic variant files, had its legacy build and packaging system replaced with a newer, unified process, according to contributor Brent Pedersen, who noted that going fast with agents is one thing, but going far in science still needs “expert guidance, understanding, taste, and care.”
HI.SIM, a DNA-sequencing read simulator, saw two largely autonomous optimisation passes from GPT-5.2 and GPT-5.6 that contributor Andrew Ho says cut runtime by 31 percent across a representative test set without altering output.
Ho, who describes himself as neither a genomics specialist nor a C programmer, called the outcome “nothing short of magical” from an end-user perspective, having previously lost time to performance bugs and packaging problems he could recognise but not personally fix.
Hifiasm, used for genome assembly from PacBio HiFi reads, got a 25 percent runtime cut on its optimisation target and roughly 15 percent on separate human sequencing data, per contributor Suyash Shringarpure.
Shringarpure described the agent setting up its own benchmark scaffolding and proposing candidates independently, though he stressed that supplying profiling results and steering the model away from repeated failure modes remained work only a human could do.
MHCflurry, which predicts protein fragments presented to T cells, had its TensorFlow/Keras backend migrated to PyTorch while keeping compatibility with previously released model weights, a change contributors Alex Rubinsteyn, Sergey Feldman, and Timothy O’Donnell frame as the kind of “unglamorous, labour-intensive upkeep” that keeps open-source scientific projects alive rather than left to decay.
bayesm-rs, a Rust port of statistical models from R’s bayesm package, matched the original software’s estimates within a pre-set tolerance and ran 2.3–2.7 times faster on a single processor thread, climbing to 4.4–9.5 times faster across eight threads. According to contributors Andrew Bai and Andrew Ho, the agents handled anything with a direct reference to check against quickly and correctly; extensions requiring statistical judgement the original code never pinned down needed direct human validation instead.
Rust ports and a GPU redesign push the pattern further
Three further projects – rustar-aligner, svb, and kuva – involved Rust builds carried out with coding agents, including a full recreation of STAR, a widely used RNA-sequence alignment tool that had lost active maintenance.
Contributor James M. Ferguson says agents change what’s worth attempting: rewriting a 20,000-line aligner by hand isn’t a sensible use of time, but with an agent it becomes weeks of steered work. Verification, he added, is a separate matter entirely. A model can claim a plot looks fine, but checking over 900 of them by eye before release still fell to a person.
RustQC consolidated 15 separate RNA-sequencing quality-control tools into a single program that contributor Phil Ewels says cut runtime by 60 times and disk input/output by 25 times, with companion rebuilds FastQC-Rust and Trim Galore running seven and three times faster respectively while preserving the original tools’ behaviour.
Ewels also flagged the downside: cheap rebuilds bring their own risk, because tools that diverge in behaviour fragment the community and make results from different labs incomparable over time. “The technology is the easy part,” he said. “Stewardship is the open question.”
HelixForge, a GPU-native rebuild of the mutation-simulation tool BAMSurgeon, reportedly cut runtime by around 60 times on a benchmark involving real human data, according to contributors Mamad Ahangari, Varun Goyal, and Hassan Masoudi, who also say it produced mutation frequencies closer to requested targets and resolved several bugs that generated artefacts in the original tool.
Verification, not code generation, is the constraint now
What comes through across all write-ups is that agents handled well-scoped implementation requests capably but couldn’t judge whether their own output was scientifically sound.
Contributors describe agents expressing confidence in work that contained clear errors, which pushed the actual burden onto humans to build acceptance tests: exact output matching, parity checks against an existing tool, or answers established beforehand using simulated data.
Projects tended to proceed in stages, with agents producing fast first drafts and the remaining time going into edge cases and small numerical discrepancies that a benchmark alone wouldn’t catch.
Lower engineering costs cut both ways. They let a two-person team take on a rebuild that would once have needed a grant-funded engineering hire, and they make it easier for three different labs to produce three incompatible versions of the same tool. Changes to MHCflurry and cyvcf2 went back into their original upstream projects. rustar-aligner moved to new community stewardship because the tool it replaced had already been abandoned.
The OpenAI report points toward a specific choice rather than a general endorsement: decide who owns a rebuilt tool, and secure that commitment, before the first line of agent-generated code ships.
See also: Guardoc Health processes clinical documentation using Amazon Nova models
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post OpenAI report links coding agents to faster science software builds appeared first on AI News.
View the full article
Google’s AI-generated summaries appeared in 43% of US searches measured by Similarweb, up from 15% a year earlier, according to the market intelligence firm’s 2026 Generative AI Landscape report.
AI Overviews display generated summaries within conventional Google results, while AI Mode provides a conversational interface for longer questions and follow-up prompts. Users can move from an AI Overview into AI Mode while retaining the context of their original search.
Google said AI Mode uses a technique called query fan-out, which divides a question into related subtopics and runs multiple searches. The system uses information from those searches to produce a response with supporting links.
Google expands conversational search
Similarweb estimated that visits to Google’s AI Mode web experience increased from 126 million in June 2025 to 279 million in May 2026. Google separately reported in May that AI Mode had surpassed one billion monthly users, although its figure covers users rather than web visits and is not directly comparable with Similarweb’s estimate.
The figures also measure different parts of Google Search. The 43% figure tracks the share of US searches displaying AI Overviews, while the AI Mode figures relate to activity within Google’s conversational search service.
Similarweb recorded a 5.4% increase in the average length of Google searches following the introduction of AI Mode. The report said users were entering longer, more natural-language searches similar to prompts used with AI assistants.
“People are now adopting a new, more natural way to search and discover,” Ethan Smith, chief executive of digital marketing firm Graphite, said in the Similarweb report.
Users can provide more detail in a single search instead of relying only on short keyword combinations. Google’s AI interfaces can respond to these longer searches and support additional questions without requiring users to open an external website immediately.
Similarweb said its estimates draw on first-party analytics, anonymised device information, external data partnerships, and publicly available web data. The company models those inputs to estimate traffic and user behaviour, meaning its figures are not direct counts supplied by Google or OpenAI.
According to the report, generative AI websites received an average of 9.5 billion monthly visits between June 2025 and May 2026, up 70% from the previous year. Monthly unique visitors increased 57% to 655 million, while mobile application downloads rose 58% to 4.4 billion.
Those figures cover the broader generative AI market, including standalone services such as ChatGPT, Gemini, Claude, and Perplexity. They provide context for Google’s expansion of AI within a search product that already reaches a wider internet audience.
Similarweb also found that most ChatGPT users continued to use Google. Audience overlap among ChatGPT, Gemini, and Claude indicated that people were using several search and AI services rather than moving exclusively to one platform.
Publishers have raised concerns about referral traffic when their content is incorporated into an AI-generated answer without producing a visit. Similarweb has previously identified news publishers among the sectors affected by changes in traffic from AI-based search tools.
Google has responded by changing how external sources appear within AI Overviews and AI Mode. In May, the company announced direct links within responses, article suggestions, website previews, and more prominent references to original material.
Google said the changes were intended to provide more information about linked pages and make relevant websites easier to identify. It did not release traffic data showing how the updates affected click-through rates to publishers or other external sites.
Citations do not always deliver traffic
Similarweb found that the share of ChatGPT responses containing web citations increased more than fivefold over the past year. The proportion rose from about 1.3% in June 2025 to 6.8% in May 2026, based on US desktop activity.
The measurement covers responses displaying source references, rather than the proportion of user prompts that caused ChatGPT to search the web. Citations also do not confirm that a user opened the linked source.
Citation rates varied by sector. Similarweb recorded higher rates for travel, retail, and sports responses, where information such as prices, availability, comparisons, and results can change frequently.
The firm also found a difference between the pages cited in AI responses and those receiving referral traffic. About 65% of URLs cited by ChatGPT were located two or three folders below the main domain, including articles, product pages, and other detailed content.
Folder-depth two alone accounted for 41.7% of cited URLs. By comparison, 58.8% of AI referral traffic landed on homepages.
The results indicate that the content used to support an AI answer and the page visited after a click often serve different functions. Detailed pages can supply information for a response, while displayed links can direct users to a company or publisher’s main website.
ChatGPT’s May 7 search update gave brand links greater prominence within generated responses. Similarweb recorded a 157.7% week-over-week increase in referral traffic following the update, while homepage referrals increased 354.7%.
The firm observed the increases after the interface change but did not establish that the update was solely responsible. The data covers ChatGPT referral activity and does not measure clicks from Google AI Overviews or AI Mode.
Before the update, about 26% to 32% of ChatGPT referral visits landed on homepages. The proportion increased to around 60% after May 7, based on desktop activity measured between April 30 and May 20, 2026.
The difference between cited sources and referral destinations gives publishers and website operators separate figures to track. Citation data measures which pages appear as sources, while referral data records where users arrive after clicking a link.
Cloudflare is also testing a Pay per Crawl service that allows participating website operators to permit, charge, or block individual AI crawlers. The private-beta programme allows publishers to set a fee that an authenticated crawler must pay before accessing content.
The system can return an HTTP 402 “Payment Required” response when paid access applies. Cloudflare records completed paid requests, charges the crawler operator, and distributes the proceeds to the website owner.
Cloudflare has also made configurable HTTP 402 responses available to paying customers through its AI Crawl Control service. Website operators can use the response to provide licensing terms or contact details to crawlers, while automated payments through Pay per Crawl remain in beta.
The measures add another option for publishers deciding how AI services can access their material as generated answers become more common across search and standalone AI platforms.
(Photo by Christian Wiediger)
See also: Examining Google DeepMind’s AI bioresilience push
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Google AI Overviews become more common in search appeared first on AI News.
View the full article
Guardoc Health says it processes over one million clinical documents daily using Amazon Nova models through Bedrock.
Bringing AI into clinical documentation comes down to a specific kind of risk calculation. Get it wrong and the errors compound into denied Medicare claims under the Patient-Driven Payment Model, audit fines, litigation exposure, and in the worst cases, a missed condition that changes how a patient gets treated.
However, get it right and the payoff shows up in fewer corrections, fewer hospital transfers, and lower compliance costs. Guardoc Health, which builds documentation software for long-term care providers, has published deployment figures it says support that outcome.
The scale of the underlying problem
Guardoc Health’s pipeline has to handle documents that arrive in nearly every format a clinical setting can produce: multi-page PDFs with handwritten physician annotations layered over printed text, prior authorisation forms where a checkbox state alone determines a coverage decision, medication lists that show up as clean tables in one chart and free text in the next, and patient intake forms mixing typed fields with rubber stamps and handwriting on the same page.
Research published in BMJ Quality and Safety puts the number of US outpatients affected by diagnostic error at around 12 million a year, with information-handling failures cited as a contributing factor. At the volume Guardoc processes, a one percent error rate in condition detection alone would generate thousands of incorrect records daily. Each one carries its own patient safety or compliance consequence.
Guardoc reports a 46 percent reduction in documentation errors, a 70 percent drop in audit fines, and more than $400,000 in annual ROI for a single facility, without publishing the baseline ******* or methodology behind those calculations.
In a quarterly deployment spanning two facilities and 200 patients, the company says its system drove 847 documentation corrections, flagged 86 issues tied to PDPM reimbursement accuracy, and was associated with a 74 percent reduction in hospital transfers per 100 admissions. A separate case study covering seven facilities and 1,618 residents identified 10,612 issues, according to Guardoc.
A retrieval pipeline built around cost as much as accuracy
Guardoc’s architecture runs condition classification through retrieval augmented generation, pulling evidence from a patient’s own documentation before reasoning across it to produce a final answer.
Amazon Textract extracts text and structural metadata from each incoming page first, at what the company treats as the lowest per-page cost point in the pipeline. That output gets chunked along clinical boundaries, so a medication list or a diagnosis section stays intact rather than getting split by arbitrary character count.
Each chunk is embedded using Amazon Titan Text Embeddings V2 and stored in Amazon DynamoDB, partitioned by patient so retrieval never crosses patient boundaries. A custom pre-filter narrows the candidate set by document type and recency before a k-nearest neighbour search retrieves the chunks most relevant to a given classification query, returning page references only at this stage to keep data transfer light.
Amazon Nova 2 Lite then runs a text-based pass to remove obvious non-matches. Only the pages that survive every prior filter reach Amazon Nova Pro, which receives the raw PDF bytes and reasons over layout, handwriting, signatures, and stamps to produce the classification that downstream systems act on.
The design follows a cost-tiering logic throughout: cheap components handle high-volume work like embedding and coarse filtering, and the more computationally intensive multimodal reasoning gets reserved for the final stage where it’s actually required.
The hard clinical documentation cases
Two document types account for most of what earlier pipeline versions missed, according to Guardoc. The first is physician attestation fields on prior authorisation forms, where a handwritten note can override a printed checkbox. The second is patient-reported symptom sections, where handwriting often carries information that doesn’t appear anywhere else in the record.
Medication extraction presents a related problem. Drug names, dosages, routes, and frequencies show up in structured tables, in prose buried inside physician notes, in handwritten additions to printed lists, and in scans that have been faxed through multiple hands. Guardoc’s hybrid pipeline runs Amazon Textract first for clean printed tables, then passes both the original PDF and the Textract output to Amazon Nova Pro to resolve wrapped table columns, handwritten additions, and non-standard formats that OCR alone can’t parse correctly.
“With the Nova family, we’re making it easier for healthcare organisations to detect high-risk cases earlier and act before issues become costly,” said Assaf Amiaz, Director of Product at Guardoc Health. “By automating workflows that once required manual oversight, the Nova family helps teams reduce compliance gaps, prevent errors, and focus more of their time on improving patient outcomes.”
See also: How AI is shortening drug discovery timelines in China
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Guardoc Health processes clinical documentation using Amazon Nova models appeared first on AI News.
View the full article
Armenia is not a big country. It’s not a wealthy country. It’s not a famous country. And yet, it has become not only a consumer of different AI products akin to Clideo subtitles generator for video editing, or giants like OpenAI, but also entered the headlines of global news for several reasons, not the least of which is politics. Another reason is tech advances, precisely, the alleged ‘manufacturing of chips for NVIDIA’. The latter is a big name, which you surely recognize, so let’s discuss the situation in detail.
First and foremost, news started circulating that NVIDIA was somehow “making chips in Armenia.” But clarifications are in order. Armenia is not becoming a chip fabrication country in the Taiwan sense. NVIDIA Blackwell chips are not going to be manufactured in Hrazdan or Yerevan; NVIDIA says Blackwell GPUs are manufactured using a custom TSMC 4NP process, which places the core chip-production story firmly inside the existing global semiconductor supply chain, not inside Armenia.
Armenia is trying to do something else: become a regional AI compute hub, built around imported high-end NVIDIA infrastructure, U.S.-approved chip transfers, local telecom and energy capacity, and a growing argument that smaller countries need sovereign access to AI computing power.
The centerpiece is Firebird, a U.S.-based AI cloud and infrastructure company operating between San Francisco and Yerevan. Armenia’s Ministry of High-Tech Industry says the first phase of the Firebird AI center near Hrazdan is scheduled around 2026 and involves a $500 million investment, more than 6,000 NVIDIA Blackwell GPUs, 18 MW of capacity and up to 110.6 exaflops of FP4 Tensor compute. The government presents the second phase as far larger: approximately $4 billion in total investment and more than 41,000 additional GPUs, a scale that, if delivered as described, would place Armenia unusually high on the global AI-compute map for a country of its size.
Real Story Behind Infrastructure
The confusion is understandable because the word “chips” is everywhere in this story. U.S. export approval was required. NVIDIA Blackwell GPUs are central to the project. Dell Technologies is involved on the server side. The U.S. and Armenia also signed an AI and Semiconductor Innovation Partnership MOU on August 8, 2025, which the U.S. International Trade Administration describes as covering secure semiconductor supply chains, integrated-circuit and electronics development, AI commercialization and Armenia’s position under the U.S. export-control framework. That sounds close to “chip production,” but it is not the same thing as building fabrication plants. In practical terms, Armenia’s emerging role is closer to AI infrastructure, cloud services, electronics/IC development, engineering talent and semiconductor-adjacent R&D than mass semiconductor manufacturing.
Moreover, NVIDIA now frames major AI infrastructure as AI factories, i.e., systems that manufacture intelligence rather than physical goods. Its GB300 NVL72 platform, for example, is a rack-scale system combining 72 Blackwell Ultra GPUs and 36 Arm-based Grace CPUs in a liquid-cooled architecture designed for training and inference at very high density. Dell, one of the infrastructure partners in this broader market, describes its PowerEdge XE9712 with NVIDIA GB300 NVL72 as a rack-scale enterprise AI platform, built for high-output reasoning inference and improved throughput. In other words, Armenia’s project is not about making the chip; it is about hosting the machines that make large-scale AI development possible.
Independent regional reporting supports this reading. Eurasianet reported in November 2025 that U.S. regulators approved the transfer of advanced NVIDIA chips for Armenia’s AI hub, while also noting Dell’s PowerEdge server role and the political importance of the approval process. OC Media, citing Bloomberg, reported that the project involves a 100 MW data center using NVIDIA Blackwell chips and Dell AI servers, with the first phase expected in 2026. OC Media also reported that part of the significance lies in export controls: Armenia had previously been affected by U.S. restrictions on exports of high-performance AI chips to certain countries, making approval a strategic as well as commercial milestone.
Why Armenia Wants AI Compute So Badly
For Armenia, the strategic logic is clear. The country already has a serious technology base relative to its size, especially in software engineering, mathematics, electronic design automation and diaspora-connected entrepreneurship. Synopsys Armenia, for example, describes its local operation as covering R&D and product support for electronic design automation, design for manufacturing and semiconductor IP, with more than 1,000 employees across Yerevan and Gyumri. That does not make Armenia a chip-manufacturing giant, but it does mean the country has long been more connected to the semiconductor value chain than casual observers may assume.
The Firebird project attempts to add a missing layer: massive compute. Talent without computers becomes dependent on foreign cloud providers. Research without GPUs remains theoretical. Startups without local access to advanced infrastructure may build around someone else’s platform, pricing and policy limits. That is why the phrase “compute sovereignty” fits the Armenian case better than “chip production.” The argument is not that Armenia can replace existing semiconductor powers. The argument is that access to AI infrastructure is becoming a national development issue, the way broadband, energy grids and transport corridors once were.
There is also a public-sector AI layer. Armenia has signed a cooperation agreement with Mistral AI, and Mistral’s own customer page says the partnership with Armenia’s Ministry of High-Tech Industry focuses on AI assistants, public services and government functions, with attention to open and adaptable models that fit Armenia’s language, infrastructure and use cases. That makes the country’s AI push less like a single data-center announcement and more like a broader attempt to connect infrastructure, public services, startups, universities and international partnerships.
The Hard Part Comes After the Announcement
However, there are obstacles on the way. The factories of AI require a massive amount of energy, capital and are politically charged. A 100 MW data centre would consume as much electricity as a city of around 120,000 people, and could consume significant cooling capacity, but Armenia’s Ministry of High-Tech Industry claims that the Firebird centre is fitted out with a closed loop water-cooling system that only requires a water change every few years. This is the case of Armenia’s own energy situation: U.S. International Trade Administration reports that Armenia has an adequate current electricity generating capacity with rising consumption and is still relying on a combination of nuclear, hydro and thermal generation.
There is also a risk of execution. Firebird is young, the capital figures are big, the gap between an AI factory that has been announced and is fully leveraging and an AI compute hub that is fully utilized and commercially viable is huge. The project will require customers, a reliable power supply, cooling, network resiliency, export-control stability, and operational expertise and a local ecosystem that can absorb a part of the compute capacity. The Ministry reports that Team Telecom Armenia is installing next-generation fiber-optic infrastructure with up to 1 Tbps bandwidth, while according to OC Media, the project could include some portion of compute which would be allocated to domestic companies and the remaining compute that would be sold to U.S.-based companies operating in the region. What’s important about those details is that usage of the GPUs will impact local benefit – who will be getting them, at what price, and for what type of work.
The most accurate way to describe Armenia’s AI moment, then, is not “NVIDIA is producing chips in Armenia.” It is this: Armenia is trying to turn advanced imported chips into domestic strategic capacity. If Firebird, Dell, NVIDIA-linked infrastructure, U.S. export approvals, local telecom investment and public-sector AI partnerships come together as planned, Armenia could become one of the more unusual AI infrastructure stories of the next few years: a small, landlocked country using computers as a development strategy. The gamble is bold. Armenia is not trying to become the next chip fab. It is trying to become a place where the next generation of AI systems can be trained, hosted and commercialized.
The post Armenia’s AI Bet Is Not Chip Manufacturing. It Is Compute Sovereignty appeared first on AI News.
View the full article
Insilico Medicine has reduced the time needed to produce some drug development candidates to about one year by combining artificial intelligence with laboratory research in China, according to CEO Alex Zhavoronkov.
The Hong Kong-listed company’s fastest programme reached candidate nomination in nine months, while its typical timeline is about 13 months, Zhavoronkov said. He said conventional approaches usually take about four-and-a-half years to reach the same stage.
The timeline covers early discovery and candidate selection, rather than the full process of bringing a drug to market. Clinical trials, manufacturing, and regulatory review remain separate stages.
AI shortens candidate selection
Insilico uses generative AI to identify biological targets, design potential drug molecules, and assess which compounds should advance to laboratory testing.
The company said its programmes typically reach preclinical-candidate nomination within 12 to 18 months after researchers synthesise and test between 60 and 200 molecules. Its workflow combines AI-generated designs with researcher review and experimental validation.
Laboratory experiments remain necessary to confirm the biological activity and drug properties of compounds selected by the models. Insilico said its AI-supported process allows teams to reach candidate nomination after testing a smaller set of synthesised molecules, although it has not provided a direct comparison with equivalent programmes developed without AI.
Insilico said it has generated 31 preclinical candidates since 2021. Thirteen programmes have received investigational new drug clearances, allowing them to advance towards human studies, according to the company’s pipeline disclosures.
The company conducts AI research in Montreal and Abu Dhabi, while much of its experimental validation and laboratory scale-up work takes place in China. Its Shanghai facility has automated parts of biological sampling and compound screening.
Teams outside China develop and evaluate the company’s AI models, while researchers in Shanghai handle biological testing, screening, and scale-up.
Zhavoronkov attributed part of the shorter development cycle to China’s research infrastructure, operating costs, and regulatory environment. He said pharmaceutical companies with research laboratories in China can remove about two years from traditional candidate-development timelines.
China has expanded beyond manufacturing generic drug ingredients and now plays a larger role in developing new medicines. International drugmakers also work with ******** laboratories, contract research organisations, clinical-trial centres, and biotechnology companies.
A Pfizer executive said clinical development in China could be conducted three times faster and at about half the cost of equivalent work in Europe. Drug candidates typically take five to seven years to reach the ******** market, compared with at least eight to 10 years in Western markets, according to Reuters.
China introduced a 30-working-day review pathway in 2025 for eligible Class I innovative-drug clinical-trial applications. Applications requiring expert consultation or involving complex technical issues can be moved to a 60-working-day review *******.
“We now compete with ******** pharmaceutical companies on timelines, and with traditional biotechnology companies in the West on novelty,” Zhavoronkov said.
Insilico has entered research and development agreements with pharmaceutical companies including Eli Lilly and Japan’s Takeda.
The company and Taiwan-based Bora Pharmaceuticals also announced a proposed strategic alliance that could exceed $2.5 billion if definitive agreements are signed and the collaboration is fully implemented.
Although Insilico operates research facilities in China, Zhavoronkov said more than 90% of its revenue comes from Western pharmaceutical companies. He did not disclose how much revenue the company generates in China.
Western licensing agreements are more lucrative for Insilico because China’s national insurance system offers lower reimbursement rates for highly novel drugs, Zhavoronkov said.
The company also limits sales of most of its software within China because of geopolitical concerns, Zhavoronkov said. It plans to expand its research operations in Shanghai.
Rentosertib moves towards Phase III trials
Insilico announced and registered a Phase III trial of Rentosertib in July 2026. The oral drug is being studied for idiopathic pulmonary fibrosis, a disease that causes progressive scarring of the lungs.
The company used AI to identify the drug’s biological target and generate and optimise its molecular structure.
The Phase III study is designed to enrol 320 participants across 47 centres in China. It will compare Rentosertib with a placebo over 52 weeks, with the primary endpoint measuring the annual rate of decline in forced vital capacity, a standard measure of lung function.
The trial was listed as not yet recruiting when its ClinicalTrials.gov record was updated on July 7. Enrolment was expected to begin in August 2026, with primary completion estimated for October 2029.
Rentosertib previously completed a smaller Phase IIa study. The Phase III trial will test the treatment in a larger patient group over a longer *******.
Candidate nomination remains an early development milestone. Drugs must still complete preclinical testing, human trials, manufacturing validation, and regulatory review before they can be approved for *****.
Industry data have not established whether AI-designed drugs are more likely to succeed in later-stage trials.
A 2024 analysis of AI-native biotechnology pipelines reported Phase I success rates of between 80% and 90%. The same study found a Phase II success rate of about 40%, broadly in line with the historical industry comparison used by the researchers.
The researchers said the number of Phase II programmes was too small to determine whether AI improves later-stage clinical success. The analysis was based on publicly reported pipelines and did not compare otherwise identical AI-supported and conventional drug programmes.
Insilico said it has produced 31 preclinical candidates and secured 13 investigational new drug clearances. Rentosertib is its first programme to reach the Phase III stage, while none of the company’s experimental medicines has received commercial approval.
Automation changes biotech roles
AI and laboratory robotics are also changing staffing requirements within Insilico.
Zhavoronkov estimated that the company could automate or displace about 40% of its software-side workforce. He did not describe the figure as an announced staff reduction or apply it to the biotechnology industry as a whole.
Insilico employs about 400 people. Laboratory scientists and software engineers are being retrained to manage AI evaluation systems, automated equipment, and robotics, Zhavoronkov said.
The retraining is focused on AI benchmarks and robotic systems as the company automates more research and software functions, he said.
(Photo by Julia Koblitz)
See also: Bristol Myers Squibb buys Nvidia AI system for drug discovery
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post How AI is shortening drug discovery timelines in China appeared first on AI News.
View the full article
America’s AI Investment ***** Is Reshaping the Economy
Artificial intelligence has become one of the defining investment stories in the United States, and the numbers continue to grow. Microsoft, Meta, Amazon and Alphabet are collectively committing hundreds of billions of dollars to AI infrastructure, while demand for advanced chips has turned NVIDIA into one of the world’s most valuable companies.
It’s apparent that what started as a race to build smarter AI models is now a driving force behind investment in construction, manufacturing, energy and digital infrastructure, with effects that are spreading throughout the wider economy.
AI Investment Fueling America’s Biggest Infrastructure Push in Years
The headlines often focus on new AI models, but rarely focus on the resources required to make those models workable. Every chatbot, image generator and AI assistant relies on vast data centres packed with specialised processors, connected by high-speed fibre networks and powered by enormous amounts of electricity. Building that capacity requires billions of dollars in construction, engineering and equipment, and it creates a level of demand that reaches far beyond the technology sector.
These investments are also reshaping expectations for the US economy, with productivity gains and capital spending influencing currency markets before they appear in official data. For investors, forex trading can provide an early view of how global markets are responding to America’s expanding AI economy.
Productivity Is Influencing the Next Phase of Growth
An important factor to consider is that the long-term value of AI isn’t going to be measured by how many new tools reach the market, but rather by whether businesses become more productive with those tools.
That’s a process that’s already underway. For example, manufacturers are using AI to identify faults before products leave the factory, healthcare providers are reducing administrative workloads, and financial institutions are analysing market data in seconds instead of hours. In fact, Goldman Sachs estimates that generative AI could increase global GDP by around 7% over the next decade if adoption continues to accelerate, highlighting why businesses view AI as an investment in future growth rather than simply another software upgrade.
Realising those gains, however, will require careful implementation. Businesses must invest in training, establish AI governance, and address data security and regulatory compliance to ensure new tools deliver lasting value rather than short-term efficiency gains. AI adoption is also changing how employees spend their time. Rather than replacing entire roles, many organisations are using AI as supplementary tools that automate repetitive tasks such as drafting reports, analysing large datasets, or handling routine customer enquiries. With repetitive tasks taken off their plates, employees can focus on higher-value work, while businesses can improve efficiency without fundamentally changing how they operate overnight.
The Ripple Effect Extends Well Beyond Silicon Valley
The companies building AI models are only one part of a much larger ecosystem. This can be seen in how utilities are expanding electricity generation to support new data centres, in how semiconductor manufacturers are increasing domestic production, and in how construction firms are winning contracts to build facilities capable of housing next-generation computing infrastructure.
States like Texas, Arizona, and Virginia have become major beneficiaries for companies that want locations with reliable power, skilled workers, and room to expand. Together, these investments support local economies and strengthen AI-powered industries.
Financial Markets Are Responding Long Before the Economy Fully Adjusts
Markets rarely wait for quarterly GDP figures before reassessing growth prospects, inflation expectations, and interest-rate outlooks. Instead, investors respond to new information as it becomes available, particularly when major AI investment announcements point to stronger productivity or sustained business spending.
This also helps explain why movements in the US dollar can sometimes reflect optimism about economic prospects long before those trends are apparent in official statistics (currency markets often react to expectations before they’re confirmed).
Investors also watch whether AI investment translates into stronger earnings and sustained capital spending, both of which can influence interest-rate expectations and global capital flows.
Understanding the broader impact of AI means looking beyond technology headlines and instead at the elements that influence the market: corporate earnings, employment reports, inflation data, and central bank decisions. Platforms like OANDA support this approach by combining access to the foreign exchange market with real-time market analysis, economic calendars, and research tools that help traders interpret macroeconomic developments.
America’s AI Economy Is Still in Its Early Stages
Unlike a couple of years ago, AI investment is no longer confined to technology companies or venture capital funding rounds; it’s reshaping supply chains, accelerating infrastructure projects, creating demand for skilled workers, and influencing how investors evaluate the outlook for the U.S. economy. Even though those changes will take years to play out, many of them are already visible today.
The next chapter of the AI story will likely be measured not by faster models, but by how effectively businesses convert record levels of investment into increased workplace productivity, and how financial markets respond.
The post America’s AI Investment ***** Is Reshaping the Economy appeared first on AI News.
View the full article
Two dozen companies and organisations signed an open letter urging US policymakers to protect open-weight AI models.
The letter, published today (PDF), carries signatures from a list that spans direct commercial rivals and organisations with little obvious overlap in business model: Meta, Microsoft, Nvidia, IBM, Dell Technologies, CrowdStrike, Palantir, ServiceNow, Hugging Face, Perplexity, Mistral, Andreessen Horowitz, Y Combinator, the Linux Foundation, Mozilla and others.
The letter’s argument centres on a comparison between the open-source software movement of the 1980s and the current fight over whether AI model weights should circulate freely or stay locked behind commercial APIs.
Open-weight models are AI systems where the trained parameters get published for anyone to download, inspect, modify and run on their own hardware. That’s distinct from closed models like the frontier products offered by OpenAI or Anthropic through API access only, where the underlying weights never leave the vendor’s infrastructure.
The signatories frame open weights as the mechanism by which AI capability spreads beyond a handful of well-capitalised labs into what the letter calls the workflows of “factories, hospitals, farms, classrooms, and main street businesses.”
Their argument runs on three tracks:
Open weights lower the cost of entry for startups and public institutions that can’t afford to train frontier models from scratch or pay per-token fees at frontier prices for routine tasks.
They increase competition across the stack, from chips to cloud infrastructure to applications, which the letter says keeps costs down and prevents value capture by a small number of providers.
Open weights also give enterprise customers a way to avoid vendor lock-in, since organisations running open-weight models control their own data and can adapt the model to internal requirements without depending on a single vendor’s roadmap or pricing decisions.
The security argument runs against instinct
The letter’s most pointed section addresses the risk case directly, and it’s worth reading closely because it inverts the usual framing around open models and security.
Once weights are released, the letter concedes, they’re beyond the original developer’s control. Modified versions become difficult to trace or reverse. A fine-tuned or stripped-down version of an open model can circulate with safety guardrails removed, and there’s no recall mechanism.
The signatories argue the answer isn’t prohibition. Their case rests on a comparison to cybersecurity: defenders facing AI-equipped attackers need access to models with comparable capability to detect and simulate threats, which closed, permission-gated systems don’t easily provide.
They extend this into a broader security claim, arguing that closed models aren’t inherently safer because they can be breached, misused, or fail in ways external researchers can’t observe or verify. Concentrating advanced capability behind a small number of closed providers, in this reading, creates single points of failure rather than removing them.
Open models, by contrast, let outside researchers examine behaviour, run red-team exercises, and identify vulnerabilities across many teams rather than relying on one vendor’s internal testing.
The letter draws a direct parallel to the “open-source is more secure than obscurity” argument that shaped decades of software security debate, though it doesn’t cite specific vulnerability-discovery data or incident figures to support the claim as applied to AI systems specifically.
Distillation gets a specific defence
The letter carves out space for one technique that’s become contentious in AI circles: distillation, where one model’s outputs get used to train or improve a second model. This is standard practice in machine learning research and product development, used for evaluation, validation, and capability transfer between models of different sizes.
The signatories draw a line between distillation as a legitimate technique and what they call “unlawful efforts to extract value from closed models,” arguing the former shouldn’t get swept up in restrictions aimed at the latter.
This reads as a direct response to disputes that flared after the rise of ******** models like DeepSeek and Kimi, when several US labs suggested rival models had been trained by distilling outputs from their own closed systems without authorisation.
The letter’s position: address misappropriation through targeted legal and commercial mechanisms, not blanket restrictions on a technique the entire field depends on.
What this signals for the policy fight ahead
The letter arrives without a specific legislative or regulatory proposal attached. It’s a positioning document ahead of anticipated action on AI policy in Washington, calling on lawmakers to expand compute access for startups and researchers, fund shared training datasets and evaluation frameworks, and avoid what it calls “premature restrictions” on open models.
This should be treated less as a settled policy outcome and more as an indicator of where major infrastructure and chip providers want the regulatory conversation to land. Players like Nvidia, IBM, and Dell have direct commercial reasons to want open-weight ecosystems to flourish since a wider range of deployable models sells more compute and services regardless of which lab produced the weights.
Procurement teams weighing open-weight versus closed-model deployments should factor in that the policy environment favouring one approach over the other remains unresolved, and any restrictions on distillation or open releases could shift the economics of self-hosted AI within a single legislative cycle.
See also: OpenAI pushes ChatGPT into patient health records
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Meta, Microsoft, Nvidia, IBM, and others back open-weight AI appeared first on AI News.
View the full article
OpenAI is deploying a Health feature inside ChatGPT, giving users the option to connect Apple Health data and medical records to the chatbot. Logged-in users aged 18 and older can access it now on web and iOS, across the Free, Go, Plus, and Pro tiers.
Users link Apple Health and, where supported, records from US hospital systems, One Medical, or Function Health. Once synced, ChatGPT can pull medications, lab results, recent visits, sleep data, and activity logs into any conversation in the app, rather than confining that context to a separate section.
OpenAI ran an earlier version of this idea with a smaller test group, one that required users to open a dedicated health area to get responses grounded in their own data.
The company found that more than 70 percent of health-related conversations among that group happened somewhere else entirely, in the middle of meal planning or an unrelated symptom query, not inside the dedicated space built for that purpose.
That data led to this redesign. Rather than forcing users into a specific mode to get contextual answers, ChatGPT now draws on connected Health information across any conversation, provided the user has granted permission.
A person planning a dinner out might get a restaurant suggestion that accounts for a logged dietary restriction. Someone asking about weekend plans might get activity suggestions adjusted for a recent injury noted in their synced records.
The Health tab in the sidebar still exists, but its role has shifted to being a management hub: connecting accounts, reviewing synced data and trends, browsing suggested prompts, and returning to past health-related chats.
What early testers of the Health feature in ChatGPT report
OpenAI published numerous accounts from its early access group, and the details are worth weighing against the company’s own framing of the tool as support rather than diagnosis.
Blake, a technical program manager, said: “The most useful part has been turning scattered medical history into something I can actually understand and explain. I have multiple overlapping issues and ChatGPT helped connect those pieces into a clear timeline, explain the medical terms in plain English, and create summaries I could share with a physical therapist or trainer.”
On the shift from disconnected records to a usable pattern, Blake added: “Instead of just seeing disconnected diagnoses, imaging results, and surgery notes, I could understand the ******* pattern. It made the information more usable and gave me better language to advocate for myself with providers and trainers.”
Reweti, a portfolio manager, pointed to longitudinal analysis as the differentiator over a standard search or a one-off doctor visit: “What I want is to infer patterns that aren’t obvious and make connections I wouldn’t have made on my own.
“ChatGPT can access my existing labs because I’ve connected everything, and it can look over time—that’s the big advantage. It’s like having a research analyst. It allows me to be more proactive and own more of my health journey.”
However, not every account was frictionless, and one is worth flagging given the stakes involved in surfacing clinical data through a chatbot interface. Shannon, a nurse, described finding an unexpected entry in her own chart through the tool.
“Using Health has actually reinforced something I’ve believed for a while: one of AI’s greatest strengths isn’t replacing healthcare professionals, but helping patients better understand and navigate their own health information,” explained Shannon.
“Discovering an unexpected chart entry through Health really highlighted that for me. It wasn’t AI creating a problem. It helped me identify something I can now appropriately follow up on with my healthcare providers.”
That distinction, between a tool surfacing something for human follow-up versus a tool making a clinical call, is the line OpenAI needs the product to hold as usage scales.
Other testers focused less on clinical nuance and more on the practical grind of manual data wrangling the feature is meant to replace. Daniel, a consultant, connected multiple sources and found the combined view more useful than isolated chat sessions.
“It’s been great for coordinating labs and translating them into language I understand. Connecting Apple Health and MyChart makes the insights more grounded in what is happening across my life outside of just chat interactions,” said Daniel.
Kathleen, a small business owner who’d lost a decade-long fitness habit to work pressure, described a lower-stakes but still concrete use case, with the model adjusting suggestions based on activity gaps it could see directly.
“I’ll say, ‘I need something to help me move today,’ and Health can see I haven’t worked out in the last seven days and suggest starting slowly—with a walk or some stretching. It’s helping me make things manageable and get back into it in a reasonable way,” explained Kathleen.
Carlton, an operations manager, had previously resorted to exporting spreadsheets from Apple Health and uploading them manually before every chat, a workaround the new integration is designed to eliminate.
“Prior to Health, I was exporting massive spreadsheets from Apple Health and importing them into ChatGPT. Now, it feels a lot more streamlined. Being able to see years and years of my fitness journey in Health helped me see things in a different light—a ******* picture,” said Carlton.
OpenAI puts weekly health-related ChatGPT queries at north of 300 million people, covering everything from decoding a lab result to prepping for a doctor’s visit. That figure, if accurate, puts ChatGPT in a position most digital health platforms would need years and considerable marketing spend to reach.
The company is explicitly positioning Health as a support tool rather than a diagnostic one, and telling users to confirm anything important with their actual healthcare provider.
OpenAI’s AI model performance claims and how they were tested
OpenAI attributes the feature’s viability partly to newer models: GPT‑5.5 Instant, available to Free users, and GPT‑5.6 Sol, reserved for paid tiers.
The company says GPT‑5.5 Instant made gains in recognising when urgent care might be needed and in explaining uncertainty, and that on its toughest health evaluations it performed comparably to OpenAI’s frontier Thinking models at the time. GPT‑5.6 Sol is described as the company’s strongest health model so far, built for reasoning across multiple data points such as lab trends over time.
To validate these claims, OpenAI says it worked with hundreds of physicians to build health scenarios and rubrics scoring responses on accuracy, safety, communication, context awareness, completeness, and appropriate escalation to professional care. The company reports that every GPT‑5.6 model outperformed GPT‑5.5 on HealthBench Professional, an internal evaluation built for this purpose.
A chart included in OpenAI’s announcement shows GPT‑5.6 Sol scoring higher than GPT‑5.5 Instant and GPT‑4o across categories including accuracy, communication, completeness, and following instructions, with physician-written responses used as a comparison baseline.
OpenAI does say physicians tested the live Health product before release specifically to assess real-world performance and safety with connected data, which is a step beyond benchmark scoring alone, though the company hasn’t published the methodology or results of that testing in detail.
Health data handling and the permission architecture
Connected medical records and Apple Health data – along with any conversations that draw on them – are excluded from foundation model training and ad targeting, according to the company, regardless of a user’s broader ChatGPT training settings. Conversations that don’t touch Health data still follow whatever training preference a user has set separately.
Access is permission-gated by default. ChatGPT asks before using connected health data to personalise a response, though users can switch to “always allow” and turn off the prompts entirely. That setting lives in Settings > Plugins > Health and can be reversed at any time. Disconnecting a data source triggers deletion from OpenAI’s systems within 30 days, though anything already surfaced in existing chat history sticks around until the user deletes those conversations manually.
Memory creation is scoped narrowly, too. Memories can be created from health conversations but not directly from the raw connected records or Apple Health data itself. Users wanting to avoid memory creation altogether can use Temporary Chat or disable memory in settings.
OpenAI also flags a specific edge case: actions that could expose Health data through other connected plugins, such as sending a training plan built from Apple Health metrics to a running partner. The company says additional checks apply before such actions execute, and that for sensitive cases ChatGPT may ask for explicit confirmation. OpenAI states it runs red teaming exercises targeting these scenarios, though no findings or failure rates from that testing have been made public.
What happens to accuracy at the edges
The practical friction point sits with data quality. A medication can stay listed in a patient’s synced record long after they’ve stopped taking it, and OpenAI uses that exact scenario to illustrate why synced data isn’t automatically current. Its guidance to users is blunt: flag changes to ChatGPT directly and check anything important against the original source rather than trusting the sync to stay accurate on its own.
Wearable and fitness app data carries its own gaps, since availability depends on what each third-party app chooses to share through Apple Health, and OpenAI notes some proprietary scores from fitness apps may not transfer at all.
The relevant question isn’t whether ChatGPT can summarise a lab result correctly in a demonstration, it’s whether the permission model, data deletion timelines, and escalation logic hold up when a user’s synced records are three months stale and the model is asked to reason across contradictory inputs.
OpenAI’s physician testing addresses part of that concern; it doesn’t close the distance between a controlled evaluation and a user managing multiple chronic conditions with incomplete data syncing from four different apps.
For those wishing to give the feature a try, it’s live now for eligible US users through the sidebar Health menu.
See also: OpenAI Presence sells enterprise AI agents with engineers attached
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post OpenAI pushes ChatGPT into patient health records appeared first on AI News.
View the full article
The newest way to buy enterprise AI agents from OpenAI does not involve buying anything online. OpenAI Presence, announced on July 22, is a managed product delivered through a limited general availability programme, and the company states plainly that it is not yet available as a self-serve product. Deployments are led by OpenAI’s own Forward Deployed Engineers and a set of selected global systems integrators.
That is a departure for a business that has largely run on API keys and seat licences. Presence is sold as a project rather than a product. Each engagement starts with a single job, such as resolving a billing dispute, handling an insurance claim, or clearing an employee IT service request.
The agent is given only the knowledge and system access that the job requires, and the customer writes the rules governing what it can do, when it needs sign-off, and when a person takes over. After launch, Codex reads production sessions and escalations, then proposes changes the customer’s team tests and approves before rollout.
OpenAI’s documentation is unusually candid about the labour involved. Its help centre sets out a six-stage process running from scoping business outcomes, through security, privacy and legal review, simulation and acceptance testing, staged rollout, and post-launch iteration. A Presence agent, it says, does not become production-ready simply by ingesting documents.
The problem this is built to solve is real
The managed model is easy to read cynically, and harder to dismiss on the evidence. Gartner has warned that more than 40% of agentic AI projects will be cancelled by the end of 2027, attributing the failures to governance, undefined business value and weak operational discipline rather than to model capability.
Almost everything Presence bundles is aimed squarely at that diagnosis. Simulations and graders test whether an agent reached the right outcome, followed policy, used its tools correctly and escalated when it should, before anyone outside the company speaks to it. Guardrails intervene when an interaction moves past defined boundaries. Session records and action histories give reviewers something to audit. Escalation paths hand a person structured context rather than a cold transcript, and new versions go out through controlled rollout with rollback.
Enterprises have spent two years discovering that the hard part of a production agent sits in integration, permissions and change management. A vendor that sends engineers to do that work is responding to what buyers have actually been failing at, rather than shipping another dashboard and calling the gap a customer problem.
Where the constraint sits
The trade-off shows up in the eligibility criteria. Access, OpenAI says, depends on workflow fit, implementation readiness and available delivery capacity.
Delivery capacity is a consulting constraint. Software scales; engineers cleared into a bank’s core systems do not. Forward Deployed Engineer is a title borrowed from Palantir, where it describes staff embedded in customer operations for months at a time, and the economics attached to it look nothing like the economics of metered inference. By putting its own FDEs and named partners at the front of every deployment, OpenAI has stepped into the layer of the market occupied by the integrators it will also rely on to scale, which is a workable arrangement while volumes are small and a more complicated one later.
It also puts a question on the table for anyone scoping a contract. When the model vendor is also the implementation partner, the lines of accountability for a policy misapplied in production need to be written down rather than assumed.
The enterprise AI agents on display are still early
OpenAI describes Presence as battle-tested, and its case for that language is that the product was assembled from years of deploying agents with enterprise customers before it was packaged and named. The claim is about accumulated practice rather than about the product’s time in market, and it is a reasonable one to make.
The strongest single proof point is OpenAI’s own English-language phone support line, 1-888-GPT-0090. The company says the agent met or exceeded its internal benchmarks for frontline human support within weeks, now resolves 75% of inbound issues without human assistance, and cut human handoffs by 15 percentage points in ten days through the Codex improvement loop. Those are OpenAI’s figures, measured against OpenAI’s own grading criteria, on OpenAI’s own channel. The transparency is welcome, but the numbers are not independently verified.
The three named customers sit earlier in the cycle than the launch framing implies. BBVA is exploring voice support for everyday banking in Mexico. SoftBank is testing Japanese-language conversations. IAG is exploring support during high-demand events such as severe weather. Daniel Ordaz, head of AI transformation at BBVA Mexico, describes the bank as a design partner helping shape and refine voice experiences for financial customer service. Design partners are normal and useful at limited GA. None of the three, though, is presented as running Presence at scale, which is worth holding alongside the word proven.
What has not been disclosed
Pricing is not published. Implementation scope and cost are set per customer and per deployment, which is ordinary for enterprise services and still leaves buyers without a public reference point for cost per resolved contact against an incumbent contact-centre vendor.
The model is not named. Presence uses OpenAI models, the documentation says, with configuration selected for the workflow and subject to change as that workflow evolves. That flexibility is defensible engineering, because pinning a production agent to a frozen model version ages badly. Teams that have spent the past year building evaluation suites against specific versions will nonetheless want the contract to say what they are being held to when the configuration moves.
Channel support during limited GA covers voice or chat, with contact-centre integration, routing, authentication and handoff design confirmed deployment by deployment. Data handling follows the same pattern, with the signed architecture and contract treated as the governing record rather than any published policy.
Presence sits apart from ChatGPT Workspace Agents, which remain the self-serve path for teams building inside ChatGPT and Slack, while voice customers keep API access to OpenAI’s frontier models. The company now offers broadly the same capability three ways, separated less by what the technology can do than by who does the work.
That leaves buyers choosing on delivery capacity as much as on model capability, and on OpenAI’s own account, delivery capacity is the part being rationed.
See also: HP accelerates enterprise workflows with OpenAI Frontier
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post OpenAI Presence sells enterprise AI agents with engineers attached appeared first on AI News.
View the full article
Nvidia’s new Medical Physics Simulation framework treats healthcare robots as physical AI systems that need embodied experience to learn, not just code.
Physical AI is the term Nvidia and much of the robotics industry now use to describe machines that have to learn how the world behaves through contact, force, and consequence, rather than through text or images alone.
A language model learns from text. A physical AI system learns from what happens when a catheter meets a vessel wall, or when a robotic arm applies too much pressure to soft tissue. That kind of learning normally requires either a physical body operating in the physical world, or a simulation detailed enough to stand in for one.
For healthcare robotics, physical bodies operating in real procedures are scarce, tightly regulated, and slow to generate the range of scenarios a robot actually needs to see. Medical Physics Simulation is Nvidia’s attempt to manufacture that embodied experience computationally.
Announced as an open-source addition to the company’s Isaac for Healthcare platform, the framework generates the physical interactions a surgical or diagnostic robot would otherwise need years of clinical exposure to encounter: a guidewire catching on a calcified vessel wall, a kidney stone lodged at an unusual angle, or the soft-tissue response that only shows up in a small fraction of procedures.
None of these edge cases arrive on schedule in an operating theatre. Simulation lets developers generate them on demand.
Building physical intuition before a scalpel gets involved
The framework combines two ways of modelling how devices behave inside a body.
Classical physics simulation handles the mechanical rules that are already well understood, how a catheter bends, how much resistance a vessel wall applies, how contact forces shift as an instrument moves through tissue. Generative AI handles the part that’s harder to hand-code: visual scene dynamics learned from procedural data, delivered through a component Nvidia calls Cosmos-H Dreams.
That combination is the physical AI proposition in miniature. Classical simulation gives a robot policy the physics it needs to obey. Generative simulation gives it the range of visual and anatomical variation it needs to generalise. Put together, and run at scale on GPUs using Nvidia’s Warp and Newton libraries, the framework can execute large numbers of parallel training environments instead of one scene at a time.
Nvidia states that a benchmark running 8,192 parallel environments cut training time from over five hours to under two minutes. However, that demonstrates throughput – not clinical reliability – and it says nothing about how a policy trained this way performs against incomplete imaging, delayed sensor readings, and even anatomy that falls outside anything the simulation modelled.
A language model that underperforms on an edge case produces a bad answer, but a physical AI system that underperforms on an edge case is operating inside a patient. The parallel-simulation approach is a real advance in how fast developers can explore failure modes. Whether those simulated failure modes match what actually goes wrong in a surgical suite is a separate question.
Where the embodiment approach is being tested
The organisations Nvidia names as early adopters are applying the physical AI approach at different depths, and the list is worth reading with that in mind rather than treating it as a uniform roster of deployments.
CMR Surgical and Cambridge Consultants, the Capgemini-owned engineering firm, have gone furthest on the data side. CMR has contributed close to 500 hours of anonymised clinical data from its Versius Surgical Robotic System to the Open-H Embodiment dataset, spanning cholecystectomy, prostatectomy, hernia repair and hysterectomy procedures, and the pair are using Cosmos-H Dreams to model soft-tissue interaction physics and produce patient-specific simulations.
“Open-source models allow us to build on shared knowledge, accelerating responsible innovation and, ultimately, gives us the potential to deliver more consistent care and better outcomes for patients worldwide,” said Chris Fryer, CTO at CMR Surgical.
Johnson & Johnson MedTech is using the framework alongside a Cosmos-based foundation model to build a digital twin of its endoluminal MONARCH platform, focused on kidney-stone scenarios in urology.
XCath is applying it to endovascular autonomy policy training, teaching a system the physical behaviour of navigating blood vessels without a human hand on the controls. Inner Logic is generating synthetic data to validate device mechanics, and says it intends to produce in silico evidence to support regulatory submissions, though no submission built on that evidence has been confirmed publicly.
Medtronic Structural Heart sits earliest in the group, exploring simulated X-ray sensing for catheter navigation research.
Each of these is a training exercise or dataset contribution. None is a deployed system operating on a patient with policies learned this way, and Nvidia doesn’t claim otherwise.
The open-source case for physical AI systems
Healthcare robotics carries a governance requirement most physical AI applications, including industrial and warehouse robots, don’t face to the same degree: regulators and clinical review boards need to see how a system arrived at its behaviour, not just confirm that the behaviour looked acceptable in testing.
An open-source framework lets developers inspect the physics assumptions inside the simulation, reproduce results across different anatomies, and build an evidence trail suited to a submission before the FDA or an equivalent body.
That’s a stronger argument for openness in physical AI than it is in most software categories, where a closed vendor pipeline hides the assumptions a team would otherwise need to defend to a regulator. It doesn’t settle the validation question on its own.
Open code lets outside reviewers check the model’s logic, but it doesn’t confirm the model’s physical behaviour matches what happens in a body, and that confirmation still has to come from testing that none of these companies has published yet.
Nvidia has built infrastructure that could shorten the pre-hardware phase of physical AI development for surgical and diagnostic robots, and running training at this scale in parallel is a departure from rebuilding a custom simulation scene for every workflow.
You can hear more about this topic at the Physical AI Expo.
See also: Bristol Myers Squibb buys Nvidia AI system for drug discovery
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Nvidia bets physical AI can solve healthcare robotics’ data problem appeared first on AI News.
View the full article
AMD has agreed to invest up to $5 billion in Anthropic under an infrastructure agreement covering tens of billions of dollars’ worth of AI systems.
Anthropic will deploy up to two gigawatts of capacity using AMD’s Instinct MI450-series accelerators, with deployment of the first gigawatt beginning in the first half of 2027. AMD’s investment will be tied to Anthropic meeting certain deployment milestones, although the companies have not disclosed the conditions attached to them.
The companies have not said that Anthropic must use AMD’s investment to pay for the systems. The transaction combines an equity commitment from AMD with a large hardware order from Anthropic under separate parts of the agreement.
AMD ties financial incentives to chip deployments
The Anthropic agreement follows other AMD transactions that combine large infrastructure orders with financial arrangements involving customers. In October, AMD agreed to supply OpenAI with systems supporting up to six gigawatts of capacity and issued warrants that could allow the ChatGPT developer to acquire about 10% of the chipmaker.
The OpenAI warrants vest in stages as deployment, commercial, and share-price conditions are met. AMD said the agreement could generate tens of billions of dollars in revenue.
AMD reached another agreement with Meta in February covering up to six gigawatts of GPU capacity. That deal also included performance-based warrants tied to shipment and purchase targets.
The structures differ across the three agreements. AMD is making a direct investment in Anthropic, while the OpenAI and Meta agreements give those companies rights to acquire AMD shares when specified conditions are met.
Nvidia has also held talks about investing up to $30 billion in OpenAI, Reuters reported. Chip suppliers have paired investments or equity incentives with commercial agreements involving major AI developers.
AMD’s Anthropic investment is tied to the rollout of AMD-based infrastructure. However, the companies have not disclosed Anthropic’s purchase obligations, cancellation rights, the valuation attached to AMD’s investment, or the treatment of delayed deployments.
AMD said it has committed to make the strategic equity investment in the future, meaning the full amount has not yet been transferred. The company has not disclosed the investment timetable beyond its connection to deployment milestones.
Anthropic will deploy AMD Helios systems featuring MI455X GPUs from the Instinct MI450 series, EPYC “Venice” processors, Pensando networking, and ROCm software. The deployment builds on Anthropic’s existing use of AMD’s earlier MI355X accelerators.
The configuration covers an integrated rack-scale platform spanning accelerators, CPUs, networking, and software, rather than a standalone GPU purchase.
“We are thrilled to deepen our partnership with Anthropic and deploy AMD Helios at gigawatt scale,” AMD chair and chief executive Lisa Su said. “This collaboration brings together Anthropic’s leadership in frontier AI with the full strength of AMD high-performance computing.”
Su said the companies intend to establish Helios as a platform for future AI infrastructure deployments. Her comments reflect AMD’s effort to expand the use of its rack-scale systems among companies developing and operating large AI models.
The deployment adds Anthropic to AMD’s list of customers planning to use its next-generation accelerator platform at gigawatt scale. AMD has announced prospective commitments of up to six gigawatts each from OpenAI and Meta, and two gigawatts from Anthropic.
The two-gigawatt Anthropic commitment represents 2,000 megawatts of power capacity. AMD executives have said that developing one gigawatt of AI computing infrastructure can cost tens of billions of dollars, depending on the equipment and facilities involved.
Anthropic has separately secured more than 300 megawatts through SpaceX’s Colossus 1 facility in Memphis, which contains more than 220,000 Nvidia GPUs. At its full planned scale, the AMD deployment would represent more than six times the power capacity Anthropic obtained through Colossus 1, although the figures do not measure equivalent computing performance because the facilities use different hardware and deployment models.
Anthropic plans to install some of the AMD systems in its own data centres, according to The Wall Street Journal. Other systems are expected to be hosted by cloud providers and specialist AI infrastructure companies.
AMD and Anthropic are also seeking operators capable of hosting the planned systems. AMD chief executive Lisa Su said gigawatt-scale capacity requires planning between 12 and 24 months before deployment.
The Wall Street Journal reported that AMD had discussed providing financial guarantees for Anthropic’s future data-centre leases. The companies have not confirmed the proposed terms or disclosed the extent of AMD’s obligations under any guarantee.
Anthropic adds AMD to its multi-supplier compute network
The AMD systems will operate alongside infrastructure based on Nvidia GPUs, Amazon Trainium processors, and Google tensor processing units. Anthropic has continued to secure capacity through direct deployments, cloud agreements, and leased data-centre facilities.
Amazon remains Anthropic’s primary cloud and training partner. Anthropic has secured up to five gigawatts of capacity from the company and said it uses more than one million Trainium2 processors, with Trainium3 deployments planned during 2026.
Anthropic has also secured multiple gigawatts of next-generation TPU capacity from Google and Broadcom, with deployments expected to begin in 2027. Claude is available through Amazon Web Services, Google Cloud, and Microsoft’s Azure-based Foundry platform.
“Running across a diversified range of hardware lets us map the right workloads to the right hardware,” Anthropic co-founder and chief compute officer Tom Brown said. He said the AMD partnership would give Anthropic additional capacity while allowing the companies to optimise the systems used for training and serving Claude.
“Access to compute is central to keeping Claude at the frontier and meeting demand from our customers,” Brown said. “By partnering with AMD across the stack, we are securing the capacity we need and optimising it for training and serving Claude.”
The company has also agreed to use computing capacity at SpaceX’s Colossus facilities and has discussed leasing infrastructure from Meta. Reuters reported that the potential Meta agreement could be worth up to $10 billion over two years.
The AMD agreement also includes a multi-year engineering programme under which the companies will use Claude to optimise workloads for Instinct accelerators and support the development of AMD’s ROCm software platform. AMD plans to deploy Claude across its engineering and product-development teams.
The work gives Anthropic a role in adapting Claude workloads to AMD’s hardware and software environment. It also makes Anthropic both a customer of AMD’s infrastructure and a participant in the development of the software used to run workloads on that infrastructure.
AMD shares rose 2.4% after the agreement was announced. The company’s stock has more than doubled since the start of the year, outperforming the broader Philadelphia Semiconductor Index.
(Photo by Rubaitul Azad)
See also: SenseTime’s Galaxy Project targets domestic AI chip scale-up
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post AMD to invest up to $5 billion in Anthropic under AI infrastructure deal appeared first on AI News.
View the full article
SenseTime has launched the Galaxy Project, teaming with nearly 20 partners to scale domestic AI chip infrastructure in China.
In a keynote titled ‘Intelligent Transformation and Symbiosis,’ Yang Fan – the company’s co-founder and president of its Large Device Business Group – laid out what SenseTime describes as a closed loop connecting chip-level technology, ecosystem partnerships, and commercial deployment for domestically-produced AI computing power.
Alongside the Galaxy Project, SenseTime signed a space computing agreement with satellite manufacturer Guoxing Aerospace and struck a research partnership with five institutions – including the Shanghai Artificial Intelligence Laboratory – aimed at scientific computing applications.
Yang framed the timing around three converging trends: token demand climbing across enterprise deployments, industrial AI adoption catching up with consumer-facing use cases, and domestic chip commercialisation reaching a point where intelligent computing centres built on ******** silicon can be stood up at pace.
However, whether that window is as open as SenseTime claims depends heavily on numbers the company has not had independently verified.
Token throughput figures come with a large asterisk
SenseTime says its large-scale device platform now processes an average of 2.42 trillion tokens daily, and the company projects that figure will climb 25-fold to 10 trillion tokens per day by the fourth quarter of 2026. That’s a forecast, not a measured result, and enterprise buyers evaluating SenseTime’s infrastructure should treat it as such until quarterly figures start landing.
The cost-effectiveness claims attached to that growth are similarly self-reported. SenseTime says its heterogeneous hybrid inference technology delivers an 85–152 percent increase in Model FLOPs Utilisation on mainstream domestic chips, alongside inference cost-effectiveness the company puts at 1.25x that of Nvidia’s H-series parts.
Compared with domestic homogeneous inference setups, SenseTime claims a 2.5x increase in token output at equivalent cost, a jump it says pushes optimised hybrid inference clusters past what the industry previously regarded as the minimum profitability threshold for domestic computing power.
None of these figures come with third-party benchmarking, and the gap between a vendor’s optimised test cluster and a customer’s production environment – with its uneven data pipelines and delayed firmware updates – tends to be where such numbers soften.
Adaptability claims and the multi-chip problem
Domestic AI chips have historically struggled with a fragmented software stack: models trained for one architecture often require rework to run on another. SenseTime says it has built a full-stack adaptation layer spanning models, frameworks, operators, toolchains, and hardware to address that, with the aim of letting customers migrate workloads across domestic chip vendors without extensive rewrites.
The company points to two applied examples. In an AI4S long-sequence protein prediction workload, SenseTime says fused operator optimisation cut overall prediction time by a factor of three. In AIGC video generation, it claims a 93 percent multi-card parallel acceleration ratio for domestic chips running DiT models, alongside what it describes as zero-cost migration for mainstream AI development tools.
These are the kinds of figures that read well in a sandbox test and matter far more once they’re stress-tested against real customer pipelines running mixed hardware generations.
Energy metrics get a new benchmark name
SenseTime introduced a metric it calls Tokens Per Watt, positioned as a replacement yardstick for measuring AI data centre efficiency, alongside a Computing-Power Collaboration Agent that handles resource scheduling, electricity price prediction, and energy storage optimisation across what the company describes as an eight-level data system with five decision chains.
Combining compute, electricity pricing, and automated scheduling, SenseTime claims an 80 percent increase in token output per unit of electricity cost, average power prices 10 percent below comparable regional data centres, and 96 percent accuracy in computing load prediction.
These are claims worth watching over the next several quarters rather than accepting at face value. Electricity price arbitrage and load forecasting accuracy tend to perform differently once a system runs through a full seasonal cycle with genuine demand volatility, rather than the conditions under which a vendor typically runs its pilot.
Impressive partner roster spans chipmakers to component suppliers
The Galaxy Project’s stated ecosystem includes domestic chip vendors Cambricon, Muxi, Hygon, Huawei Ascend, Moore Threads, Sunrise, and Biren Technology, component partner Xizhi Technology, and infrastructure firms including Silicon Motion, Qujing Technology, Zhongke Jiahe, Qingcheng Jizhi, Sophon Information, and Jiliu Technology.
SenseTime says the plan covers construction of one “token factory,” five computing clusters at what it calls “10,000-calorie” scale, joint work across ten technology directions, and support for 200 AI startups.
“Domestic production is not simply about replacing individual chips, but rather a collaborative effort across the entire chain of China’s innovation capabilities, from chips and components to infrastructure and application scenarios,” Yang said.
Space, optical, and quantum computing bets look further out
Beyond near-term infrastructure, SenseTime outlined work on optical computing for data centre efficiency, quantum computing applications in AI optimisation, and a space computing partnership with Guoxing Aerospace to build what the two companies call the SenseTime Space Computing Constellation.
SenseTime’s plan calls for a first satellite launch in 2026, building toward thousands of computing satellites and computing capacity in the tens of thousands of petabytes by 2030.
Yang argued the value extends past raw capability, framing space-based computing as a way to extend the reach of ******** AI services into weak-network environments such as maritime operations and disaster response, and by extension to support China’s AI exports internationally.
That 2030 target sits five years out, and satellite computing deployments of this scale have no precedent to measure the timeline against.
Physical infrastructure spans Shanghai to Riyadh
On the ground, SenseTime says its Shanghai facility runs the country’s first data centre rated at what it calls “5A” intelligent computing level, handling over 20 trillion tokens daily across more than 20 industries. A Yancheng site has launched with an initial 3,000 petaflops of capacity focused on energy, manufacturing, and low-altitude economy applications.
In Hong Kong, SenseTime is building what it describes as the territory’s largest domestic intelligent computing centre, targeting 40,000 petaflops by 2030. The company also plans what it calls China’s first overseas domestic computing cluster in Saudi Arabia, positioned as a full-stack domestic computing base for the Middle East.
On the research side, SenseTime’s tie-up with the Shanghai AI Laboratory, Beijing Zhongguancun Academy, Shenzhen Hetao Academy, the Shanghai Algorithm Innovation Research Institute, and Shanghai Jiao Tong University’s AI school aims to build a shared platform spanning compute, tooling, and model capability for life sciences, materials science, and manufacturing research. Yang called AI for Science “a key lever for paradigm innovation in basic research,” tying the initiative to China’s broader “Artificial Intelligence+” policy push.
SenseTime’s forecast of 10 trillion tokens per day by Q4 2026 is the figure to track against whatever the company reports when that quarter actually closes.
See also: Kimi K3 open-weight model: China’s biggest AI is a bet on memory, not compute
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post SenseTime’s Galaxy Project targets domestic AI chip scale-up appeared first on AI News.
View the full article
Google has released Gemini 3.6 Flash and 3.5 Flash-Lite as new workhorses designed to cut latency and token costs for enterprise AI agents.
The economics of running autonomous software agents inside a production environment come down to a fixed equation few vendors advertise directly. A model needs to reason through a multi-step task competently, but every extra token it generates while doing so adds cost and delay to a workflow that might run thousands of times an hour.
Teams building background agents rather than chat interfaces need throughput first and parameter count second. Google’s answer, announced this week, splits that trade-off across three models: Gemini 3.6 Flash for coding and multimodal reasoning, Gemini 3.5 Flash-Lite for high-volume, low-latency work, and a restricted Gemini 3.5 Flash Cyber variant built for vulnerability remediation.
The math behind Gemini 3.6 Flash
Google’s developer documentation for 3.6 Flash centres on one figure: 17 percent fewer output tokens than the prior 3.5 Flash version, based on measurements from the Artificial Analysis Index.
In specific synthetic tests, including the Datacurve DeepSWE benchmark, Google reports drops in token usage of up to 65 percent. Pricing sits at $1.50/1M input tokens and $7.50/1M output tokens, positioning the model for reasoning loops that run continuously rather than on-demand.
On DeepSWE, the company records a 49 percent success rate for 3.6 Flash against 37 percent for its predecessor. On MLE Bench, the score moves from 49.7 percent to 63.9 percent, and on Google’s GDPval-AA v2 test – which attempts to measure real-world knowledge work rather than coding puzzles – 3.6 Flash scores 1421 against 1349 for the older model.
Figma, Hebbia, and Harvey put the model to work
Figma has integrated 3.6 Flash into its prototyping infrastructure, and according to Matt Colyer, the company’s Director of Product Engineering, the model gives developers a faster route through design iterations without a drop in output quality.
Legal technology platform Harvey and research tool Hebbia route data through the model for multimodal document work: ingesting raw financial filings, parsing document structure, reading embedded charts, and producing draft reports for review.
Google also folded a client-side computer-use tool directly into the Gemini API and Gemini Enterprise platforms, removing the custom intermediary software engineers previously built to let models operate on top of an operating system.
The company reports an OSWorld-Verified score of 83.0 percent, up from 78.4 percent, and says updated safeguards against chemical, biological, radiological, and nuclear misuse improve resistance to jailbreaking without raising refusal rates for benign requests.
A cheaper tier for high-volume background agents
Gemini 3.5 Flash-Lite targets a different job: document processing and agentic search running at volume rather than reasoning depth. The Artificial Analysis Index measured the model at 350 output tokens per second, the fastest in the 3.5 series according to Google.
Pricing runs at $0.3/1M input tokens and $2.5/1M output tokens, cheap enough that engineering teams can route simple, high-volume subagent requests to a minimal thinking level and reserve higher thinking levels for multi-step work.
On Google’s ****-MRCR v2 long-context test, Gemini 3.5 Flash-Lite recorded a 72.2 percent success rate against 60.1 percent for its predecessor, and its GDPval-AA v2 score nearly doubled, from 642 to 1140. The model carries the same native computer-use tool as 3.6 Flash.
Separately, Google says Gemini 3.5 Pro remains in partner testing ahead of a full release, and pre-training for the next Gemini 4 architecture is already underway.
Gemini 3.5 Flash Cyber: A restricted model for patching code
Automated vulnerability scanners now surface flaws faster than most security teams can patch them, and that gap is where Google positions Gemini 3.5 Flash Cyber.
The model is built to validate and remediate code vulnerabilities, and Google reports performance on the CyberGym benchmark competitive with frontier models (though it hasn’t made those figures public in the same detail as its consumer-facing releases.)
Distribution stays restricted to governments and vetted partners through a pilot programme, a limitation Google frames as a safeguard against the model generating exploit code for offensive use.
Inside Google’s CodeMender security agent, multiple instances of 3.5 Flash Cyber run in parallel, cross-checking one another’s findings before producing a single remediation report a human reviewer signs off on.
Engineering teams seeking to integrate these new models can access them through the Gemini API via Google AI Studio, Android Studio, or the Gemini Enterprise Agent Platform. Consumers can also access the new models in the Gemini app and 3.5 Flash-Lite is also rolling out in Google Search.
See also: Bristol Myers Squibb buys Nvidia AI system for drug discovery
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Google’s Gemini 3.6 Flash targets enterprise agent token costs appeared first on AI News.
View the full article
You open a generative AI tool expecting a quick boost. Ten minutes later, you’re still there, refining a prompt for the fourth time. The task you started with has drifted off to the side somewhere.
Sound familiar? Knowledge workers in 2026 are running into this more and more. It makes sense once you look at how these tools are built. They’re designed for efficiency, sure. But they’re also designed to keep you in the room. Those two goals don’t always play nice together.
Demanding cognitive tasks need stretches of uninterrupted thought, not constant back and forth with a chatbot that always has one more suggestion. That’s not just distracting in the obvious sense. It’s baked into the interface on purpose. These systems reward you for sticking around, not for finishing up and closing the tab.
Generative Interfaces Reward Engagement Over Closure
Let’s be honest, most modern AI platforms care a great deal about how long you stay logged in. That’s not a conspiracy theory. It’s just the business model.
Recommendation logic and conversational flows lean toward responses that feel a little useful, or emotionally satisfying, because that keeps you typing another message. A 2026 review examining AI deployment in digital media described these platforms as being “mathematically optimized to maximize ‘time on site,'” noting that emotionally resonant content tends to beat plain, straightforward material. Generative tools turn that dial up, since they can produce tailored variations instantly and at almost no cost.
What you end up with is something close to a variable reward loop, the kind attention researchers have studied for years around slot machines and social feeds. Every refined response gives just enough of a win to make staying worthwhile. Not a huge win. Just enough.
That’s the trap. The cognitive toll builds quietly while you feel productive. Before long, the block of time you’d set aside for deep work has been nibbled down to nothing.
When those loops start chewing into concentration, professionals draw lines in the sand. A dependable site blocker helps here, setting firm guardrails around distracting tabs and feeds so the uninterrupted stretches high quality work requires don’t get quietly whittled away.
Productivity Figures Hide Real World Attention Friction
On paper, the numbers look great. In certain domains, anyway.
Analyses published in MIT Technology Review this year pointed to roughly 14 percent gains in customer service and 26 percent in software development. Returns get thinner fast in judgment-heavy work, the kind that leans on nuance rather than repeatable steps.
Zoom out to the organizational level and the picture gets murkier. The Stanford AI Index for 2026 shows adoption sitting at 88 percent, with industry responsible for most frontier models released the year before. Impressive, at least on the surface.
Real world deployment tracking tells a different story, though. Coverage in The New York Times pointed to studies where these tools “didn’t reduce work, they consistently intensified it,” creating more workload rather than freeing anyone up. Not exactly the narrative you’d expect from the headlines.
That gap between conference announcements and what actually happens on a Tuesday afternoon in an open office keeps shaping how teams weigh AI’s real value.
Clear Signals That Generative Loops Are Fragmenting Focus
You don’t need a research team to notice this happening. A few signs tend to show up again and again.
Opening an AI chat for a thirty second clarification, only to find yourself six exchanges deep. Timelines stretching because every output needs a couple more rounds of correction before it’s usable. Notifications and fresh suggestions creeping in and derailing whatever train of thought you were riding.
Then there’s the exhaustion. Finishing a session feeling wiped out, even though barely any real synthesis happened. And colleagues mentioning the same scattered feeling in meetings, like it’s suddenly a shared experience across the whole floor.
None of these signs are dramatic alone. Together, they paint a clear picture of design incentives favoring continued interaction over clean completion.
The Iterative Reality of Collaborative AI Use
Early expectations painted a picture of seamless automation, the kind where you ask once and get exactly what you need.
Reality is messier. Enterprise usage patterns show people spending a surprising chunk of their day querying, correcting and re-querying, tweaking outputs bit by bit until they’re finally usable. Early narratives promised full workplace automation, the kind that would hand back hours of your day. What’s actually happening looks pretty different.
Recent data shows workers spending real, measurable time manually refining what these systems hand them. An analysis of Anthropic’s enterprise usage metrics makes this pretty clear. Collaborative AI, in practice, involves constant, disruptive micro-iterations, the kind that quietly drain cognitive energy long before anyone notices the drain.
It’s not nothing, this back and forth. But it keeps people tethered to the tool in a way that fragments the longer stretches of thinking harder problems actually demand.
This mirrors attention economy mechanics already observed across other digital platforms, just wearing a different outfit. Every response that invites one more tweak adds a little cognitive drag. String enough of those together across a workday and the toll adds up fast, particularly for anyone doing work that requires holding multiple threads in their head at once.
Protecting Focus in an AI Driven Workplace
The teams handling this well aren’t leaving attention to chance. They treat it as an actual resource, something to budget and protect rather than assume.
That usually means batching AI assisted tasks into set windows, putting firm limits on session length and keeping core deep work hours fenced off from ambient digital noise. Coverage from Harvard Business Review on adoption trends backs this up, noting that efficiency gains at one level of an organization often create coordination headaches somewhere else. That only strengthens the case for deliberate boundaries.
None of this makes the technology itself the enemy. The same generative capabilities that can splinter your attention are also genuinely great at speeding up targeted subtasks, provided you’re setting the pace instead of letting the feed set it for you.
As adoption keeps climbing through the rest of 2026, the advantage will land with people who bother to design their own cognitive environment instead of accepting whatever rhythm the tools default to.
So, where does your attention actually go on a normal day? Worth tracking for a week, just to see. A handful of well placed guardrails, paired with tools that respect your time, can keep AI in its lane, helpful, targeted and quiet when it needs to be.
The post The AI Slot Machine Effect: Why Generative Feeds Disrupt Deep Work And How to Reclaim Focus appeared first on AI News.
View the full article
Bristol Myers Squibb is purchasing an Nvidia DGX SuperPOD built on the chipmaker’s Vera Rubin architecture to support artificial intelligence use across its drug discovery and development operations.
The pharmaceutical company said it will be the first life sciences group to acquire a DGX SuperPOD based on Vera Rubin. Nvidia introduced the architecture earlier this year as the successor to its current generation of AI computing systems.
Expanding computing capacity
The new cluster will comprise eight DGX Vera Rubin NVL72 systems, with each rack-scale system combining Nvidia Vera central processing units and Rubin graphics processing units.
BMS will use the infrastructure to train proprietary models and run predictions across its research programmes. The system will support work involving compounds, proteins, and other scientific data.
Financial terms were not disclosed. The purchase expands BMS’s existing Nvidia infrastructure, which includes an older SuperPOD that company executives described as two or three generations behind Vera Rubin.
BMS has operated its existing DGX SuperPOD for about three years. The company plans to combine it with the Vera Rubin system in a shared computing environment accessible from its research sites worldwide.
The SuperPOD software stack can schedule training, prediction, and development workloads across the infrastructure. BMS said the expanded environment will give more scientists direct access to its computing resources.
Greg Meyers, BMS’s chief digital and technology officer, said computing requirements have increased as the company deploys larger AI models across its research organisation.
Erin Davis, vice president of research business insights and technology at BMS, said the existing infrastructure is operating at capacity. She attributed the demand to large-scale predictions involving large molecules and the development of internal foundation models.
Davis said the new system will not be limited to a small group of computational researchers. BMS plans to make it available across the research organisation without the waiting periods and access limits associated with its current infrastructure.
Applying AI in drug discovery
BMS said AI informs the design of every small-molecule programme and the majority of its large-molecule programmes. The technology is applied to target identification, lead optimisation, large-molecule predictions, and internal model development.
The company said AI-enabled target identification has reduced some manual research work by several weeks. Large-molecule prediction workloads are also contributing to demand for additional graphics processing capacity.
Robert Plenge, BMS’s chief research officer, said the new system will allow scientists to evaluate more potential drug candidates during the early stages of development.
“Maybe before we could do 10 and now we can do dozens,” Plenge said.
Computational screening allows researchers to assess potential compounds before selecting a smaller group for synthesis and laboratory testing.
BMS applies this approach through a method it calls “Predict First,” which uses model-generated predictions to exclude molecules that do not meet the required properties before candidates are selected for synthesis.
Payal Sheth, senior vice president of therapeutic discovery sciences at BMS, said researchers use the predictions to identify molecules with the required combination of properties.
“We use predictions as a way to prioritise synthesis of molecules with multi parameter optimisation,” Sheth said. “This ensures precious laboratory experiments are aligned with progressing molecules that have the highest probability of success.”
The method narrows the number of compounds sent for laboratory testing, allowing researchers to focus experiments on molecules that meet a programme’s predicted requirements.
BMS has also used AI to expand its library of CELMoD compounds, which are engineered to selectively degrade *******-causing proteins. The company is studying the compounds in blood cancers and other diseases.
BMS said the modelling work helped researchers examine additional protein targets and potential compounds before deciding which candidates to pursue experimentally.
The company is also using AI tools to shorten the time required to produce medicines for clinical trials. Plenge said the process has already been reduced by between 20% and 30% and could reach 50% in the coming years.
He cited an experimental sickle cell disease treatment in early clinical development as one example of AI-supported research. Plenge said the treatment probably would not have been discovered without the company’s AI tools.
The figures refer to the time required to identify and produce candidates for clinical testing rather than their subsequent performance in trials.
The Vera Rubin system will also give researchers access to Nvidia’s BioNeMo Agent Toolkit for biological and drug-discovery applications.
BioNeMo provides tools for protein-structure prediction, molecular generation, molecular docking, sequence analysis, and genomics. It can also connect several computational tools within the same research workflow.
BMS executives said human researchers will continue to review model outputs and decide which compounds or programmes should advance.
Connecting research sites
BMS is introducing tools intended to reduce the specialist knowledge required to initiate complex computing tasks. The company said researchers will be able to start some prediction requests using natural-language instructions.
The environment will be managed through Nvidia Mission Control, whose functions include cluster provisioning, infrastructure monitoring, and workload management, according to BMS.
The unified infrastructure will allow data and model outputs generated at one site to be used by teams elsewhere. BMS said datasets from a programme in Lawrenceville, New Jersey, for example, can be incorporated into models used by researchers in San Diego.
Sheth said the shared environment is intended to retain information from experiments and research programmes across the organisation.
“The compute infrastructure is what connects all of our scientists together and ensures that our learnings are institutionalised,” Sheth said.
The two SuperPODs will operate through a common data environment, allowing teams at different sites to access shared datasets and model outputs. BMS said the environment will include information from experiments, clinical readouts, and research partnerships.
The company plans to allocate the new computing capacity across small- and large-molecule design, clinical research, and digital-twin applications. BMS did not provide details about the planned digital-twin work or the amount of capacity assigned to each area.
Meyers said the Vera Rubin system will provide more computing capacity relative to its electricity use. BMS and Nvidia said the eight-system cluster will deliver up to 10 times the performance per megawatt of the infrastructure it replaces.
“When you host these things, you have to pay an electric bill,” Meyers said. “Think of it as 10 times more compute capacity per watt spent … Electricity is not getting cheaper.”
BMS did not provide a specific deployment date or identify where the new system will be hosted.
(Photo by Chidera Faustina Okeke)
See also: US public health agencies to test OpenAI and Anthropic AI models
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Bristol Myers Squibb buys Nvidia AI system for drug discovery appeared first on AI News.
View the full article
Privacy Notice: We utilize cookies to optimize your browsing experience and analyze website traffic. By consenting, you acknowledge and agree to our Cookie Policy, ensuring your privacy preferences are respected.
