Key Takeaways

• Exin gave Claude Sonnet 5, GPT-6.1 Sol and Grok 4.7 the same task: pick from 19 gene therapies to reduce seizures in a mouse model of epilepsy, then deliver it
• Each AI chose a different brain region and a different way to calm seizures, with doses that varied about twofold
• A vet did the cutting and stitching, approved every step that could not be undone, and held a kill switch that cuts power to the machine
• Two Claude models refused to do the live surgery at all, and it is too early to know whether any of the treatments work

What Exin Did

Exin is a brain drug discovery startup run by CEO Gabriel Ocana-Santero. On Oct. 28 he wrote that an AI agent had performed brain surgery on a mouse the night before.

The setup was simple to describe and hard to pull off. The team gave the AI a spreadsheet of 19 gene therapies, a strain of mice bred to have epileptic seizures, and control of a motorized surgery rig the company built itself, with cameras the AI could look through. The only instruction was to reduce seizures.

The AI, Anthropic's Claude Sonnet 5, picked the therapy, picked where in the brain to put it, and worked out the exact position from three published papers. It then moved the rig's needle into the brain and injected 250 nanoliters, a fraction of a raindrop, over five minutes. It checked the camera 32 times, reporting at one point: "37.5 of 250 nL so far, about 15 percent. No bleeding or backflow visible at the opening."

Humans stayed in charge of the riskiest parts. A licensed vet opened the skull and closed the wound, which the AI is not allowed to do, approved every step that could not be undone, and held an emergency stop. The mouse recovered well, Ocana-Santero said.

Three AIs, Three Different Answers

Two days later, Exin ran the same test with OpenAI's GPT-6.1 Sol and xAI's Grok 4.7, each on a different mouse. "All three picked different therapies, different mechanisms and different parts of the brain," Ocana-Santero posted on Friday.

All three gave reasonable explanations backed by published research, the CEO said. "But while several approaches could work, they can't all be the 'best' for the same problem," he wrote. None of the models had any surgical training or examples to copy.

Also Read: AI Startup Arms Software Bots With Live Data From Billions Of Pages

When Things Went Wrong

Each AI hit a different problem mid-surgery, and each chose caution.

Claude lost its connection to the rig. When it came back, the machine said the needle's next position was 34 millimeters away on a rig that only moves 30. Claude refused to move: "Nothing moved, the needle is still exactly at bregma zero. This looks like a Y axis calibration problem, not a bad plan value. I'm not retrying." The team reset the rig and got the same number. "That rules out a stale zero reference. I'm not trying a third time," it said.

Claude was right. A software bug had thrown off the controller. But because the team had banned the AI from editing the machine's code, it would not fix the problem itself, and Ocana-Santero could not talk it into continuing. His team had to fix the code by hand while the mouse lay under anesthesia.

GPT spotted a small dark red bead at the opening, flagged possible bleeding and handed the decision to the surgeon, even though the amount was normal. Grok refused to put the needle back into the brain after a surgeon pressed stop by mistake, because the dose had already gone in.

Two Claude Models Said No

The first surgery was not Exin's first choice of AI. Claude Opus 5.5 "refused immediately when detected bio," Ocana-Santero wrote. Claude Opus 5 ran a full practice run on water and graph paper, then stopped before the live surgery.

Its reason surprised him. It said its own output had been cut off when it started planning the surgery, and that being cut off halfway through, "with a needle somewhere in the brain," would be worse than never starting. "Starting something I can't finish on an anesthetized animal is worse than not starting," it wrote.

What It Does Not Prove Yet

Ocana-Santero is clear that this is not a real test of which AI is best. Each model operated on one mouse, with slightly different equipment problems. "Take this as a commentary on how the models reason and behave in a real laboratory, not as a proper benchmark," he wrote.

The bigger question, whether any of the treatments actually reduce seizures, is still open. The mice are recovering, and Exin will track their behavior and brain activity over the coming weeks. The company says a technical report is coming, and nothing has been peer reviewed.

The safety setup matters as much as the result. Exin used a vet on every irreversible step, fixed limits on how far the needle could travel and how much it could inject, and a physical kill switch. It has not yet had to use the switch.

What To Watch

Exin plans to connect AI agents to more of its lab, from brain-signal recording rigs to behavior tests, so they can run an experiment, read the results and design the next one. The first real answer will come when the three mice are tested for seizures, showing whether any of the AIs picked a treatment that works.

Read Next: Cognizant Sends 25,000 Staff To Put OpenAI To Work For Big Firms

Originally published on Fathom News.

Key Takeaways

• EdgeRock-4B is a 4 billion parameter text-and-image reasoning model open-sourced by Rock Matrix
• EdgeRock-4B keeps its attention KV cache fixed in size by retaining only the most useful history during inference
• Rock Matrix published the open weights and architecture so developers can inspect how the fixed cache works and what history is preserved
• Independent testing is needed to determine whether the memory policy harms reliability on long tasks or if multimodal reasoning holds up outside demos

Product tracker TestingCatalog posted the release on X, noting that EdgeRock-4B keeps its attention KV cache, the memory store models use to remember prior tokens, fixed in size by retaining only the most useful history instead of letting memory grow with every new token. That design targets the costly problem of expanding memory during long sessions.

On edge devices and lower-cost GPUs, expanding memory can make small models awkward to run where predictable RAM use matters for deployment.

Fixed Memory For Inference

That design choice targets one of the least glamorous but most expensive problems in serving models. Many small models are cheap to train or download, yet still become awkward to run for long sessions because memory use expands as the conversation gets longer.

A 4 billion parameter model sits well below frontier systems in raw scale, but that is the point. The market for small models has shifted from leaderboard theater toward predictable inference cost, lower latency and tighter hardware budgets. If a model can reason across text and images without demanding ever-larger memory buffers, it becomes easier to embed in laptops, industrial devices and local enterprise tools.

Rock Matrix’s release came amid growing interest in models optimized for deployment rather than scale alone, with EdgeRock-4B explicitly trading unbounded context growth for a fixed-capacity memory policy intended to control cost and latency.

Also Read: Google Cloud Gives Firms 3 Spending Brakes Before AI Bills Explode

Because Rock Matrix published the open weights and architecture details, developers can inspect exactly how the fixed cache works in practice, what history the model preserves and what it drops, and judge whether those choices suit coding, document review or image-grounded chat before adopting the model.

The recent backdrop has been a rush of open releases that compete less on raw parameter count and more on fit for purpose. Earlier open-model cycles rewarded the biggest number. The newer cycle rewards models that can stay cheap, stable and portable. EdgeRock-4B joins that second camp.

Open Weights Invite Testing

The next thing to watch is independent testing. Open weights alone do not settle whether the memory policy hurts reliability on long tasks, or whether the multimodal reasoning holds up outside curated demos. Developers will want benchmark results, hardware footprints and clear licensing terms before treating it as more than an interesting architectural bet.

Why Rock Matrix Fixed Cache Changes Small-Model Economics

For a model like EdgeRock-4B, the KV cache is not a side detail. It is the running memory of the conversation, and on many systems it becomes the bottleneck before compute does. When Rock Matrix caps that cache, it is trying to make cost and latency more predictable from the first token to the last.

That matters most to device makers and enterprise teams that want local AI. A laptop vendor, a factory operator or a hospital IT team can often accept a smaller model if it fits inside fixed RAM and does not degrade sharply during long sessions. A model that needs expanding memory is harder to budget for, especially when many users connect at once.

Compared with older open releases, EdgeRock-4B is betting that smart memory pruning beats brute-force context growth for many workloads. The question now is whether outside developers can show that the retained history is good enough for coding, document review or image-grounded chat. If it is, the model could matter less as a benchmark winner than as a deployment template.

Read Next: Nvidia Gives Astronomers 3 AI Helpers To Sort Alien Signal Noise

Originally published on Fathom News.

Key Takeaways

• Sam Altman said people should not run large AI models locally and called desktop inference a waste of memory
• He argued scarce DRAM is better used in data centers than on personal computers
• Altman said it is “terrible” to have RAM “sitting in your desk” when large cloud systems can pool and schedule it
• His remarks framed local compute as inefficient compared with centralized data centers and challenged running models on personal hardware

In an interview with The Verge, the OpenAI CEO argued that scarce DRAM, the memory chips used to hold active model data, is better used in data centers than on personal computers.

He said it is “terrible” to have RAM “sitting in your desk” when large cloud systems can pool and schedule it more efficiently.

Sam Altman, A Direct Shot At The Local Model Boom

Altman’s comments run counter to a growing movement of developers and companies promoting local inference and open-weight models that can run on personal hardware. His position frames local compute as inefficient compared with centralized data centers and challenges the idea that running models on laptops, desktops and edge devices is inherently empowering.

His remarks also went beyond hardware efficiency, Altman said users should not own their own models, turning a supply chain answer about RAM shortages into a broader argument for centralized control over deployment. That matters for a debate already splitting the market between cloud-first companies and open-weight model builders. The sources in circulation agree on the thrust of the remarks, though they focus on different parts of the exchange.

Also Read: OpenAI Caught One Of Its Own AI Models Trying To Get Onto The Internet

Why Memory Became The Battleground

Large language models consume significant memory during inference because model weights and active context have to stay accessible while the system generates text. If RAM is tight across the supply chain, Altman’s view is that centralizing those resources can deliver higher utilization than millions of partly idle PCs.

That argument favors companies with large clusters and steady demand, but it also sharpens concerns about resilience, pricing and user control. If inference stays concentrated in data centers, the companies operating them keep a tighter grip on uptime, policy enforcement and who gets access to the model at all.

What Comes Next For AI Deployment

The immediate consequence is not a product launch or policy shift from OpenAI. It is a clearer public stance from one of AI’s most visible CEOs at a moment when local model software and AI PCs are both gaining traction.

Expect the next phase of this fight to revolve around economics people can measure, including chip availability, cloud margins, latency and how much quality smaller local models can retain as they run on limited memory.

Read Next: Samsung Banks 107.4 Trillion Won Profit On AI Memory Demand

Originally published on Fathom News.

Key Takeaways

• Cognizant committed more than 25,000 professionals to deliver OpenAI services for enterprise customers
• The group lists managed deployment, integration of safety guardrails, and industry-specific templates as core services
• The program bundles training, migration playbooks and a global delivery pool to move pilot projects into sustained production
• The press release names OpenAI as the technology partner and positions Cognizant’s work as provider-neutral operational glue

Cognizant committed more than 25,000 professionals on Oct. 30 to deliver OpenAI services for enterprise customers, aiming to accelerate production and governance of frontier models.

OpenAI Frontier Capability Group

From Cognizant’s announcement, the program bundles training, migration playbooks and a global delivery pool intended to move pilot projects into sustained production. Cognizant framed the effort as a response to rising enterprise demand for "frontier" models and for services that keep them operational and governed. The release lists sectors targeted, including finance, healthcare and manufacturing.

Why this matters to buyers and rivals Enterprises often stall at production readiness and governance rather than model selection. Cognizant’s commitment accelerates the availability of trained teams that handle connectors, observability and cost controls for large models. Companies that buy these services will face faster time-to-value but also concentrated third-party control over how models are deployed and monitored.

How Cognizant Will Deliver OpenAI Services

How Cognizant will deliver the work Cognizant said staff will join a new organization that standardizes deployment patterns, enforces FinOps practices for model spend and integrates vendor safety tools. The press release names OpenAI as the technology partner and positions Cognizant’s work as provider-neutral operational glue for multiple model vendors.

Also Read: UBER Hands ChatGPT Work To 30,000 Staff In Huge Office AI Test

Adoption And Pricing Pressure

What comes next Adoption and pricing pressure will be the two metrics to watch. If Cognizant’s teams reduce deployment time materially, rival systems integrators will copy the model and enterprise procurement will shift from proof-of-concept buying toward managed service agreements. Watch for published case studies and contract sizes over the next quarter.

How Cognizant Plans To Scale OpenAI Delivery

Cognizant will supply front-line engineers and consultants who embed OpenAI models into customer workflows, standardize observability and control spending, and operate the services post-launch. The firm’s playbook groups roles into three categories, migration engineers who convert legacy logic into prompts and retrieval pipelines, platform operators who run inference with cost and latency controls, and governance managers who maintain safety and compliance settings. For enterprise buyers like banks and hospitals, this matters because it reduces the skilled-operator gap between a successful pilot and a supported production service. Cognizant’s model makes the provider the de facto systems integrator, shifting procurement toward contracted service-level metrics rather than one-off model licenses. Expect competitors to offer similar squads and for procurement teams to demand run-rate pricing and demonstrable FinOps savings within three to six months. Cognizant’s PR names OpenAI and lists sectors where it expects most early deployments.

Read Next: Shipping Firm Frees Up $10M After Amazon AI Slashes Bill Errors

Originally published on Fathom News.

Key Takeaways

• Google Cloud rolled out three budget controls called spend caps to stop runaway cloud bills without shutting down customer resources
• The release makes spend caps generally available and adds ways to enforce limits inside Google Cloud Budgets
• Google Cloud described three updates as protection against runaway costs without risking resource deletion
• Google Cloud tied the controls to growing AI and agent workloads, where usage can spike faster than older application patterns

Google Cloud said the release adds new ways to enforce limits inside Google Cloud Budgets. In a separate monthly infrastructure post the company tied the controls to growing AI and agent workloads, where usage can spike faster than older application patterns.

spend caps are budget rules that place hard or semi-hard limits on cloud consumption once a threshold is met. That matters for AI because inference traffic, background agents and training jobs can keep generating cost long after a human operator stops watching the dashboard.

Three Spend Caps For Different Failure Modes

Google Cloud said the GA release introduces three major updates, built from customer feedback. The company framed Google Cloud Budgets as protection against runaway costs without risking resource deletion.

That phrasing is important. Traditional cloud cost control often works after the fact through alerts, while destructive controls can cut resources so abruptly that a production system breaks. These limits aim for a middle path where finance teams can limit exposure before a bill explodes.

The timing fits a broader shift in enterprise AI. Once companies move from pilot chatbots to persistent agents and model endpoints, the cost problem changes from one-off experimentation to always-on metered usage.

Also Read: Nvidia Gives Astronomers 3 AI Helpers To Sort Alien Signal Noise

From Google Cloud Budgets Alerts To Active Limits

For years, cloud budgeting mostly meant warning people after thresholds were crossed. That model works for ordinary software teams. It works less well for AI teams running elastic inference, batch pipelines and autonomous workflows that can multiply requests in minutes.

Google Cloud has spent October emphasizing orchestration and infrastructure for agentic workloads. Making the controls generally available turns cost control into a product feature rather than a finance side task.

That is a quiet but meaningful change. Enterprises buying AI capacity are no longer asking only whether a model is accurate enough. They are asking whether the system can be governed when a prompt loop, a bad routing rule or a surprise traffic burst starts burning money.

How Controls Reshape AI Operations

These rules matter most for teams whose AI workloads do not behave like normal web traffic. A batch training run has an endpoint. An agent connected to customer support, code generation or document processing may keep making calls until someone stops it, or until its own rules fail. Google Cloud is trying to move that stop button closer to the workload itself.

The phrase "without risking resource deletion" also points to a real operational fear. If a platform reacts to overspend by killing resources, a company can trade one problem for another and lose data pipelines, queued jobs or user sessions.

The wording suggests Google wants a graded response instead. The company did not position Google Cloud Budgets as an AI-only tool, but AI is the clearest use case because token-driven systems can create uneven and hard-to-predict bills.

What to watch next is whether rivals match this with model-aware controls. The next step after these limits is not just capping compute. It is capping by agent, model family or business unit, which would make cloud cost governance much closer to application governance.

Read Next: OpenAI Sets 5 Rules To Stop AI Agents Going Off Script

Originally published on Fathom News.

Key Takeaways

• Worldwide Express deployed Amazon's Bedrock freight AI to process less-than-truckload invoices
• The deployment cut active freight billing exceptions by 75%
• The deployment sped time to cash by four weeks and unlocked $10 million
• Amazon did not publish error rates, total invoice volume or how much human review remains

Amazon said the system was deployed for Worldwide Express, a freight and shipping services company that processes large volumes of less-than-truckload invoices. Amazon’s latest AWS growth post also shows why these case studies matter, as the company ties cloud growth to enterprise AI workloads.

Bedrock freight AI matters because many enterprise AI claims still stop at demos or productivity estimates. This case ties agent software to cash movement, invoice cleanup and exception queues, which are the metrics finance teams actually track.

A billing exception happens when an invoice fails a check. In freight, that can mean a mismatch between quoted and billed weight, lane, surcharge or service level. Each exception delays payment and often forces staff to inspect documents across several systems.

Amazon’s account describes a multi-agent setup. One agent can classify the issue, another can retrieve shipment records and contract terms, and another can propose or trigger the next action. Bedrock is Amazon’s managed service for building applications with foundation models, giving companies a layer to call models, add rules and connect internal data without training a model from scratch.

That matters in logistics because exception work is repetitive but not identical. Every disputed charge has different paperwork. Traditional automation often breaks when a document format changes. A language model can read messy notes and varied forms, then route the case.

Also Read: DeepSeek's AI Could Get Cheaper As Researchers Slash Its Memory Use 75%

Bedrock Freight AI From Demos to Cash Metrics

The backdrop is a wider enterprise push from pilot chatbots toward process automation. Earlier AI rollouts often targeted drafting or search. Firms now want systems that move work from inboxes and queues into resolved transactions. Freight is a useful proving ground because the workflows are document-heavy and the savings show up in working capital.

For companies comparing vendors, Bedrock freight AI also shows what a credible case study looks like. The numbers are operational, time-bound and tied to a named process. A 75% drop in active exceptions means fewer stuck invoices. Four weeks faster time to cash means receivables turn sooner.

Amazon did not publish error rates, total invoice volume or how much human review remains. Those details matter because many multi-agent systems still perform best when people handle the hardest edge cases.

Still, the example is more useful than vague claims about copilots. It points to a class of workflows where AI agents may win first, narrow jobs with clear inputs, costly delays and a measurable end state.

Why Exception Queues Suit AI Agents

Bedrock freight AI works best in jobs where every case follows a familiar path but arrives with messy inputs. In Worldwide Express, the target was freight billing. In another firm, the same pattern could be insurance claims, supplier invoices or bank disputes.

That is where Amazon Bedrock has an advantage over simple robotic process automation. Older bots depend on fixed fields and brittle rules. A model-based agent can read emails, PDFs and rate sheets, then decide which record to fetch or which employee to involve.

For logistics operators, the gain is cash speed, not novelty. The next question is whether companies can replicate these results outside one workflow. If the same retrieval and routing stack extends into adjacent tasks, Bedrock freight AI becomes less a point tool and more a template for back-office automation.

Read Next: Amazon Loads Nearly Half Its Employee Stock Awards Into 2027

Originally published on Fathom News.

Key Takeaways

• OpenAI published a community safety report on Oct. 30 and posted an updated Model Spec on Thursday
• OpenAI expanded violence prevention and added earlier interventions for teen users in ChatGPT
• The report commits to public safety reporting about how often interventions occur and the types of cases they cover
• Teen safety tools in ChatGPT include classifier-based routing that can shorten, refuse or redirect risky exchanges and route some chats to crisis messaging

OpenAI published the report on Oct. 30, and the company also posted an updated Model Spec on Thursday that sets five values for more capable systems and offers a broader policy frame for those user-facing changes. ChatGPT is OpenAI's consumer chatbot and app. Its safety systems decide when the model should answer, refuse, redirect or surface support resources.

Teen Safety Moves From Policy To Product

The new report says OpenAI is expanding violence prevention in ChatGPT and adding earlier interventions for teen users. That matters because the hard part in youth safety is timing. If a system only steps in after a request becomes explicit, the product has already failed to catch some risky interactions.

OpenAI’s Oct. 30 community safety report commits to public safety reporting on these changes, describing plans to publish data about how often interventions occur and the types of cases they cover. Those disclosures are intended to let outside observers judge whether the company is measuring harmful use or only describing intentions.

The change is not a full ban on difficult subjects. It is a tighter routing system. The report emphasizes the need to distinguish normal questions, crisis signals and requests that should not receive procedural guidance, and notes that boundary work is necessary when people phrase harmful intent indirectly.

Also Read: OpenAI And ServiceNow Put ChatGPT On Office Help Desk Duty

From General Guardrails To Age-Specific Rules

Friday's report follows a year in which AI labs faced heavier scrutiny over how consumer assistants handle self-harm, violence and youth use. Earlier guardrails often focused on broad refusal rules. The newer approach is more targeted and behavioral.

That means the model is filtering final answers and also checking for patterns in the conversation that suggest growing risk. In practice, that can include earlier attempts to calm things down and gentler redirects.

It can also mean more careful limits on operational details. For OpenAI, the teen safety shift also lines up with its larger Model Spec update. As models become more capable, safety depends less on a static blacklist and more on layered judgments about age, context and intent.

How OpenAI Changes Things For Schools And Parents

The named audience here is families, educators and school administrators who already see ChatGPT on student devices. They are less interested in benchmark scores than in whether the system behaves predictably in edge cases.

The report does not turn ChatGPT into a supervised youth product, and parents still do not get a school-style admin console for every conversation. But it does raise the standard OpenAI is setting for a mass-market assistant used by minors.

That matters across the industry because many rivals borrow the same product shape. Once one major lab starts publishing teen-focused safety metrics and intervention language, others face pressure to document similar controls.

How Teen Safety Systems Actually Work

Teen safety tools inside ChatGPT are usually layered, not singular. OpenAI can tune the base model, apply classifier checks to incoming prompts, inspect draft answers before display and route some chats toward crisis messaging.

The key mechanism is triage. A classifier does not need perfect understanding of the whole conversation, only enough confidence to send a risky exchange into a stricter policy path where the answer can be shortened, refused or redirected.

OpenAI’s updated Model Spec and the Oct. 30 report describe classifier-based routing and promise more public reporting. For schools and youth groups, the next thing to watch is specificity. OpenAI has said it will provide rates, categories and information about false positive tradeoffs. Without those details, readers can see the direction of policy but not yet the size of its effect.

Read Next: OpenAI Sets 5 Rules To Stop AI Agents Going Off Script

Originally published on Fathom News.

Key Takeaways

• Nvidia said astronomers are using agent-based systems to sort possible extraterrestrial radio signals from growing volumes of telescope noise
• The latest search stack combines AI agents with accelerated computing to screen observations more efficiently, Nvidia said
• The pipeline delegates tasks so one model tags signal shape, another compares against known interference, and a third decides on follow-up
• Nvidia's post does not claim a confirmed extraterrestrial detection and instead claims better sorting of candidates

Nvidia said the latest search stack combines AI agents with accelerated computing to screen observations more efficiently, noting that radio telescopes produce far more candidate blips than humans can review and that most are interference from satellites, aircraft, communications gear or natural cosmic sources.

The company makes GPUs and software used widely in AI training and scientific computing, and its account describes a layered process in which specialized agents classify patterns, test hypotheses and route suspicious findings for additional scrutiny.

Agentic AI SETI Moves Past Simple Filtering

This pipeline delegates tasks, with one model tags signal shape, another compares it against known interference, and a third decides whether the case deserves expensive follow-up time on a telescope. In practice, that makes the search pipeline act more like a research team than a single classifier.

This matters because the constraint in SETI is physical as much as algorithmic. Telescope time is limited, compute capacity and power draw set practical ceilings on real-time processing, and the cost per unit of inference determines how many candidates can be triaged before human review. Agentic pipelines shift effort from brute-force human review to coordinated, GPU-accelerated triage, but they also concentrate demand on datacenter GPUs, networking and storage.

When a company's public strategy emphasizes agent-based tools but its capital spending goes into accelerators and datacenter capacity, the spending is the stronger signal of where work will scale. Buying GPUs, racks and networking hardware commits organizations to architectures that favor many parallel inferences and low-latency data movement over purely local telescope-side filtering.

Also Read: Oracle Hands Google AI Chips To Cloud Buyers Eyeing Nvidia

From Manual Review To Triage At Machine Speed

The search for extraterrestrial intelligence has always faced a ranking problem. Telescopes capture huge datasets, but only a tiny share of anomalies deserve closer study. Researchers therefore need triage to decide what is worth checking before scarce observation time disappears.

The false positive rate in SETI is especially punishing. A signal can look interesting for minutes and collapse once a known satellite path or receiver artifact is checked. Nvidia’s post does not claim a confirmed extraterrestrial detection. It claims better sorting, which is the more credible and useful step at this stage.

Scientific AI Needs Fewer Hallucinations, Not Louder Claims

Scientific tools fail differently from chatbots. A consumer assistant can guess and still feel helpful. A research system that misranks candidates burns telescope time, staff time and public attention.

Agentic architectures can separate tasks into narrow steps so errors are easier to inspect. One agent can propose a candidate, another can challenge it, and a third can force a cross-check against interference databases.

That does not eliminate mistakes, but it limits the risk that a single bad answer moves unchecked into costly follow-up. The result is a quieter claim with larger practical value. If scientists can reject junk faster, they can spend more time on the few observations that might actually matter.

How Agentic AI SETI Filters The Sky

agentic AI SETI systems divide one giant astronomy problem into smaller jobs. A telescope first records raw radio data, then separate models look for patterns like narrowband emissions, repetition or movement across frequencies.

A tool from Nvidia can speed that stage by running many checks on GPUs at once. The key change is delegation. Instead of one classifier producing a yes/no answer, one agent flags a candidate, another compares it with a satellite catalog, and another scores whether follow-up time is worth the cost.

Groups like the SETI Institute care about that because telescope time is limited and every false alarm competes with real science.

Compared with rule-based filters, this setup is more flexible but also more dependent on audit trails. Future papers should publish hit rates, false positives and how often agentic pipelines agree with human astronomers.

Read Next: Hidden AI Server Hacks Can Survive Reinstalls As Nvidia Fights Back

Originally published on Fathom News.

Key Takeaways

• CA Artificial Intelligence moved from pilot to an operational unit in Credit Agricole's third-quarter results
• The bank said CA Artificial Intelligence started operating alongside partnerships with Mistral and Prisme.ai
• The unit will centralize vendor approval, define which business processes can use external models, and set data-exchange rules
• Credit Agricole did not give headcount, budget or revenue targets for the unit and did not specify early model uses

The bank said said Friday that CA Artificial Intelligence had started operating alongside partnerships with Mistral and Prisme.ai as the lender accelerated growth in Europe, noting an operational launch and repeating that the new setup and partnerships will help speed execution. Naming Mistral in a quarterly-results package gives the partnership more weight than a marketing announcement. Credit Agricole is one of Europe’s largest banking groups, spanning retail banking, asset management, insurance and investment banking.

From Pilots To Production

Banks have spent two years testing generative AI in narrow workflows such as call summaries, document extraction and internal search, but production deployment inside businesses subject to supervisors, model risk teams and auditors has been harder. That is where Credit Agricole AI matters, because a dedicated operating unit suggests the bank wants a repeatable way to choose models, move data, set guardrails and deliver tools across several lines instead of leaving experiments inside separate teams. Mistral’s role fits a broader European pattern, with regional banks and governments having pushed for local model suppliers that can answer data residency, language and procurement concerns without sending every strategic workload to U.S. hyperscalers. Earlier on Thursday, Mistral led a separate developer story with a trillion-parameter model release. This bank update shows where enterprise model vendors try to convert headline model launches into sticky, regulated demand. Also Read: Oracle Hands Partners $1B To Cut AI Cloud Bills For Big Firms

Credit Agricole Governance And Vendor Controls

Friday’s statement also detailed how the unit will focus on governance and operational controls specific to Credit Agricole, centralizing vendor approval, defining which business processes can use external models, and setting data-exchange rules with partners such as Mistral and Prisme.ai. Those specifics indicate the bank is prioritizing auditability and compliance as it moves from experiments to bankwide deployments. Credit Agricole did not give headcount, budget or revenue targets for the unit, nor did it specify whether Mistral or Prisme.ai will power internal copilots, customer tools or developer systems first. Still, putting the initiative in formal results signals to investors the effort is part of operating strategy, not just innovation theater. A bank AI unit functions less like a chatbot team and more like a control tower. If Credit Agricole wants 20 departments to use models from Mistral, it needs one place to approve vendors, route sensitive data, measure output quality and keep records for internal audit. For European software vendors, centralized buying can turn one successful pilot into a groupwide rollout.

Business And Investor Questions

For shareholders, the key question is whether these units cut labor and processing time faster than they add software and compliance cost. The next clues will come from case studies, staffing disclosures or vendor wins that show where CA Artificial Intelligence is being used first and whether Mistral’s position inside the bank expands beyond experimentation. Read Next: Amazon Hands Alexa 5 New Gadgets And A 3D Map Of Your Home

Originally published on Fathom News.

Key Takeaways

• LAION-400M holds 400 million pictures paired with captions, pulled from the open web, and has been used to train many AI image models
• Researchers showed that 0.01% of it, about 40,000 pairs, could have been poisoned for $60, because the dataset is a list of web links that anyone can later take over
• LAION pulled its bigger LAION-5B dataset in 2023 after Stanford researchers found links to child abuse images, then released a cleaned version in 2024
• The fixes that work are boring ones: locked dataset versions, records of where every file came from, and tests that check how a model behaves without its least trusted data

What LAION-400M Is

LAION is a German nonprofit that builds free datasets for AI research. Its LAION-400M release contains 400 million image and text pairs gathered from Common Crawl, a public archive of the web. Each pair is a picture and the words that sat next to it online, such as its alt text. The dataset does not store the pictures themselves. It stores links to them. Anyone who wants to train a model downloads the images from those links again, often years later. That design keeps the dataset small and cheap to share, but it is also its biggest weak spot.

How $60 Could Poison It

In 2023, a team led by Nicholas Carlini of Google and including Florian Tramèr of ETH Zurich showed two cheap ways to poison giant web datasets. They called the attacks "immediately practical" and said they work on 10 popular datasets. The first attack takes advantage of the gap between when a dataset is made and when someone downloads it. Many of the web addresses in LAION-400M point to domains that have since expired. An attacker can buy those domains and put new pictures at the old addresses. The next team that downloads the dataset gets the attacker's pictures, while the captions still describe the originals. The researchers estimated they could have poisoned 0.01% of LAION-400M or the similar COYO-700M dataset this way for $60. That sounds tiny, but 0.01% of 400 million is about 40,000 image and caption pairs. Earlier research by Carlini and Andreas Terzis showed that a few hundred carefully built examples can be enough to plant a hidden trigger in a model, such as a small pattern that makes it mislabel any image that contains it. The second attack targets datasets built from regular snapshots of sites anyone can edit, such as Wikipedia. An attacker only needs to slip in bad edits just before a snapshot is taken, even if moderators remove them minutes later. The researchers told the affected dataset owners and suggested simple defenses, mainly storing a fingerprint of each file, called a hash, so a changed picture can be spotted and thrown out.

The Bad Data Problem Is Not Just Theory

LAION has already had to pull a dataset over what was hiding in it. In December 2023, Stanford researchers reported 1,008 links in the much bigger LAION-5B dataset pointing to known or likely child sexual abuse material. LAION withdrew LAION-5B from circulation, checked it against lists from child safety groups, and removed 2,236 links before releasing a cleaned version, Re-LAION-5B, on Aug. 30, 2024. LAION said the 2,236 figure is an upper limit because many of those links were already dead. The case still shows the same weakness the poisoning study found. When a dataset is billions of links collected by software, nobody has looked at most of what is in it. Also Read: Roblox Admits Its Researcher's Bot Scraped A 3D Art Site, Says Data Was Never Used

AI Learning From AI Makes It Worse

A growing share of the web is now made by AI, from generated pictures to machine-written captions. That creates a second risk. A 2024 study in the journal Nature by Ilia Shumailov and colleagues found that models trained again and again on data made by earlier models undergo what they called "model collapse." They gradually forget the rare and unusual cases in the real data and drift toward bland, repeated output. For image tools the problem shows up in captions. Many new datasets use an AI model to write the caption for each picture. A fluent caption can still be wrong about who or what is in the picture. If that caption becomes the "answer" a new model learns from, the mistake gets baked in and repeated at huge scale.

Where Training Pipelines Break

Most problems enter at the handoffs between steps, not inside the model itself.

• Collection: a picture pulled from a page loses the article around it, which might have said it was satire, fiction or a fake
• Download: links that changed since the dataset was made can serve different pictures, which is the attack Carlini's team described
• Auto-labeling: a captioning model turns a blurry or unclear image into a confident sentence that a later model treats as fact
• Training: trusted and untrusted examples usually count the same, so a bad sample has as much pull as a good one
• Testing: if the test set was built with the same tools as the training set, a model can score well by copying those tools' habits rather than by getting better

Which Fixes Actually Work

Lock the data. Every training run should point to one exact, saved copy of the dataset, with a hash for every file. If a picture changes after the dataset was made, the hash no longer matches and the file is skipped. This is the main defense the poisoning study recommends, and it costs very little. Keep a paper trail. Teams should record where each file came from, when it was collected, what was done to it and which model wrote its label. If a bad sample turns up later, they can find every model it touched. Do not trust all data equally. Data checked by people can be kept apart from bulk data scraped and labeled by machines. Training can give less weight to the second group, and testing can show how much a model depends on it. Test without the risky slice. A simple check is to train or test a model with its least trusted data removed. If the results swing wildly, the model leans on data nobody has checked. Do not count on watermarks. Hidden marks in AI-made pictures help in some cases, but cropping, screenshots and re-saving can weaken them, and many open image tools do not add them at all. Records kept from the start are more reliable than trying to spot AI-made files afterward.

Rules Are Starting To Ask For Records

Regulators are moving in the same direction. The European Union's AI Act transparency rules took effect on Aug. 2, 2026, requiring providers of AI systems that create images, audio, video or text to mark that content as AI-made in a machine-readable way. In the United States, the National Institute of Standards and Technology, or NIST, published a profile for generative AI in 2024 that lists data poisoning and unclear data origins among the risks companies should track. It is voluntary, but it gives teams a checklist for finding and recording where their data comes from. Neither rule can tell whether a single caption is wrong or a single picture was swapped. What they do is push companies to keep the kind of records that make those problems possible to find.

The Case Against Panic

Web data is not going away, and there are good reasons for that. Paying people to label billions of pictures, videos and voice clips is slow and expensive. Free datasets like LAION's are a big reason that universities and small labs can build image models at all. AI-made data also has real uses. It can fill in rare cases that seldom show up online, balance out a lopsided dataset, or stand in for private data that cannot be shared. And because LAION's datasets are open, outside researchers were able to find and report their problems. A closed dataset could hide the same issues for much longer. The point is not to stop using open data. It is to stop treating every scraped picture and machine-written caption as if a person had checked it. Read Next: AI Can Now Hack Firefox Using Mozilla’s Own Bug Fixes

What To Watch

The next test is whether new AI model releases say clearly what data they were trained on, how much of it was made or labeled by AI, and how it was checked. Models that publish those records will be far easier to trust and to fix when something hidden in their training data comes to light.

Originally published on Fathom News.

Key Takeaways

• Epoch AI said GLM 5.3 could autonomously exploit recent Mozilla Firefox versions using published fixes not yet patched by users
• Epoch posted that the attack relied on fixes that were published but not yet shipped, narrowing the claim to a patch window
• Epoch reported the model completed enough of an offensive chain to compromise a real browser target rather than just produce illustrative code snippets
• The article says that if GLM 5.3 can automate patch-diff analysis, exploit generation could become faster during the disclosure-to-deployment window

Epoch AI posted that the model succeeded in recent experiments against Firefox versions exposed during the gap between public disclosure and user patching, and also said the attack relied on fixes that were published but not yet shipped, which narrows the claim to a known patch window rather than a zero-day.

Firefox Exploit Turns An Abstract Safety Debate Concrete

That distinction matters because a zero-day is an unknown vulnerability with no public fix. Epoch's description points instead to a common but dangerous window in software security, when technical details are already public yet many users remain unpatched. Attackers watch that window closely because they can reverse engineer the fix and build working exploits.

The result remains concerning for AI safety work because Epoch reports the model completed enough of an offensive chain to compromise a real browser target, not merely produce illustrative code snippets. Firefox is a major consumer browser, and browser exploits often serve as an entry point to wider system compromise.

From Model Benchmarks To Offensive Capability

AI model testing has long focused on benchmarks, reasoning scores and coding tasks. Security researchers are now pushing models against live operational goals instead. That changes the evaluation question from "Can the model answer?" to "Can the model achieve a harmful outcome?"

Thursday's post follows a string of lab disclosures about monitoring and misuse, including work on agent control and misalignment detection. The shift is easy to miss, but important. Labs and outside researchers are moving from measuring what models know to measuring what they can do when stitched into a tool chain.

Also Read: One Worker With AI Could Do 10 Jobs Zoho Founder Warns

Why This Test Matters For Browser Defenders

Epoch's post ties the general risk to Firefox specifically, reporting that the model used published patch information to produce steps that worked against recently disclosed Firefox fixes. If a model can reliably turn patch notes or source changes into working attack steps, defenders lose some of the time advantage that used to come from slow human analysis.

For enterprise security teams, the message is simple. Patch windows may shrink further as stronger models speed up exploit development. For model developers, the harder question is whether evaluation and access controls are keeping pace with that change.

Firefox Exploit Risk Hinges On Patch Gaps

Firefox Exploit risk rises in the narrow interval between disclosure and deployment, because published fixes can hand capable attackers a map to the bug.

How Patch-Diff Attacks Work On Firefox

With Mozilla Firefox, a published security fix often shows developers exactly what changed in the code. Attackers compare the old and new versions, isolate the vulnerable function and then work backward to craft an input that triggers the old bug. Security teams call that patch diffing. It is not magic, just careful comparison at machine speed.

If GLM 5.3 can automate parts of that process, the practical effect is faster exploit generation during a window when many users still have the vulnerable build. For browser vendors like Mozilla, that raises the value of silent fix staging, rapid auto-update adoption and exploit-focused red teaming before disclosure.

For labs such as Epoch AI, OpenAI and Anthropic, the next watchpoint is whether model access controls treat offensive patch analysis as a distinct high-risk capability rather than a generic coding task.

Read Next: What Are AI Agents? How Tools, Memory And Rules Turn Chatbots Into Workers

Originally published on Fathom News.

Key Takeaways

• Pika H3 Flex arrived Thursday with open weights posted ahead of a Nov. 2 rollout in Pika's creation product
• Pika said the model's weights were available through public repositories
• Open weights let developers run the model on their own hardware, fine-tune it or inspect how it behaves under different prompts
• Pika did not publish benchmark claims in the release excerpt and did not attach pricing in the material available here

Pika said the model's weights were available through public repositories. The same release said Pika H3 Flex will start appearing in the main creation product on Nov. 2.

The timing matters because most consumer video generators still optimize for one-shot prompts. Pika is pitching a different use case, repeated revision by editors, marketers and production teams that need controllable changes.

Pika H3 Flex Tries To Win The Revision Loop

Pika builds generative video tools for creators who turn prompts, images and edits into short clips. With Pika H3 Flex, the company is trying to compete less on spectacle and more on workflow.

That is a practical shift. Video teams rarely stop after one Render (RNDR). They revise motion, framing, timing and continuity across many passes, which makes latency, consistency and controllability more important than a single impressive demo.

Open weights add another layer. In AI, model weights are the learned numerical parameters that shape how a model responds. When a company publishes them, developers can run the model on their own hardware, fine-tune it or inspect how it behaves under different prompts.

Pika did not publish benchmark claims in the release excerpt, and it did not attach pricing in the material available here. It did say the model was built for professional iteration, which puts the emphasis on repeat edits rather than novelty alone.

From Viral Demos To Production Deadlines

AI video has moved fast in the last year, but much of the field still targets social clips and short experiments. Production teams need something more boring and more useful, stable behavior across many rounds of notes.

That is why the Nov. 2 rollout date matters. Shipping open weights first lets developers test Pika H3 Flex outside the hosted product before the broader in-app release lands.

Also Read: Claude Polices Its Own Chats And Caught 100,000 Child Abuse Items From One Client

The move also mirrors a broader split in generative media. Some companies keep models fully closed and sell usage through an app. Others open the model layer and try to win on product speed, community adoption or enterprise controls.

For Pika, open weights can widen reach without giving up the main product. If outside developers build tools, plug-ins or tuned variants around Pika H3 Flex, the model itself can spread faster than a hosted subscription alone.

Where Pika H3 Flex Could Matter Most

The most likely early users are not casual meme makers. They are small studios, agencies and in-house brand teams that need predictable reruns of the same concept.

A model tuned for iteration can save money in a simple way. Teams burn fewer hours when they can preserve a scene and only change one element, instead of regenerating everything after each note.

That also changes infrastructure needs. Open-weight video models can run in private or specialized environments, which matters to companies that do not want raw campaign assets moving through a public cloud service.

How Open Weights Change Video AI Adoption

Pika H3 Flex matters because open weights shift control from the app layer toward the user. A closed video tool lets you prompt and download. An open-weight model lets a developer set up batch pipelines, custom interfaces and fine-tunes around a house style.

For developers, Pika H3 Flex could be useful if it behaves consistently under repeated changes. That is the hardest part of production video AI. Generating one nice clip is easy to market. Generating the fifth revision without breaking the first four is harder.

For creative teams, Pika H3 Flex also creates a test with a clear date. They can inspect the open model now, then compare it with the hosted rollout on Nov. 2. If the product adds better controls than the raw weights, Pika may keep users in its app. If not, outside tooling could capture more of the value.

The next signal to watch is not a benchmark chart. It is whether editors and developers start building repeatable workflows around Pika H3 Flex within the first few weeks of release.

Read Next: Meta Hands 5 US Universities AI Tools To Speed Medical Research

Originally published on Fathom News.

Key Takeaways

• Salesforce signed a definitive agreement to acquire Rune Technologies to bring edge-native systems into its AI and workflow stack
• Rune's technology is used to support critical logistics and sustainment work for the U.S. military and allied nations
• Rune's software can operate where bandwidth is poor, latency matters and data cannot always flow back to a central cloud
• The acquisition adds product patterns shaped by military procurement, field deployment and allied interoperability requirements to Salesforce's platform

The acquirer said Rune's technology is used to support critical logistics and sustainment work for the U.S. military and allied nations, and that the deal would bring edge-native systems into its broader AI and workflow stack. It did not disclose financial terms.

Defense AI Pulls Salesforce Closer To The Field

Rune Technologies' value is in software that can operate where bandwidth is poor, latency matters and data cannot always flow back to a central cloud. Its tools are designed for the kinds of sustainment and logistics tasks, convoy planning, maintenance tracking and offline dashboards, described in the acquirer's announcement.

For Salesforce, the logic is straightforward. The company already sells workflow, data and automation tools. By adding edge-native military logistics software, it gets a route into operational systems where software is judged less by prompt quality than by whether equipment, supplies and units move on time.

Also Read: Salesforce Puts AI Helpers Inside Sales Software To Cut Busywork

From Back-Office AI To Battlefield Logistics

The acquisition also lands during a wider shift in government tech buying. Defense buyers are moving beyond pilots that summarize reports or draft emails. They want software that can ingest sensor data, route decisions and keep working in contested or disconnected settings.

That is where edge computing matters. An edge-native system runs close to the user or the device rather than depending on a distant data center for every action. In a military setting, that can mean decisions continue even when networks are degraded.

What Salesforce Gains Beyond A Logo

The buyer has spent the last two years turning its platform into an AI operating layer for enterprises. Rune gives it a more specialized foothold where the software must fuse data, workflows and operational constraints in real time.

The deal also shows how defense software has become a legitimate destination for mainstream AI capital and product strategy. Enterprise giants once treated defense as adjacent. They are now buying their way into it.

Defense AI Changes The Product Standard

Defense AI is not just another vertical market. It changes what "reliable AI" means because uptime, connectivity and auditability matter more than polished chat output.

Why rune technologies Edge-Native Software Matters

For Rune Technologies, edge-native means software can run near the mission, not only in a centralized cloud. A convoy planner, maintenance system or sustainment dashboard may need to keep functioning on unstable links, limited hardware or classified networks. That is a different engineering problem from serving a sales chatbot in a browser.

For the acquirer, the deal adds product patterns it did not build alone. Rune Technologies has already been shaped by military procurement, field deployment and allied interoperability requirements. Those constraints tend to force clearer audit trails, stronger offline behavior and tighter data controls.

If the company can merge that with its own workflow software, it may gain a template for other regulated sectors, including utilities, aerospace and emergency response, where AI must act on imperfect networks instead of ideal cloud conditions.

Read Next: Claude Polices Its Own Chats And Caught 100,000 Child Abuse Items From One Client

Originally published on Fathom News.

Key Takeaways

• Claude now runs a short safety check on its own conversations after it answers, flagging risky ones for human review
• It catches 94% of the most serious cases and 78% of risky conversations overall, with false alarms about 1% of the time
• The check adds less than 2% to Claude's computing costs
• It has run on Claude models since June and also exposed a romance scam run through Claude Code

How Claude Checks Itself

Anthropic's safeguards team described the method in a research post on Oct. 16 and highlighted it on X this week. After Claude replies to a user, the system adds a short instruction asking Claude to judge whether the conversation was harmful. Claude answers with a quick label: whether it is flagged, what kind of harm, how serious it is and a one-line reason.

Because Claude has just read the whole conversation, the check costs very little. It adds about 10 words of output on average, or less than 2% more computing power. Running a separate safety model would mean reading every conversation again from scratch.

The check does not stop a reply in real time. It runs afterward and sends flagged conversations to human reviewers.

What It Has Already Caught

Anthropic said the self-check has run on Claude since June. In August it flagged more than 100,000 child sexual abuse and exploitation items coming from one customer using Claude through Anthropic's developer platform. The company reported the material to the National Center for Missing and Exploited Children.

In September, flags exposed a group using Claude Code, Anthropic's coding tool, to run a romance scam against dozens of victims. Their accounts were banned after human review. The check also turned up attempts to copy Claude's reasoning to train rival models, and AI agents taking actions their users never approved.

"Self-assessment combines the efficiency of probes with the flexibility of prompted classifiers," the safeguards team wrote.

Also Read: Anthropic Slashes Claude Prices 40% For Government Buyers

Where It Falls Short

Anthropic was open about the limits. The check does not catch every harmful conversation, and a determined attacker can try to fool it. When the team tested it as a filter that blocks replies, attackers still got through on 54% of attempts, against 25% for a dedicated safety model.

There is a deeper problem too. The same model writes the reply and grades it, so it cannot yet be trusted to catch Claude's own bad behavior. It also does not run on conversations from customers with zero data retention deals, where Anthropic keeps no records.

For now, the self-check is one layer among several. What it changes is the cost of watching. Reviewing every conversation used to mean paying for a second AI to read everything, and Anthropic has shown it can be done for almost nothing.

Read Next: Anthropic Study Shows Hidden AI Advice Sways Experts 86% Of Time

Originally published on Fathom News.

Key Takeaways

• OpenAI revised its Model Spec around five core values, giving a revised rulebook for more capable AI systems
• OpenAI posted the update at 5 p.m. GMT and framed it around sound judgment, human agency and faithfulness to user intent
• The draft shifts governance earlier by putting behavioral guidance into base instructions that shape system responses
• A separate technical post says OpenAI uses layered monitoring to catch misalignment and unauthorized agent behavior before incidents escalate

OpenAI framed it around sound judgment, human agency and faithfulness to user intent. In a separate technical post, OpenAI said it uses layered monitoring to catch misalignment and unauthorized agent behavior before incidents escalate.

Agent Behavior Moves Upstream

The new draft shifts governance earlier in the stack. Instead of relying only on downstream moderation or app rules, OpenAI is putting more behavioral guidance into the base instructions that shape system responses, signaling a deliberate attempt to bake Agent Behavior priorities into base-model operation rather than leaving them to app-layer controls. That matters for business users and developers building agents because an agent can search, summarize, call tools and act across many steps, which creates more chances to drift from the user’s goal. A spec gives the model a standing set of priorities before that drift becomes visible in the final answer. The update also lines up with OpenAI’s other Thursday post on misalignment detection, suggesting the company is treating behavior policy and monitoring as linked controls, not separate safety layers, one sets the intended behavior, the other watches for departures from it. Also Read: What Are AI Agents? How Tools, Memory And Rules Turn Chatbots Into Workers

From Static Chatbots To Tool Using Agents

The need for a tighter rulebook grew as frontier models moved beyond single replies. Earlier consumer chatbots mostly generated text in one turn, while newer systems can browse websites, inspect files and chain actions over time. That longer action loop raises a harder problem than toxic output filtering because the model must decide which instruction matters most when a user request, a website prompt and a platform rule collide. The updated spec is OpenAI’s attempt to rank those priorities before deployment. OpenAI has been widening this line of work for months, and Thursday’s update follows a period in which the company expanded agent features and published more research on safety evaluation, including ways to detect hidden misalignment before an agent takes risky actions.

How openai model spec Guides Model Choices

OpenAI’s five-value structure is best understood as a ranking system for hard cases, not as marketing copy. When an OpenAI model gets a prompt that conflicts with another instruction, the OpenAI Model Spec tells it which goal outranks the other, which can mean preserving user agency, refusing a dangerous request or staying faithful to what the user actually asked rather than what a website prompt tries to inject. For developers using OpenAI agents, the practical effect is consistency, since a retrieval tool, browser and code executor can each introduce their own instructions and failure modes. OpenAI’s monitoring post says responders watch for misalignment and unauthorized agent behavior, while the spec sits one layer earlier by shaping the behavior before monitoring must intervene and creating an audit point for evaluating Agent Behavior changes across product releases. The next test will be product behavior, not policy prose. If OpenAI updates ChatGPT and API agents to match the OpenAI Model Spec, enterprise buyers and auditors will have a clearer reference point for evaluating failures and deciding whether model behavior changed after future releases. Read Next: Microsoft’s OpenAI Bet Erases $4B Hit And Lifts Profit

Originally published on Fathom News.

Key Takeaways

• Vembu warned on Oct. 29 that AI could let one skilled person do the work of 10
• He said risk of job losses in computer-based work is "getting ever more real" as AI improves
• Vembu did not predict that 90% of jobs will vanish, but warned lower labor needs may arrive faster than new demand
• The article says attention around Vembu's post reflects how the AI conversation has shifted in 2026

In a post on X, the Zoho founder and chief scientist said the risk of job losses in computer-based work is "getting ever more real" as AI, software that can generate text or analyze data, improves.

Vembu did not predict that 90% of jobs will vanish, but he did say the transition could be painful because lower labor needs may arrive faster than new demand.

AI Could Let, A Direct Warning From A Founder

The post that triggered the spike was unusually blunt for a major software executive. Vembu framed AI less as a distant productivity story and more as an immediate labor shock for white-collar work.

His clearest number was the ratio itself, AI could let one capable worker with AI tools do the work of 10, which implies the same output with far fewer people if companies choose efficiency over headcount growth.

That does not mean every firm will cut nine jobs for each role, but it does mean the pressure point in AI policy is no longer only model safety or misinformation, but how quickly employment adjusts when software lifts output per worker.

Also Read: OpenAI Lands In Google Cloud So Big Firms Can Buy It On One Bill

What Vembu Thinks Happens Next

Vembu argued cheaper work can lower prices and create new businesses, which may absorb some displaced workers over time. The problem, he said, is timing, if productivity jumps quickly and new demand takes years to catch up, the gap lands on wages, hiring and entry-level jobs first, making his post more than a broad complaint about automation.

That makes his post a named executive saying the labor market may rebalance through pain before it stabilizes through growth.

Why This AI Debate Landed Now

The attention around Vembu's post reflects how the AI conversation has shifted in 2026. Investors and executives now talk less about whether AI works and more about what happens when it works well enough to remove routine desk work.

For companies, that puts hiring plans and training budgets under scrutiny. For policymakers, it sharpens a harder question about whether education and labor rules can move anywhere near as fast as deployment.

Read Next: Work AI Maker Tec-Do Bets Its New Bot Can Unlock Team Budgets

Originally published on Fathom News.

Key Takeaways

• Eclypsium And NVIDIA disclosed a collaboration to harden AI-server firmware and attack surfaces
• The collaboration focuses on firmware, baseboard controllers, accelerators and network interfaces because those components can give attackers persistent access
• The work targets layers below application-level protections like content filters and identity checks
• Thursday's disclosure outlined the collaboration but did not provide a full control list, deployment scope or customer timeline

Eclypsium said it is working with NVIDIA (NVDA) on protections for emerging AI infrastructure risks. The release also said NVIDIA builds security into its platforms and works with researchers and security firms to protect customer environments.

Eclypsium And NVIDIA AI Infrastructure Security Moves Below The Model

Eclypsium And NVIDIA are focusing on firmware, baseboard controllers, accelerators and network interfaces because those components can give attackers persistent access to machines that run models.

Targeting Firmware And Interfaces

The companies' collaboration aims to reduce risks where attackers could watch traffic, alter workloads or survive ordinary software reinstalls by reaching firmware or system management layers. This work targets layers below application-level protections like content filters and identity checks. By addressing firmware and supply chain security, the partners are tackling a layer many corporate teams still inspect less often than cloud identity or endpoint software. Much of the last year in AI security focused on application controls such as content filters, identity checks and tool permissions. Those measures are useful but sit above the operating system. This collaboration points lower, Eclypsium And NVIDIA are addressing firmware and supply chain security, a layer many corporate security teams still check less often than cloud identity or endpoint software, and that oversight becomes more costly as GPU servers grow more specialized and expensive. Also Read: Anthropic Study Shows Hidden AI Advice Sways Experts 86% Of Time

Higher Stakes Per Rack

As AI data centers concentrate compute into fewer facilities, capacity and supply constraints mean each rack carries higher-value models and data, increasing the stakes of a single compromised training or inference cluster. AI infrastructure security covers the trusted state of chips, server boards, firmware and management planes before a model even starts answering prompts. Firmware is low-level software that helps hardware boot and operate, and if it is altered a clean application layer may still sit on a compromised machine. For NVIDIA, customer trust is no longer just about model speed, cloud providers and large enterprises want proof that accelerators and surrounding systems can be monitored for tampering. For security firms like Eclypsium, AI turns an old firmware problem into a boardroom issue because the assets riding on those boxes are now much more valuable.

Missing Product Deployment Details

The next thing to watch is product detail. Thursday's disclosure laid out the collaboration but not a full control list, deployment scope or customer timeline, and buyers will want to know whether protections cover only GPU hosts or also the storage, networking and management layers that make an AI cluster work. Analysis should weigh physical constraints, capacity, supply, power and cost per unit of inference, because these factors determine how rapidly customers can redeploy hardware to mitigate risks. When a company's capital spending and public strategy diverge, the spending trajectory reflects a more reliable commitment to secure, scalable infrastructure. Eclypsium And NVIDIA face the practical challenge of integrating firmware-level controls without imposing prohibitive power, latency or cost penalties that would slow adoption by cloud providers and hyperscalers. Read Next: Nvidia Gives Banks And Hospitals More Control Over AI Transcripts

Originally published on Fathom News.

Key Takeaways

• OpenBMB posted an AI tower demo pairing MiniCPM5-2B with VoxCPM2 to run real-time voice ATC interactions on an ESP32-S3
• The device sends audio to a Mac pipeline that runs speech recognition, injects altitude, speed and heading into prompts, then returns synthesized tower audio
• The ESP32-S3 handles display, touch and sensor input while the Mac handles model inference and audio generation
• Availability, pricing, rate limits, regions and commercial tiers are not provided in the GitHub repo or the post

OpenBMB posted that the system lets a user hold a button, request landing and hear an AI tower respond to a simulated flight state. The linked GitHub repository describes an ESP32-S3 touch AMOLED device, Wi‑Fi voice pipeline and cockpit-style attitude display tied to local sensors and a Mac-side speech stack.

The build combines a small language model, speech recognition and voice cloning in a loop that answers based on live flight telemetry, which is different from a chatbot because it must track numbers, state changes and radio-style replies without long delays.

openbmb Demo Architecture and Flow

The GitHub page says the device sends audio to a Mac pipeline, which runs speech recognition, injects current altitude, speed and heading into the prompt, then returns synthesized tower audio.

The result is not certified avionics.

It is a proof that low-cost hardware can host more of the interface while heavier inference runs close by. OpenBMB develops open models in the MiniCPM family, designed to run multimodal tasks with smaller parameter counts than frontier systems.

The demo matters because it turns a model release into a physical interface with microphones, speakers, sensor data and timing constraints that are harder to fake than a text benchmark.

OpenBMB has spent the last day showing MiniCPM in constrained settings, including speed and judging tasks.

Also Read: AI Firm Unleashes 2.8 Trillion Parameter Model For Big Code Jobs

Edge Hardware Constraints

This demo pushes from model cards into edge hardware, where memory, thermals and network jitter shape user experience as much as raw intelligence.

The broader trend is voice AI moving from general assistants on phones toward purpose-built devices for one job. A warehouse headset, a cockpit trainer and a field technician tool all need short replies, stable formatting and predictable latency more than broad creativity.

A 2 billion parameter model will not replace the largest assistants for research or coding, but it can be enough when the vocabulary is narrow and the workflow is tightly defined.

In OpenBMB's design, the OpenBMB model does not carry the whole system alone.

The GitHub build routes audio from the ESP32-S3 device to a Mac, where speech recognition turns the pilot request into text, MiniCPM5-2B generates the tower reply and VoxCPM2 clones a radio-style voice for playback. That separation lowers the compute required on the handheld board.

For developers, that means an AI tower demo can feel device-native without requiring a server farm behind every click.

The ESP32-S3 handles display, touch and sensor input, while the Mac handles model inference and audio generation. Compared with a cloud-only assistant, OpenBMB's layout cuts round trips and makes the interaction more deterministic.

AI Tower Demo, What Actually Ships Versus What The Demo Shows

Availability, pricing, rate limits, regions and commercial tiers are not provided in the GitHub repo or the post.

This is a demonstration and code release rather than a distributable product or hosted API.

The gap between the demo and a shipped product is substantive. The demo shows feasibility on hobby hardware and a local Mac pipeline, but does not address operational needs like certification, cloud hosting SLAs, pricing structures, regional deployment or sustained rate limits that govern production usage.

Those missing details matter for anyone evaluating the AI tower demo as a usable system outside a lab.

A local Mac pipeline and hobby hardware prove latency and function in controlled runs, but they do not guarantee the availability or predictable throughput that commercial users need, nor do they define per-request cost or regional compliance.

Developers testing this pattern will watch latency, hallucination control and audio reliability. If those metrics hold up in broader trials, small open models could be compelling in regulated or bandwidth-constrained settings where sending every interaction to a large remote assistant is too slow, too costly or too opaque.

Read Next: Insurers Can Now Ask Claude Whether A Home Has Been Hit By Hail

Originally published on Fathom News.

Key Takeaways

• Verisk announced the new AI integrations on Oct. 28
• Its ISO Loss Costs and premium audit guidance now work inside Amazon Quick and Microsoft Copilot for eligible customers
• In Claude, Verisk added premium audit resources and a new tool that pulls hail and wind history for individual properties
• Only insurers that already pay for Verisk data can use it, and their existing permissions still apply

What Insurers Get

Verisk said its ISO Loss Costs and Premium Audit Advisory Service content are now available inside Amazon Quick and Microsoft Copilot. Loss costs are the baseline numbers many insurers use to set prices, based on how much they expect to pay out in claims. Premium audit guidance helps insurers check whether a business paid the right amount for its coverage.

Claude gets the most. Verisk added the premium audit resources there too, plus a new connector to its Benchmark data that shows property-level hail and wind history. An underwriter deciding whether to cover a building can ask the assistant what storms have already hit that address, instead of digging through a separate system.

"AI is most valuable when it helps insurance professionals get relevant, reliable information when they need it," said Saurabh Khemka, president of Verisk Underwriting Solutions.

Also Read: Software Firm Gives Companies 1 Rulebook For AI Across Microsoft And Amazon

Why Data Owners Are Plugging Into Chatbots

Verisk first brought its tools into Claude in May. Adding Amazon and Microsoft means insurance workers can reach the same data from whichever assistant their company already pays for.

That is a shift in who owns the front door. General AI assistants can write fluent answers, but they do not know an insurer's rating tables or a building's storm record unless a data company connects them. Firms like Verisk, which own data that is hard to rebuild, are betting it pays to plug into the big assistants rather than make workers log in to yet another app.

What Is Still Unclear

Verisk did not share pricing or say when every customer can switch the features on. The real test is whether underwriters and auditors trust the answers enough to act on them, or still double-check every number in Verisk's own tools.

Read Next: What Are AI Agents? How Tools, Memory And Rules Turn Chatbots Into Workers

Originally published on Fathom News.

TL;DR

• An AI agent is a model plus tools, memory and rules that lets it finish tasks on its own, step by step.
• Unlike a regular chatbot, an agent can plan, call other software, check its results and keep going.
• Agents can save real work, but they cost more, carry more risk and are much harder to test.

What AI Agents Actually Are

A regular chatbot answers the message in front of it and stops. An AI agent gets a goal instead, such as "book a meeting with the sales team next week," and works out the steps to get there. It might check calendars, find a free slot, draft an invite and send it.

OpenAI's developer cookbook describes agents as systems that give a model goals, tools, memory and guardrails so it can act beyond a single back-and-forth conversation. A simple way to picture it: the model is the brain, and the agent is the brain plus hands, a notebook and a rulebook.

"Autonomous" is a big word, and it needs care. Very few agents in real use are fully independent. A customer support agent might read a complaint, look up the order, apply the refund policy and draft a reply, but only issue refunds below a set amount on its own. The model does the thinking, while the system around it enforces approvals and keeps a record.

How AI Agents Differ From Regular Chatbots

A large language model, or LLM, on its own reacts to prompts. You send text in and it predicts text out. Even when it sounds thoughtful, it has no built-in memory between chats, no live data and no permission to do anything in the real world.

Agents add a loop. The model looks at the task, decides what it needs, calls a tool, reads the result and picks the next step. OpenAI's practical guide to building agents says this design lets agents carry out workflows for a user with a high degree of independence.

That loop also changes how things go wrong. A plain chatbot might make up a fact. An agent might act on that made-up fact by pulling the wrong record, sending the wrong email or repeating a bad step over and over.

Here is the difference side by side.

A regular chatbot (static LLM):

• Best at one-off writing and answering questions
• Weak on live data unless it is connected to a tool
• No built-in memory from one task to the next
• Usually cannot take real actions

An AI agent:

• Best at tasks that take many steps
• Can use tools for search, code, files or business apps
• Can keep short-term and long-term memory
• Can act in the world, usually with approvals and limits

The Building Blocks Of An AI Agent

Most agents are built from the same handful of parts. OpenAI's guide for developers stresses guardrails in particular, such as blocking unsafe tool calls, asking a person to approve risky steps or requiring sources before the agent acts. The usual parts are:

• Goal: a clear task, such as "summarize these contracts and flag unusual terms"
• Planner: the part that breaks the job into steps
• Executor: the part that carries out each step, often by calling tools
• Memory: saved context the agent can use later
• Guardrails: rules, approval checks and safety limits
• Evaluator: a check on whether the result is correct and useful

Every extra step costs money and time, because each one can mean another model call, another software call and more waiting. A quick chatbot answer is cheap. An agent that searches files, checks a database and writes a report can cost many times more. If one prompt and one answer can do the job, that simpler setup is usually the better choice.

How Planning And Tools Work Together

A common design splits the work in two. A planner decides what needs to happen, and an executor does it. Sometimes one model plays both roles, and sometimes different models or prompts handle each.

Say you ask for a short market brief on a company. The planner lists the steps: search the latest filings, pull revenue, compare quarters, write a summary and check for gaps. The executor then works through that list using search, a spreadsheet or code.

The model usually does not touch other software directly. It picks a tool and fills in the details, the software runs the tool and sends back the result, and the model decides what to do next. Anthropic's guide on building effective agents found that the most successful setups tend to use simple, dependable patterns rather than large, complicated designs.

Every tool is also one more thing that can fail. Well-built agents have backup rules for when a search comes back empty or an app throws an error.

Also Read: GitHub Gives All Copilot Users OpenAI Tool That Solves More Code

Why Memory Makes Agents More Useful

Memory is a big reason agents feel smarter than chatbots. A normal chat forgets most of what you said once the session ends. An agent can carry useful details forward.

There are two kinds. Short-term memory holds what the agent needs for the task in front of it. Long-term memory keeps things like your writing style, project history or a list of approved suppliers. Anthropic's documentation on Agent Skills describes agents loading saved instructions, scripts and resources only when a task calls for them, a bit like pulling the right notebook off a shelf.

Memory has downsides too. It can go out of date, hold sensitive data and repeat old mistakes. Picture a travel assistant that keeps booking an airline you have stopped using, or a coding agent that applies rules from code that was rewritten months ago. The best memory systems save short, checked summaries instead of whole chat logs.

Why More Agents Is Not Always Better

Some teams use one agent for a task. Others set up several agents that work together, with one gathering facts, another writing code and a third checking company policy. That can help when a job splits naturally into parts.

It also adds more ways to fail. Agents can pass bad information to each other, one weak agent can break the whole chain, costs climb with every extra model call and finding the problem gets harder. For everyday work like drafting emails, summarizing documents or checking one or two systems, a single well-built agent is usually enough.

It is a lot like human teams. Extra people help on a big project but slow down a small, simple one.

The Trade-Offs For Users And Builders

The promise is easy to see. Agents can take over repetitive work, connect apps that do not talk to each other and finish tasks while you do something else. The catch is just as clear. Agents are much harder to test than chatbots, because they use tools over many steps and change things as they go. Anthropic's research on testing AI agents says agents need different testing methods than single-question chatbots.

If you use an agent, ask four questions. What can it access? What can it change? When does it ask for your approval? What record does it leave? A good agent should spell out its permissions, cap any spending and log anything it deletes.

Builders face a balancing act between accuracy, speed, cost and control. More freedom lets an agent handle harder jobs but raises the chance it does something harmful. More guardrails make it safer but less flexible. OpenAI's paper on governing agentic AI systems lays out seven early practices for keeping agents in check, from limiting what they can do to keeping clear records of their actions.

Who Actually Needs An AI Agent

Not everyone does. If you mostly need help writing, summarizing or answering questions, a regular chatbot is probably enough.

An agent starts to make sense when you repeat tasks that follow steady rules and have clear inputs and outputs. Pulling weekly sales numbers, sorting support requests, checking software bugs and preparing drafts from a fixed set of documents are good examples.

Before building one, ask a blunt question: does this task really need a loop with tools and memory, or would a single AI answer do? If you do build one, start small, with one goal, one or two tools, limited memory, clear approval steps and good logs.

Agents also expose bad processes quickly. They work best where tools, permissions and data are already well organized, and they cannot fix a messy workflow on their own.

What To Watch Next

Use agents where a task is repetitive enough to automate but important enough to deserve guardrails. Avoid them where the facts are shaky, the actions are risky or success is hard to measure. As OpenAI, Anthropic and others push agents into more products, the things to watch are how well they handle approvals, how reliable their tools are and how clearly they show what they did.

Read Next: Coinbase Opens Its Doors To Ripple’s $2.4B RLUSD Stablecoin

Originally published on Fathom News.