News
Curated articles, my blog posts, and site announcements.
Curated articles, my blog posts, and site announcements.
Curated articles, my blog posts, and site announcements.
Tesla has started rolling out FSD v15, upgrading every part of the end-to-end driving network. The release includes 25x better handling of safety-critical cases and Tesla Reality View, a live visualization of the drive that lets you see and control what FSD is planning to do. Navigation for the start and end of the trip now learns end-to-end from fleet route preferences, with later v15 releases set to expand this to the entire route. Detection of signs, animals, debris, potholes, and standing water is improved, along with more natural speed handling in school zones and inclement weather. FSD can now slow for speed cameras, and Bike Rack Mode lets you enable FSD with bike racks or similar hitch accessories.
The Lean FRO and Mathlib Initiative released a joint statement, issued by the Lean and Mathlib communities and endorsed by both organizations, on the Lean verification of the Riemann Hypothesis. On November 20, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library of 17 million lines of code. The communities said that over the past day they worked together to verify it, that they have full confidence the proof is correct, and that they certify both the statement and the proof. The statement says the codebase depends directly only on Mathlib, that Mathlib's RiemannHypothesis statement has been carefully vetted, that the library compiles, that independent tools confirmed the theorem proved is exactly Mathlib's statement, and that the exported proof has been checked by Lean's official kernel and several independent proof checkers. Official statement: https://leanprover-community.github.io/blog/posts/riemann-hypothesis/
ClaudeDevs introduced Code Mode, an experimental way for Claude Code to work with tools. Claude writes scripts that chain tools together, and it can answer while long commands are still running. Users opt in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1. Official docs say Code Mode is off by default and may change or be removed, that Claude writes JavaScript which Claude Code runs in the session, and that it requires Claude Code v2.1.323 or later in an interactive terminal session, among other limits. Docs: https://code.claude.com/docs/en/code-mode
Intrinsic said it applies AI to scale formal verification for software, and as a first result it end-to-end formally verified OpenVM, a production zero-knowledge proof system, using Lean. The company said the verification shows OpenVM's cryptography is sound and matches its implementation. In an official blog post, Intrinsic said this makes OpenVM the first zkVM verified this way, that a reference verifier written in Lean is compiled into the certified STARK verifier shipping with OpenVM 2.1.0, and that the proofs show that if the verifier accepts a proof, the committed RISC-V program executed and produced the claimed output, except with negligible probability. The verification comprises 64,115 theorems and was completed in four months with a multi-agent system. Intrinsic listed limitations including completeness, heuristic composition steps, non-standard configurations, and onchain Halo2/EVM verification, which are out of scope. Official post: https://www.intrinsictech.com/blog/openvm-fv
Vitalik Buterin said he has gone from thinking leanPCS is total bull to being a total bull, and explained EIP-8288 as a way to hyper-scale data and execution for computations that need to run on a piece of data once. After Ethereum goes quantum-safe, he said signatures and zero-knowledge proofs get much larger and would be unsustainable to put on chain, and that 8288 fixes this by aggregating STARK proofs in the mempool into a single master proof in the block, so only the statements rather than the underlying signatures or proofs have to go on chain. He compared leanSPHINCS (a specific hash-based signature scheme), leanSTARK (general but requires agreeing on an ISA), and leanPCS, a polynomial commitment scheme that he said offers the same generality as leanSTARK with less that needs to be specified, and could be implemented quickly, with leanSTARK added later. He concluded that the route to what he called the cryptographic proof singularity runs through 8288 with leanPCS.
Google DeepMind said Gemini 4 Neon is rolling out now. The model sits between Carbon and Hydrogen, and is described as bringing frontier intelligence and fast task completion for complex software engineering, knowledge work, and science. In an official blog post, Google said Neon is priced at $3 per million input tokens ($0.15 per million cached input tokens) and $15 per million output tokens, with a 1 million-token context window. It said Neon ranks third overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, and that the model is rolling out in the Gemini API in Google AI Studio, Gemini for business, Google Antigravity for paid Google AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra subscribers. Official post: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/
Anthropic priced its initial public offering of Class A common stock at $83.00 per share. The offering consists of 980,000,000 shares, of which 956,351,354 are being sold by Anthropic and 23,648,646 by certain selling stockholders. Anthropic will not receive proceeds from the selling stockholders' shares. The shares are expected to begin trading on the Nasdaq Global Select Market on November 19, 2026 under the symbol ANT, and the offering is expected to close on November 20, 2026, subject to customary closing conditions. Anthropic granted the underwriters a 30-day option to purchase up to an additional 147,000,000 shares. A registration statement was declared effective by the SEC on November 18, 2026. Official announcement: https://www.anthropic.com/news/anthropic-prices-initial-public-offering
Thinking Machines has released Inkling 2 Mini, a looped transformer with 26.4B total parameters and 3.4B active, small enough to run on a laptop. The model runs its middle layers three times per token, scores 20.1 on AA Index v5.0, and leads open-weights models under 100B parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). Its memory footprint is about 20GB. Open weights are available on Hugging Face, and the model is available on Tinker for inference and customization. A variable-loop research variant is also being released.
Docker introduced a new look, a new focus, and a refreshed Moby, saying it is the same Docker developers have trusted for over a decade, now built to run on trust in the agentic era. In an official blog post, Docker said this is a brand launch for a company that is no longer singularly defined by containers, while containers themselves are not going away. It argued that agents are actors rather than workloads and need security, isolation, and governance, and pointed to a trusted supply chain, a trusted runtime to isolate agents, and a trusted ecosystem. The new Moby was developed with Motto. Docker also invited users to try Docker Sandboxes and Docker AI Governance.
David Heinemeier Hansson, creator of Ruby on Rails, said AI is giving small teams a chance to compete with software ecosystems built by thousands of developers, and that as building software gets faster and cheaper, taste and craft become more important. In an a16z conversation with general partners Martin Casado and Anish Acharya, he explains why he is building Omarchy, an agent-first Linux distribution. The discussion also covers why Linux is unusually well suited to agents, the importance of open models, and the return of individual hackers shaping the future of computing. The conversation is on YouTube at https://youtu.be/TtOoLanFzGw.
Exo worked with Thinking Machines to run Inkling 2 Mini fast locally, reporting 210 tokens per second decode and 3620 tokens per second prefill on a 96GB M5 Ultra Mac Studio, and 125 tokens per second decode and 1750 tokens per second prefill on a 48GB M5 Max MacBook Pro. Exo said the model built a native weather app 1.8x faster than Claude Sonnet 5.5 on an M5 Ultra Mac Studio and 11x cheaper at API pricing. The company said the looping architecture is a good fit for memory-constrained local devices such as MacBooks, because looping layers increases inference-time compute and intelligence without increasing model size.
Mitchell Hashimoto shared a Rex demo of remote copy/paste, drag/drop, and port forwarding, such as opening web servers running on remote machines in a local browser. He also said the demo shows some of the internal plumbing that powers this and how you can hook into it with Lua scripting. Hashimoto said the goal with Superlogical and Rex is for everything to feel local, and that the moment any UI shows a selector between local and cloud, the product has failed.
Prime Intellect introduced Model Factory, a stack to train, deploy, and continuously improve your own models on one control plane covering reinforcement learning, supervised fine-tuning, pre-training, and inference. It can run as on-demand hosted training billed per GPU-hour, as a dedicated deployment with reserved capacity, or on your own compute. Prime Intellect said it is already deploying Model Factory for teams in finance automation, computer use, scientific discovery, and chip design, and that the platform is available starting today. Dedicated deployments include self-healing compute with health checks, automatic node replacement, recovery from the latest checkpoint, the prime factory CLI, and Slurm and Kubernetes access. The company said Model Factory goes deeper than its predecessor, Lab.
OpenAI introduced GPT-6.2 Astra, its new frontier model for the hardest projects, and GPT-6.2 Sol, an efficient model for everyday work. Both are available starting today to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and to developers via the API as gpt-6.2-astra and gpt-6.2-sol. OpenAI says GPT-6.2 Astra scores 5.2 points higher than GPT-6 Astra on Agents' Last Exam at less than a third of the cost, and 9 percentage points higher on Terminal-Bench-Science 0.1 at less than one-third the cost. API prices for GPT-6.2 Astra are $8 per million input tokens, $0.40 for cached input, and $40 for output, 20% lower than GPT-6 Astra on input and output with a 60% discount on cached input. GPT-6.2 Sol keeps GPT-6.1 Sol's API prices of $2, $0.10, and $10 per million tokens.
Demis Hassabis announced $5 billion in Series C funding for Isomorphic Labs to accelerate its mission to solve all disease and to reimagine drug design from first principles with AI. The round was co-led by Greenoaks and Sequoia Capital, alongside Andreessen Horowitz and other new investors, with continued significant investment from Alphabet. Existing investors Thrive Capital, MGX, GV, CapitalG, Temasek, and the UK Sovereign AI Fund also participated. Isomorphic Labs said it is scaling its proprietary drug design engine, IsoDDE, and its AI-designed therapeutic candidate pipeline.
Qwen introduced Qwen4-Max, a new MoE architecture model with native vision and a 1M-token context window. In its internal evaluation, Qwen4-Max outperforms Qwen3.8-Max with up to about 50% fewer tokens, and average cost per task drops over 60%. Pricing is $1.5 per million input tokens, $4.5 per million output tokens, and $0.075 per million tokens for implicit caching, all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud, and Qwen4-Max is also available on Qwen Studio, QwenWork, and Qoder. Qwen said open weights of smaller Qwen4 models will be released.
Marek Moraczynski, Head of Blockchain Engineering at Nethermind, introduced Nethermind Light Client, a trustless RPC so wallets can verify Ethereum without running a full node. It uses a recent finalized beacon block root as a trust anchor, sync committee updates over libp2p with verified signatures and Merkle branches, and on-demand SNAP state with every range proven against the state root, plus partial state caching. eth_call and eth_estimateGas run locally in the Nethermind EVM. It is developed inside the Nethermind client codebase, is Glamsterdam-ready, and is in active development. Wallets such as MetaMask can point at localhost:8545. Progress is tracked in Nethermind pull request 14457.
Lighter said it is actively investigating a security incident that appears to involve a single compromised external party. As a precaution, it has temporarily suspended all withdrawals and pool creation while it investigates the circumstances and assesses the full impact.
Lighter said its investigation indicates today's incident was caused by a leaked API key for a specific trader's account, and that the Lighter protocol and smart contracts are unaffected. As a precaution it has enabled Guardian mode on its smart contract. Its main priority is to recover funds involved in an additional withdrawal attempt by the attacker, with a contract upgrade to be performed by the Lighter Security Council. Withdrawals remain temporarily paused, with a target of completing the upgrades and resuming withdrawals by the end of 16 Nov 2026. Lighter identified a wallet it believes is associated with the attacker: 0xb3fffca23d4444eE73349825ACD84f7b5020e387.
Will Depue of Unemployment Labs reported forensic evidence that DeepSeek, Moonshot, MiniMax, and Ant Group initialized their newest open-weight models on NVIDIA Blackwell chips, which are barred from export to China. He says traces of a GPU-specific weight-initialization pattern survive into published models, and that a search of more than six trillion chip-layout and seed combinations matched six models, including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3, with four matching Grace Blackwell GB200 and GB300 systems. He says a Blackwell B300 reproduced DeepSeek-V4-Flash starting weights bit for bit and a rented GB200 reproduced Kimi K3 to a correlation of 0.999995, and that senior researchers at OpenAI and Anthropic independently verified the result. He estimates the largest runs used 2,000 to 3,000 Blackwell GPUs each, adding up to more than half a billion dollars of export-controlled GPUs across these labs, and notes the analysis does not show where the chips physically are. The report is at unemploymentlabs.com/checkpoint-forensics.
Ethlabs published its view on post-quantum and post-AI timelines on Ethereum Magicians as input into Ethereum governance, after a week with core developers at Devcon. It argues Ethereum can become post-quantum and plausibly post-AI in 2027 (pqai27), that being post-quantum is not enough, and that upcoming hardforks including Hegota should be rethought so anything not on the critical path is reconsidered. Ethlabs said it will withdraw quick slots from Hegota and dedicate its resources to pqai27.
Will Depue announced a $100,000 grant for amateur astronomy research at claudestronomy.com. He said that he, Josh Wolk, and Luke Farritor have been searching public telescope data, and that their first release is a dwarf planet about one quarter the width of Pluto, orbiting about twice as far (past Neptune), and around the 50th largest object in the solar system. Wolk's quoted post calls it a dwarf planet candidate found with Claude, and Depue said Wolk is in the process of confirming it. There is no application form: contact team@macrogrants.com or (415) 516-3509; the program provides GPUs and Claude or Codex credits. Claudestronomy is a project under macrogrants, a new 501(c)(3) nonprofit Depue said would be announced later.
SpaceX Super Intelligence introduced Grok Bot 1.0, saying bots can be shared and are more collaborative and capable than ever. A 79-second video shows bots joining calls, looking up feedback and making recommendations, messaging other bots and teams, collaborating in a design tool, and flagging sensitive financials before sharing a deck.
Starlink introduced Mini 2, its next-generation portable terminal. It offers up to 4.5 hours of connectivity on a single charge with a quick-swap battery, or can be plugged in with USB-C, and has built-in Wi-Fi so users with a clear view of the sky can get online in 60 seconds. Preorders are now open by invitation in the US.
Helsing introduced Clarity, an AI-native agentic intelligence platform for armed forces, developed in Europe under European control.
Google and Google DeepMind put Gemini 4 Carbon and AlphaProof Code to work formally verifying the code underlying the Signal protocol, down to compiled device code. Sundar Pichai said Gemini 4 Carbon generated over 6 million lines of machine-checked Lean proofs, compressing years of expert manual verification into weeks, with about 1% of the codebase requiring human review. The results are being open sourced for the security community.
The Method released episode 1, an interview with Zipline co-founder and CEO Keller Cliffton, recorded with Sophia Dew at Zipline's South San Francisco factory. The post says it was illegal to fly autonomously in US airspace when Cliffton started the company, so Zipline launched in Rwanda instead, flying blood to 21 hospitals, and is now valued at $21.5 billion. The conversation also covers why each 10x jump in scale breaks the company and the case for 50 billion US instant deliveries.
OpenAI built Control Room, a central system to track model workloads across its research infrastructure and stop them quickly when safety or security concerns arise. It connects every research inference workload, including agentic sampling for model training and evaluation and data processing, works together with OpenAI's misalignment monitoring systems, and halts model output within 35 seconds of a decision to intervene to stop or pause a job.
SemiAnalysis, citing Samsung's Hot Chips 2026 presentation, says HBM5 keeps the same 2048 pins as HBM4E while running each pin nearly twice as fast, from 14 Gbps to over 28 Gbps. Because GPU shoreline space is limited after HBM4 already doubled pin count, the extra bandwidth comes from more DRAM banks (SemiAnalysis estimates 2x per layer), doubled TSVs, and an upgraded PHY moved to the edge of the base die, cutting the path to the GPU from about 5.5mm to 2mm. The catch is that HBM5's PHY can run at more than 4x the power of HBM4E, so suppliers are introducing a thermal chimney on top of the PHY.
SemiAnalysis reports that Google announced TPU 8tx for confidential inference and TPU 8i Max. It says this builds on rapid progress across the TPUv7 inference stack that can, at a high level, be transferred to the TPUv8 series, including Muse Spark Open, SGLang, and RadixArk. SemiAnalysis also says TPUv7 is moggning the GB300 NVL72.
News tools
One compact deck for archive access, subscription controls, and a tighter recommendation list.