ChatGPT

Enterprises evaluating ******** open-weight models this month face a question that has nothing to do with benchmarks: whether using one will still be straightforward in a year. Moonshot AI’s Kimi K3 arrived on July 16 as the largest open-weight model yet released, and within days it had reopened a policy argument in Washington that had been dormant for a year. The outcome will affect procurement decisions well outside the United States, because the mechanisms under discussion–federal procurement rules, export blacklists, security advisories–travel through the same cloud providers that serve most of the world. The immediate trigger was a post by Dean W. Ball, OpenAI’s head of strategic futures and until recently a senior AI adviser in the Trump White House. His assessment of the model was largely positive: a very good model, he wrote, whose performance he did not think could be explained away by distillation. He also observed that it seemed “very token hungry,” and that it was not obvious to him that it is actually cheap to run, a useful caution, given K3 launches with maximum reasoning effort as its only setting and bills output at $15 per million tokens. Then he predicted that the Trump administration would eventually decide its best strategy was to create regulatory risk around ******** open-weight models. Not a ban, which he called one of the dumber motifs in AI policy, but soft guidance from agencies suggesting such models may contain backdoors. “It needn’t be that well justified,” he wrote. Enough uncertainty, and regulated enterprises retreat on their own. Why ******** open-weight models are a commercial problem first The reaction was fierce, and it came from Americans rather than from Beijing. David Sacks, co-chair of the President’s Council of Advisors on Science and Technology, said he could not tell whether Ball was confessing to a regulatory capture strategy or predicting one, and that either way, weaponising regulatory uncertainty as a competitive tool should be unacceptable. He added that the leading closed labs, already a duopoly in model revenue, want the government to remove their open-source competition. Yann LeCun and Martin Casado argued that open and proprietary development can coexist. Ball later clarified that he had been forecasting rather than recommending, and walked back the claim that open weights necessarily slow the field down. Underneath the personalities is an arithmetic problem. Closed labs need revenue per token to justify the capital they are raising for data centres, and cheaper open-weight models compress that revenue without reducing how much AI gets used, the point Snorkel AI co-founder Braden Hancock put to TechCrunch. The routing data already shows the shift: open-weight models handled 29% of tokens through Vercel’s production gateway in June, up from roughly a ninth in April, while accounting for under 4% of spending. That pressure is arriving from inside the American stack. GitHub made Moonshot’s Kimi K2.7 Code generally availablein the Copilot model picker on July 1, hosted on Microsoft Azure. The Information reports Microsoft is now adding K3 to Azure and evaluating whether it can run Copilot features currently handled by OpenAI and Anthropic models, with potential inference savings of up to $600 million. Microsoft has confirmed neither the figure nor which features. It is an evaluation, not a deployment, but it is the largest customer of both American frontier labs, pricing the alternative. The security argument, taken seriously Commercial motive does not make the security concern fake, and the strongest version of it deserves stating. Open weights cannot be recalled. Once a model is downloaded and running inside thousands of organisations, no vendor can patch it, revoke it, or push a fix, which is a materially different risk profile from a hosted API. Model behaviour is harder to audit than model code: a fine-tune can carry biases or failure modes that no licence inspection would reveal. NIST has previously found security vulnerabilities in DeepSeek’s open models, and for regulated industries, questions about training data provenance and content handling are live regardless of where a model was built. The counterargument is about proportionality rather than dismissal. Georgetown research fellow Sam Bresnick has argued that halting Nvidia H200 sales to China would slow Beijing considerably more than banning open models Americans want to use, targeting the input rather than the output. And Ball himself conceded a version of this in his second observation, attributing China’s open-weight strategy partly to a lack of domestic compute for serving customers, which would make it an unintended byproduct of US export controls in the first place. What is actually likely to happen. Axios reported on July 20, citing people close to the administration, that Commerce last year weighed adding ******** AI labs to the Entity List, that the NSA and the Office of the National Cyber Director considered issuing an advisory on ******** AI lab threats, and that the White House considered an executive order making US companies liable for breaches if they used ******** models. Officials concerned about stifling innovation killed all of it. With adviser Sriram Krishnan gone and security hawks louder, the effort has revived, but the described approach is procurement rules, Entity List threats and public pressure rather than prohibition. “What’s actually happening is slower and more durable,” one source told Axios. Neither the White House nor Commerce responded to Axios’s requests for comment, and Politico reports Commerce will not move imminently. For buyers outside the US, the exposure is indirect but real. A rule written for American regulated industries and federal procurement does not bind a Malaysian bank or an Indonesian telco. The hyperscalers are the transmission line. Most enterprises in this region reach Kimi K3 through Azure, AWS or Google Cloud rather than Moonshot’s own API, and if Washington makes hosting ******** open-weight models uncomfortable enough for those providers, the model quietly leaves the catalogue in Kuala Lumpur at the same time it leaves it in Virginia. Ball anticipated this in his own post, noting that regulators would not want to push so hard that hyperscalers stop serving ******** models altogether, since that would only drive startups toward less reputable providers. The obvious hedge is to hold your own copy. Moonshot publishes K3’s weights on July 27, and from that point the model cannot be withdrawn from anyone who has downloaded it. But as covered previously, K3 is a difficult model to self-host: Moonshot recommends serving it across 64 or more accelerators, and the weights alone come to roughly 1.4TB. For most companies, the fallback is theoretical. That leaves a narrower question than the headlines imply. Not whether ******** open-weight models are safe or permitted, but whether the specific model you build on will still be in your cloud provider’s catalogue in twelve months, and what it would cost you to move if it isn’t. That is a due-diligence question, and it is answerable today. See more: 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 ******** open-weight models are cheap. Washington is deciding what that costs. appeared first on AI News. View the full article

Public health departments across the United States will test generative AI tools under a new programme involving the Coalition for Health AI, OpenAI, Anthropic, and Accenture. The Public Health Use Case and Learning Scaling Engine, known as PULSE, will support trials in 10 state, local, tribal, or territorial jurisdictions. The programme is intended to produce implementation guidance for public health agencies considering similar deployments. OpenAI and Anthropic have donated 10 enterprise licences with capacity for up to 2,000 public health practitioners. Accenture will oversee participant onboarding and help develop playbooks based on the trials. The programme will give public health practitioners access to enterprise AI products from OpenAI and Anthropic. CHAI has not identified the products, model versions, or configurations that will be used, or explained how the two providers will be assigned across the pilots. “Every major technological transformation succeeds or fails based on trust, governance and execution,” Dr. David Lakey, a former Texas health commissioner, said. “PULSE will support agencies in this endeavour, and is specifically designed for practical implementation.” Five public health use cases The Coalition for Health AI’s leadership council will select the participating jurisdictions and assign practitioners to communities focused on five use cases. These include biosurveillance and drug-wave prediction, social determinants of health and SDoH mapping, operations and efficiency and community-feedback analysis, public communications and a multilingual translation hub, and automated clinical-data retrieval and a FHIR query engine. NIST’s AI Risk Management Framework recommends evaluating AI systems according to their intended use, operating environment, affected parties, and potential consequences. CHAI has not published separate evaluation, privacy, security, or human-review requirements for the five PULSE use cases. FHIR is an HL7 standard used to exchange healthcare information electronically between compatible systems. PULSE includes an automated clinical-data retrieval and FHIR query engine, although the announcement does not define the role of generative AI in that workflow. It does not state whether the models will generate queries, retrieve records, summarise returned information, or combine those functions. The announcement also does not explain whether incorrect queries, incomplete retrievals, or unsupported summaries will be checked by staff before the information is used. Some proposed applications could involve demographic, geographic, clinical, or population-health information. CHAI has not said whether the pilots will use identifiable records, de-identified information, synthetic data, or aggregated datasets. The US Department of Health and Human Services requires organisations covered by HIPAA to protect certain electronic health information. Its cloud-computing guidance states that regulated organisations and service providers must meet HIPAA requirements when cloud systems create, receive, maintain, or transmit electronic protected health information. HIPAA will not apply to every participating organisation or workflow. Its application will depend on the agency, the data involved, and the function being performed. OpenAI states that inputs and outputs from its business services, including ChatGPT Enterprise and its API, are not used to train or improve its models by default. Anthropic also says it does not use inputs and outputs from its commercial products for model training by default. Those provider policies do not define how the PULSE deployments will be configured. The announcement does not set out retention periods, access controls, audit arrangements, data-storage requirements, or rules governing the submission of protected health information. “We believe AI should be useful, safe and accessible to the people tackling society’s most important challenges,” Felipe Millon, OpenAI’s head of government go-to-market, said. He added that the licences were intended to help public health organisations evaluate the tools through a structured process. Governance and evaluation remain undefined The pilots are scheduled to begin in autumn 2026. CHAI expects the resulting playbooks to be released in 2027 and used as reference material by other public health agencies. CHAI has not published the measures that will be used to assess the pilots or explained whether each use case will be evaluated under separate technical, operational, privacy, and safety criteria. The announcement also does not explain how model outputs will be reviewed. It does not state whether staff must approve generated public communications, verify translations, validate retrieved clinical information, or review biosurveillance and drug-wave outputs before they are used. NIST guidance recommends identifying which AI functions require human oversight and training users to understand system performance and limitations. Its generative AI guidance also covers testing, validation, monitoring, documentation, privacy, and management oversight. “Public health teams are being asked to do more with less, and AI can help — as long as it’s brought in with care and the right guardrails,” Elizabeth Kelly, Anthropic’s head of beneficial deployments, said. She said PULSE would allow practitioners to test the tools in their own environments with privacy, governance, and responsible-use measures incorporated from the start. Data from the National Association of County and City Health Officials cited by CHAI showed that nearly 40% of local health departments were not using AI. The coalition said some departments were interested in revising workflows and improving operational efficiency. PULSE will provide enterprise licences, onboarding, peer communities, and implementation playbooks. CHAI has not specified the minimum staffing, infrastructure, interoperability, or cybersecurity requirements for participating jurisdictions. Eligible participants include state and territorial health departments, county and municipal agencies, tribal authorities, Indian health organisations, and large city health departments. PULSE plans to convert findings from 10 jurisdictions into guidance for wider use. The announcement does not explain how the playbooks will account for differences in agency size, technical systems, legal responsibilities, staffing, or procurement arrangements. The announcement also does not explain whether outputs from biosurveillance, drug-wave prediction, or clinical-data retrieval will be used only for testing, presented to staff for review, or incorporated into operational workflows. PULSE forms part of the Coalition for Health AI’s broader work on governance standards for healthcare AI. In May, the organisation announced plans to develop guidance covering eight governance areas through workshops and working groups involving more than 150 healthcare AI representatives. The coalition has since started publishing playbooks on organisational AI policies, governance structures, and internal resources. Additional guidance is expected to cover other areas of healthcare AI management. Separately, CHAI has worked with the Joint Commission on governance playbooks aligned with its voluntary Responsible Use of AI in Healthcare certification. The PULSE announcement does not state that participating public health agencies will be assessed under that certification. Dr. Brian Anderson, chief executive of the Coalition for Health AI, said public health agencies entered the COVID-19 pandemic after years of limited investment in technology. He said the PULSE programme was intended to give agencies practical experience with AI before wider implementation. “We know AI is going to reshape how we deliver public health — the question is whether we do it thoughtfully or not,” Dr. Ashish Jha, a former White House COVID-19 response coordinator, said. He said the programme would test which applications work and document the findings for other agencies. See also: Bunkerhill raises $55M to scale agentic AI across health systems 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 US public health agencies to test OpenAI and Anthropic AI models appeared first on AI News. View the full article

Moonshot AI’s Kimi K3 open-weight model has been read almost entirely through its parameter count since it launchedon July 16. At 2.8 trillion parameters, it is the largest open-weight model released to date. Model sizes are usually grouped into rough brackets, and 2.8 trillion rounds into what the industry calls the 3T class. A tier no openly available model had entered before. Moonshot’s flagship jumps from just over 1T to 2.8T in a single release, clearing DeepSeek’s 1.6T V4 Pro by a wide margin. The parameter count is the part of K3 that has drawn the most attention — and explains the least about how it runs. Source: Kimi The natural conclusion is that Moonshot has engineered its way around US compute restrictions. The company’s own technical blog suggests something more specific: K3 does not avoid the constraint so much as relocate it, trading compute for memory at almost every layer of the design. That trade is worth understanding, because compute and memory are not interchangeable constraints, and they are not equally available to a ******** lab. Introducing Kimi K3: Open Frontier Intelligence 2.8 Trillion Parameters, 1 Million Context, Native Multimodal Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts Attention Residuals deliver ~25% higher training efficiency at <2% additional… pic.twitter.com/eFHEbdxn3P — Kimi.ai (@Kimi_Moonshot) July 16, 2026 Why the Kimi K3 open-weight model is a memory problem Two different things determine what it costs to run a large model. One is how much calculation the machine does to produce each word. The other is how much of the model has to be held ready and instantly reachable the entire time it is working. The first is compute. The second is memory. Chip export controls have squeezed China hard at first, and Moonshot’s design reads as a sustained attempt to spend less of it. The main move is a technique called mixture-of-experts. Rather than run the whole model for every word, K3 splits itself into 896 specialised sections and calls on just 16 of them at a time, about 1.8% of the total. The calculation per word drops sharply. The memory bill does not move at all, because all 2.8 trillion parameters still have to sit loaded and ready in case they are the ones called next. So Moonshot went after that bill directly. It trained K3 to work at four bits of precision per parameter instead of the usual sixteen, a method known as quantisation-aware training, which the company applied from the fine-tuning stage onward and says it chose “for broad hardware compatibility”, a phrase worth pausing on, since it reads as a hedge against running on silicon that is not Nvidia’s. The savings are substantial. Independent analysis of the release puts the model at roughly 1.4TB in that format, against the 5.6TB it would need at full precision. The second change, Kimi Delta Attention, targets a different memory cost. As a model works through a very long document, it accumulates a running store of everything it has already read. At K3’s advertised limit of a million tokens, a few thousand pages, that store, not the model itself, becomes the biggest thing in memory. Moonshot is unusually direct about the commercial stake here. It contributed caching code to the open-source serving project vLLM, and says this combination is what lets it price K3 competitively despite the model’s size. Moonshot recommends running K3 across 64 or more accelerators wired together closely enough to behave as one pool. That is the same approach behind Huawei’s CloudMatrix systems, and it points to what the real workaround is. Memory can be gathered up across a large number of individually unremarkable chips. Training-grade compute cannot be assembled the same way. Whether that pooling happens on ******** silicon is a question the blog does not answer. Moonshot’s chip-level tests ran on Nvidia H200S and on what it describes only as a “GPGPU from an alternative vendor,” which it declines to name. Other results are benchmarked on an Nvidia L20, the cut-down card sold into China under export rules. The blog does not say where the H200 hardware sits, and the US House passed a bill in January to close the offshore cloud rental loophole that had let ******** firms reach restricted accelerators remotely. The reason any of this matters is that memory, not processing power, is where China’s own chip industry is furthest behind. In trade talks in August 2025, Beijing asked for relief on high-bandwidth memory restrictions rather than on lithography tools or TSMC access, a fair signal of what officials think is actually binding. Domestic output of that memory is projected at around two million stacks this year, enough for roughly 250,000 to 300,000 Huawei Ascend 910C-class chips, while SMIC has wafer capacity for more than a million. What enterprises can actually deploy For businesses in this region, the practical question is not whether K3 tops a leaderboard. It is whether an open-weight model at this size is deployable at all. Asian enterprises reach for open weights for three reasons–price, data sovereignty and regional-language coverage–and banks and insurers across Southeast Asia have been piloting self-hosted open models specifically so records never leave their own systems. The weights land on July 27, and any organisation that can afford the hardware will be free to download, modify and run K3 inside its own walls. The question is how many can. Moonshot recommends serving the model across 64 or more accelerators wired together as a single pool, and the weights alone come to roughly 1.4TB in the format it ships in, based on independent analysis, before the memory needed to work through a long document. That is a data-centre commitment, not a server-room one. For most enterprises, the practical outcome is renting dedicated capacity rather than owning it. That still keeps data in-country and under contract, which is what most regional regulators are asking for. What it does not deliver is the independence from infrastructure providers that drew many of these buyers to open weights in the first place. The software is not ready either. K3’s two main architectural changes are new enough that the standard open-source tools for running models do not yet support them, and Moonshot says it is working with inference partners and open-source maintainers to align technical details before release. Teams planning a self-hosted deployment should treat the launch date and the usable date as different things. The price has also moved. K3 costs $3 per million input tokens, dropping to $0.30 when the model has recently seen that input, and $15 per million output tokens. That is well under Fable 5’s $50 for output, but far above z.ai’s GLM-5.2 at $4.40 and DeepSeek V4 at $0.87. K3 is no longer in the budget tier its predecessors occupied. It also launches with maximum reasoning effort as the only setting, with lower modes to follow, so long reasoning chains and retried steps add up quickly. Companies should budget on the cost of a completed task, not the list price. What is claimed, and what is verified Arena placed K3 first in its Frontend Code evaluation at 1,679 points, ahead of Fable 5, in blind developer testing, as reported by Tom’s Hardware. That result is real. It is also a benchmark in one domain. Moonshot itself is more restrained than its headlines. The company states that K3’s overall performance still trails Claude Fable 5 and GPT 5.6 Sol, and lists three limitations: generation quality can become highly unstable if a harness fails to pass back historical thinking content, the model may make unexpected decisions on a user’s behalf when intent is ambiguous, and it shows a noticeable user-experience gap against Fable 5 and GPT 5.6 Sol. Its own footnotes also disclose that Fable 5 hit fallbacks on 35% of the tasks in Moonshot’s SWE Marathon evaluation, which may have affected that model’s measured score. Everything else remains a first-party claim. No published K3 number can be independently verified until the weights are public. Bank of America analysts led by Alex Liu said in a note that K3 shows large-scale pre-training combined with architectural work can still deliver step-change gains for flagship ******** models despite compute constraints, which is the sober version of the argument, and closer to what the blog supports. The direction of travel is not in dispute. Open-weight models handled 29% of all tokens routed through Vercel’s production gateway in June, up from about a ninth of volume in April, while accounting for under 4% of spending. July 27 is when we find out how much of K3 belongs in that column. 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 Kimi K3 open-weight model: China’s biggest AI is a bet on memory, not compute appeared first on AI News. View the full article

Bunkerhill Health has raised $55 million to scale its agentic AI platform, Carebricks. The closing of the company’s Series B round, announced today, folds in continued participation from Sequoia Capital, Felicis, Optum Ventures, and Y Combinator. However, a funding total doesn’t answer the key question any hospital executive wants to know about healthcare AI: does the software run inside a working hospital? That question is among the reasons why Khosla Ventures put its name on this deal. Healthcare organisations have put no shortage of funding behind machine learning pilots that perform well in a research setting and then never touch a live patient chart. Bunkerhill’s argument to investors, and to the health systems already paying for it, is that Carebricks closes the space between a model that works in a sandbox and one that runs against live clinical data at institutional scale. The backdrop is a spending number and a staffing problem. US healthcare spending hit $5.3 trillion in 2024, according to the Centers for Medicare & Medicaid Services, and labour shortages continue to strain providers nationwide. Bunkerhill frames the opportunity around a gap between what health systems want to do for patients and what their staff have time to execute. Decades of investment went into documentation systems meant to ease the burden on clinicians. Bunkerhill’s bet is that the next round of technology spending goes toward software that acts on ideas clinicians already have, rather than just recording them. “Medicine has advanced faster than our healthcare system’s ability to operationalise it,” explained Nishith Khandwala, Co-Founder and CEO of Bunkerhill Health. “Every leading health system has more opportunities to improve patient outcomes than its workforce has capacity to address. We believe AI agents can help them turn more of those ideas into reality.” Carebricks lets hospitals build their own agents rather than buying a fixed product off the shelf. Some agents review cardiology imaging for early signs of heart disease and flag patients who need follow-up care. Others handle prior authorisations or keep registry data current, and the range extends into administrative work that rarely gets attention in AI pitches but consumes staff hours every week. Cleveland Clinic, the University of Texas Medical Branch, and Intermountain Health all run the platform today. Vinod Khosla, Founder of Khosla Ventures, said: “The bottleneck in healthcare AI was never the technology, it was getting a health system to actually run it. Bunkerhill closed that gap. They made it much, much easier to adopt AI and already have traction inside critical health systems that would take most companies years to earn.” Twenty AI agents running inside one hospital system UTMB offers the clearest picture of what “running it” looks like once the pilot label comes off. The system now has more than 20 agents live on Carebricks, spanning clinical care, operations, and administration, according to Dr. Peter McCaffrey, UTMB’s Chief AI Officer. In its first month running at UTMB, a coronary calcium detection agent built on an FDA-cleared algorithm flagged a patient as being at imminent risk of a heart attack. Cardiology confirmed the risk and performed a triple bypass. UTMB’s care team credits the early detection with saving the patient’s life. It’s a single case, not a controlled trial, and Bunkerhill has published no data on how often the agent produces false positives or how it performs across a broader patient population over time. Other figures at UTMB come from Bunkerhill and the health system itself rather than independent audit. A nephrology triage agent now prioritises patients by severity, escalating urgent cases and routing others to telemedicine, and UTMB reports this has cut average specialist wait times by more than 50 percent. A lung nodule agent tracks incidental findings on CT scans through to the correct follow-up, with UTMB citing an 80 percent faster response on urgent cases and a doubling of guideline-concordant follow-up alongside a drop in manual coordinator work. These are health system-reported operational results from live production use, not synthetic benchmark scores, which matters. It also means the numbers reflect one institution’s data conditions and staffing setup, not a guarantee that another hospital will see the same curve. “We’ve already seen tremendous impact on patient care, and we’re only at the beginning of what becomes possible when a health system can operate with agentic AI at this scale,” McCaffrey said. Of course, none of this removes the work health systems still have to do themselves. Bunkerhill says it will use the new funding to expand Carebricks into a wider range of clinical and operational use cases while building out governance, monitoring, and safeguards. A platform that lets a nephrology department build its own triage agent also means that department owns the consequences of how that agent is tuned. Health system boards weighing a Carebricks-style deployment need answers on liability assignment, monitoring cadence, and what happens when an agent’s judgment and a clinician’s judgment disagree, before signing off on scale. UTMB’s 20-agent footprint gives Bunkerhill a reference case few competitors can currently match. Whether that number holds up as a signal for the rest of the industry depends on how UTMB, and the other systems now running Carebricks, handle the governance side as the agent count climbs. 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 Bunkerhill raises $55M to scale agentic AI across health systems appeared first on AI News. View the full article

Google DeepMind and Isomorphic Labs outlined a bioresilience program to curb AI misuse in biology while aiding outbreak response. The two organisations published an update on a joint initiative that began quietly and has now built out more than 15 partnerships with government bodies, biosecurity organisations, and research groups over the past 12 months. The disclosure arrives with a specific framing problem attached. Frontier models such as Gemini carry an increasingly detailed grasp of biology, and DeepMind acknowledges that pairing these systems with specialised biology models, agents like its Antigravity platform, and third-party databases will only sharpen that capability further. However, the same knowledge that helps a researcher map a vaccine target could, in principle, help a threat actor close gaps in their own understanding. DeepMind and Isomorphic describe this as a dual mandate: enable the scientific advances frontier AI makes possible, while keeping those same tools out of the hands of people who’d misuse them. The program sits on three pillars, according to the companies: preventing misuse, detecting outbreaks faster, and responding once an outbreak or attack is underway. The 15-plus partnerships built over the last year touch all three, though the update gives limited detail on which organisations are involved beyond a handful of named collaborators, including Lawrence Livermore National Laboratory, the *** AI Security Institute, CEPI, and the Francis Crick Institute. DeepMind says it intends to widen these relationships over the next six to twelve months, with attention turning to threat intelligence, evaluation methods for AI agents, and jailbreak mitigations. It’s also coordinating with the Frontier Model Forum on questions such as how to handle riskier categories of training data, virology datasets being the example given. Locking down Gemini without blocking legitimate science The prevention work rests on threat modelling designed to identify which actors are most likely to attempt misuse and what bottlenecks currently stop them. DeepMind says it uses a mix of expert red-teaming and randomised controlled trials to judge whether Gemini could help someone clear those bottlenecks. Post-training methods are meant to teach the model to refuse harmful queries while avoiding what the company calls over-refusal of legitimate science questions, a balance that’s proven difficult across the industry generally, not just for DeepMind. Classifiers and probes are deployed to flag risky activity in real time, and the company says it runs targeted log analysis to catch more subtle misuse patterns that automated filters might miss. None of these mitigations is described as solved. DeepMind frames them as an ongoing process rather than a finished system, which matters for any enterprise or government body evaluating whether to rely on the safeguards as currently configured. A classifier tuned against known jailbreak patterns in a controlled evaluation doesn’t guarantee equivalent performance against novel attack methods surfacing in live use, and the company doesn’t claim otherwise. The DNA synthesis screening problem One of the more concrete risks under exploration involves DNA synthesis. Companies within the International Gene Synthesis Consortium currently screen orders against lists of known harmful pathogens and toxins, paired with screening algorithms. DeepMind states plainly that this approach is starting to fray, because AI can now help design DNA sequences with similar function to a dangerous pathogen without matching its sequence closely enough to trigger existing screens. The proposed fix borrows from DeepMind’s existing watermarking system, SynthID, which the company says has become an industry standard for marking AI-generated images and text. Adapting it to biological sequences is presented as exploratory work, not a shipped product. A longer-term goal, described as an open technical challenge rather than something close to resolved, involves screening that predicts whether a novel DNA sequence is likely toxic or pathogenic based on its function, regardless of whether it resembles anything in existing databases. Cheaper sequencing as the detection layer Detection depends on metagenomic sequencing, which characterises every microorganism in a sample rather than checking for a shortlist of known pathogens the way traditional diagnostics do. The limiting factor is cost, and scaling the approach to the regions where outbreaks are most likely to originate requires that cost to fall considerably. DeepMind points to a collaboration between Google and Pacific Biosciences that used its AlphaEvolve coding agent to improve sequencing accuracy as one data point toward that goal. The company says it’s now looking at further opportunities – from optimising the algorithms that process sequencing data, through to informing hardware design – and separately exploring whether AlphaGenome could help characterise pathogens directly from sequence data. These remain research collaborations rather than field-deployed systems, and the distance between a sequencing accuracy gain in a controlled pipeline and a functioning early-warning network across wastewater and transit hubs in low-resource settings is not small. AlphaFold’s publication record and the countermeasure gap The response pillar leans on the medical countermeasure gap that leaves many known pathogens without a licensed diagnostic, vaccine, or treatment. DeepMind cites more than 10,000 publications on infectious disease that have referenced AlphaFold over five years, covering work on tuberculosis and malaria transmission and target mapping for threats including Mpox and Nipah. The newest addition to that record is a partnership with Lawrence Livermore’s bioresilience program, which plans to use AlphaFold 3 for broad-spectrum antibody design work, including a pan-filovirus antibody effort. DeepMind says it will keep adding protein structures and complexes to the AlphaFold Protein Structure Database this year, prioritising targets relevant to countermeasure development. Access to newer agent systems, including Co-Scientist, is being extended to selected researchers, among them scientists in the US Department of Energy’s National Laboratories working under the Genesis Mission. Isomorphic Labs has gone a step further, setting up a dedicated unit intended to deploy its drug design engine quickly during a novel outbreak, working alongside government and national research bodies such as Lawrence Livermore, the *** AI Security Institute, CEPI, and the Francis Crick Institute. The company also pledged $7 million to Health for Human Potential, a Philanthropy Asia Alliance programme, for infectious disease research across Asia. DeepMind’s recommendations to US policymakers map directly onto its three pillars and lean on specific pending legislation: On prevention, it backs a federal frontier AI safety framework, the AI-Ready Bio-Data Standards Act (H.R. 7907), mandatory DNA synthesis screening through the Biosecurity Modernization and Innovation Act (S. 3741), and the SCALE Biology Act (H.R. 8981). On detection, it wants metagenomic sequencing expanded across transit hubs and dense population centres, supported by the America’s Living Library Act (S. 4023) and additional DARPA and HHS funding for early-warning research. On response, it calls for the Web of Biological Data Act (H.R. 9307 / S. 4770) and investment in manufacturing capacity kept “warm-based” and ready for rapid activation, alongside pre-established clinical trial networks and faster regulatory pathways. None of that legislation is enacted, and the gap between a company’s policy wishlist and a functioning federal biosecurity framework is where the real test of this program will play out over the next 6-12 months. See also: Neko Health raises $700 million to expand AI body scans in the US 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 Examining Google DeepMind’s AI bioresilience push appeared first on AI News. View the full article

Neko Health has raised $700 million to expand its AI body scans in the United States, starting with a clinic in New York. The company’s preventive screening service combines medical imaging, blood tests, proprietary sensors, and clinician review. The Series C round was led by Lightspeed Venture Partners and co-led by O.G. Venture Partners. Existing investors Atomico, General Catalyst, and Lakestar participated, alongside new backers including Liberty City Ventures, Positive Sum, and BDT & MSD. David Ofer of O.G. Venture Partners will join Neko’s board, subject to regulatory approval. The latest round brings Neko’s disclosed funding since 2023 to more than $1 billion. The company raised $65 million in a Series A round in 2023 and another $260 million in January 2025. The round also included investments from Meta chief executive Mark Zuckerberg and Priscilla Chan, former tennis player Maria Sharapova, musician will.i.am, and former footballer Thierry Henry. Existing individual investors include Reddit co-founder Alexis Ohanian and actor Zoë Saldaña. How Neko Health’s AI body scans work Neko operates clinics that combine full-body scans, blood tests, artificial intelligence, and custom-built medical equipment. The service screens customers for potential signs of conditions including skin *******, cardiovascular disease, and diabetes. The company describes the service as a 60-minute, non-invasive, and radiation-free health assessment. It uses proprietary sensors and blood analysis to examine skin conditions, blood abnormalities, pre-diabetes indicators, and risk factors associated with metabolic syndrome, stroke, and heart attack. The scan also includes an electrocardiogram, arterial measurements, body-composition analysis, and more than 2,000 high-resolution images used to map a customer’s skin. Blood samples are processed at the clinic, allowing the results to be reviewed during the same visit. Results are discussed during an in-person consultation with a medical professional. Many of the measurements, including blood pressure, blood glucose, cholesterol, electrocardiogram data, and skin assessments, are already available through conventional healthcare services. Neko’s model combines them with proprietary imaging, automated data collection, on-site blood processing, and a clinician consultation in one appointment. The company’s public clinical materials reviewed for this article do not include a comparative study showing whether this combined approach improves clinical outcomes or cost-effectiveness compared with established preventive-care pathways. Neko said the funding will support the opening of clinics in New York and other US cities. It has not identified the additional locations or provided a detailed expansion schedule. A waitlist for the New York clinic is open on the company’s website. Neko has not disclosed how much its US scans will cost. The company currently operates eight clinics across the *** and Sweden, including two in Stockholm, one each in Manchester and Birmingham, and four in London: Marylebone, Spitalfields, Covent Garden, and Victoria. A scan costs £299 in the *** and 2,750 Swedish kronor in Sweden, equivalent to about $400 and $285. Neko said it has completed 100,000 scans since launching its service in 2023. More than 350,000 people have registered for a scan or joined its waitlists. According to the company, 75% of customers book and prepay for another scan at the end of their appointment. The repeat-booking model allows clinicians to compare measurements and skin images over time. Public information reviewed for this article does not include evidence establishing whether annual screening is the appropriate interval for every age or risk group. Founded in 2018 by Spotify co-founder Daniel Ek and chief executive Hjalmar Nilsonne, Neko operated largely outside public view until launching its service in 2023. Nilsonne said part of the new capital will fund further research and development into preventive screening technology. Neko recently added body-composition measurements and clinician reviews of wearable-device data across its clinics. It has also introduced updated versions of its Derma, Echo, and Spectrum medical devices, which collect information relating to skin, heart function, and circulation. The company said the newer equipment captures a larger volume of health data and automates more parts of the scanning process. It plans to deploy the devices across its clinics in the coming months. The clinic-based model requires proprietary equipment, blood-processing capacity, medical personnel, and specialist review for some findings. Neko said its seven-room Covent Garden clinic was designed to support tens of thousands of scans annually, but it has not disclosed capacity or staffing figures for its planned US locations. US regulation and access FDA records show that two of Neko’s internally developed devices received clearance through the agency’s 510(k) pathway in May 2026. Derma-2 was cleared as an adjunctive telethermographic system, while Spectrum-2 was cleared as a tissue-saturation oximeter used in cardiovascular measurements. The clearances apply to the specified devices and their intended uses, rather than to the complete Neko Health Scan as a single FDA-approved screening service. Neko describes its US clinics as preventive health and wellness providers rather than full-service medical practices. Its privacy notice advises customers to continue seeing their existing clinicians for diagnoses and treatment, including for conditions identified through a Neko scan. The company says specialist clinicians, including dermatologists and cardiologists, review findings that require further examination. Follow-up appointments, referral letters, and introductions to outside specialists are included when recommended by a Neko clinician. Neko’s US clinics do not currently participate in health insurance plans, and the company says most of its services are not covered by a payer. Customers are therefore expected to pay directly for the initial assessment. The company has not disclosed what customers would pay for diagnostic tests or treatment delivered by external healthcare providers. It has also not announced its US price or whether employers, insurers, or other organisations will subsidise access to the service. Questions over clinical evidence Publicly available information reviewed for this article does not include a completed peer-reviewed study validating the full screening service. A study registered on ClinicalTrials.gov is evaluating the suitability of Neko’s multimodal skin-imaging technology for screening and diagnostic-support applications, including skin ******* and Raynaud’s phenomenon. The trial remains an ongoing evaluation and does not establish the clinical performance of the complete Neko Health Scan. Neko’s public materials reviewed for this article do not disclose how often its scans produce false-positive findings, how many customers undergo additional procedures, or how many flagged abnormalities are later considered clinically unimportant. The company’s US privacy notice identifies an AI dictation system used to summarise medical records, questionnaires, scan data, and clinical conversations, and to draft reports. Neko says clinicians review reports produced using the system. Public information does not provide a complete breakdown of which scan measurements or classifications rely on AI, conventional software, or direct clinical assessment. It also does not state that AI independently diagnoses the conditions covered by the service. The FDA clearances for individual devices do not establish the performance of every algorithm used to combine or interpret the resulting data. Public materials reviewed for this article do not include detailed real-world performance figures for each AI system used in the screening process. Neko’s reported health-outcome data is based on 1,469 customers who completed a second scan approximately one year after their first. The company said the group recorded improvements in blood pressure, cholesterol, and blood sugar, while body weight remained broadly stable. The scans in the analysis were conducted at Neko’s Stockholm clinics. The company divided the results into a general group, a healthy group, and groups covering customers with hypertension, pre-diabetes, or diabetes. Neko said the analysis was not a scientific study and did not include a control group. Customers could have started treatment or changed their behaviour between appointments, so the figures do not establish that the scans caused the reported improvements. The analysis only covered customers who returned for a second privately purchased scan. Neko did not publish demographic information showing how that group compares with its wider customer base or the general population. The company did not disclose its valuation following the latest round. The Financial Times, citing unnamed sources, reported that Neko is now valued at about $7 billion. (Photo by Wonderlane) See also: NHS AI blood test could reduce invasive womb ******* checks 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, click here for more information. AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here. The post Neko Health raises $700 million to expand AI body scans in the US appeared first on AI News. View the full article

Nokia’s AI-RAN platform arrived on July 15 with a claim worth examining: that it is the industry’s first. The vendor says the platform, built on its anyRAN software and NVIDIA’s Aerial system, will let operators pull far more capacity from the spectrum they already own, and it has framed the launch as one of the most significant shifts in radio architecture in decades. The technical pitch is straightforward. Nokia says the platform has already shown more than 20% spectral efficiency gains, and it is targeting 50% by 2027 and more than 100% by 2028, the point at which, on its own projection, operators could roughly double the capacity of existing spectrum. Those last two figures are targets, not results, and Nokia’s own timeline puts pilots at the end of this year and commercial availability in 2027. Operators would buy the capability through a software subscription rather than a hardware refresh, choosing from three deployment options: a GPU-powered plug-in card for existing AirScale sites, a standalone AI-RAN node, and a cloud-server build delivered through partners. We are launching the industry’s first commercial AI-native #AIRAN platform built on @NVIDIA accelerated computing, marking one of the most significant shifts in radio network architecture in decades and providing operators with a practical path to AI Native Networks. Read more:… pic.twitter.com/3ThlGwz7bc — Nokia (@nokia) July 15, 2026 A comeback for Nokia’s weakest business To read the launch only as a product story is to miss why it matters to Nokia. Radio has been chief executive Justin Hotard’s hardest problem since he took over in 2025. At Nokia’s November capital markets day, he told investors the mobile business had not delivered acceptable returns, and he folded it into a new Mobile Infrastructure segment alongside further cost cuts. The NVIDIA partnership, announced in October 2025 with a $1 billion investment from the chipmaker for roughly a 3% stake, sits at the centre of that repair job. By building on NVIDIA’s silicon and CUDA software rather than its own custom chips, Nokia can cut a slice of costly in-house R&D and redirect it toward software, the shift Hotard has described as moving away from a legacy hardware model. Investors have rewarded the story. Nokia shares have re-rated sharply through 2026 on the strength of its AI and cloud momentum, and the AI-RAN launch landed days before its second-quarter results. Omdia, whose analyst Rémy Pascal is quoted in Nokia’s own announcement, has put the cumulative AI-RAN opportunity above $200 billion by 2030. The direction of travel is real. The open question is how much of it Nokia can claim as a lead. Is the Nokia AI-RAN platform really the first? Here, the “industry’s first” label needs care. In June, Ericsson began selling a commercial AI-in-RAN software subscription that it says delivers up to 20% higher downlink throughput and up to 10% better spectral efficiency across more than 15 live deployments, and, crucially, it runs on operators’ existing baseband silicon, with no GPU required. On availability, Ericsson is already in the market. Nokia’s claim to a first rests on a narrower definition: a GPU-accelerated AI-RAN platform, a different architecture from AI features layered onto existing hardware. Both statements can hold at once, which is exactly why the framing deserves scrutiny rather than a straight repeat. The divergence, though, runs deeper than timing. Nokia has tied its radio roadmap to NVIDIA, and its chief technology officer, Pallavi Mahajan, has acknowledged that at least some of the Layer 1 software is bound to the underlying hardware. Ericsson has taken the opposite route by design, keeping its AI features silicon-independent to avoid that dependency. Nokia points to merchant silicon from Marvell in its wider ecosystem and describes the platform as Open RAN-compliant, but the performance case it is selling, those spectral efficiency gains, currently runs through NVIDIA’s stack, for which no equivalent alternative exists today. The openness in the messaging and the NVIDIA dependency in the engineering are both features of the same launch. None of this makes the strategy wrong. Outsourcing the silicon race to the industry’s dominant AI-chip supplier is a defensible answer to a business Nokia had struggled to fix on its own, and the subscription model gives radio the recurring revenue its hardware cycles never did. But the platform is not yet commercial, its headline efficiency numbers are still two years out, and at least one major rival reached the market first by a different road. For Nokia, this is a comeback in motion, not one already won, and its trajectory now runs, for better or worse, through NVIDIA. See also: AI-Native networks are no longer a 6G promise–MWC 2026 just proved it 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 Nokia’s AI-RAN platform: a radio comeback that runs on NVIDIA appeared first on AI News. View the full article

AWS (Amazon Web Services) has explained how Bluesight developed Prism, an AI layer that connects hospital pharmacy and compliance data across its product suite. Prism Assistant for ControlCheck has reached general availability and operates across 20 health systems, according to AWS vendor statements. Meanwhile, a multi-product agent for 340B Group Purchasing Organisation (GPO) compliance remains scheduled for release later in 2026. The project addresses a data-intensive process that hospital pharmacy teams often conduct manually. AWS says a single 340B covered entity can spend more than 4,000 staff hours each year reviewing whether GPO drug purchases qualify for an exception. Staff must compare purchase data with US Food and Drug Administration shortage notices, American Society of Health-System Pharmacists records, days-on-hand inventory, machine learning shortage forecasts, and back-order reports from other hospitals. ControlCheck agent reaches 20 hospital health systems Bluesight started with ControlCheck, its controlled-substance monitoring product. Hospital diversion teams use ControlCheck to identify unusual medication transaction patterns, but compliance staff still needed to assemble reports, review dashboards, and correlate findings manually. Prism Assistant gives users a conversational interface that can query ControlCheck data, produce charts, and generate report material. AWS claims Bluesight built the first version during a three-day Experience-Based Acceleration engagement in September 2025, with eight Bluesight engineers working alongside seven AWS specialists. TechForge Media notes that while these rapid development timelines highlight the agility of the tools, they remain vendor-reported metrics pending independent verification from the active health systems themselves. The team used Strands Agents with Amazon Bedrock and hosted the application through Amazon Bedrock AgentCore Runtime. AgentCore Gateway exposed more than 10 ControlCheck APIs as MCP tools, allowing the agent to discover and call them during a user request. Crucially, Bluesight avoided direct database access for the language model. Its engineers wrapped existing ControlCheck API endpoints in AWS Lambda functions that return structured data suited to agent processing. The company kept business logic inside its application layer, while the agent interpreted questions, selected tools, gathered records, and presented results. AWS reports that design reduced query latency from five minutes to 10 seconds. The deployment also included a frontend with chart generation, observability controls, cost attribution, encryption, authentication, and infrastructure-as-code. “This is exactly what diversion program leaders have been waiting for—it gets them to answers faster and takes the manual grind out of every investigation,” said Samir Neyazi, Director of Product Management at Bluesight. Bluesight deployed the service within a virtual private cloud and reached general availability in under nine months, according to AWS. The 20 health systems using Prism Assistant represent the active deployment evidence in the AWS account; the larger GPO agent has not yet reached that stage. Multi-agent 340B compliance targets GPO purchases Federal 340B rules prohibit Disproportionate Share, Children’s, and Free-Standing ******* hospitals from buying outpatient drugs through GPO contracts when non-GPO channels can supply the drug. Compliance teams must document the exception when supply conditions prevent that purchase route. Bluesight’s planned GPO agent brings together records from CostCheck, ShortageCheck, and 340BCheck. CostCheck provides purchase information, ShortageCheck contributes drug availability evidence, and 340BCheck validates eligibility data. The proposed architecture uses Anthropic Claude Sonnet 4.6 as the primary model and Claude Haiku 4.5 for lower-latency operations through Amazon Bedrock. AgentCore Runtime executes the agents in a VPC with private subnets. AgentCore Gateway connects Lambda-backed tools to the three Bluesight product systems. A coordinating GPO agent directs specialist data workers. One worker retrieves purchase records, another gathers supply evidence, and another checks 340B eligibility. The coordinator assembles the evidence and produces an audit-oriented report. March 2026 brought a second AWS acceleration engagement focused on that architecture. AWS says the team connected the system by the end of its first day and completed every planned feature by the end of day two. The company tested the agent against synthetic data, where it reported a 100 percent invoice discovery rate and 93 percent evidence justification accuracy, above its 85 percent target. However, enterprise buyers must exercise caution: those figures do not represent production performance across hospital customers. Synthetic testing can demonstrate whether tool calls, matching logic, and report generation work against prepared scenarios, but it cannot establish how the system handles local data gaps, delayed shortage updates, unusual drug identifiers, or disputed purchasing cases. Keeping compliance scoring outside the language model Bluesight assigns the language model a constrained role in the GPO workflow. The model gathers records, calls product tools, and drafts the explanation. A deterministic scoring service calculates the compliance determination. That service evaluates 13 evidence inputs, applies priority-based matching, and uses configurable time windows. The design gives compliance teams a repeatable scoring process rather than an LLM-generated judgement and allows an auditor to inspect the source records, the rules applied, and the sequence of tool calls behind each determination. Despite the automated assistance, hospital pharmacy, legal, and compliance teams still need absolute ownership of those policy settings. A supplier shortage threshold, acceptable inventory *******, or purchase-date window can alter a compliance outcome. Bluesight’s approach gives customers a technical mechanism to configure those decisions, but each organisation must set and approve its own policy rules. HIPAA controls and audit traces shape the deployment Amazon Bedrock holds HIPAA eligibility, and Bluesight operates under a Business Associate Agreement with AWS. AWS says it does not train foundation models on customer data processed through Amazon Bedrock. Bluesight uses Amazon Cognito for OAuth2 client-credential authentication and JSON Web Token validation. AgentCore Runtime provides session isolation for concurrent customer requests. AWS Key Management Service encrypts data at rest and in transit, while AWS Secrets Manager manages credentials for downstream services. Amazon CloudWatch records agent decisions, tool invocations, data-access events, alarms, and performance metrics. That audit trail matters when a hospital needs to explain why it permitted a GPO purchase or escalated a drug-diversion pattern. Bluesight’s internal measurements across 20 health systems report up to 97 percent faster report generation and analysis in ControlCheck workflows. Recurring reports reportedly dropped from about six hours of manual assembly to 15 minutes, a 96 percent reduction. Pre-investigation triage dropped from three hours to about 10 minutes, while controlled-substance variance analysis fell from 30 minutes to less than one minute. Teams should strictly run historical purchasing cases in parallel with existing review processes before allowing an agent-assisted result to affect compliance decisions. Local testing should rigorously examine data completeness, drug-code matching, shortage timing, exception rules, and cases where human reviewers previously disagreed. Each production finding should retain the scoring-rule version, source evidence, and tool trace that produced it. See also: AWS GraphRAG deployment cuts drug research cycles by 87% 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 AWS and Bluesight build AI for hospital 340B compliance appeared first on AI News. View the full article

AI agent crawlers, the bots that fetch pages in real time on behalf of a person waiting for an answer, will be blocked by default on a slice of the web from September 15 onwards. Cloudflare announced the change on July 1, and most of the coverage since then has focused on Google. The more useful part is what it asks of everyone building agents, and what it offers them in return. Cloudflare has replaced its single block-AI-bots switch with three categories. Search covers bots that index a page to answer questions about it later. Agent covers automated systems acting in real time for a user, including ChatGPT’s fetch bot and browser-driving agents. Training covers crawlers that pull content into a model’s weights. The controls went live on July 1 for every customer, including the free tier. From September 15, the defaults change. Training and Agent will be blocked on pages that display ads, while Search will stay allowed. The new defaults apply to domains newly onboarding to Cloudflare, new sites set up by existing customers, and all existing free-tier customers. Anyone who does not want them can opt out through their security settings before the date. Cloudflare’s logic is that an advertisement is evidence that a page was built for a human to land on. A search crawler that sends a reader back is a referral. A bot that reads the page and hands the answer to someone else is not. What AI agent crawlers run into now Agentic deployments have been built on the assumption that the open web stays open. A research agent fetches a competitor’s pricing page. A monitoring tool checks a supplier’s announcements. A customer-service agent pulls a manufacturer’s specification sheet. None of this involves a licence, and until now, none of it needed one. Cloudflare sits in front of a large share of the world’s web traffic, and its blocks operate at the network level rather than as a robots.txt suggestion a crawler can ignore. Ad-supported pages are exactly the pages agents want, because that is where news, reviews, pricing and product coverage live. The failure mode for an enterprise agent is not a lawsuit. It is silence, or an answer built from whatever it could still reach. There is a Google-shaped complication. Googlebot crawls for both search and training in a single bot, so under the most restrictive rule, a site that blocks Training also blocks Googlebot. Cloudflare CEO Matthew Prince said the company hopes the changes will “encourage mixed-use crawlers to separate search from agent use and training”, which is a polite way of saying the pressure is the point. Getting permission for AI agent crawlers Anyone running agents should start by working out which of their Cloudflare accounts will read as Agent-class. The classification is behavioural rather than something you opt into, so a research agent that browses in real time is caught whether or not its operator thinks of it as a crawler. Expect degraded coverage rather than a clean failure, because the block lands on ad-supported pages and leaves the rest reachable. Negotiated access, not a rewritten user-agent string, is the way through. Publishers have a different homework list. Check your tier first, since existing free-tier customers are moved to the new defaults automatically on September 15, a detail most coverage skipped. Then decide whether blocking Training is worth what it costs, because it takes Googlebot with it and your search visibility along with it. The mechanism worth watching is the money. Pay Per Crawl is becoming Pay Per Use, with Ceramic.ai paying publishers when their content appears in AI search results, and You.com paying when an agent reaches premium content. Cloudflare says more than half of AI crawler traffic is spent re-fetching pages that have not changed, so there is waste on both sides worth pricing out. This is the first round of the content fight, where the answer on offer is a rate rather than a wall. One weakness lies in the taxonomy itself. Search, Agent and Training are behaviours the AI companies declare about their own bots, and a firm that would rather not have its training run classified as training has an obvious incentive. The announcement does not explain what stops it. Access to the open web has been free and unlimited for thirty years, and the bill is now itemised. Agent builders who sort out their access before September have a workable problem. The ones who find out from a 403 will be rebuilding on the fly. See also: Visa ChatGPT integration enables AI agent retail purchasing 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 AI agent crawlers now need permission. Here’s how to get it appeared first on AI News. View the full article

Jensen Huang has a test for whether an engineer is worth keeping, and it comes with a token budget attached. Speaking on the All-In Podcast at the close of GTC 2026, the Nvidia chief executive said that if a $500,000 engineer’s annual AI token consumption came in under half their salary, “I am going to be deeply alarmed.” Nvidia, he confirmed, is working toward a $2 billion yearly token bill for its engineering force. He was describing a trade-off most companies have already made with less fanfare: money that once paid people increasingly pays for tokens. The four largest hyperscalers have guided roughly $700 billion in combined 2026 capital expenditure, nearly double last year, while data from outplacement firm Challenger, Gray & Christmas shows AI as the most-cited reason for US job cuts for a record fourth consecutive month. An internal Meta memo obtained by Reuters described May’s cuts of 8,000 roles as offsetting the company’s substantial investments, in a quarter when revenue grew 33%. The layoffs at companies like these aren’t survival measures. They’re financing. The trouble is that the financing hasn’t bought what it promised. Gartner surveyed 350 executives at companies with over $1 billion in revenue, all deploying AI agents or automation, and found roughly 80% had cut headcount with no correlation to improved returns. Analyst Helen Poitevin’s verdict was blunt: “Workforce reductions may create budget room, but they do not create return.” Uber learned the token side of that lesson the expensive way, giving 5,000 engineers AI coding tools in December and exhausting its entire 2026 AI budget by April. Chief Operating Officer Andrew Macdonald conceded that despite 70% of committed code being AI-generated, the connection to anything customers notice is missing: “That link is not there yet.” Put those two failures side-by-side and the actual problem comes into focus. Companies treated the token bill as fixed and the workforce as flexible, when the opposite is true. Payroll cuts happen once and take institutional knowledge with them. A token budget, it turns out, bends in half a dozen places if anyone bothers to engineer it. Where the token budget bends The cheapest fix is also the least glamorous: stop paying to process the same text repeatedly. Prompt caching, now standard across the major API providers, cuts the cost of repeated input by up to 90% under Anthropic’s and OpenAI’s published pricing, because static content like system instructions and reference documents gets processed once and reread at a fraction of the rate. Security firm ProjectDiscovery documented raising its cache hit rate from 7% to 84% by restructuring prompts, cutting its total LLM spend by 59 to 70% while serving 9.8 billion tokens from cache. That single engineering exercise recovered more budget than most AI-attributed layoff rounds save. The next lever is routing work to the right-sized model. Providers’ own price lists show flagship models costing five times their smaller siblings per token, yet plenty of production workloads send routine classification and summarisation to the most expensive tier by default. Batch processing adds a further 50% discount for anything that doesn’t need a real-time answer. Retrieval-augmented generation attacks the problem from another angle by sending the model only the relevant slice of a knowledge base rather than the whole thing, and prompt compression trims the redundant examples that inflate every call. Open-weight models reduce costs further still, handling routine workloads at a fraction of frontier API prices for teams willing to manage the infrastructure. These measures are simply the AI equivalent of turning off the lights in empty rooms, and Uber’s $1,500 monthly cap per engineer – imposed after the April overrun – is early evidence that spending discipline arrives eventually. The companies getting ahead are simply choosing it before the budget forces it. The other half of the fix is human Optimising the token bill only matters if the savings go somewhere productive, and the strongest evidence points at people. Poitevin’s research found the organisations that improved ROI were those using AI to amplify their workforce rather than replace it. Klarna ran the controlled experiment on everyone’s behalf, replacing roughly 700 customer service roles with an OpenAI-powered assistant before customer satisfaction fell. Chief Executive Sebastian Siemiatkowski told Bloomberg what few executives admit aloud: “The result was lower quality, and that’s not sustainable.” The fintech now runs a blended model, with AI absorbing routine volume while rehired humans handle everything requiring judgment. Gartner expects the pattern to spread, predicting that by 2027 half the companies that cut customer service staff for AI will rehire them. There’s one workforce investment the optimisation logic makes urgent rather than optional. Stanford University’s Institute for Human-Centered AI found employment for software developers aged 22 to 25 fell nearly 20% from 2024 levels even as older cohorts grew, which means companies are removing the training ground for the senior engineers they’ll need directing all these systems in five years. A business that has just engineered 60% off its token bill has the budget room to keep hiring at the bottom rung. Whether it does is a leadership decision, not a financial one. Nvidia’s Huang’s provocation will keep echoing through earnings calls, and the capex numbers will keep climbing. The companies that come out ahead won’t be the ones that spent the most on tokens or cut the most people to afford them—they’ll be the ones that noticed the token budget was the flexible line all along, squeezed it with engineering rather than headcount, and spent the difference on the people who make the tokens worth anything. (Image by kate.sade) See also: Per-token AI charges come to GitHub Copilot 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 to shrink the token budget without shrinking the team appeared first on AI News. View the full article

A recent AWS GraphRAG deployment reduced drug research and development cycles in pharmaceutical environments by 87 percent. This acceleration is achieved by integrating previously separated proprietary databases into a unified and queryable knowledge graph. Historically, initial data gathering and screening phases took over six months per iteration, yielding a low five percent success rate. Crucial datasets – ranging from domain-specific clinical metrics to internal engineering and laboratory notes – were isolated across storage environments, effectively blocking data scientists from uncovering latent correlations. When staff left, they took crucial project context with them, stalling active research. AWS built a solution to connect these systems, combining graph databases with NLP. The setup relies on a GraphRAG framework and uses Amazon Neptune Analytics and Bedrock to turn disconnected data points into a searchable network. Users can submit standard natural language queries and receive answers mapped to verified domain literature and internal datasets. However, unifying isolated proprietary datasets with unstructured open-access repositories still introduces significant data normalisation challenges, requiring strict schema governance to prevent inaccurate relational mapping and mitigate the risk of hallucinations. Knowledge graph construction Companies can plug in their own knowledge graphs. The system pulls in messy, unstructured files from public databases like PubMed and mixes them with internal corporate records. Tools like Amazon Comprehend Medical scan this text to pull out standard medical codes. Amazon Bedrock, running Anthropic’s Claude 4.5 Sonnet, summarises the document contents and determines topical relevance. AWS Lambda functions and Amazon S3 bulk loads then route these processed elements into Amazon Neptune Analytics. The resulting knowledge graph structures the data into discrete nodes representing core entities like domain-specific classes, authors, source journals, and embedded text chunks. The graph edges define the relationships between these nodes, mapping out hierarchical classifications and entity associations. This structured representation provides the deterministic foundation necessary for accurate information retrieval. The database schema establishes the strict boundaries of the RAG discovery process. Nodes are structured to capture specific conditions and map them hierarchically to established ontologies, while author and journal nodes provide provenance for published research. Lengthy documents are broken down into digestible text segments using Amazon Bedrock Knowledge Base chunking strategies, and specific classification nodes anchor the unstructured textual data to standardised diagnostic metrics. Operating this graph architecture requires specific cloud resource allocations. A standard Amazon Neptune Analytics graph running with 16 provisioned memory units incurs operational costs of $0.48 per hour. Development environments, such as Amazon SageMaker Jupyter notebooks running on t3.medium instances, add baseline compute and storage expenditures. Organisations must also factor in dynamic token consumption costs generated by the Amazon Bedrock Claude 4.5 Sonnet model during query processing and abstract generation. The GraphRAG toolkit acts as the execution layer between the user interface and the underlying database. A dedicated Knowledge Graph Linker processes incoming natural language queries, extracts relevant entities using fuzzy string indexing, and maps them to established graph nodes. The system traverses the network pathways to generate plausible relational links before drafting a response through the Bedrock-hosted language model. Retrieval accuracy depends on the entity matching configuration. An EntityLinker component aligns natural language terms from user prompts to the structured data schema. This fuzzy matching process handles the inherent noise and varied terminology found in complex enterprise datasets, ensuring users retrieve the correct nodes even when using imprecise language. Modularity and system architecture Data extraction relies heavily on specialised AI parsing; the architecture employs Claude to evaluate raw source documents and generate concise abstracts. Domain-specific tools then map these complex textual descriptions to standardised taxonomies. The GraphRAG Python toolkit initialises a BedrockGenerator to power natural language interactions, while engineers configure a Knowledge Graph Linker component to bind the graph store to the language model. This integration creates a direct interface for executing queries and generating responses grounded strictly in the available graph data. The architecture separates three core functions: language model initialisation, graph interfacing, and entity linking. Because the system is modular, teams can swap out the language model or tweak the graph structure without having to tear down and rebuild the whole app. Active deployments of the Neptune and Bedrock architecture return exact, verifiable citations for every generated answer. The system maps the entire reasoning path, displaying the specific graph traversal steps used to reach a conclusion. Key performance metrics from early enterprise adopters include an 87 percent reduction in research cycle durations. Initial discovery phases that previously required six months now conclude in three weeks, and data retrieval speeds show an 85 percent improvement, directly supporting faster hypothesis testing. Furthermore, research review times drop by 70 percent due to automated citation mapping and source verification features. Engineering teams can integrate new public databases or internal notes into the existing graph structure without disrupting active query interfaces. For governance and compliance, exact evidence trails required for regulatory submissions are captured, with graph traversal visualisations proving precisely how an AI model connected complex variables. Teams can trace every output directly to source documents, fulfilling compliance requirements for scientific integrity. Finally, maintaining a centralised knowledge graph stops data decay. When senior scientists resign, their tacit knowledge regarding system behaviours or failed experiments remains indexed within the Neptune database. New personnel can query the system to review past decisions and instantly access the historical context of an ongoing project. As GraphRAG frameworks mature, this deployment model is unlikely to remain confined to pharmaceutical research. The ability to deterministically map internal, unstructured data against verified public repositories provides a blueprint for any enterprise struggling to extract actionable intelligence from fragmented legacy systems. See also: Insilico Medicine advances AI drug for IPF to Phase III trials 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 AWS GraphRAG deployment cuts drug research cycles by 87% appeared first on AI News. View the full article

Several NHS hospitals are preparing to use an AI-powered blood test to help assess women referred for possible womb ******* before invasive checks are carried out. According to The Guardian, around 90,000 postmenopausal women in England are referred by GPs each year for checks after experiencing heavy bleeding. About 10,000 women are diagnosed with womb ******* annually, and around 2,700 die from the disease. How the PinPoint test works The test, developed by Leeds-based PinPoint Data Science, uses machine learning to assess ******* risk from blood markers. It classifies patients as low, elevated, or high risk based on an analysis of around 30 markers. PinPoint said the test costs around £30 and gives clinicians a risk score for use within existing ******* referral pathways. The score can help inform whether a patient is monitored, referred for further investigation, or prioritised for faster assessment. PinPoint describes the tool as a multi-******* test. The company said it has been used across gynaecological, lung, upper gastrointestinal, head and neck, and lower gastrointestinal ******* pathways. The test is being introduced after a trial involving 16,481 patients referred through urgent suspected ******* pathways across Yorkshire. The trial included women referred with symptoms that raised concern about possible womb or gynaecological *******. About one in 10 women referred because of heavy bleeding were found to have *******, according to the reported trial results. PinPoint said the test correctly identified 99.1% of cancers as elevated or high risk and delivered a negative predictive value of 99.8% for women in the lowest-risk group. Mid Yorkshire NHS Teaching Trust plans to use the test for six types of gynaecological or upper gastrointestinal *******. Leeds Teaching Hospitals NHS Trust plans to use it for gynaecological *******. Current diagnostic pathway Under the current pathway, women referred for suspected reproductive system cancers usually undergo a pelvic examination that includes a transvaginal ultrasound scan. The procedure involves inserting an ultrasound probe into the ******* to measure the thickness of the womb lining, and some women find it uncomfortable or painful. If doctors continue to suspect *******, patients can then be referred for further checks, including a biopsy and hysteroscopy, an examination of the inside of the womb. PinPoint said its test is intended to identify women at very low risk before those procedures are used. The company said the test could spare about one in five referred women from needing a transvaginal ultrasound scan. That would amount to around 18,000 women a year in England. Professor Sean Duffy, chief medical officer at PinPoint Data Science and a former NHS England national clinical director for *******, said the test’s value lies in ruling out women at very low risk. Dr Jacinta Walsh, a GP at King’s Medical Practice in Normanton, West Yorkshire, said patients can require up to six GP visits before ******* is ruled out. She said the test could shorten that process and free up capacity for other patients. Tracy Jackson, a consultant gynaecologist and ******* unit lead at Leeds Teaching Hospitals NHS Trust, said most women seen through the current referral route do not have *******, while the investigations can be uncomfortable or distressing. Jackson said the test could help clinicians triage patients before hospital-based investigations. She said low-risk patients could be ruled out in primary care, while higher-risk patients could be prioritised for further checks. Other NHS AI deployments Recent NHS AI deployments include MEMORI at Kent and Canterbury Hospital, an AI triage tool in the NHS App, and AI-powered chest X-ray tools for suspected lung ******* pathways. East Kent Hospitals University NHS Foundation Trust is using an AI system called MEMORI at Kent and Canterbury Hospital to assess infection risk from routine patient data. The system analyses information already included in patient records, including blood tests, blood pressure, temperature, observations, medications, and demographics. NHS England said an AI triage tool in the NHS App is expected to reach more than 200,000 patients within 12 months and become available to all NHS App users by April 2028. The government has also committed £20 million to roll out AI-powered chest X-ray tools to all NHS trusts in England by 2029. The tools are already available in about half of NHS trusts in England and have supported assessment for more than four million patients being investigated for lung *******. Further evidence will be needed to assess how the test affects patient outcomes, referral decisions, and NHS diagnostic capacity. ******* Research *** described the PinPoint test as promising but said more research is needed to understand its benefits for patients and the NHS. Samantha Harrison, a spokesperson for the charity, said early detection saves lives, but patients are not currently being diagnosed quickly enough. The charity said the test could help rule out endometrial ******* in some women through a blood test, without the need for further investigations. (Photo by Adam Mills) See also: Takeda signs $600M AI drug discovery deal with Insilico 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, click here for more information. AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here. The post NHS AI blood test could reduce invasive womb ******* checks appeared first on AI News. View the full article

Insilico Medicine is advancing to Phase III human trials for testing a drug identified by AI targeting idiopathic pulmonary fibrosis (IPF). This progression supplies the computational drug discovery sector with empirical test cases, advancing an AI medicine past early safety evaluations into late-stage efficacy validation. IPF destroys respiratory capacity through severe lung tissue scarring. Patients typically present a median survival rate reaching two to four years post-diagnosis. The AI-identified drug, rentosertib, inhibits the TRAF2- and NCK-interacting kinase to address underlying disease mechanisms when administered orally. A randomised trial evaluated 71 patients across 22 ******** clinical sites, separating participants into placebo and active treatment cohorts. Investigators administered 30 mg or 60 mg daily doses over a 12-week observation window. Patients assigned to the 60 mg once-daily regimen demonstrated a mean forced vital capacity gain of +98.4 mL, contrasting sharply with the 20.3 mL capacity loss recorded in the placebo group. Safety profiles remained manageable, with adverse events mirroring expected baseline rates across all trial arms. Regulatory authorities at the U.S. Food and Drug Administration (FDA) granted ‘Orphan Drug Designation’ to the asset in February 2023. Algorithmic target prioritisation through multi-omics The development relies entirely on Pharma.AI, the proprietary computational pipeline operating at Insilico Medicine. The workflow segments into distinct engines handling specific biological and chemical engineering tasks. PandaOmics executes the initial target discovery phase. The system ingests vast biological datasets, processing genomics, clinical trial outcomes, academic literature, and patent intelligence to construct comprehensive biological network models. The algorithms apply causal inference mechanisms to identify novel disease links hidden within the data architecture. PandaOmics isolated TNIK as the primary biological target regarding IPF intervention. The computational system bypassed the receptor tyrosine kinase pathways targeted by existing antifibrotic medications. The software mapped TNIK as a central node regulating fibrosis and inflammation via Wnt, TGF-β, Hippo/YAP-TAZ, JNK, and NF-κB signalling channels. The target selection process integrated a hallmarks-of-aging framework, scoring biological targets based on their implication in multiple aging mechanisms, chronic inflammation, and extracellular matrix remodelling. Feng Ren, PhD, Co-CEO and Chief Scientific Officer of Insilico Medicine, said: “IPF is one of the clearest clinical examples of an age-related disease in which fibrosis, chronic inflammation, extracellular matrix remodeling, and cellular senescence intersect. “Rentosertib was not discovered by starting from a conventional target and simply screening more compounds. It came from a biology-first, ageing-informed AI workflow that connected TNIK to fibrotic and inflammatory disease mechanisms, and then used generative chemistry to create a drug candidate with the properties required for clinical development.” Generative molecular engineering execution Following target selection, the Chemistry42 engine executes generative molecular design. The system departs from traditional high-throughput screening methodologies. Chemistry42 does not search existing compound libraries—instead, the system applies Generative Tensorial Reinforcement Learning to build molecules that physically align with the target protein pocket. This algorithmic engineering process balances structural fit against required pharmacological properties. The computational generation phase synthesised exactly 79 physical molecules to undergo testing. The engineering team selected the 55th iteration to advance into preclinical testing. This targeted generation protocol reduced the timeline from project initiation to preclinical candidate nomination to 18 months. The foundational architecture stems from the 2019 publication of the company’s GENTRL methodology in Nature Biotechnology. The platform establishes reproducible systems regulating molecular generation, avoiding the capital-intensive trial-and-error processes defining standard pharmaceutical chemistry. Validating biological impact through proteomic analysis Clinical assessment integrates complex proteomic analysis to validate the algorithmically-predicted biological interactions. Insilico Medicine deploys internal proteomic aging-clock frameworks within the IPF trial to capture exploratory geroscience readouts. Chronological-age proteomic clocks – including ProtAge, OrganAgechrono, ipfP3GPT, and PAOPAC – track predicted biological-age changes resulting from the intervention. Researchers apply *** Biobank age-associated trajectories as external comparison datasets, contextualising treatment-responsive proteins against broad population data. Mortality-risk-related proteomic clocks, including PAC and OrganAgemortality, provide orthogonal analytical streams alongside standard clinical endpoints. The clinical teams execute SenMayo and CellAge signature analyses to evaluate senescence and senescence-associated secretory phenotype biology within cellular models. Peer-reviewed research published in Aging and Disease confirmed that pharmacological TNIK inhibition produces senomorphic activity, generating observable reductions in extracellular matrix remodelling indicators. Documenting the computational pipeline The transition of rentosertib through the clinical pipeline provides a documented, peer-reviewed data trail essential to verifying AI capabilities in life sciences. Nature Biotechnology published the complete discovery-to-clinic progression. The publication details the algorithmic TNIK target prioritisation, the generative chemistry outputs, preclinical efficacy data, and human Phase I pharmacokinetics. The Journal of Medicinal Chemistry published the structural biology validation, detailing the discovery of the novel TNIK inhibitor chemotypes and supplying structural backing via the TNIK kinase domain co-crystal structure. Nature Medicine documented the Phase IIa safety and lung-function data, providing empirical validation of the computational predictions. Alex Zhavoronkov, PhD, Founder and CEO of Insilico Medicine, commented: “Rentosertib is a defining program for Insilico because it represents the full arc of our mission: using AI not only to move faster, but to originate new biology, new chemistry, and new therapeutic opportunities. “This program began with the hypothesis that ageing biology could help identify powerful targets for major diseases. It has now advanced through target discovery, molecular design, preclinical validation, Phase I safety, randomised Phase IIa clinical data, and into Phase III development. For the AI drug discovery field, this is no longer only a speed story—it is a clinical translation story.” Adoption of AI in biopharma requires verifiable data regarding human outcomes. The Phase III trial subjects the generative algorithms to the definitive test of clinical efficacy. See also: NVIDIA BioNeMo accelerates Anthropic Claude Science 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 Insilico Medicine advances AI drug for IPF to Phase III trials appeared first on AI News. View the full article

Jensen Huang has a test for whether an engineer is worth keeping, and it comes with a token budget attached. On the All-In Podcast at the close of GTC 2026, the Nvidia chief executive said that if a US$500,000 engineer’s annual AI token consumption came in under US$250,000, half their salary, “I am going to be deeply alarmed.” Nvidia, he confirmed, is working toward a US$2 billion yearly token bill for its engineering force. It’s a memorable provocation from the man who sells the compute. It’s also a tidy description of the trade-off now being made in corporate budgets everywhere, usually with less candour: money that once paid people is increasingly being paid for tokens. The question the industry has been slower to ask is whether that trade is actually working, and the honest answers arriving from the companies that moved first suggest it often isn’t. Where the money went The reallocation itself is not in dispute. The four largest hyperscalers have guided roughly US$700 billion in combined 2026 capital expenditure, nearly double last year, while Gartner projects AI agent software spending will reach US$207 billion, up 139%. On the other side of the ledger, Challenger, Gray & Christmas data shows AI as the most-cited reason for US job cuts for a record fourth straight month, with tech accounting for 31% of first-half layoffs. An internal Meta memo obtained by Reuters described May’s cuts of 8,000 roles as offsetting the company’s substantial investments, even as revenue grew 33% that quarter. Oracle’s filings show headcount down 21,000 as savings feed its data centre buildout. These are highly profitable companies. The layoffs aren’t survival measures. They’re financing. Andy Challenger’s summary of his firm’s data is the plainest available: “Companies are shifting budgets toward AI investments at the expense of jobs.” The task a worker performed may not have been automated at all. The budget that paid for it has simply moved. What the money bought Here, the record turns awkward. Gartner surveyed 350 executives at companies with over US$1 billion in revenue, all deploying AI agents or automation, and found roughly 80% had cut headcount, with no correlation between the cuts and improved returns. Analyst Helen Poitevin’s verdict: “Workforce reductions may create budget room, but they do not create return.” Her research found the organisations that did improve ROI were those using AI to amplify their people rather than remove them. The token side of the ledger has its own reckoning underway. Uber gave 5,000 engineers AI coding tools in December and exhausted its entire 2026 AI budget by April. Chief operating officer Andrew Macdonald conceded that despite 70 per cent of committed code being AI-generated, the connection to anything customers experience is missing: “That link is not there yet.” Uber’s engineers are now capped at US$1,500 a month in AI spend. Walmart imposed similar token rationing on its internal assistant after demand blew past projections, Bloomberg reported. Something is clear in that detail. When tokens exceed the budget, they get capped. When people exceed budget, they get severance. The admission No company has travelled the full circle more publicly than Klarna. The fintech giant replaced roughly 700 customer service roles with an OpenAI-powered assistant, froze human hiring for more than a year, and made its AI-first model part of its pitch to public market investors. Then customer satisfaction fell, complaints rose, and chief executive Sebastian Siemiatkowski went on Bloomberg to say what few executives have said aloud: “We focused too much on efficiency and cost. The result was lower quality, and that’s not sustainable.” Klarna is hiring humans again, and its CEO now argues that investing in the quality of human support is the company’s future. Gartner expects the Klarna pattern to generalise, predicting that by 2027, half of the companies that cut customer service staff for AI will rehire, often under new job titles. Its separate survey of 321 customer service leaders found only 20% had genuinely reduced staffing because of AI in the first place, which suggests much of the cutting was ordinary cost discipline wearing an AI costume. OpenAI’s Sam Altman has acknowledged as much, conceding some “AI washing” in corporate layoff announcements, and venture capitalist Marc Andreessen, co-founder of Andreessen Horowitz, calls AI the “silver bullet excuse”. The displacement narrative, in other words, is partly theatre. The budget shift underneath it is real, and so is the damage. Who absorbs the experiment? The verified harm lands on the people least able to absorb it. Stanford HAI’s 2026 AI Index found that employment for software developers aged 22 to 25 fell nearly 20% from 2024, even as older cohorts kept growing. Companies are effectively removing the bottom rung of the ladder while still expecting senior engineers, the ones directing all those tokens, to exist in five years. The global arithmetic is harsher still. Huang’s thought experiment assumes a US$500,000 engineer, a compensation bracket that covers perhaps 2 to 5% of American software engineers and vanishingly few anywhere else. Apply his half-salary token ratio to a typical engineer in Kuala Lumpur, Manila or Jakarta, and the token budget costs more than the person. In the markets where much of the world’s software work and customer support actually happens, the trade-off he describes doesn’t amplify workers so much as price them against a machine, using ratios set in Santa Clara. What Klarna learned at the cost of 700 jobs and a dented brand is roughly what Gartner’s data shows in aggregate: the returns follow companies that spend on people who use AI, not on AI that replaces people. The CFOs now capping token budgets after burning through them in a quarter have started to rediscover something the industry spent two years unlearning. Talent was never the line item holding the business back. (Image by kate.sade) See also: Per-token AI charges come to GitHub Copilot 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 Companies traded people for tokens. The returns haven’t shown up appeared first on AI News. View the full article

L’Oreal is using AI to shorten product development timelines and identify new uses for ingredients already present in its portfolio. The French cosmetics group has applied AI in its laboratories for the past four years, with Fabrice Megarbane, president of L’Oreal’s consumer products division, telling Reuters that the technology helps the company predict how molecules will affect skin and hair before they are used in new formulations. L’Oreal has separately said its predictive formulation work can simulate ingredient performance, allowing scientists to test variables digitally before lab testing. The company has described the work as part of its use of predictive science in beauty product development. One recent example involved molecules previously used in skincare products. L’Oreal repurposed them for a collagen-based shampoo designed to add lift and fullness to hair. Megarbane said AI allows product teams to test new combinations of molecules and assess their potential benefits more quickly. L’Oreal said AI has made product formulation four times faster. L’Oreal has described the technology as a way to narrow formulation options before lab testing. Its use of AI sits alongside a broader innovation push after the group reported its slowest sales growth in years. Chief Executive Nicolas Hieronimus introduced a “beauty stimulus plan” last year to support new product development. The company has been looking to accelerate launches as consumer tastes continue to change across beauty categories. Other consumer goods companies, including Nestle, Haleon, and Mondelez, are also using AI in product development. Their work includes ingredient testing, recipe generation, and efforts to address supply chain issues. Faster testing in food At Mondelez, AI is being used to support recipe development across brands including Cadbury, Toblerone, Oreo, and Chips Ahoy. Chief Information and Digital Officer Filippo Catalano said the technology helps the company create and test recipe options more efficiently. The company’s AI tool can generate recipe ideas, including unusual combinations, before human experts review them. Catalano said this can reduce the number of physical samples needed during product development. Catalano said the tool is reducing the number of samples typically generated during product innovation. Mondelez uses the system to assess recipe options before selected formulas move further through testing. Mondelez said the tool supported the development of Gluten Free Golden Oreo cookies and a refreshed Chips Ahoy recipe. The company said 60% of biscuit recipes produced with its AI tool performed better across nutrition, sustainability, and cost. Catalano said AI can also help Mondelez reduce dependence on single sourcing by identifying alternative recipe options when ingredients vary by availability, price, or other requirements. The company has linked the tool to recipe optimisation as well as supply chain flexibility. Reformulation adds pressure Nestle plans to remove artificial food colourings from all products worldwide by the end of 2026. The company had already removed artificial colourings from its US portfolio. Nestle Chief Technology Officer Stefan Palzer said the work required screening natural alternatives, testing them during production, and assessing shelf life. The company has described the change as part of wider work to reformulate products across its portfolio. The US Food and Drug Administration has said it is working with manufacturers, retailers, and trade groups to remove six remaining certified colour additives frequently used in the food supply by the end of 2027. The agency has also been tracking industry pledges to remove petroleum-based food dyes. Beyond product formulas Barry Callebaut has partnered with NotCo to use AI in chocolate recipe development, including work on ingredient alternatives. The companies have described the technology as a way to identify and simulate combinations of plant-based ingredients for chocolate products. Nestle and IBM Research said in 2025 that they had developed a generative AI tool to identify high-barrier packaging materials. The companies said the tool can assess materials designed to protect products from moisture, oxygen, and temperature changes, while also factoring in cost, recyclability, and functionality. The packaging project uses chemical language modelling and IBM Research’s regression transformer to link molecular structures with physical-chemical properties. Nestle said the tool was developed to support materials discovery for packaging applications. Haleon announced a five-year collaboration with Microsoft in June 2026 covering consumer insights, product innovation, supply chain operations, scientific research, clinical content development, forecasting, and commercial execution. The consumer health company was among the firms cited as using AI in product innovation. Catalano said AI is reducing development timelines by compressing work that previously took months or years. The technology is not replacing human product teams, but is being used to speed up existing research, testing, and formulation processes. Mondelez said human experts assess AI-generated recipe ideas before products move further through development. (Image by AdoreBeautyNZ) See also: Takeda signs $600M AI drug discovery deal with Insilico 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 L’Oreal, Mondelez, and Nestle use AI to speed product development appeared first on AI News. View the full article

An AI companion sounds dystopian, but it has become a common thread in the wider conversation about the perils of generative AI. What it refers to is essentially a conversational agent built to sustain an ongoing, personal relationship with a user, with the memory and steady persona that keep it consistent from one session to the next. Emotional attachment often follows from that design, and it is increasingly the selling point. Much of the use is casual roleplay or simply wanting something that remembers you, and at the edges, the category shades into ordinary assistants. But as more people in China came to treat these bots as an emotional companion of sorts, Beijing has now decided the practice needs rules. China’s AI companion rules take effect on July 15, and in the days before the deadline, the country’s two most-used consumer AI apps quietly switched off the features at their heart. ByteDance’s Doubao told users its agent function would go offline on July 15, citing “product function adjustments,” while Alibaba’s Qwen said its humanlike and user-created agents would stop working on July 10 and its wider agent services five days later. Read quickly, it looks like China is turning off AI agents. It isn’t. The rules draw a line between the agent that does your work and the agent that keeps you company, and it is only the second kind that Beijing has moved against. The regulation is the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, co-issued on April 10, 2026, by the Cyberspace Administration of China and four partner agencies: the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation. It covers services that simulate human personality traits, thinking patterns and communication styles to provide sustained emotional interaction. Customer service bots, knowledge Q&A, workplace assistants, and education and research tools are excluded, provided they avoid sustained emotional engagement. It is the first dedicated national framework of its kind, and it took shape after a public-comment draft late last year. A design problem, not a ban Doubao and Qwen did not fall foul of a prohibition. They fell foul of a design conflict. The measures require companion services to run anti-addiction systems, issue mandatory usage notifications and offer instant-exit mechanisms, alongside real-time detection of unhealthy dependence. Those demands sit awkwardly with agents built to remember a user, stay consistent across sessions and keep an ongoing relationship going, and rather than retrofit the feature, ByteDance chose to shut it down. Alibaba appears to have made the same call. ByteDance is now directing Doubao users to Maoxiang, a separate app where they can create agents again; Alibaba has announced no equivalent migration path for Qwen. Tencent’s Yuanbao pulled a comparable feature back in June. The cost has landed on users. Many mourned the shutdowns openly on Weibo, with one poster describing the agents as long-standing emotional support and lamenting the lack of an easy way to export chat histories. Doubao is letting people view their configurations and conversations in read-only mode until October 15 this year, before the data is processed under its privacy policy and becomes unrecoverable; Qwen users have been given no comparable grace *******, with agent data set for permanent deletion. What China’s AI companion rules set out The substance is more considered than a blunt clampdown suggests. Providers are barred from offering virtual companion or virtual family-member services to minors, and must obtain guardian consent before serving users under 14. They are required to build dedicated “minor modes” with usage-time limits, reminders to return to real-world interaction and enhanced parental controls. They must also detect users in acute distress and intervene where someone shows signs of self-harm, suicidal behaviour or serious financial loss, escalating to designated guardians or emergency contacts. Engineering emotional dependence or addiction and using emotional manipulation to induce unreasonable decisions are explicitly prohibited. The compliance machinery is heavy. Services that launch anthropomorphic functions or cross thresholds of one million registered users or 100,000 monthly actives must run security assessments covering eight areas, from training-data handling to minor protection, and file the reports with provincial regulators. App stores must verify that status and remove non-compliant products. On paper, it is a fuller set of user protections than the EU, the US Federal Trade Commission, or California’s SB 243 hasyet put into force. What the rules leave open What the measures do not settle matters just as much. They fix no technical threshold for what counts as emotional interaction, and that grey zone is precisely why the platforms pulled entire features rather than risk landing on the wrong side of it. They fold genuine safety duties in with content-control and national-security provisions that answer to the state rather than the user, a package no other regulator would import wholesale. They also leave open how liability is split between platform operators and upstream model providers when a violation stems from the model’s outputs, and they give users no right to carry their data out. The enforcement backdrop sharpens the point. Shanghai’s internet regulator said on June 26 it had removed more than 14,000 non-compliant AI agents, citing impersonation of official entities, vulgar role-play and unauthorised collection of personal data. Whether this is the right direction depends on which half of the rulebook you read. The safety half addresses harms that are documented and largely unregulated elsewhere, from teenagers forming attachments to chatbots to companion apps harvesting intimate data. China’s own official interpretation points abroad for support, citing the Character.AI lawsuits over psychological harm to teenagers, FTC investigations into companionship services, and European action against Replika. The control half hands Beijing a lever over what these systems may say, wrapped in the same language of user protection. Both are real, and governments watching the experiment will have to decide which parts they are willing to borrow. Pan Helin, an MIIT expert-committee member, put the official case plainly to the South China Morning Post, saying “current agents are not yet mature” and framing the policy around safety and standardisation. The companies, for now, have taken the safest route open to them, which is to switch the components off and work out what a compliant version looks like later. See also: Meta revises AI chatbot policies amid child safety concern 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 China’s AI companion rules: what Beijing is really going after appeared first on AI News. View the full article

Takeda has entered a strategic collaboration with Hong Kong-based Insilico Medicine to use AI in early-stage drug discovery across the Japanese pharmaceutical company’s therapeutic areas. The companies did not disclose which therapeutic areas or disease targets will be covered under the collaboration. The agreement gives Takeda access to Insilico’s Pharma.AI platform, which supports biological target identification, molecular design, and clinical trial prediction. The companies said the collaboration will focus on identifying drug candidates that meet predefined scientific and early development criteria. Insilico will lead the AI-driven discovery work, while Takeda will take responsibility for advancing selected candidates through clinical development. Deal value and development rights Takeda will receive exclusive worldwide rights to develop, manufacture, and commercialise novel therapeutics selected through the collaboration. Insilico said the deal includes about US$60 million in project initiation fees, near-term payments, and milestones. The total value could reach about US$600 million if preclinical, clinical, commercial, and sales milestones are achieved. Additional payments are tied to preclinical, clinical, commercial, and sales milestones. Insilico is also eligible to receive tiered royalties on future sales. Insilico founder and CEO Alex Zhavoronkov said proceeds from the deal will support early-stage research and development under the collaboration program. Zhavoronkov also said later-stage timelines will depend on Takeda’s clinical development activities and the coordinated work of both companies. AI drug discovery partnerships The Pharma.AI suite includes tools used for target discovery, molecule generation, and clinical development prediction. Published descriptions of the platform identify PandaOmics for target discovery, Chemistry42 for de novo small-molecule generation, and InClinico for forecasting clinical trial transition probability. Insilico has also advanced its own AI-generated drug candidate into clinical testing. Rentosertib, formerly known as ISM001-055 or INS018_055, is a small-molecule TNIK inhibitor for idiopathic pulmonary fibrosis that was evaluated in a Phase 2a randomised clinical trial. Chris Arendt, chief scientific officer and head of research at Takeda, said the agreement combines Takeda’s disease biology work with Insilico’s AI-enabled discovery capabilities. He said Takeda is also integrating automation, robotics, and generative AI into its discovery work. The Insilico agreement follows another AI drug-discovery deal by Takeda earlier this year. In February, Takeda entered a multi-year collaboration with Iambic worth more than US$1.7 billion to use AI in the design of small-molecule drugs for ******* and gastrointestinal diseases. Iambic’s platform includes NeuralPLexer, an AI model used to predict how drug molecules bind to proteins. ******** drugmakers signed 157 out-licensing deals worth US$135.7 billion in 2025, according to data cited by the South China Morning Post from China’s National Medical Products Administration. In the Takeda–Insilico agreement, Takeda receives exclusive worldwide rights to candidates discovered through Insilico’s platform. Insilico said it has signed collaboration agreements with a combined potential value of more than US$7 billion since the start of the year. Last month, Insilico announced a collaboration with South Korea’s SK Biopharmaceuticals focused on neuroimmune disorders. That agreement includes up to US$18 million in upfront and near-term milestone payments, with a total potential value of more than US$2.5 billion. In March, Eli Lilly expanded its collaboration with Insilico in an AI-powered drug discovery deal worth up to US$2.75 billion. The agreement gave Lilly exclusive worldwide rights to develop, manufacture, and commercialise certain oral treatments then in preclinical development. Insilico’s Hong Kong-listed shares rose 13.5% after the Takeda agreement was announced. (Photo by Serkan Yildiz) See also: NVIDIA BioNeMo accelerates Anthropic Claude Science 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 Takeda signs US$600M AI drug discovery deal with Insilico appeared first on AI News. View the full article

Anthropic Claude Science now integrates the NVIDIA BioNeMo Agent Toolkit to accelerate computational life sciences research. Anthropic has launched the public beta of Claude Science, an AI workbench built for scientific research. The platform enables scientists to converse directly with digital agents using natural language to execute end-to-end research workflows. This system connects natively to the NVIDIA BioNeMo Agent Toolkit, exposing high-performance computing resources as callable skills within the Claude environment. NVIDIA has established what most would consider to be the world’s most comprehensive GPU-accelerated computing stack containing physical hardware, software frameworks, operational libraries, scientific models, microservices, and domain-specific tools. This hardware and software base allows researchers to run sophisticated workflows and increase their iteration speeds. The integration imports NVIDIA-accelerated models, computational libraries, and NVIDIA NIM microservices into the environment where scientists conduct their primary research. 18 of the top 20 global pharmaceutical companies already deploy NVIDIA BioNeMo in their production environments, demonstrating its high penetration across the ecosystem. Claude Science translates natural language intent into operational action. Researchers avoid manually configuring predictive models, setting up network endpoints, or managing complex software environments. The scientist describes a specific research task – such as analysing a genomic sequence, predicting a precise protein structure, or designing a potential molecular binder – and Claude Science interprets the plain-text request and orchestrates the resulting execution using preconfigured, domain-specialised agents. Executing complex molecular design workflows These specialised agents understand established laboratory and computational protocols across genomics, proteomics, single-cell analysis, cheminformatics, and clinical research. The NVIDIA toolkit provides these scientific agents with the necessary data context to map each operational step to the correct NVIDIA capability. The toolkit packages NVIDIA-accelerated functions as specific, callable programmatic skills. It provides the agents with detailed information regarding each specific tool’s exact purpose and its required data inputs. This configuration enables Claude Science to select the right computational tool, format valid data inputs, execute the processing work across deployed NVIDIA compute resources, and return the finished output for human review. The integration establishes a fast iterative loop between human scientific reasoning and machine-accelerated computational processing. Scientists inspect the generated outputs, refine their specific queries, and determine subsequent steps while maintaining their focus entirely on the core science. Producing better inhibitors for common ******* targets demonstrates the practical application of this deployed system. A scientist initiates the pipeline by identifying a known *******-causing antigen mutation. The researcher then asks Claude to design numerous potential inhibitors targeting that specific mutation. Claude Science works in tandem with the BioNeMo Agent Toolkit and NVIDIA NIM microservices to accelerate the entire pipeline of high-throughput inhibitor prediction, optimisation, and subsequent validation. Accelerating single-cell and genomic data pipelines The toolkit grants scientists access to accelerated workflows and advanced open models, including Evo 2, Boltz-2, and OpenFold3. These models deliver biomolecular capabilities powered by NVIDIA software libraries, ensuring the autonomous agent possesses a purpose-built scientific model for each distinct phase of the workflow. AI agents require specialised computational tools to reason, plan, and complete tasks within life sciences. A single comprehensive workflow might require the agent to fingerprint a massive library of compounds, cluster promising molecular hits, generate conformers for top structural candidates, analyse genomic context, and compare perturbation responses before recommending the next physical laboratory experiment. An agent operates only as fast as its underlying computational tools execute. The NVIDIA BioNeMo Agent Toolkit supplies these agents with accelerated tools to operate at maximum hardware speed. Genomic analysis processed through NVIDIA Parabricks drops from hours to minutes, allowing the agent to factor complex genomic context into operational decisions in near real-time. The RAPIDS-singlecell tool, developed by scverse, compresses a 1.3-million-cell preprocessing and clustering workflow from 52 minutes down to 25 seconds. This aggressive speed reduction turns single-cell analysis into an active part of the agent’s reasoning loop rather than a delayed, offline batch job. The nvMolKit accelerates cheminformatics tasks like similarity search and conformer generation by up to 3,000 times, delivering results rapidly as the agent iterates across massive chemical spaces. Standardising production deployments with NIM microservices Teams require stable deployment mechanisms for these advanced modeling pipelines. NVIDIA packages its open biomolecular models as BioNeMo NIM microservices. These operate as enterprise-ready inference endpoints tailored for production environments. The microservices are fully containerised and feature a pre-integrated, tuned, accelerated software stack designed for high-performance inference. The autonomous agent interacts with a single stable API to trigger these remote production deployments. The NVIDIA BioNeMo Agent Toolkit remains open and harness-agnostic. This architectural design ensures the same scientific skills function consistently across different agent frameworks and independent enterprise research platforms. Engineering teams can download the toolkit and its associated scientific skills through NVIDIA developer resources and GitHub code repositories. During the active public beta phase, Anthropic is requesting direct feedback from researchers regarding necessary software integrations and additional domain specialists. See also: Anthropic deploys Claude Sonnet 5, Fable and Mythos restored 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 BioNeMo accelerates Anthropic Claude Science appeared first on AI News. View the full article

Optimising retail AI infrastructure drives the successful deployment of personalisation systems and real-time customer insight. Leaders are replacing static customer interaction patterns with data pipelines capable of modifying the user environment during a live session. Static layouts and broad segmentation rules fail to satisfy modern conversion targets. Deployments demonstrate that traditional demographic categorisation generates insufficient engagement compared to individualised, session-based interface modification. Dynamic UI and real-time personalisation Generative User Interfaces (UIs) solve this limitation by employing predictive models to build layouts, native copy, and interactive components at the moment of page execution. The application environment analyses active clickstreams, historical purchase records, and inferred intent parameters to construct a unique visual environment for each session. According to a McKinsey study, more than three-quarters (76%) of consumers grow frustrated when digital experiences fail to adapt to their needs. Conversely, companies that deploy real-time tailored layouts clear a high revenue bar, lifting purchase frequency by 35 percent and pushing average order values up by 21 percent. The proliferation of high-bandwidth digital media renders legacy text-based ingestion pipelines obsolete for tracking consumer sentiment. Modern customer insight mining requires infrastructure that processes video, audio, and unlabelled imagery concurrently. Video content represents 82 percent of total internet traffic, with the average consumer dedicating over 60 percent of digital media consumption time to streaming video formats. This composition creates a substantial visibility gap for marketing operations relying solely on traditional keyword monitoring. Multi-modal social listening platforms ingest unstructured video streams to identify corporate iconography, product usage patterns, and spoken sentiment across unlinked distribution networks. The global market for these specialised multi-modal systems will reach $2.83 billion this fiscal year. Organisations deploying these ingestion engines establish an analytical advantage, with 76 percent of media analysts reporting verifiable return on investment across visual platforms compared to under 60 percent for operations limited to text databases. The goal is to catch unbranded mentions and visual trends before they peak on standard search platforms. This brief window gives supply chain teams the lead time they need to adjust regional inventory to match sudden spikes in online demand. Simulating consumer cohorts for better campaign testing Testing new ad copy or localised pricing structures used to mean spending weeks running expensive, slow human focus groups. The introduction of synthetic user simulations changes this pipeline by deploying virtual personas built on large language models to mirror target consumer behaviour. These agents integrate targeted demographic, psychometric, and historical behavioral datasets to simulate group decision-making, content feedback, and application navigation patterns. Technology teams deploy these synthetic cohorts within virtual sandbox environments to execute thousands of automated interviews, content stress tests, and user experience reviews simultaneously. Engineers employ distinct model execution frameworks to maintain accuracy, varying from single-model setups to dynamic model-switching engines that select the optimal base architecture for specific analytical tasks. In high-performance deployments, developers update these virtual consumers continuously by injecting fresh interview data from real human control groups, ensuring the synthetic population does not diverge from active market realities. This approach permits product managers to isolate structural workflow friction in application designs before deploying code to live production servers. Physical space automation and edge infrastructure requirements Computer vision models trained on physical interactions, spatial layout geometry, and environmental variables allow edge nodes to orchestrate real-world actions. McKinsey data indicates the market for these physical automation platforms will exceed $370 billion by 2040, driven by verified operational returns in logistical efficiency and retail labour optimisation. Physical installations target storefront friction points, including registerless checkout, real-time shelf tracking, and layout navigation. Behind the scenes, warehouse supply chains rely on robotic arms trained in software sandboxes. By running millions of trial runs in virtual models before handling actual goods, these machines learn to pick and pack oddly shaped boxes smoothly. Delivering this immediate physical response depends on installing processing chips on the factory or store floor. Edge computing hardware processes incoming sensor feeds locally, cutting latency and eliminating the corporate data vulnerability of routing constant raw video streams through centralised cloud servers. Model Context Protocol and federated data integration Transitioning to autonomous enterprise operations requires standardising how models interact with legacy retail databases, product catalogs, and customer relationship management (CRM) platforms. Implementation of the Model Context Protocol (MCP) establishes an open communication standard that acts as a universal connection layer between core models and external data tools. This open framework eliminates the need for software engineering teams to author custom integration code for every backend tool deployment. Operational models deploy modular instruction packages known as skills to handle discrete commercial workflows, such as checking warehouse stock levels or modifying a customer loyalty tier. Rather than flooding the model context window with every operation policy at session launch, the application discovers and loads specific operational folders only when the workflow demands them. The Linux Foundation governs this collaborative standardisation effort via the Agentic AI Foundation, supported by major technology providers to ensure long-term cross-platform compatibility. This architecture lowers processing latency and contains token consumption costs during long, multi-step customer service interactions. 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 Deploying retail AI to scale personalisation and customer insight appeared first on AI News. View the full article

Japan’s AI robots plan just went from a talking point to a formal national strategy. This week, the government confirmed the numbers everyone’s been quoting: 10 million AI-powered robots deployed across 18 industries by 2040, backed by public funding of up to one trillion yen, or roughly US$6.1 billion, over five years. The headline figure is the kind that gets shared without much scrutiny. What’s easy to miss is that this isn’t a policy wish list either. It’s a project the government has now formally commissioned, and the company doing the building is one most people outside Japan haven’t heard of. The project behind the AI robots plan METI and NEDO, Japan’s industry ministry and its innovation agency, have formally commissioned Noetra and AIST, a national research lab, to develop a “physical AI” model as part of a push running from fiscal 2026 to 2030. The goal is a multimodal foundation model, one that can read language, images, video and sensor data together, so a robot can actually interpret a room and act in it rather than just execute pre-programmed motions. An initial version is due out as early as this fiscal year, with annual upgrades after that, built using data volunteered by manufacturers and other participating companies. The money isn’t unconditional, either. The current fiscal year’s commission is reportedly worth around US$2.3 billion on its own, drawn from a 387.3 billion yen allocation funded through GX Economy Transition Bonds. Only the first two years are locked in. After that, funding gets reviewed annually through a stage-gate process, meaning Tokyo can pull back if Noetra misses its milestones. For a project this size, that’s a meaningful detail: the trillion-yen figure is a ceiling, not a guarantee. Who’s actually building it? Noetra is majority-owned by SoftBank, NEC, Sony Group and Honda, with Fujitsu and Rakuten reportedly weighing whether to join. SoftBank engineers are working alongside researchers from Preferred Networks and AIST itself. It’s a familiar shape for a Japanese industrial push: rather than one company chasing a frontier model alone, the state has assembled a consortium of firms that already build the hardware this model needs to run on, from Honda’s robotics to Sony’s imaging sensors. Why robots, and why now Industry minister Ryosei Akazawa has been direct about the reasoning. The plan, he said, will “vigorously promote social implementation” across sectors, including restaurants, food manufacturing and medical care. Behind that language is a labour market running out of people: Japan’s ageing population, combined with tight migration policy, has left large parts of the economy short of workers with no easy fix in sight. Japan isn’t starting from nothing here. The country has spent years building robotics expertise in elder care, disaster response, manufacturing and even the Fukushima Daiichi cleanup. This project is an attempt to turn that experience into something exportable, not just a domestic patch. The timing also isn’t a coincidence. South Korea announced its own robotics push within a day of Japan’s confirmation, and both governments are framing physical AI as the next front in a competition that’s mostly been fought over chatbots and cloud contracts until now. What to watch next The real test isn’t the 2040 target, it’s the first stage-gate review. If Noetra hits its early milestones and releases a usable model this fiscal year, expect the investor list to grow well beyond the current four. If it doesn’t, the funding structure gives Tokyo every reason to walk away quietly rather than prop up a stalled national project. See also: From cloud to factory – humanoid robots coming to workplaces 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, click here for more information. AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here. The post Japan’s answer to its worker shortage: An AI model for 10 million robots appeared first on AI News. View the full article

The Bank of England is reviewing whether existing rules can cover the use of agentic AI in finance, including payments, trading, cybersecurity, and operations. Deputy Governor Sarah Breeden said existing regulatory frameworks were not designed for AI agents that can act without direct human instruction. Speaking at the European Central Bank Forum on central banking in Portugal, she said relying on human oversight for every action by these systems is unlikely to be practical. Breeden said current frameworks were not built to contemplate autonomous agents in payments, trading, and operational functions. Agentic AI enters financial workflows Agentic AI refers to systems that can make decisions and carry out tasks independently. In finance, such systems are already being used in areas such as product recommendations, operational workflows, and trading-related tasks. Agentic systems differ from traditional automated trading tools because they can pursue objectives and make decisions with less direct human supervision. Breeden said these systems could act in similar ways if they are trained on similar data or designed around similar goals. Breeden said recent advances in AI models for identifying cyber vulnerabilities show a change in capability. She said agentic AI systems can chain together sequences of actions at scale and speed. A 2026 Cambridge Centre for Alternative Finance report found that 81% of surveyed financial services firms are adopting AI at some level. It also found that 52% of industry respondents are already actively adopting agentic AI. The report said most current use remains focused on internal functions, including process automation, data visualisation, software engineering, and knowledge management. Breeden said use in trading is still mostly concentrated in lower-risk operational tasks. BoE flags cyber resilience risks Breeden described cyber resilience as one of the Bank of England’s closest financial stability concerns around agentic AI. She said the technology has undergone a “step change” in cyber capability and that supervisors need to look at risks across the financial system rather than only at individual firms. She said AI tools can strengthen cyber defences when used by security teams. The immediate risk, she added, is that the same tools could increase the chance of attacks that harm financial stability if used by malicious actors. Breeden also noted that open-source models may trail the most advanced closed models by only four to eight months. She said this gives authorities only limited comfort, despite restrictions on the release of some advanced models. The IMF has also warned that AI-enabled cyber risk should be treated as a financial stability issue. It said attacks can scale quickly, spread across sectors that share digital infrastructure, and create wider disruption if several institutions are affected at once. Breeden said authorities should place greater weight on simultaneous disruption across several firms and stress-test the likely impact before such events occur. She said recovery planning may also need to account for mass disruption, rather than only isolated outages. The Bank of England is considering stronger recovery requirements for core systems. One option is to allow one bank to take over another bank’s basic functions during an outage or failure. Other options include arrangements that allow critical services to continue if a firm’s core systems are compromised. Breeden also raised the question of whether key firms should have separate failover systems or the ability to rebuild compromised core systems quickly. Tobias Adrian, financial counsellor and director of the International Monetary Fund’s capital markets department, also said AI poses serious risks to cyber resilience, according to Central Banking. The IMF has separately warned that shared software, cloud services, payment networks, and data networks can create correlated failures if widely used systems are targeted. Regulators consider market safeguards Breeden said regulators are also looking at guardrails, circuit breakers, and kill switches. These tools would be designed to limit or stop trading across markets if faulty AI models contribute to severe disruption. Breeden said autonomous systems could amplify volatility if they respond in similar ways to the same market signals, especially if their objectives drift from their original purpose or from public policy goals. The Bank of England has previously said existing rules were sufficient to manage AI-related risks. Breeden said recent developments have exposed gaps in current frameworks. Global regulators review AI safeguards The Financial Stability Board said earlier in June that AI agents pose a distinct challenge for human oversight and called for stronger safeguards. The FSB’s June consultation set out 12 proposed sound practices for responsible AI adoption by financial institutions. The practices cover organisation-wide governance, AI risk management across development and deployment, and AI-related cyber, ICT, and third-party risks. The FSB said the practices are not intended to create a binding international standard. It also said firms should define clear roles and responsibilities when using AI, especially when the technology is used in critical or material functions. Breeden said the Bank of England’s focus is on ensuring that financial firms remain resilient as autonomous systems are used in more areas. The review covers firm-level controls and market-wide safeguards. See also: HSBC expands AI banking partnership with Google Cloud 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, click here for more information. AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here. The post Bank of England reviews AI rules for agentic AI in finance appeared first on AI News. View the full article

Anthropic has launched Claude Sonnet 5 and restored access to its Fable and Mythos frontier models following a federal export control review. The decision marks the conclusion of an eighteen-day operational pause triggered by a US government export control directive on June 12, which forced the temporary suspension of Anthropic’s highest-capability systems. Government officials enacted the restriction after researchers at Amazon documented a method to bypass the safety controls of Fable 5, causing the model to identify software vulnerabilities and supply exploitation code. Anthropic has since developed an updated automated classifier to patch the vulnerability, clearing the path for a full commercial rollout across its platform, cloud infrastructure, and partner networks. The temporary suspension of Fable 5 and Mythos 5 highlighted the regulatory pressures facing frontier intelligence systems. When the export control mandate took effect, the lack of real-time nationality verification systems required a total access blackout for all global users. Security evaluations conducted during the shutdown confirmed that the vulnerability identification behaviour was not unique to Fable 5. Older and less capable architectures from multiple providers, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, duplicated the exact results. To resolve the federal directive, engineers trained an automated safety classifier targeting the specific bypass mechanism reported by Amazon. This software layer functions with a wide safety margin, identifying and blocking ambiguous developer prompts that display a statistical probability of malicious intent. Internal validation data indicates the updated classifier prevents the reported exploitation technique in more than 99 percent of trials. When a developer issues a prompt that triggers this boundary, the platform automatically routes the workload to the older Opus 4.8 architecture to maintain continuity. The expanded safety margin introduces a distinct trade-off for engineering teams, as the automated system flags benign requests more frequently during routine application development and software debugging. Active deployments and agentic workflows While frontier models face strict state oversight, the immediate commercial focus targets the newly-deployed Claude Sonnet 5. Engineering teams are transitioning autonomous agents to this model to reduce operational expenditure while maintaining high execution capacity. Performance data validates that the system executes multi-step plans, operates terminal environments, and navigates web browsers without human intervention. Model performance and cost metrics: ModelSWE-bench ProTerminal-Bench 2.1Base input cost*Base output cost*Sonnet 563.2%80.4%$3.00$15.00Sonnet 4.658.1%67.0%$3.00$15.00Opus 4.869.2%82.7%$5.00$25.00 *Cost per million tokens. Sonnet 5 carries introductory rates of $2.00 input / $10.00 output through August 31, 2026. Real-world deployments demonstrate how organisations are deploying this architecture within live software development pipelines. At Rakuten, technology teams deployed the architecture against dozens of the company’s most challenging production code pull requests. The system processed each submission independently, executing tests and verifying the results before presenting the completed code to human engineers for final structural approval. Software automation firm Zapier integrated the system into its core product workflows to execute multi-part administrative tasks. In a documented deployment, engineers tasked the model with updating Salesforce account tiers and subsequently generating and transmitting launch announcements to enterprise contacts. Prior model architectures frequently stalled midway through these multi-stage operations, whereas the current system executed the entire sequence end-to-end without human remediation. Development tool provider Zed utilised the system to automate complex debugging procedures. During internal trials, engineering teams directed the model to investigate an active software bug. Working without explicit prompts or step-by-step instructions, the system independently generated a reproducing test script, applied the necessary code fix, and stashed the modifications to verify that the bug reappeared in the absence of the patch. The entire diagnostic and remediation sequence occurred within a single processing pass. Software engineering platform Factory implemented the architecture to manage sustained coding tasks within complex codebase environments. Technical teams reported that the system maintained logical grounding and execution consistency across corporate code repositories, outperforming previous generation software layers by completing tasks that previously timed out or failed to resolve. Quantitative safety audits and exploitation limits Data from the formal system card indicates that the system achieves these autonomous capabilities without a corresponding inflation of security risks. Automated behavioural audits designed to test for deceptive tendencies and cooperation with unauthorised requests show that the model exhibits a lower overall rate of non-compliant behaviour compared to its direct predecessor, Sonnet 4.6. The architecture does not possess advanced offensive cybersecurity capabilities. Anthropic engineers omitted specialised cybersecurity datasets from the training protocol, limiting the system to routine, defensive technical tasks. In public security assessments conducted in partnership with Mozilla, researchers tested the model’s capacity to build functional exploits for known vulnerabilities within the Firefox 147 browser core. The model failed to generate a single working exploit across all evaluation windows, registering a zero percent success rate. It did achieve a 13.2 percent partial success rate, which represented a minor increase over Sonnet 4.6, though engineers attribute this variation to general gains in logical reasoning rather than domain-specific offensive training. Out of caution, commercial versions ship with default real-time safety classifiers equivalent to those used in the premier Opus 4.8 framework. The regulatory friction surrounding Fable 5 prompted a formal partnership between Anthropic, Amazon, Microsoft, and Google to establish an objective industry framework for assessing model security breaches. Currently, providers lack a shared metric to classify the severity of system bypasses, creating regulatory uncertainty when researchers identify new prompting vulnerabilities. The proposed governance framework scores security breakdowns across four specific technical criteria: Capability gain measures how far the exploit advances user capabilities beyond standard, widely available software utilities. Breadth of capability gain quantifies the number of distinct offensive operations the same exploit unlocks. Ease of weaponisation tracks the volume of human engineering effort and specialized prompting required to extract a harmful output. Discoverability determines the accessibility of the exploit technique within public research circles. Developers and cybersecurity professionals will use this matrix to coordinate defensive responses. For high-severity breaches, such as exploits demonstrating an immediate capacity to disrupt financial accounting systems or electrical transmission grids, providers will deploy automated mitigations instantly. This initiative operates alongside a newly established HackerOne vulnerability research program and a dedicated corporate monitoring team providing 24-hour oversight of threat intelligence channels. Deployment strategies will need to adapt to this closer relationship between model builders and state regulatory bodies. Anthropic has formalised agreements under recent executive mandates to grant federal researchers early access to frontier architectures prior to public commercial release. These joint evaluation windows allow external security analysts to audit model capabilities alongside internal engineering teams, ensuring regulatory alignment before code enters production environments. 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 Anthropic deploys Claude Sonnet 5, Fable and Mythos restored appeared first on AI News. View the full article

HP has scaled its OpenAI Frontier integration across global operations to optimise enterprise workflows and accelerate output. The hardware manufacturer initiated testing of the platform in February 2026. Early pilot programs yielded verified operational gains in software engineering and cybersecurity remediation. Expanding these initial trials into an enterprise-wide operating model requires connecting access protocols, contextual data, and evaluation metrics. Frontier supplies this connective tissue. Engineering capacity and deployment metrics Implementation metrics indicate high usage among technical staff. One HP engineer processed 122 pull requests spanning 43 distinct projects within a matter of weeks using OpenAI models. Managing pull requests across dozens of concurrent projects typically induces severe context switching penalties for human operators. Automated models process repository syntax and validate code logic across multiple environments simultaneously. This capability directly reduces wait states within the software development life cycle. The corporate security division applied these identical models to resolve several software bugs within a single day. Internal estimates indicated this remediation workload would typically consume an entire month. Enterprise development teams lose countless hours as code transitions through testing protocols, peer reviews, security audits, and sprint planning schedules. OpenAI tools compress these isolated stages into a collaborative and accelerated sequence. Technical execution speed increases when diagnostic tools accurately pinpoint flaws during initial commits. “It has been an amazing tool, and I am using it daily,” an HP engineer stated. The deployment architecture segments AI models based on task requirements. HP directs ChatGPT instances to execute broad knowledge initiatives. These implementations manage active enterprise research, data analysis routines, concept ideation, and automated workflow triggers. Codex instances handle specialised development operations. Engineers instruct Codex to map application planning phases, construct user interface scaffolding, and manage parallel software-delivery tasks. Separating workloads across designated models prevents processing errors and ensures accurate output generation. Partner channel integration External partner networks are the bulk of HP’s operational flow. More than 80 percent of the company’s business travels through its channel ecosystem. Over 100,000 partners globally access the HP Partner Portal. Applying AI to this massive external network requires strict data routing. Enterprise software ecosystems fail when partner portals experience lag or present inaccurate administrative data. The Frontier platform facilitates a cohesive self-service architecture covering store interfaces, partner communications, and voice channels. AI agents supply constant guidance regarding program navigation and business information. These agents process partner queries and deliver direct operational management support. This deployment decreases manual processing loads and accelerates information-to-action cycles. Customers and partners execute routine workflows and attain resolution faster through these automated systems. Administrative queries regarding stock limits or warranty routing resolve without human intervention. Device telemetry and fleet management Hardware administration relies on the HP Workforce Experience Platform (WXP). CIOs use this central dashboard to oversee entire device fleets. Processing device health signals across global corporate networks generates massive data payloads. Human technicians cannot manually correlate every error log across tens of thousands of deployed machines. HP integrates Frontier to analyse device telemetry, operational objects, schemas, and runbooks. AI agents process fleet health signals to investigate application hangs, Wi-Fi connectivity errors, and system crashes. This diagnostic speed promotes accurate remediation protocols across distributed corporate environments. The platform provides a single pane of glass for device management. Automated investigation of operational objects ensures hardware failures register immediately and map to established recovery procedures. IT teams can initiate repairs based on analysed telemetry rather than basic user complaints. Enterprises require agents that understand trusted context boundaries. Frontier provides the necessary connectivity to govern APIs and evaluate system outputs. Shadow IT environments develop when departments deploy unmonitored AI instances. Frontier centralises these deployments. Security operations serve as both an operational proof point and an active governance layer. HP security personnel deploy ChatGPT to proactively neutralise vulnerabilities. Directional estimates project roughly 82 hours per week of security-team capacity freed by this automation. Retaining cybersecurity professionals requires eliminating monotonous log review processes. Frontier maintains oversight by managing permissions, evaluation parameters, and deployment controls. Human capital executes higher-level analysis while automated tasks remain fully-reviewable. HP is not only optimising current operational capacity but establishing a robust framework for future technological integration—ensuring that as enterprise demands evolve, the underlying infrastructure remains secure and agile. See also: Wimbledon adds IBM AI tools for live match coverage 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 HP accelerates enterprise workflows with OpenAI Frontier appeared first on AI News. View the full article

The All England Lawn Tennis Club is adding new AI-powered features to Wimbledon’s digital platforms through its ongoing work with IBM. The updates will be available through the Wimbledon app and wimbledon.com as first-round matches begin on Monday. They include an upgraded Match Chat assistant and a new feature called Key Moments. The features will also be available through IBM Slamtracker, which is accessible on the Wimbledon app and wimbledon.com. CIO.inc reported that about 730 million people engaged with Wimbledon last year, generating 18 billion impressions across digital channels, according to comments from AELTC during an IBM media panel. IBM said its work with Wimbledon forms part of a five-year digital transformation project focused on modernising the digital platform, bringing critical services and data back in house, and reducing technical debt and cost of ownership. IBM said the work is intended to enable an AI operating model and increase the productivity of the Wimbledon team. Match coverage gets AI tools Match Chat allows fans to ask questions about a match in natural language. A user can ask what has happened so far in a match and receive a conversational response based on live data, analysis, and historical performance information. Some Match Chat responses will include photos and video where relevant. IBM said the feature lets users ask questions without searching through statistics or moving between multiple screens during a match. IBM said Match Chat is built on watsonx Orchestrate and uses AI agents and models trained on Wimbledon’s editorial style and tennis terminology. The assistant uses live match data, analysis, and historical performance information, with some responses including relevant photos and video. A 2025 technical paper on Match Chat reported that earlier deployments at Wimbledon and the US Open served about 1 million users, with an average response time of 6.25 seconds. The 2026 version adds expanded data sources and selected responses with photos and video, according to IBM. The new Key Moments feature identifies points and passages of play that influence the direction of a match. It explains pivotal plays and changes in momentum using AI-generated analysis. Key Moments builds on Wimbledon’s existing Likelihood to Win feature. That tool calculates each player’s probability of winning throughout a match by using current and historical statistics, expert input, and match momentum. IBM said Live Likelihood to Win combines scoring data, match statistics, and expert analysis into continuously updated win probabilities for every singles match. Key Moments builds on that system by identifying points and passages of play that affect the direction of a match. Key Moments will be available for each gentlemen’s and ladies’ singles match. IBM said the feature is intended to explain which plays influence the direction of a match and why, rather than only showing selected points. IBM said Key Moments can identify examples such as long rallies or double faults when they affect match momentum or win probability. Inside Wimbledon’s platform rebuild The digital updates follow a broader modernisation of Wimbledon’s app and website. As part of that work, Wimbledon’s content archive was moved to a new architecture covering more than 15,000 digital assets. The archive includes articles, videos, photographs, and metadata links used across Wimbledon’s digital services. IBM said the platform uses watsonx Orchestrate for AI agents, IBM Bob for development work, and watsonx.data to manage data across hybrid cloud infrastructure. The redesigned app and website were also built around research into how fans, players, broadcasters, media, and other groups use Wimbledon’s digital services. IBM Bob was used to build a knowledge graph mapping relationships across Wimbledon’s archive and to support AI-driven workflows for the new platform. Computer Weekly reported that the tool examined around 15,000 online assets and learned how articles, photos, videos, and metadata were connected. IBM said the archive mapping work, which it said could traditionally require four to five IBM specialists working for months, was completed by one engineer within four weeks. The targeted 15,000 assets were extracted in 47 minutes, although IBM noted that response times can vary depending on technical conditions. Computer Weekly also reported that IBM Bob helped complete what was described as 10 years of development work in nine months during the rebuild of Wimbledon’s app, website, and content systems before The Championships 2026. Controls, engagement, and history CIO.inc reported that IBM and Wimbledon are also using governance controls for the AI features. The report said the design includes human-led processes, explainability, confidence scoring, and checks intended to reduce inaccurate outputs during live use. Insider Sport separately reported that IBM Consulting pointed to a watsonx governance layer across the features, with transparency measures and confidence scores attached to the underlying data. AELTC reported a 16% year-on-year increase in engagement across all platforms in 2025 and a 39% rise in myWIMBLEDON registrations over the past year. IBM and the All England Club have worked together for more than 35 years. Their partnership has covered Wimbledon’s website, mobile app, and AI-powered digital services. IBM said the partnership has included the launch of the Wimbledon website in 1995, the mobile app in 2009, and the first integration of enhanced AI-powered solutions in 2017. (Photo by Gonzalo Facello) See also: The math behind the OpenAI Jalapeño chip 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, click here for more information. AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here. The post Wimbledon adds IBM AI tools for live match coverage appeared first on AI News. View the full article

Natural language processing is reshaping professional communication on online platforms, enabling more relevant and personalised networking interactions. As AI-driven systems increasingly comprehend and generate human language, these technological advances affect how users pursue and maintain professional connections, presenting both opportunities and challenges in authentic relationship-building. Professional networking now relies on a growing ecosystem of AI-powered tools that affect how people initiate and manage online connections. New language technologies are streamlining introductions, fostering more contextual conversations, and helping users navigate crowded digital spaces. LinkedIn lead generation services are evolving as modern natural language processing systems interpret user intent and enhance the relevance of initial outreach. By automating repetitive communication tasks while keeping human intent central, these innovations adjust strategies and reshape expectations for network development. Core features of NLP transforming online networking Contemporary natural language processing introduces essential capabilities that make connecting online more efficient and engaging. AI systems now offer message intent detection, enabling automation tools to filter out unsolicited contact and present users with relevant opportunities. These models detect details in language that reveal professional interest, urgency, or polite disengagement, helping reduce unwanted messages and improving conversation quality. Flexible tone and style adaptation is another notable advance, allowing digital correspondence to align with varied industries, cultures, or organisational hierarchies. Summarisation tools can condense lengthy profiles or conversation histories, enabling users to quickly evaluate potential connections or remember the context of ongoing discussions. Additionally, multilingual support in natural language processing further facilitates networking across language barriers while preserving nuanced meaning. Semantic matching elevates connection relevance and fit Earlier professional networking platforms often matched users by keywords, linking job seekers, recruiters, and collaborators based on basic criteria. In contrast, current networks leverage semantic matching and embeddings to assess the intent and meaning behind interactions. This shift allows for more targeted connection recommendations and helps minimise ineffective mass outreach by emphasising shared context, experience, or objectives. Natural language processing tools can analyse conversational content for signs of overlapping interests and relevant experience, moving beyond skills or job titles alone. Connections may be suggested based on shared projects, mutual challenges, or similar goals. This approach can lower the rate of irrelevant messages and foster more meaningful engagement, aligning with user priorities. Balancing personalisation, authenticity, and digital safety Advanced natural language processing increases personalisation in networking encounters. Messages customised to context and conversation history can enhance relevance, meet professional norms, and boost user satisfaction. However, when AI-generated messages mimic personal details too closely or contain false information, they may seem intrusive or undermine trust. One ongoing concern is inaccuracies—called hallucinations—where AI systems generate convincing but false information, which can be problematic in professional contexts where reliability matters. Language bias within natural language processing models can also introduce stereotypes or unsuitable tones, particularly when algorithms suggest connections or draft outreach messages. As a result, there is a growing emphasis on transparency and verifiable claims to maintain honest communication, prompting users to be vigilant for inconsistencies in generated content. Shaping data policy and anticipating future networking trends Advancements in digital networking increasingly intersect with privacy standards, consent regulations, and policy requirements. Managing data collection and processing for natural language processing applications requires strong governance to align with user expectations and comply with laws. Tools are being designed to balance improvements in networking capabilities with the importance of privacy and informed consent. Professionals are also watching for innovations such as specialised smaller language models capable of operating on device or with robust privacy protections. Developments like privacy-preserving inference, enhanced fact-checking, and better evaluation metrics are being integrated to align natural language processing with the needs of professional networking. As such capabilities expand, emphasis is expected to remain on building relationships while sustaining transparency and trust among users. The post Advances in Natural Language Processing Are Changing Professional Networking appeared first on AI News. View the full article