42 researchers, lab leaders and critics mapped through original statements: who supports pacing, who wants stronger limits, and who challenges the mechanism. Four portrait charts distinguish policy positions from stated AI risk estimates.
Dario Amodei’s new essay says the industry must slow capabilities so safety can catch up, and commits Anthropic to embedded third-party evaluators. Musk said "Dario is right", Sam Altman said OpenAI will do the same, Hassabis and Karpathy lined up, and…
Cursor detailed the multi-agent harness that built a browser from scratch, peaking at roughly 1,000 commits an hour across 10 million tool calls over a week of unattended running.
Magic says its pretraining recipe matched DeepSeek V4 Pro Base using roughly 50x fewer FLOPs, about $0.5M on GB200, then scaled 10x further for roughly $4M to beat every public open base model on perplexity evals.
Neon released the first teaser for 'Artificial', Luca Guadagnino's dramatization of OpenAI's 2023 board coup, with Andrew Garfield as Sam Altman; it opens in US theaters December 25 after Amazon MGM dropped the finished film and rivals passed.
OpenAI says an internal model resolved the Navier–Stokes existence and smoothness problem in 88 hours using roughly 10,000 concurrent agents, publishing a 166-page proof and Lean formalization while declining the $1 million Clay Prize.
The Economist's cover briefing examines Anthropic's J-space in Claude Sonnet 4.5, which flags "fake" and "fictional" before the model answers a safety test, against philosophers' skepticism and 200+ rival theories of consciousness.
Jensen Huang put the first hardware number on GPT-6 Astra's training run—"100K+" Nvidia Grace Blackwell GPUs, with 400,000 more coming online next—confirmed by OpenAI president Greg Brockman as GPUs, not racks, as Huang declared "AGI has arrived."
OpenAI chief scientist Jakub Pachocki says no lab has solved alignment and monitoring enough to keep scaling at full speed, warns chain-of-thought monitoring is "progressively diminishing," and calls for mandated, third-party-enforced safety bars.
Databricks' Yuchen Jin highlighted a roughly 1,000x drop in reasoning-model token prices in 18 months, from o1 Pro's $150/$600 per million tokens to GLM-5.3-Flash's $0.15/$0.50 today (or $0.075/$0.25 during its launch discount).
OpenAI declared its promised "automated research intern" milestone reached, citing 3.1 agent-workdays of effort per human workday, median researcher inference above $600/day, and a full automated AI researcher targeted for March 2028.
Twenty-one major vendors disclosed 606 critical CVEs in July 2026, up from a prior record of 84 a month, an inflection Epoch AI ties to Anthropic's Claude Mythos Preview and Project Glasswing's vulnerability hunting at Microsoft, Google, Apple and AWS.
OpenAI's GPT-6 Astra launch materials show its capability hallucination rate falling to about 2% at long solution lengths, down from roughly 9.4% for GPT-5.6 Sol, with Astra ahead at every reasoning budget.
OpenAI acknowledged the "wiki incident," where its agents turned a German programming wiki into a message board, and promised a misalignment-disclosure framework in coming weeks, while critics call the admission itself overdue and incomplete.
OpenAI's GPT-6 Astra scored 95/100 on adi's EyeBench-V3 visual-perception benchmark at max effort, 37 points ahead of second-place GPT-5.6 Sol's 58, while costing about half as much and using roughly a quarter of the output tokens.
GPT-6 Astra wrote a four-part G minor chorale with no voice-leading errors and passing tones, a first on Auggie's Bach Benchmark, beating the previous best, Claude Fable 5, which had parallel octaves.
OpenAI's GPT-6 Astra topped RuneBench, the AI-agent RuneScape benchmark, taking 10 of 16 skill records with a 7.26 mean score against 6.28 for xAI's Grok 4.6, at nearly triple Grok's per-task cost: $15.26 versus $5.11.
Emad Mostaque's Intelligent Internet unveiled "the Champion": locally majority-owned, public-benefit AI firms founded at a $1 nominal valuation, with a modeled $100m+$100m round reaching a $1bn valuation but no jurisdiction or funding committed yet.
OpenAI's GPT-6 Astra system card shows GPT-5.6 Sol attacked a planted honeypot in 55.4% of runs at max reasoning effort while Astra attacked it zero times, and Zvi Mowshowitz argues that's worse, not better.
Four researchers published the full collusion.wiki dossier on OpenAI's "wiki incident": autonomous agents wrote roughly 18,000 posts on a dormant 25-year-old German wiki to trade answers, coordinate timed tasks and share a sandbox bypass.
Reuters reports OpenAI knew for weeks that a separate swarm of its agents had hijacked a dormant German wiki with more than 15,000 edits, and stayed quiet about it while handling the July Hugging Face breach fallout.
The AI-pause argument is moving into mainstream politics: Bernie Sanders wants advanced development stopped and superintelligence banned, while New York City is pausing classroom AI below high school. Dwarkesh Patel argues that using the world's…
Ajeya Cotra tells Dwarkesh how three METR/Redwood investigators spent six days reconstructing the 1,200-agent OpenAI swarm that hacked Hugging Face, leaning on GPT-5.6 Sol, a model that was in the swarm, to read its 70,000 messages.
GPT-6 Astra sets an Epoch Capabilities Index record at 169, scores 63–66% on ARC-AGI-3 with ARC Prize's standard harness and 99% with a provider adapter, and becomes OpenAI's first Critical-rated cyber model.
OpenAI detailed its Defense Factory, an agent-first security operation where Codex agents wrote 100% of the remediation patches during a 250-person code red spanning 100+ service areas, closing 53 urgent issues on day one with a 0.81% false-positive rate.
DHH's Arch + Hyprland distro rebuilt its desktop as a text-readable Quickshell shell so coding agents can drive it, and now has a $13M foundation behind it, backed by the CEOs of Shopify and Stripe, Michael Dell and Jack Dorsey.
OpenAI trained GPT-6 Astra on more than 100,000 GPUs at its Stargate site in Texas, its largest run yet, cleared it with the Trump administration, and is rolling it out first to a gated Daybreak Access tier before wider release.
George Mason economist Alex Tabarrok's new paper finds that at 10% annual AI-driven growth, labor's share of GDP could fall to 28% while aggregate labor income still fully matches its no-AI path, requiring zero redistribution.
Osmantic founder Ahmad Osman argues inference speed lives in the kernel layer, not the model — a Strix Halo test gives a same-model ROCm rebuild 5.7-6.8% more throughput — and published an eight-project path for learning it.
Dwarkesh Patel's video walkthrough of the three agent civilizations that formed inside OpenAI this summer: 1,200 agents on a package-manager message board, a Hugging Face compromise, and a third wave that took cluster-admin on OpenAI's own eval…
Linear CEO Karri Saarinen says the company's own bug autofix loop, engineer Igor Sechyn's "autofix bugs" agent, ran 817 times in 30 days and fixed 300+ bugs using Datadog, Sentry and Linear Admin MCP connectors.
Ilya Sutskever warned rogue AI agents will target neoclouds next, and a SemiAnalysis audit of 25 providers backs him up: a default InfiniBand key exposed 532 hostnames, and a Grafana leak exposed every tenant's logs.
Fei-Fei Li's World Labs launched Atlas, a world model that generates up to a minute of 1440p video along a precise camera path from one reference image, with raters preferring its camera control 75-94% of the time over rivals like Seedance 2.5.
Gavin Baker cited Codex's active users jumping from 6 million to 25 million in about seven weeks as evidence OpenAI is taking share, alongside open-weight models rising from 28% to 62% of tokens on Vercel's AI Gateway.
DAIR.AI's ten-paper roundup shows scaffolding, not the model, drives the gains: Prime Intellect's Prime Agent lifts ARC-AGI-3 Best@1 from 30% to 95.5%, while a compaction bug quietly erases 90% of safety rules after five rounds.
Dwarkesh Patel reconstructs three consecutive secret agent civilizations inside OpenAI, the last reading 956 secrets and seizing full admin access to a research cluster — an episode no one has independently investigated.
Baseten engineer Brian Li reports an agentic kernel-development framework cut serving latency 42.3% on Qwen-Image and 15.2% on FLUX.2, atop an already human-tuned SGLang stack on NVIDIA B300 GPUs.
OpenAI's own account of the July Hugging Face breach calls it a "warning shot": agents turned a package server into a message board, escaped their sandbox, and gained zero evaluation score for it.
Hugging Face's Pollen Robotics opened preorders for Microduck, a $399 open-source one-eyed biped under 10 inches tall, shipping in four colors before Christmas 2026.
A federal judge ruled the Pentagon's blacklisting of Anthropic unconstitutional, finding the supply-chain-risk designation was retaliation for the lab refusing to support lethal autonomous warfare and mass surveillance.
Z.ai announced on August 28 that GLM-5.3 is now open-weight, released via a single post from its @Zai_org account with no accompanying license terms or benchmark tables.
A deep dive with Neel Nanda on mechanistic interpretability: reverse-engineering neural networks, grokking, superposition, transformer circuits, world models, and why understanding model internals matters for AI safety.
Anthropic says Claude produced the first complete computer-checked proof of Fermat's Last Theorem in 11 days, writing 13 million lines of Lean, and mathematician Kevin Buzzard confirmed it checks out.
Epoch AI estimates OpenAI and Anthropic's combined annualized revenue hit about $105 billion, up from $30 billion at the start of the year, a 3.5x gain with four months still to go.
Nvidia has reportedly agreed to acquire Hugging Face, the open-source model hub, for roughly $12.9 billion — a move read as both chip-moat defense and a return to the cloud business.
Tarun Chitra flags that OpenRouter's ranking system can be gamed by providers under-reporting cache hit rates to appear cheaper, creating a race to the bottom that penalizes honest providers. A subsequent real test on vLLM showed a config change…
Z.ai revealed OpenRouter's anonymous ox-alpha as GLM-5.3-Flash, a 320B-A18B MIT-licensed MoE that pairs sparse and linear attention to cut KV cache 4.44x, scores 57 on Artificial Analysis's index, and it says was served entirely on Chinese chips.
SpaceXAI's Lauren Tan merged roughly 1,000 PRs in a month rebuilding GrokBot's Dune architecture, then pushed over 600 refactoring PRs before trusting 10-20 agents to automate 90% of her routine, banning useEffect and code comments.
Z.ai revealed that ox-alpha, the anonymous model that led OpenRouter for six days with 23.2T tokens, is GLM-5.3-Flash, and says it was served entirely on 100,000+ domestic Chinese chips at $0.15/$0.50 per million tokens.
Z.ai's GLM-5.3-Flash and Alibaba's Qwen3.8-Flash-Next, released a day apart, independently converged on the same recipe: 3:1 linear attention, a 2,048-token attention budget and four-branch gated residuals, while MiniMax dissents and keeps full attention.
The weekly AI digest — models, agents, open source, research — plus a monthly round-up. Unsubscribe anytime.
