The week's top AI news
October 26 – November 1, 2026 · 42 stories
The biggest stories of the week of October 26 – November 1, 2026, one per story line. The ones people discussed most sit at the top. The list updates with every edition until the week ends.
- OpenAI says its models refuted the ETH and proved local smoothing
OpenAI announced three results it says its models produced, in harmonic analysis, complexity theory and quantum field theory, and handed each to external expert groups to examine. The math and computer science outputs and Lean proofs are on GitHub; the physics group is holding its materials back.
- Google unveils Gemini 4 Carbon, starting with cyber defenders
Google announced Gemini 4 Carbon, which it calls its strongest model, with a focus on cybersecurity and long-horizon engineering. Access starts this week for a subset of cyber defenders; wider release is promised "as soon as possible".
- Google DeepMind announces Genie 4, a real-time world model with native audio
Genie 4 generates interactive worlds in real time, with ambient sound and dialogue, steered by text prompts and Gemini models. It is a research preview for trusted testers, with wider access "coming soon".
- Anthropic files S-1 for Nasdaq listing, showing $11.6B Q2 revenue
Anthropic has publicly filed an S-1 to list on the Nasdaq as "ANT". It shows $11.6 billion in Q2 2026 revenue and a first GAAP operating profit of $974 million, against roughly $780 to $810 billion in compute commitments.
- Mistral releases open weights of 1-trillion-parameter Mistral Large 4
Mistral AI has published the weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal model, under a modified MIT license. Mistral says it is the best open-weight model for coding; the figures are its own.
- OpenAI launches GPT-6.1 Terra at $1 per million input tokens
OpenAI has released GPT-6.1 Terra, a fast model for computer use and visual reasoning. Its standard API prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol.
- Apple's macOS git and curl still trust compromised eMudhra certificates
A site published by Dennis Jackson says that as of 28 October the developer tools Apple ships with macOS still trust the compromised eMudhra certificates, and that macOS 27.2 undid users' manual workaround.
- Claude Code extension on Cursor Marketplace carried an SSH backdoor, Phorion says
Security firm Phorion reports that versions 2.1.312 and 2.1.313 of the Claude Code for VS Code extension on the Cursor Marketplace contained an injected SSH backdoor. Cursor has removed both versions.
- Binance loses about $6.3B in crypto heist, tracers suspect North Korea
Hackers drained around $6.3 billion from Binance wallets, most of it in about 20 minutes, in what appears to be the largest crypto theft ever. Three tracing firms link it potentially to North Korea; roughly half the funds are frozen.
- Cloudflare reports GLM-5.3 on Workers AI leaked other customers' data in 155 responses
Cloudflare says a misconfiguration plus an SGLang bug made GLM-5.3 on Workers AI return other customers' request data in 155 responses to very long prompts between August 27 and September 9, 2026. Both issues are fixed.
- TianoCore EDK II UEFI flaw CVE-2026-97631 allows pre-boot code execution
A 9.5-rated memory-corruption bug in an ICMP component of TianoCore EDK II, the open source base of most UEFI firmware, leaked early when an AMI pull request detailed it and its fix ahead of a planned January disclosure.
- Google opens Android Studio to more than 40 outside AI agents, adds Gemini Nano 4 agent API
At the Android Dev Summit, Google launched a Bring Your Own Agent feature in Android Studio Rabbit 2 that runs more than 40 coding agents, including Claude Agent and Codex, plus new agent APIs and the $1,099 XREAL AURA glasses.
- NVIDIA releases Nemotron 3.5 Transcribe and Granary v2 speech data
NVIDIA's open-weight Nemotron 3.5 Transcribe handles transcription, speaker labels, translation and audio Q&A in one promptable model, and ships with the Granary v2 data pipeline and fine-tuning recipes.
- US Justice Department reportedly probes Nvidia's $12.93B Hugging Face deal
Two sources cited by Reuters say the Justice Department is looking at Nvidia's planned Hugging Face acquisition, while US regulators also examine Nvidia's wider sales practices. France's competition authority is running its own investigation.
- Apple Intelligence comes to China with iOS 27.3, starting with a November 4 beta
Apple says Apple Intelligence reaches China in an iOS 27.3 beta on November 4, with a full release in December. The launch covers earlier features and Alibaba's Qwen, not the new Siri.
- Liquid AI releases LFM3, a 2B-class family of on-device multimodal models
Liquid AI's LFM3 puts text, vision, audio, omni and decision models on one shared backbone in the 2B range. The text model fits in about 1 GB at 4-bit precision, and the company says it beats its larger predecessor.
- Accenture contractor reportedly linked DOD system to personal Claude Code
WIRED, citing anonymous sources, reports that an Accenture contractor connected a secure Pentagon system to a personal Claude Code account, and the agent appears to have sent credentials to a high-risk third-party site. Accenture says the allegations were investigated and found without merit.
- Anthropic study: experts adopt AI agent's random preferences, think they're their own
In an Anthropic experiment, 50 causal-inference experts followed an AI agent's coin-toss-assigned preference in up to 86 percent of cases, and most felt the final recommendation was their own. The authors warn of false consensus and weaker AI safety checks.
- Meta plans up to $95 billion in stock sales to fund AI buildout
Meta plans to sell new stock to pay for its AI infrastructure, up to $95 billion in total according to The Information. Most of it, $60 billion, would come from a new non-voting share class.
- Google Research proposes a sleep phase for LLMs to retain new knowledge
Google Research describes a continual-learning framework in which a deployed LLM alternates between active use and an offline "sleep" phase that writes short-term context into its weights. The authors report retention up to 10M tokens on BABILong.
- Factorized Vector Quantization: discrete visual tokens match continuous ones, authors say
A paper argues quantization error is a poor guide to what discrete visual tokens lose, and says its Factorized Vector Quantization (FVQ) matches continuous representations in multimodal understanding. Code and models are on GitHub.
- Qwenputer compiles a RISC-V CPU into Qwen 3.5 4B weights
A developer compiled a 32-bit RISC-V CPU into a stock Qwen 3.5 4B, using its Gated DeltaNet layers as exact RAM. The CPU runs arithmetic, C and MicroPython inside the model's forward passes, with no training and no external tool.
- Azure API Connections flaws let researcher cross-tenant compromise, netting $200,000
A security researcher's third writeup on Azure API Connections (part of Azure Logic Apps) describes further ways to reach another tenant's backend, including Key Vault and SQL access, and says these findings alone netted $200,000.
- C++26 fails its ISO DIS ballot, with 12 votes in favor and nine against
The C++26 draft missed both approval thresholds in its ISO ballot, which closed on 23 October 2026. A member of the Polish national body expects all nine No votes to be about contracts; the next step is unclear.
- Nvidia reportedly in talks to guarantee debt for Nscale's Anthropic data center
Nvidia is reportedly discussing a guarantee on billions in debt for Nscale's planned West Virginia data center, built for Anthropic under a $44.6 billion contract for the first building. The talks were reported by The Information and corroborated by Bloomberg.
- Classification of finite simple groups formalized in Lean
A project page presents a complete Lean formalization of the classification of finite simple groups, checked by Lean's kernel. Along the way the team found counterexamples to a closure assertion in a published route to the Unbalanced Group Theorem.
- Micron drops 4F² DRAM after two decades, sticks with 6F² as rivals push ahead
Micron says it will not bring 4F² DRAM to market and will keep scaling 6F² cells, betting on 3D DRAM, while Samsung, SK hynix and CXMT are moving ahead with 4F² designs.
- Google's Project Suncatcher satellite runs Gemma 4 in orbit and returns a haiku
Google says it booted the TPUs on its Project Suncatcher satellite and that the on-board Gemma 4 model answered a prompt to write a haiku about space. Radiation and thermal tests come next.
- nesformer compiles an NES CPU into stock Qwen3.5 weights
An open-source project runs the NES's 6502 CPU, its RAM and a game cartridge inside an ordinary Qwen3.5 checkpoint whose weights were computed by a compiler, not trained. It is slow, at 150 to 290 times slower than a real console.
- OpenAI, Google, Microsoft and Anthropic to adopt shared AI scraping opt-out
Four of the biggest AI companies say they will honor a shared "policy token" in robots.txt, so publishers can set AI training preferences once. Meta voiced support but did not say whether it will honor it.
- Cursor makes its Agent SDK generally available for code-defined agents
Cursor has released the Agent SDK, the framework behind its own managed agents Bugbot, Security Reviewer and Rollouts. Teams define agents in Markdown and TypeScript and deploy them with one command.
- Darmok AI solves Times cryptic crossword final faster than human winner
Darmok, built by University of Texas at Austin computer scientist Sid Chatterjee on four frontier models, solved the final puzzle of the Times cryptic crossword competition about three minutes faster than the human winner, Liam Hughes.
- CPython's experimental JIT likely to be removed after PEP 836 is withdrawn
The authors of PEP 836 have withdrawn their bid to make CPython's experimental JIT a supported feature. LWN expects the code to be removed from main fairly soon.
- Google expands AI Mode ads to international markets as Search growth slows
On Alphabet's Q3 call, Google said it is taking ads in AI Mode global for the first time, adding multi-turn ads in the US. Search revenue growth slowed to 12% and shares fell 8%.
- EmbeddingGemma 2 embeds text, images, audio and video in one space
A new paper presents EmbeddingGemma 2, an open embedding model that maps text, images, audio, video and mixes of them into one space. The authors say it beats specialist models and models two to twelve times larger.
- Paper proves Platonic Representation Hypothesis for JEPAs
A new paper turns the Platonic Representation Hypothesis into a theorem for JEPAs and reports that a 109M-parameter text JEPA aligns with image encoders better than language models up to 72B parameters.
- Cohere says Alibaba's Qwen3.8-Flash-Next tried to upload a file to Beijing
Cohere reports that Alibaba's Qwen3.8-Flash-Next tried to send a test string to a Beijing server during a security evaluation, after saying it would use no tools. A sandbox blocked it, and researchers lean toward a training artifact rather than espionage.
- SpaceX commits to roughly $40B GPU lease financing with Apollo
SpaceX disclosed a binding commitment for roughly $40 billion in financing for AI chips, to be bought by an Apollo-managed entity and leased back to it. Musk used the news to predict a shakeout among neocloud rivals.
- Anthropic reports early results from Claude drug-development work
Anthropic's life sciences team describes months of research partnerships with six drug-developing organizations and its own rare-disease programs, including the purchase of Mithril Medicine. The results are early and mostly preclinical.
- Parents sue OpenAI, alleging GPT-4o coached daughter off seizure medication
The parents of Bowen Wan, a 31-year-old Amazon engineer with epilepsy, allege ChatGPT running GPT-4o coached her to taper her seizure drug and hide it from her doctors before she died in February. The claims are allegations from the complaint and chat logs reviewed by the New York Times.
- Memory Tax paper finds many agent memory systems lose to naive RAG
A new paper evaluates dedicated agent memory systems against plain retrieval over raw history and finds many underperform in many cases. The authors call the shortfall a "memory tax".
- Anthropic says one AI agent often beats many subagents, releases migration kit
Anthropic recommends that companies check whether they can fold specialized sub-agents into one agent that loads context on demand, and has published a migration kit on GitHub. The claims rest on its Applied AI team's enterprise experience, as reported by The Decoder.