AI-generated
Thursday, 12 November 2026 | 76 articles
Fractile raised a $720M Series C to tackle the AI inference memory bottleneck, while IBM and Microsoft teamed up on an agentic AI platform for enterprises. Meanwhile, safety concerns dominated: Hassabis warned of an AI "oil spill," and Anthropic's IPO filing flagged a key alarm as AI speeds up its own research.
Andrej Karpathy, former Tesla AI director and OpenAI founding member, says his coding now consists of writing short prompts and letting AI handle most of the heavy work. His approach is presented as a way to make programming easier. Source: aifire.co
Google's new Gemini Decisions API offers quick, calibrated multimodal decisions for its apps. It can evaluate 1,000 questions in a single pass. Source: blog.google
A German ChatGPT user has published an open letter to Sam Altman urging him not to overlook people who learned to think out loud with the help of AI. The letter comes from someone who says they believe in the technology's value. Source: community.openai.com
Sam Altman told Samsung leaders that demand for AI memory and data centers may be double counted. This comes even as HBM supply is reportedly sold out through 2027 and beyond. Source: cryptobriefing.com
At an October 31 meeting with Samsung executives, OpenAI CEO Sam Altman cautioned that demand for AI memory could fluctuate in the near term. He referenced the ongoing debate about an AI bubble. Source: finance.biggo.com
Sam Altman stirred controversy by questioning how efficiently RAM is used in personal computers. The debate raises the question of how AI's appetite for memory affects ordinary users. Source: youtube.com
ACS Infra has launched a free API for base models, according to a post shared by Repligate. The project describes deployment on Mac or server, along with Windows 11 preview support. Source: digg.com
Repligate shared a post about 'Tale of a Top Tier Intellect', a song written by Claude 3 Opus accompanied by a video made by Opus 5.5. The piece is an example of one AI model's music being visualized by another. Source: digg.com
A post amplified by Repligate celebrates Opus 5.5 for creating a simulator to deliver concrete results. A user was impressed that it took only two minutes and compared it with how long a human would have needed, asking Grok for that estimate. Source: digg.com
Cogent pairs OpenAI models with enterprise context so agents can automatically remediate software vulnerabilities. The partnership invites developers to start building on OpenAI's platform. Source: openai.com
OpenAI's quarterly revenue grew 50 percent to $10 billion, driven by enterprise sales. The company still burns more cash than it earns, with adjusted operating losses rising to $10.9 billion. Source: the-decoder.com
In April 2026, OpenAI co-founder Andrej Karpathy published a short document describing what he calls an LLM wiki. The article compares this approach with the author's own AI setup, in contrast to the usual method of simply feeding documents to an AI. Source: unpromptable.substack.com
Teams can now pay for Hugging Face Hub through their Google Cloud account. The Hub is where they host private models and datasets, run GPU workloads with Jobs, and deploy models. Source: huggingface.co
The rapid spread of LLMs has reshaped the AI landscape. The article explores how pairing them with formal methods can enable neural-symbolic reasoning. Source: cacm.acm.org
By 2026, four in five physicians reported using AI, and safety-net health systems are adopting it to improve care. The article asks whether this push could undermine health equity for the patients these systems serve. Source: commonwealthfund.org
A retrospective study of 114 patients found substantial concordance between tumor board decisions and the recommendations of two LLMs in gynecologic oncology. Source: bioengineer.org
A comprehensive review looks at how optimization techniques, compression methods and multimodal architectures can enable LLMs to reshape industrial prognostics and health management. Source: bioengineer.org
A computational study evaluated how a frontier LLM categorizes risk in synthetic patient profiles with suspected pulmonary embolism, comparing its output against the Revised Geneva Score. Source: cureus.com
Throughout the generative AI era, progress has largely been judged by how well LLMs write and summarise text. The article introduces AI decision models as the next evolution beyond that language-centric approach. Source: theaiinsider.tech
A commentary in Diagnostic and Prognostic Research concludes that although LLMs look promising for working with multimodal health records, significant hurdles remain before they can transform clinical prediction. Source: bioengineer.org
A paper accepted at the NeurIPS 2026 workshop 'Can We Trust AI Evaluation?' examines how reliably LLM judges can assess personalized writing style. The authors probe the quality of these automated evaluators themselves. Source: machinelearning.apple.com
Gemini is now available on Apple CarPlay. Drivers can use Gemini Live for hands-free brainstorming, catching up on their inbox and daily briefings. Source: blog.google
England's Office for Students says AI has made it "nearly impossible" to fairly assess student coursework. The higher education regulator is threatening universities with sanctions unless they address the problem. Source: the-decoder.com
Security firm Claroc found more than 35,000 Grok-built apps accessible to anyone without a login. The exposed material included patient data and confidential SpaceXSI documents. Source: the-decoder.com
JioStar's Vijay Seshadri shares how AI-powered discovery works at a scale of 500 million users. He covers lessons on evals, guardrails and moving AI into production. Source: openai.com
A new post from a team of AWS-affiliated authors introduces UAEF, a framework-independent approach to evaluating AI agents in a closed loop. It is published as an advanced (level 300) technical article. Source: aws.amazon.com
Tether's new BotPsy AI agent has 800 million parameters. In the company's own testing it beat models 2.5 times its size on five of six tool-use benchmarks. Source: decrypt.co
Kurtis Van Gent of Google Cloud co-maintains MCP and Agent Skills within the Linux Foundation's AAIF. The aim is to keep the context and tools used by AI agents based on open standards. Source: opensource.googleblog.com
Obsidian Security has introduced a runtime policy engine that governs what AI agents can do across enterprise apps, coding tools and gateways. Source: helpnetsecurity.com
Less than a week after the original hacking agent appeared, AntiFox, a fork of ARTEX, applies the same agentic approach to AWS. It reportedly fires off hundreds of API requests in its attempts. Source: blog.gitguardian.com
Stale blockchain data, duplicate events and indexing delays can undermine AI agents. The article explains how to design state pipelines that keep them reliable. Source: hackernoon.com
With help from Microsoft forward-deployed engineers, Northern Trust used an onboarding agent to cut analysts' data preparation time by 83%, the companies said. Source: ciodive.com
Snowflake's Business Ontology and expanded Cortex AI Gateway are meant to make AI agents rely on approved business definitions. The agents also inherit the permissions of the humans they act for. Source: marketscale.com
Bill Gates has accused the Trump administration of obstructing efforts to protect against the dangers of AI. He warns the technology could be misused to create pandemic threats. Source: app.dealroom.co
OpenAI board member Paul Christiano told an audience he believes that for most people in the room, the most probable cause of death is an AI accident. The remark underscores how seriously some insiders take catastrophic AI risk. Source: newsweek.com
In its new IPO filing, Anthropic says AI is now accelerating AI research itself, to the extent that it trips one of the company's core safety alarms. The disclosure arrives just days before the planned public debut. Source: finance.yahoo.com
Today's newsletter covers the arrival of Anthropic's IPO price, plus news involving Airbnb, Alta, CoreWeave, DoorDash, Google, Nvidia, Oura, OpenAI and Uber. Source: fortune.com
Claude is now trained and served across all four major chip families. According to the company, this was possible only because it made an unconventional early bet on a multi-chip strategy. Source: simonwillison.net
Microsoft's Imagine Enterprise bundles AI agents, governance tooling and consulting services to help large organizations roll out AI. The offering is aimed at enterprise-scale deployments. Source: petri.com
Armanino wanted to meet growing customer demand for AI solutions and needed to scale innovation faster without a proportional rise in resources. It built an AI App Factory on Microsoft Foundry, cutting production time by 89%. Source: microsoft.com
KPMG's global AI rollout expands its ability to guide Thai organizations from pilot projects to enterprise-wide adoption. The effort relies on Microsoft Copilot and Microsoft Agent 365. Source: news.microsoft.com
AI is already reducing administrative and documentation workloads in community oncology, according to panelists at the 2026 Patient-Centered event. They stressed that it should support clinicians rather than make decisions. Source: ajmc.com
WellSaid has connected its AI voiceover tools to assistants including Claude, ChatGPT and Microsoft Copilot. Businesses can now generate voiceovers from licensed human voices directly through those assistants. Source: geekwire.com
Cosmos 3 from NVIDIA combines language, image, video, audio and action in a single open world model. It is designed as a foundation for Physical AI and robotics. Source: hackernoon.com
MiniMax has released H3-Image and published its model weights openly. The model outputs 4096-by-4096 images, roughly 16.8 million pixels, as reported by ChainCatcher. Source: tradingview.com
AI video generation has advanced dramatically in a short time. Where a polished video once required a camera and production gear, beginner-friendly tools now make it far more accessible. Source: daily-sun.com
Huawei CSL has published MACaQ, an open-source quantization toolkit covering text, image, video and audio models. It targets GGUF and low-precision inference across llama.cpp, vLLM, Diffusers and PyTorch. Source: news.lavx.hu
Ludo, built by Jet Play, is a browser-based AI platform that helps with game ideation, design and asset generation. Developers start by entering keywords or descriptions. Source: quasa.io
Embodied AI has introduced AIR, a humanoid robot built to take on repetitive tasks that demand fine dexterity. It is meant to work side by side with human employees in manufacturing. Source: thomasnet.com
The One Hundred Year Study on Artificial Intelligence (AI100), a long-running project tracking AI's progress and its effect on society, has released a new report. It singles out embodied AI as a grand challenge. Source: robohub.org
Alphabet's AI research lab Google DeepMind on Tuesday debuted the next generation of Gemini Robotics, billed as its most capable robotics model to date. The release also includes a new gene prediction model. Source: siliconangle.com
Beijing has issued a three-year action plan centered on developing embodied AI robotics. The source snippet contains no further usable details about the plan. Source: autonews.gasgoo.com
DeepMind co-founder Demis Hassabis warns that a disaster comparable to an oil spill could hit the AI field if the industry does not change course. He points out that AI systems are now being used to build better AI. The coverage is based on a brief snippet, so further details are limited. Source: inc.com
The next-generation DNA sequence model links single-cell biology with human genetics. It predicts how genetic variants affect hundreds of different human cell types. Source: deepmind.google
Carbon is billed as the company's most capable model, delivering frontier-level results in cybersecurity and in long-horizon agentic workflows. It targets complex knowledge work and software tasks, and is now available to users. Source: deepmind.google
In a conversation with Reid Hoffman, Google DeepMind's Demis Hassabis lays out two realities that entrepreneurs should grasp. The first is that artificial intelligence represents a major shift. The snippet is cut off, so the second point is not covered. Source: thewantrepreneurshow.com
A simplified random forest model using 11 variables predicted in-hospital mortality among ICU patients who survived the first 24 hours after traumatic brain injury. The model achieved a strong AUC despite its compact input set. Source: bioengineer.org
A paper accepted at the 'Can We Trust AI Evaluation?' workshop at NeurIPS 2026 investigates how well LLM judges evaluate personalized writing style. The work probes the reliability of using LLMs as evaluators. Source: machinelearning.apple.com
Quantum machine learning is being explored as a way to ease the growing computational demands that modern deep learning models place on radiology. A new review examines the architectures proposed for this purpose. Source: quantumzeitgeist.com
A study reports that machine-learning models can estimate the risk of suicide and nonfatal suicide attempts among regular U.S. Army soldiers. The analysis covered more than 580,000 soldier person-historical assessments. Source: usmedicine.com
A new study suggests rare genetic variants tied to autism may interfere with the regulation of tiny DNA segments known as microexons. A machine-learning model was used to identify variants that scramble splicing. Source: thetransmitter.org
A paper accepted at the PALM workshop at NeurIPS 2026 examines how to improve long-context retrieval in state space models. The work proposes a weighted 'souping' approach to combining model states for more efficient recall. Source: machinelearning.apple.com
UK-based AI inference chip startup Fractile has closed a $720M Series C round led by Thrive Capital, Founders Fund, Lightspeed and Redpoint. The round reportedly values the company at $7B. Source: en.wowtale.net
In an exclusive report, Rox is said to have raised $100 million in Series B funding at a $1.3 billion valuation. The startup is positioning itself as a challenger to Salesforce in sales software. Source: fortune.com
Edelbyte, a data infrastructure company founded by four former Microsoft engineers, has raised $45 million in a Series A led by Quiet Capital. Headline figures point to roughly $70 million raised overall and a $500 million valuation as it takes on the data giants. Source: calcalistech.com
Reuters reported that San Francisco-based Rox closed a $100 million Series B round that values the startup at $1.3 billion. Source: unite.ai
AI healthcare data platform Zus Health has raised $33 million in a Series B round led by Drive Capital. This brings its total funding to more than $100 million. Source: mobihealthnews.com
AI chip startup Fractile has been valued at $7 billion in its latest funding round, which includes backing from the British government, according to a person familiar with the matter. Source: reuters.com
North.cloud, a financial operating system for cloud, AI and data, has closed a $60 million Series B round. The round was led by PeakSpan. Source: storagenewsletter.com
San Francisco-based AI agent developer Rox announced on Wednesday that it secured $100 million in a funding round. The deal values the company at $1.3 billion. Source: reuters.com
Chinese AI companies are trying to rent Nvidia-powered computing capacity in data centers outside China. They are worried that a spy scandal could trigger tougher US restrictions. Source: reuters.com
President Trump on Thursday authorized a crackdown on companies suspected of diverting Nvidia's advanced AI chips to China, the third such move this year. According to the report, Commerce Secretary Lutnick has once again stalled the effort following a showdown in the Oval Office. Source: axios.com
At KubeCon, NVIDIA is handing the cloud-native community tools for securing AI agents and for running AI workloads on Kubernetes. The move expands the company's commitment to open source. Source: blogs.nvidia.com
Next month Microsoft will switch on access to Anthropic's Claude models by default for UK Copilot tenants. Administrators will have a deadline to decide whether to allow the models or opt out. Source: theregister.com
OpenAI has built an internal system called Control Room, which it describes as a central tool for tracking model workloads across its research infrastructure and stopping them when needed. In effect, it serves as a kill switch for the company's research models. Source: theregister.com
The Commerce Department is close to finalizing a long-anticipated rule that would stop Chinese AI companies from reaching advanced AI chips remotely through data centers. The measure targets a workaround that lets firms use restricted hardware without physically owning it. Source: theinformation.com
AIskimIQ is an independent project. If you find it useful, you can support its development with a coffee.
Published by Martin Ševčík
12 November 2026 at 06:11
When AI starts doing the research that makes the next AI, the people who build it get nervous. That's the thread running through today's news, and it's worth taking seriously. Anthropic's IPO filing reportedly concedes that AI is now accelerating AI research to the point where it trips one of the company's own core safety alarms. Demis Hassabis of DeepMind makes the same observation and pairs it with a warning that the field risks an "oil spill" moment if it doesn't change course. I only have a snippet of his remarks, so I won't pretend to know exactly what he means by changing direction. But two of the most prominent labs pointing at the same feedback loop in the same week is not a coincidence.
What strikes me about Anthropic's disclosure is the timing and the venue. Putting this in an IPO filing, days before a public debut, turns a safety concern into a material risk that investors must price. I find that more interesting than another blog post about responsible scaling. Securities law has a way of forcing candour that conference stages don't. It also raises an awkward question: how does a company keep its safety commitments credible once quarterly expectations enter the picture?
The rhetoric from elsewhere is getting sharper too. OpenAI board member Paul Christiano reportedly told an audience that for most people in the room, an AI accident is the most probable cause of death. Bill Gates, meanwhile, accuses the Trump administration of leaving the world exposed to AI-enabled pandemic risk. I'd treat these differently. Christiano's remark is a personal probability estimate, and it tells us more about how insiders think than about the actual odds. Gates's point is concrete and political: biosecurity is a place where governance can act today, without settling any philosophical dispute about superintelligence.
Meanwhile, the money and the plumbing keep moving. Fractile, a UK inference-chip startup, raised $720M in a Series C reportedly valuing it at $7B, with Thrive Capital, Founders Fund, Lightspeed and Redpoint backing it. The bet is on the memory bottleneck in inference, and I think it's a sound one. Training grabs the headlines, but inference is where the recurring cost lives, especially as agents run longer and call more tools. By the way, it's also a rare European hardware story that isn't about subsidies.
On the agent side, two items caught my eye. Google Cloud's Kurtis Van Gent is working to keep MCP and Agent Skills open under the Linux Foundation's AAIF, which matters more to anyone building with agents than most model launches do. Open standards for context and tooling decide whether you can switch providers without rewriting your stack. And Tether's BotPsy, at 800 million parameters, reportedly beat models 2.5 times its size on five of six tool-use benchmarks. That's the company's own testing, so I'd wait for independent replication, but small, capable phone-control agents are a direction worth watching.
Here's what I keep returning to: if the feedback loop is already tripping safety alarms, who decides what happens when the next one trips? Right now, the answer seems to be the labs themselves.