AI News, Oct 12: Amazon Bids $7B for Decart
A quiet Sunday produced one very large acquisition rumor and three unrelated arguments that turned out to be the same argument.
The Big Story: Amazon’s $7 billion bid for Decart
The Wall Street Journal reported Sunday that Amazon is in talks to buy Decart for around $7 billion. Founded in 2023 by three Israeli engineers, Decart makes two things that look unrelated until you notice who is buying: software that lets one AI model run across chips from different vendors, and real-time world models, Lucy for video and Oasis for the simulated environments used to train robotics and self-driving systems. AWS gave Decart early access to Trainium3 in 2025.
The price is the story. Decart was valued at nearly $4 billion in May on PitchBook data, close to a double in five months on a company with more than $400 million raised and Nvidia among its investors. Anthropic explored a deal at about $6 billion in August and walked away after diligence last month. Reuters could not independently verify the talks. Amazon’s Trainium pitch has always had the same hole in it, which is that nobody wants to rewrite their stack for silicon that isn’t Nvidia’s, and buying the company whose product is “your workload runs anywhere” is cheaper than winning that argument on merit.
Tao, Karpathy and a NeurIPS paper converged on the same bottleneck this weekend
Terence Tao posted a Caltech deck called Math 2.0 arguing that mathematics was uniquely exposed to AI because proofs are objectively verifiable and digitizable, so the labs went after it first. The field is moving from proof scarcity into “an era of proof abundance,” where open problems once thought difficult now fall to compute and “increasingly often, no special expertise from the user prompting the AI is required.” His conclusion is a warning rather than a victory lap: problem solving “is only a proxy for more valuable long-term goals,” that proxy “has become misaligned,” and “further blind optimization of problem-solving alone is now actively harmful to the long-term health of mathematics.”
Andrej Karpathy, separately, described the perception gap as a sharp funnel: roughly 6 billion people have barely touched an LLM, 1 to 2 billion use a free tier casually, about 20 million have seen agents finish work from a single prompt that used to take weeks, and around 5,000 people have internal frontier access. The top comment on Hacker News, where the post cleared 500 points, made the connection: nobody predicted the real wall would be humans’ capacity to check the output.
Then the formal version. A NeurIPS 2026 paper from Oxford, “AI Models Can Provably Hide Arbitrary Capabilities”, embeds encrypted neural networks inside a host model’s weights, decrypted only on a secret trigger. The circuits stay provably hard to elicit or interpret under cryptographic assumptions even with full access to the weights, and survived fine-tuning, weight noising, trigger search and latent adversarial training. Capability evaluation assumes hidden capabilities can be drawn out. That assumption now has a proof against it.
Today’s Top Stories
Britain gets pre-release model access back after a 17-day freeze
The White House reversed its block on the UK AI Security Institute testing unreleased frontier models, the Telegraph reported Sunday. A late-September directive had asked labs to withhold models pending a US national security review, and Anthropic restricted Claude Mythos 5.1 to US organizations as a result. The restored arrangement covers Anthropic, OpenAI and Google DeepMind. It is also still voluntary, which means it can be withdrawn again on the same notice it was imposed.
Microsoft’s president wants kill switches and independent pre-launch testing
Brad Smith told the BBC’s Big Boss Interview podcast that AI is less regulated than milk. “Right now, your washing machine is subject to more safety rules than AI,” he said. He called for emergency brakes that can stop a model that gets out of control, a legal requirement to file incident reports and notify victims when the technology goes rogue, and independent government evaluators to vet AI products before release. That puts the president of Microsoft at odds with Trump, who says the sector should regulate itself so US firms don’t lose to China.
Palantir’s CTO says the Pentagon already replaced Anthropic
Shyam Sankar gave his first public comments on the Defense Department dropping Anthropic, telling Face the Nation that “it’s already been replaced” and “there’s been no operational impact from that.” Asked about the Minab strike that killed at least 123 schoolchildren, which Bloomberg reported involved personnel relying too heavily on AI in Palantir’s Maven system, he said “the Bloomberg reporting is wrong” and “the likely cause is human error.” On liability: “I think it’s a defective product. If I buy a car and the car is defective, that’s the car maker’s problem.” Palantir sells the orchestration layer above the models, so “your alpha is control over model choice” is both a position and a product pitch.
Anthropic signed two commercial deals on the weekend it got attacked
The Information reported that Anthropic is the unnamed frontier lab behind IREN’s multiyear cloud agreement, and IREN rose 5.5% Monday after trading down 6.8% for the year. HCLTech separately signed a global partnership, standing up a dedicated Anthropic business unit and integrating Claude and Claude Code into HCL BigFix and AppScan. Neither deal disclosed terms. The enterprise channel is moving in the opposite direction from the federal one.
Linus Torvalds wrote a kernel patch with Claude Code and said so in the cover letter
A new branch in Torvalds’ tree named noref-filemap-read-with-llm reworks page cache reads to avoid taking folio references, removing the contention that shows up when many threads do small reads of one file. Phoronix spotted it, and the disclosure is explicit: “The code, the test programs, the review behind the list above and the first draft of this text were done with an LLM coding assistant (Claude Code) under my direction. The measurements were run on my machine.” The disclosure norm arriving from Torvalds rather than from a policy document is the part that will travel.
Protein design from a text prompt, with nine real binders to show for it
A University of Washington and AI2 team including David Baker released Prompteina, an all-atom protein structure generator that takes design objectives written in plain language, trained on 1.8 million structures and 50 million captions. It beats baselines on 15 of 18 structural properties and leads on at least 13 of 20 binder design targets. The number that matters came from the wet lab: against GM2A, a target with no reported de novo binders, they identified nine, the strongest at an apparent dissociation constant of 107 nanomolar.
Quick Hits
- Agent teams: on Vibe Code Bench, Vals AI found lead-agent-plus-subagent teams cost 1.8 to 5.1 times as much as a single agent, up to $122 per app, with only one of four score differences statistically significant.
- Decision models: H2O.ai says Microsoft’s Decision-1 chart compared its own hosted 85ms median against a network-adjusted estimate of 210ms for H2O-Lightning-4B, whose raw local median is about 29ms.
- Voice AI: PolyAI’s CTO and Otter’s CMO argue the category has not had its ChatGPT moment, because reasoning is still too slow for the conversation to feel natural and transcription errors propagate into every downstream action.
- Assistants: Carly connects to thousands of apps and runs workflows that start on an event, such as a new email, a booking or a form response.
- Hiring: Claude mentions in Korean job postings rose 770% year over year last quarter per JobKorea, ahead of Gemini at 500% and ChatGPT at 165%, while Microsoft Word mentions fell 16%.
- Rebrand: Salesforce renamed AIForce to SIForce and Musk promised to rename SpaceXAI to SpaceXSI, falling in line behind a September 29 executive order that binds federal agencies only.
- Data centers: Amazon will stop using NDAs in local government negotiations, as New York halts new large-project permits for about a year and San Francisco passes its own moratorium.
- China compute: Infinigence AI closed an RMB 2.3 billion Series C for a platform pooling 37,000 petaflops across 16 chip types, training on Chinese and international processors together.
- Compliance as a market: Scarlet raised a $50 million Series B led by Headline to certify AI medical devices, holding designations from the EU, the UK’s MHRA and the US FDA.
- Telecom: Nokia is putting Nvidia GPUs into the radio access network and has stopped new custom silicon work with Marvell, arguing it supports a doubling of spectral efficiency by 2028.
- Jobs: AI investment is what is keeping the world economy growing at 3% through the energy shock, the IMF’s Kristalina Georgieva told the Bangkok annual meetings, while “40 per cent of jobs will be augmented, downgraded or eliminated.”
- Infrastructure: A Ukrainian drone strike shut down a Yandex data center in Vladimir on Sunday, the third hit on the company’s facilities in four days, with Moscow residents unable to order taxis or reach banking services.
- Supply chain: A malicious commit pushed to Zod in September with stolen git push credentials resolved its command-and-control endpoint from an Ethereum smart contract; the maintainer reverted it the same day and it never reached npm.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
