Every story, at length

Mistral announced Thursday that the weights for Mistral Large 4 — nicknamed “Le Chonk” — are now available on Hugging Face and through its API. Weights are the numerical training that makes a model work; releasing them means an organization can take the full model, deploy it on its own servers, and keep its data and operations in-house. The model was first introduced earlier this month, but this is the step that turns a demo into something a company or government agency can actually own.

The pitch is aimed squarely at cybersecurity buyers. Mistral argues that organizations defending networks and code should not depend on an outside AI provider whose access could be cut off or whose terms could change. Pierre Stock, Mistral’s vice president of science, put it plainly: open weights mean access “can’t be cut off by definition.” The company says the model can detect, analyze, and respond to threats across code and IT infrastructure, and can be deployed on-premises so sensitive data never leaves a customer’s own systems.

The most concrete comparison in the announcement is economic. Mistral says its model performs comparably to GLM-5.3, a Chinese open-weight model previously considered among the best for cyber work, but uses 47 percent fewer tokens per task. Tokens are the chunks of text a model processes, and fewer tokens per job means lower running costs. Mistral also claims the model outranks all other open models on agentic coding tasks in Artificial Analysis’ Coding Agent Index.

Mistral’s CEO Arthur Mensch framed the release as a matter of who controls advanced AI: a tool “for empowerment, not dominance.” The company says it achieved state-of-the-art performance at a fraction of the compute and capital spent by closed labs. Stock also addressed worries about rogue AI agents, saying the company does not want to “fearmonger” and believes security problems are solvable.

The release, Granite Switch 4.1.1, comes just over a week after version 4.1 and is fully backwards compatible, meaning no migration work for anyone already using the earlier version. The update improves a multilingual speech-to-text model that now handles 10 languages with native punctuation and capitalization, so transcripts can feed directly into other systems without a separate formatting step. It also updates a decision-making adapter with better performance on a test set called the Jev Decision Index.

What makes the release notable is not the features themselves but the speed and cost of shipping them. IBM says that apart from training the new speech model and the decision adapter, nothing was recomputed. The base model is untouched; users simply compose the latest adapters they want and deploy. That is a meaningful shift for businesses running AI systems, because updating a model typically means retraining or redeploying large, expensive infrastructure. Here, the update arrives as a smaller, swappable component.

IBM says this pattern can extend beyond its own patch releases. The company is working on tutorials and an Adapter SDK so developers and third-party providers can build their own adapters, iterate on versions, test features, and ship updates the way they ship software. That would let outside teams add specialized capabilities to IBM's models without waiting for IBM to do it.

The most surprising detail is how little had to change to deliver the update. In a field where model updates often mean recomputing large parts of a system, IBM's claim that only the new components were trained — and that the base model required no work at all — is the detail a reader would repeat to a colleague.

The system, called WxCrop, is described in a paper posted to arXiv on 29 October. It combines weather data with agricultural records to predict crop loss for both past and forecast dates, and was demonstrated in Florida on hurricanes, freezes and dry heat. The loss model was calibrated on insurance records for each hazard, and the authors say the same procedure extends to other hazards and regions.

What sets the agent apart is that it answers questions in terms a farmer or official can act on: for each crop group, it estimates the loss as a fraction of expected season value, with every number sourced. That is a shift from raw weather forecasts, which say where a storm will strike but not what the damage will be. The paper argues this closes a gap between knowing a storm is coming and knowing what to do about it.

Three checks support the reliability claims. The hurricane index detected recorded cyclone days with an AUC of 0.98 — a score where 1.0 would mean perfect detection. Recorded losses in held-out years were reproduced to within 2%, and independent production records confirmed the county rankings with a rank correlation of 0.63, a moderate-to-strong agreement on which areas suffered most.

The work is a preprint and has not yet been peer-reviewed. But for insurers, agricultural lenders and state emergency planners, the promise is faster damage assessment after a disaster and better preparation before one. The authors frame the intended users as governments, growers and scientists responding to agricultural disasters.

Deepgram said on Thursday it is buying Daily, the maker and maintainer of Pipecat, an open-source framework that coordinates the moving parts of a voice AI system — speech-to-text, language models, and text-to-speech — while keeping delays low. The companies did not disclose the price, but said Pipecat will remain open source.

The deal matters because Pipecat has become a default starting point for developers building voice agents, the software that talks to customers or employees in real time. By owning the framework, Deepgram positions itself at the center of that ecosystem without forcing developers onto its own models. Deepgram CEO Scott Stephenson said the company wants builders to use the best models for their use case, even if that means a competitor's, and to treat openness as a commercial advantage. Daily co-founder Kwindla Kramer said he chose Deepgram partly because of Stephenson's commitment to keeping Pipecat open source and vendor-neutral "forever."

The most striking detail is the shared history: both companies were in the same Y Combinator batch in 2016. Daily has raised over $62 million from investors including Tiger Global and Nat Friedman, while Deepgram reached unicorn status earlier this year with a $130 million Series C round. Kramer said Daily began experimenting with AI-to-human voice communication around the time GPT-4 launched, which led to Pipecat.

Alongside the acquisition, Deepgram is launching the Real-Time Context Standard, an open protocol co-developed with the Pipecat team that defines how voice agents should handle audio cues like a whisper and how a language model should respond. The move comes as voice AI infrastructure companies such as LiveKit and Vapi have raised multimillion-dollar rounds this year, and model makers like ElevenLabs have pulled in significant funding — a sign that consolidation is now reaching the tools layer, not just the models.

OpenAI now offers the same deal Anthropic launched earlier in October: subscribers on $100, $200, or $500 monthly plans receive that same amount in credits to use with the company's API — the programming interface developers use to build software on top of the model. Anthropic's Max and Team plans came first, with expiring credits matching the subscription price. OpenAI's version applies the credits to one of the subscriber's API organizations.

The practical effect is that people and businesses already paying for a premium ChatGPT subscription can now experiment with or run API-based projects without an additional bill, at least until the credits expire. For developers like Simon Willison, who maintains subscriptions with both companies, the change means public-facing demos and small tools can be shipped without new infrastructure costs. The credits are tied to existing spend rather than being a discount on future purchases.

What is genuinely new here is competitive parity, not invention. Anthropic set the terms, and OpenAI followed within weeks. For customers weighing one provider against the other, the subscription-plus-credits bundle is no longer a differentiator. The most useful detail for anyone comparing plans is that the credit amount tracks the subscription tier exactly: pay $200 a month, get $200 in API credits.

The move matters most to small teams and solo developers who already pay for a top-tier subscription and want to build on the API without committing extra budget. Larger customers with negotiated enterprise contracts are unlikely to see their economics change. The credits expire, so the benefit rewards active use rather than stockpiling.

The initiative, announced on Hugging Face's blog, brings together Carnegie Mellon University, Purdue University, Georgia Institute of Technology, Arizona State University and The George Washington University, along with the Astera Institute and the Open Research Community Accelerator. The goal is to encourage research institutions to release scientific work in formats that AI tools can actually process, rather than leaving it scattered across incompatible systems or static files that lack the documentation needed for reuse.

The first test of the idea is a climate challenge hosted by Carnegie Mellon, opening November 2 and running through November 13. Participants will use historical weather data from 210 U.S. cities, with records through 2018, to predict conditions from 2019 to 2026. They will forecast four indicators: average annual temperature, total annual precipitation, the hottest day of the year, and cooling degree days — a measure of air-conditioning demand. Teams must also propose a short climate adaptation policy idea, such as preparing cities for extreme heat or managing changing rainfall patterns. Results will be judged against simple statistical baselines, with final forecasting scores published in early 2027 once full 2026 weather records are available.

The demand for reusable scientific data is already visible on Hugging Face's platform. It hosts more than 17,000 public datasets across core scientific fields including genomics, protein design, chemistry, weather and climate, Earth observation and astronomy — a figure that approaches 29,000 when medicine and mathematics are included. Researchers shared more than three times as many new science datasets in the first nine months of 2026 as in the same period of 2025. The U.S. federal government invests roughly $190 billion annually in research and development, and agencies like the National Institutes of Health and National Science Foundation increasingly require funded work to be shared publicly.

What is genuinely new here is the institutional coordination: participating universities formalize their involvement through a memorandum of understanding, and Carnegie Mellon's Open Source Programs Office is already encouraging its researchers to release suitable datasets and models on Hugging Face. The challenges are designed to build evidence about what happens when AI-ready scientific resources are actually made available for reuse.

Chief executive Andy Jassy told investors on Thursday's earnings call that the company expects 2026 cash capital expenditure of approximately $220 billion, with the majority going to AWS, and that spending will rise "meaningfully" in 2027. He said the company has "a clear line of sight" to the financial returns from that investment.

The spending is aimed at data centers and other infrastructure to support cloud demand driven by AI. AWS revenue grew 39% from the year-ago quarter, the sixth straight quarter of steady growth, and Jassy forecast growth above 40% in both 2027 and 2028. He put AWS's current annualized revenue run rate at $184 billion and said the business could reach a trillion dollars in annual revenue within five to ten years.

Investors were not fully convinced. Wall Street analysts repeatedly pressed Jassy on the spending, the signals that growth will materialize, and expected profit margins. Shares fell about 6% in after-market trading. Jassy called the AI boom a "once in a lifetime opportunity" and said Amazon would "invest to continue being the leader," adding that the plan would "forever change the size of Amazon."

Amazon is not alone in spending heavily. Microsoft also posted record revenue growth driven largely by its cloud business and has spent heavily on AI infrastructure, showing that the industry's biggest players are betting large sums that enterprise AI adoption — which Jassy said is still "in the very early days" — will justify the outlay.

The company announced the change on its third-quarter earnings call, framing it as a way to protect its data from AI scraping and to push users toward licensing deals with model providers. But the numbers behind the decision are stark: search traffic referrals dropped by more than 1 million per week on average in October, and Reddit warned that daily active users could fall by a low-double-digit percentage next quarter if the trend continues. U.S. daily active users already fell from 53.2 million in Q2 to 50.7 million in Q3, down 2% year-over-year, while global daily actives were flat at 130.3 million.

Despite the audience decline, Reddit is making more money per user than ever. Revenue rose 53% year-over-year to $895 million, ahead of Wall Street's roughly $875 million expectation, and net income climbed 74% to $283 million. Average revenue per daily active user in the U.S. jumped 53% to $13.82. Logged-in users are worth more to advertisers because they provide signals about interests and behavior that enable better ad targeting, which is central to the company's bet that a walled garden can offset shrinking traffic.

The most surprising detail is the scale of the referral collapse at a site that built its identity on the open web. Co-founder and CEO Steve Huffman wrote that the shift is "a big change for me" as someone whose life has been defined by the open internet, but called it the right decision for the platform and the business. The rollout will not be immediate: Reddit will test the logged-in experience in select markets outside the U.S. in the fourth quarter, with no timeline yet for a global rollout. Logged-out front-page access will remain important, and the company said users can still browse anonymously via a rebuilt incognito mode.

For publishers and platforms watching their own search traffic erode, Reddit's move is a signal of how the economics of the open web are being renegotiated. Google's own earnings have shown the same pressure, as AI summaries answer queries without sending users to websites.

The deal values ADT at $15.3 billion including debt, the companies announced on Thursday. Thrive Holdings is the holding-company spinout from Josh Kushner's venture firm Thrive Capital, and it counts OpenAI as both a partner and an investor. Its stated mission is to buy companies and hold them long-term, revitalizing them with AI. It currently holds more than 75 companies.

OpenAI's role is concrete: it supplies AI models, technology, and what the source calls forward-deployed engineers — people embedded in Thrive's portfolio companies to help apply the technology. Neither ADT nor Thrive said which AI-powered security products will come first, but the press release said ADT "will form the foundation of Thrive Holdings' smart home security business," with plans to invest and release new products.

The acquisition lands as AI-powered physical security is drawing attention from big consumer tech names. Apple recently entered the field with smart home and security products built with LG and Schneider Electric, and Amazon updated its smart home line with AI-powered home cameras. ADT already partners with Apple and Google, and the source says those partnerships are expected to continue. Some startups are even offering AI dog robots as roving security guards.

For ADT customers, the immediate change is ownership, not product. The company's monitoring services and existing partnerships remain in place for now. The longer-term question is how quickly OpenAI's models and engineers turn a traditional alarm and monitoring business into something more automated — and whether that changes what customers pay or what competitors must match.

The executive, Alice Huang (also known as Huang Le), worked for Megaspeed, a Singapore-based AI cloud company that the Commerce Department investigated in 2025 over concerns it was helping Chinese firms sidestep US export controls by reselling restricted Nvidia chips. US agencies had been surveilling Huang for about a year to learn more about her ties with China, the Times reported.

The concern among officials, according to the sources, was that Huang had grown close enough to Jensen Huang — she was photographed among groups at business dinners with the Nvidia CEO in the Singapore region — that she could have been gathering information about Nvidia's plans and about the CEO's efforts to persuade the Trump administration to allow more advanced chip shipments to China.

Nvidia pushed back firmly. A spokesperson told TechCrunch that the CEO's interactions with Huang were no different from those with any other regional customer: "Jensen's interactions with Ms. Huang are not unusual or unexpected. Meetings with customers are part of a C.E.O.'s job; Jensen talks with thousands of customer executives every year." The company added that regional cloud customers receive "ordinary business information — not sensitive NVIDIA trade secrets — and zero information relevant to national security."

The story lands at a sensitive moment for Nvidia, which has been lobbying Washington for room to sell more advanced chips to China while US export rules tighten. The report does not allege that Jensen Huang knew of any intelligence activity, and no charges are described. But the episode shows how export-control enforcement has expanded beyond tracking shipments to tracking relationships around the executives who shape chip policy.

David Durman, who supports researchers applying for grants at the University of Kentucky, secured the exemption after approaching the American Humanist Association (AHA), a nonprofit that advances secular humanism. Durman is a nontheistic Humanist — he does not believe in a god but holds a moral framework centered on human well-being and the active application of the mind. He argues that relying on AI to do cognitive work diminishes people's ability to act as moral agents, and he also cites the technology's environmental costs and what he calls its theft of human creativity.

The university confirmed it granted the accommodation after reviewing the request against his job responsibilities. Spokesperson Jay Blanton said the institution determined an accommodation could be provided while Durman continued performing the essential functions of his position, and noted that accommodations may be reviewed as technology and job duties evolve. Durman has not been required to use AI in his administrative role, but colleagues and leadership have recommended trying the tools, and he sought the exemption now in case suggestions become mandates.

Durman was inspired by Erin Maus, a Unitarian Universalist software engineer who earlier secured similar protections — notable because writing code without AI assistance has become rare in her field. AHA staff attorney Alexandria Huiskens said the organization wants Humanists and other nontheists to know they fall into the same protected categories as workers from Catholic, Universalist, or other faith traditions.

This is the first publicly discussed case of a Humanist winning such an accommodation, and it signals that religious-exemption claims around AI are not limited to traditional faiths. Employers now have a concrete precedent for how a secular belief system can ground a legally protected refusal to use AI tools at work.