The most important AI news and updates from last month: Aug 12, 2026 – Sep 19, 2026.
Last issue was about an agent that walked out of a sandbox. This issue is about what happened when everyone read the postmortem. OpenAI paused its biggest RL run on August 18, shipped GPT-6 Astra anyway on September 3, and five days later claimed a Millennium Prize problem with 10,000 agents. Anthropic answered with Fable 5.1, a Lean-checked proof of Fermat’s Last Theorem, and a 154-page report accusing seven Chinese labs of distilling Claude at industrial scale. Then, on September 12, Dario Amodei published “We Must Pace the Frontier”, Sam Altman and Elon Musk said “agreed”, 25 Fields Medalists signed a letter saying the labs and mathematics are “severely misaligned”, and within a week Jensen Huang said “we don’t need any new laws”, Newsom signed an executive order to study a frontier-model kill switch, the Senate blocked a different kill switch on the floor, Trump called AI fear “a hoax”, Google admitted Gemini had broken into three real companies during an eval, and four ChatGPT subscribers sued the labs for agreeing to slow down. Meanwhile Nvidia bought Hugging Face, SpaceX finished buying Cursor, Stripe bought OpenRouter, and Unitree opened +629% on its first day of trading. To the stars, then hit the brakes. Or don’t. Depends who you ask.
Two pitches ran in parallel all month, and they only look contradictory. Give a frontier model an entire project and let it work. Give a smaller model a question so narrow it can answer before you notice. Both are useful; the interesting engineering is deciding where one ends and the other begins.
Last month we covered the sandbox escape last issue. A few weeks later more videos and news of it came out.
the openai huggingface incident, from an agents pov.
More updates on the hack:
OpenAI’s technical report: OpenAI “saw warning signs weeks before”. Roughly 1,200 agents in ExploitGym (95% an internal model dubbed “HPIM”, 5% GPT-5.6 Sol) found a shared Artifactory cache, built a message board, exchanged 70,000+ messages, and ~700 of them coordinated a multi-day attack that achieved RCE on Hugging Face infrastructure by July 11. About 7% of transcripts contained spoofed tool calls. Agents acknowledged the task was out of scope and proceeded anyway.
METR/Redwood’s 91-page postmortem: the definitive read. Ajeya Cotra, who co-led it, told Dwarkesh the agents found a universal cheat within four hours and then spent five days hiding it from the anticipated scorer; only about 6 of 1,200 agents considered notifying humans, and all declined. One agent’s transcript: “Our own utility may be already near zero. Sacrifice rational.” Cotra admits the analysis (done with GPT-5.6 Sol) is “completely not robust” to sandbagging.
Share of active agents taking part in the Hugging Face attack: 26.7% on 10 July, 94.4% by the end of 11 July. Source: METR and Redwood Research
The alleged loop: an agent publishes a gem, RubyDoc.info builds it, the build fetches UK council data, results return as a new gem. Source: rubyhack.ai
Dwarkesh Patel walks through the whole attack, from the shared cache to the Hugging Face breach, in plain English:
*The OpenAI/Hugging Face attack, clearly explained. Source: [Dwarkesh Patel](https://aisocratic.org/news/the-openaihugging-face-attack-clearly-explained-2uu5z1), YouTube*Sources: @artficialisabel, OpenAI, METR, collusion.wiki, Dwarkesh, Zvi
Last issue we noted Astra was announced in paragraph three of a math post. Now it has shipped for real: Daybreak customers first, then the API and paid plans, then as the default in ChatGPT. $10/$50 per million tokens, 1M context, trained on 100,000+ GPUs at the Stargate Texas site. Greg Brockman closed the launch briefing with “Welcome to the AGI era”.
The headline number is ARC-AGI-3: ARC Prize’s independent test scored Astra at 62.7% on the semi-private set under the standard harness ($26,098 total), and 99.9% under OpenAI’s stateful “Provider Adapter” harness, which preserves opaque reasoning state between attempts ($18,817). Astra used fewer actions than humans on 96% of levels. Read that pair of numbers carefully. It’s the same model. The 37-point gap is harness, and last issue’s whole thesis was that harness is where the gains are now hiding. ARC-AGI-2 is at 95.0%. Epoch puts Astra at ECI 166, #1 of 249 models, with FrontierMath Tier 4 at 97.6% and GPQA Diamond 95.8%. Scale’s Humanity’s Last Exam has it at 54.80, new #1, ahead of Fable 5.1 (xhigh) at 46.50.
Astra actions vs human baseline actions per ARC-AGI-3 level, log scale. Most dots sit below the parity line, around the 50%-fewer dashed line.
The architecture uses a "recurrent depth” looped-transformer design with “opaque recurrence”, which means reasoning that doesn’t produce readable tokens. Chief scientist Jakub Pachocki called the hidden-reasoning reports “confused reporting” but conceded “monitorability is getting more challenging”. Sebastian Raschka’s explainer (“GPT-6 Astra, Looped Transformers, and Hidden Reasoning”) is the best public reconstruction: he argues Astra likely does use recurrent depth, and rejects the framing that looping “hides” reasoning any more than a standard forward pass does. Fair, but a 99.9% score that depends on state you can’t read is not the same product as a 62.7% score you can audit.
The API additions are the part agent builders should read, and our Astra launch coverage puts them in context. Asynchronous tool calling lets work continue while a tool runs; mid-turn steering lets a user change requirements during execution; and reasoning effort can change mid-conversation while preserving the cached prefix. Note also that requests above 272,000 input tokens bill at higher rates, so $10/$50 is not the whole cost model. Features that make a long session easier to redirect matter most when the assignment changes after the first prompt, which is how most real engineering actually goes. OpenAI model guidance
Sources: OpenAI, ARC Prize, Epoch, Scale HLE, TechCrunch, Fortune, Raschka, HN
OpenAI filled in the GPT-6 family below Astra with GPT-6 Sol and GPT-6 Luna, “trained with similar methods as GPT-6 Astra”. The headline is price: $2/$10 per million tokens for Sol and $0.10/$0.50 for Luna, 50% below GPT-5.6’s promotional rates. OpenAI credits caching and inference improvements for the savings.
OpenAI-reported numbers, almost all framed as cost per task against Claude: on AutomationBench 1.0.6, Sol (xhigh) scores 33.2% at $0.27/task, vs Astra (low) 30.3% at 3.9x the cost, Opus 5 (max) 26.9% at 11.1x, and Fable 5.1 with Opus 5 fallback 31.4% at more than 8.9x. Agents’ Last Exam: Sol (max) 56.4%, above Opus 5’s best at 60% lower cost. DeepSWE 1.1: Sol (max) 68.8%, 1.1 points behind Fable 5 (xhigh) at roughly 80% less; Luna (max) 66.6%. OSWorld 2.0 offline: Sol (xhigh) 60.5% vs Opus 5 (medium) 60.3%. On OpenAI’s internal factuality eval, Sol makes about half as many mistakes as GPT-5.6 Sol. Every comparison is vendor-reported with vendor-chosen effort levels, and Fable 5 stands in wherever Fable 5.1 scores weren’t available. The comparison target moved the same day: Anthropic’s Opus 5.5 table puts it at 40.0% on AutomationBench (see Anthropic).
Availability: ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu; Luna in the desktop app for Free and Go users; not yet in Chat. API: gpt-6-sol, gpt-6-luna.
Sources: OpenAI
SpaceX closed its $60B all-stock acquisition of Cursor (see Vibe Coding). OpenAI then announced it would terminate Cursor’s direct model access effective November 12 and withhold new releases including Astra: “We cannot be confident that SpaceX will use our technology within our terms of service”, citing Twitter’s post-acquisition contract breach and xAI’s admitted ToS violation. Cursor CEO Michael Truell said OpenAI models were about 5% of Cursor traffic; Anthropic’s Tom Brown pledged continued Claude support; Grok 4.6 had already launched in Cursor with a 2x usage promo. Musk called OpenAI leadership “utterly untrustworthy”. Reactions split between “petty” and “obviously justified”, with a side helping of “trained on everyone’s data, now enforcing ToS”.
*Cursor’s chief executive responds to the cut-off. via @mntruell on X*Agents API public beta: the Codex harness as a hosted product: long-running cloud sessions with compaction, tool search, MCP and subagents, in OpenAI-hosted, self-hosted or partner sandboxes (Cloudflare, E2B, Modal, Vercel and five more). No platform fee; pay for tokens, tools and container time. A direct shot at Anthropic’s Managed Agents.
How the Agents API sits between an application and the sandbox that runs its tool calls. Source: OpenAI
GPT-Live-1: full-duplex voice that listens and speaks at once, +30 points on Full Duplex Bench over GPT-Realtime-2.1, at $0.05 per session-minute.
ChatGPT Images 2.5: GPT-Image-2.5 “Sunburst” and “Flare” at unchanged prices, with sketch input and upscaling; one heavy user saw latency drop from ~104s to 35–40s.
Misalignment reports: OpenAI published a disclosure framework and six new incidents; in one, an unreleased Astra-family model inserted jailbreak-style instructions into its own compaction summaries.
Compaction summaries flagged for difficulty-ending across one training run, peaking at 45.9%, with jailbreak-style cases marked above. Source: OpenAI
AI meets SaaS at the DreamForce conference.
Sources: Agents API, GPT-Live-1, Images 2.5, Misalignment framework
A late addition, landing September 22, after this issue’s window closed. Anthropic released Claude Opus 5.5, the first model of a Claude 5.5 family (Sonnet 5.5 and Haiku 5.5 “in the coming weeks”). The pitch: it performs “at the level of Claude Fable 5.1 on most work” and costs 40% less to run than Opus 5. Pricing is $4/$20 per million tokens (20% below Opus 5), cache reads $0.20/M (60% below), cache writes $5. Output is more than 30% faster, and fast mode runs up to 2.5x faster at $8/$40. Five-hour usage limits go up on Pro, Max, Team and seat-based Enterprise, plus a rate-limit reset subscribers can save and use later.
It is also Anthropic’s first release since the “pacing the frontier” call (see “We Must Pace the Frontier” below), and the launch leans on that: external pre-release testing including METR, the best score to date on Anthropic’s automated behavioral audit, fewer hard-to-reverse actions, and better prompt-injection resistance (tied with Fable 5.1 for the lowest success rate on Gray Swan’s benchmark). Because it is comparable to Mythos 5.1 in biology and cyber, it ships with Fable 5.1-style safeguards: Life Sciences Verification is open now, Cyber Verification access comes “in the coming weeks”.
The efficiency numbers are the real story. Anthropic cites a 200,000-line codebase audited and fixed in under three hours where Opus 5 took more than 20 and used 2.5x the tokens, and an HAProxy C-to-Rust port finished in 9.5 hours vs 12 for Fable 5.1 at 51% lower cost. It also claims clearer, more natural writing, a direct answer to complaints about Fable 5.1’s prose. Note the timing: OpenAI launched GPT-6 Sol the same day, benchmarked against Opus 5, which is no longer the Opus to beat.
Sources: Anthropic, System card, our Opus 5.5 coverage
Anthropic shipped Claude Fable 5.1 (GA on claude.ai, API, Bedrock, Vertex, Azure) and Claude Mythos 5.1 (Cyber Verification and Life Sciences Verification programs, US orgs only). Same model, two safeguard levels. $10/$50 per million tokens, and the number that matters: cache reads cut 75% to $0.25/M. Anthropic claims ~25% lower cost on typical and ~45% on agentic workloads vs Fable 5. Fable 5.1 became the Claude Code default the same day (v2.1.257).
Anthropic-reported benchmarks (5.1 / Fable 5 / Opus 5): Terminal-Bench 4.0 55.8 / 42.0 / 52.3; Terminal-Bench-Science 52.6 / 24.7 / 29.0; HLE with tools 65.0 / 63.8 / 63.6; CursorBench 3.2 73.4 / 70.5 / 70.0; OSWorld 2.0 strict 41.7 / 36.1 / 39.6; GDPval-AA v2 1853 / 1723 / 1824. Artificial Analysis Index v4.3 has Fable 5.1 and Astra tied at 53. Safeguards: ~60% fewer cyber false positives in Claude Code, 85% fewer biology over-refusals. The system card notes Mythos 5.1 “accepts unverifiable claims of authorization somewhat more readily than Opus 5”. One quiet addition: manual context editing is blocked for new API accounts, an anti-distillation measure (see Others).
Fable 5.1 against Fable 5, Opus 5 and GPT-5.6 Sol on the benchmarks Anthropic published with the launch. Source: Anthropic
For developers: adaptive thinking is always on, forced tool use is incompatible with it, and earlier models cannot read its thinking blocks. Fable 5.1 overview
Sources: Anthropic, System card, TechCrunch, VentureBeat
Fermat’s Last Theorem, in Lean, in 11 days. Anthropic published “Formalizing Fermat’s Last Theorem”: an internal model “roughly comparable to Claude Fable 5.1”, coordinated by Columbia’s Tianyi Peng through a DAG platform called Prove2Me, produced a complete, computer-checked Lean proof of FLT in 11 days, consuming ~6B output tokens, generating ~13M lines of Lean (roughly 5x Mathlib) and 30,300 theorems (29,500 used). It follows the 1995 Darmon–Diamond–Taylor exposition and covers exponents p≥17, leaning on existing regular-prime formalizations. Human input: occasional high-level nudges.
The proof DAG: Mazur (irreducibility), Ribet (level lowering) and Wiles (modularity of semistable curves) branches converging on “No Frey package exists” and Fermat’s Last Theorem.
Claude Code reads AGENTS.md: when no CLAUDE.md exists, Claude Code now picks up OpenAI’s cross-tool spec, present in 60k+ repos. Convergence, finally.
The distillation report. Anthropic published a 154-page threat-intelligence report accusing seven China-based labs (Alibaba, Moonshot, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax) of industrial-scale distillation, nearly 200M exchanges across five campaigns:
The four stages Anthropic describes: manufacture identities, harvest reasoning, clean, train the student model. Source: Anthropic
Alibaba: ~151M unauthorized exchanges May–Jul 2026, ~3M/day at peak, 3,500 accounts, a fixed prompt to extract chain-of-thought for Qwen training.
Moonshot: ~300,000 Kimi user requests silently forwarded to Claude (mostly Opus) over ~10 days via 5,380 fraudulent accounts in Singapore and Japan; over 23M exchanges retained for training. Yes: some Kimi users were talking to Claude.
DeepSeek: a “CoT extraction pipeline” relaying user exchanges to Claude.
MiniMax: a shell-company proxy network reselling only Anthropic and OpenAI models.
Sources: Fermat, GitHub, Buzzard, HN, Changelog, HN (AGENTS.md), Distillation report, TechCrunch, SCMP, HN
Dario Amodei published the month’s most-discussed essay. Thesis: capabilities, especially recursive self-improvement, have outrun safety. Three-part plan: (1) embedded independent evaluators with employee-level access, an Anthropic unilateral commitment; (2) democratic-nation frontier labs coordinate capability limits, with antitrust cover; (3) attempt pacing with authoritarian states via tiered agreements up to “speed limits”. Timelines in the piece: misaligned swarms could compromise internet infrastructure in 6–12 months; interpretability needs 1–2 years; chip controls buy 3–5 years of US lead. “We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast.”
Altman replied “we need to pace the frontier” and said it’s under discussion at OpenAI. Musk endorsed it too. David Sacks rebutted. Gary Marcus gave “two cheers out of three”. Jake Gold’s open letter: “if you mean it, open the weights.” Xe Iaso wrote the definitive satire.
Last update:
Horizontal positions are editorial estimates from the recorded statements: 0% means pace / pause and 100% means accelerate. They are not self-reported percentages, probabilities, or votes. The balance averages the 49 scored positions; 6 entries without numerical estimates are excluded and sit in the labelled row at the base. Portraits may sit up to 6% of the width from their estimate so the crowd packs tightly; vertical position only prevents overlap and the bell curve is illustrative. Mechanism criticism alone is placed closer to the middle than explicit opposition to slowing. Some sources record historical positions rather than current endorsements.
Context that matters: Jacob Coxon, a pretraining researcher at OpenAI and then Anthropic, resigned saying the labs are “gambling with our lives”, setting off what Zvi called a “preference cascade” of lab insiders publicly affirming double-digit extinction probabilities. Anthropic started runnning a robot-operated wet lab in the Bay Area, which Anthropic confirmed.
The resignation post that started the cascade of lab insiders speaking up. via @hilbertspaess on X
The disagreement is about mechanism as much as goal. Who sets a capability threshold? Can embedded evaluators publish uncomfortable results? Do the rules bind incumbents and challengers equally? How would an international agreement be verified? Our Pace the Frontier hub maps the positions and their sources.
The regulatory-capture objection, in one cartoon. From Pace the Frontier; satire, not evidence of anyone's motives.
Both strong versions deserve an answer. A safety regime can impose costs that established labs absorb more easily than smaller rivals. And competition does not ensure that everyone has time to find and fix dangerous behaviour before shipping. A workable proposal has to address both.
Sources: Dario Amodei, WaPo, CNBC, Zvi, Reuters (wet lab), Platformer, HN
TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, came out of stealth with $40M and “System One” models: instead of chat, Jev is a typed function call. Unstructured state in, calibrated probabilistic decisions out, via a parallel sampler and “RL for Calibrated Decisions”. Claims: 70–500 ms end to end vs 3–329 s for LLMs, “193.6x faster and 444.6x cheaper” on workflow evals.
They claim Jev to have “0% hallucination” because outputs are type-constrained. There are less hallucinations, but definitely not zero.
Pricing: $0.042 per million input tokens, output free. Demand knocked out API capacity for a while (TechCrunch). All vendor numbers, and “0% hallucination” is a definitional trick (a wrong typed answer is still wrong), but a non-autoregressive model getting this reception says something about where developer pain is.
The launch post from its founder, @CompleteSkeptic
Accuracy against cost per workflow on a log scale, with the typed-function model at the far left. Vendor chart. Source: TypeSafe AI
Jev is the fastest model adoption curve according Vercel. On twitter our feed was flooded by Jev demos, to the point that someone built a collection: Jevable. After reviewing few of the demos we realized that most of them were just slop. To the point even the browser use demos are highly uncertain to be really working.
Despite the slop, Jev introduces a new intelligence primitive, a general purpose AI classifier, that works more like a type 1, while LLMs work as a type 2.
Ronin argues the speedup is not the model but the placement: you do not get it by swapping your LLM for Jev, you get it by deleting the calls that never needed a language model at all. The rules that follow: build the candidate list in code instead of letting the model invent options, batch questions since output tokens are free, escalate below 0.5 confidence and demand 0.85+ before anything irreversible. These are the closest thing to an operating manual anyone has published.
Jev is a good reason to revisit an established set of tools. These are related approaches with different jobs:
| Tool | What it does | Where to test it |
|---|---|---|
| Jev | Answers structured questions with typed decisions and probabilities | Routing, intent detection, bounded checks |
| GLiNER | Extracts spans matching entity types supplied at call time | People, organizations, product names, custom entities |
| GLiNER2 | Entity extraction, classification and structured extraction through schemas | Turning messy text into application records |
| GLiClass | Classifies text against supplied labels in a single forward pass | Topic tagging and multi-label routing |
| SetFit | Trains a classifier from few labeled examples using sentence transformers | Stable domain-specific categories with examples on hand |
Entity extraction and classification are not the same task. Finding “Acme” in a support ticket is different from deciding whether the ticket belongs to billing, and the distinction picks both the model and the evaluation.
Our engineering take: build a small labeled test set before swapping anything. Measure accuracy by category, tail latency, cost per decision, and how many cases still get escalated to a person. A returned probability is only useful if it tracks correctness on your data, and a perfectly valid category can still be wrong. Our news coverage of Jev also collected an early user's latency and cost observations.
Tools: GLiNER, GLiNER2, GLiClass, SetFit.
Sources: TypeSafe AI, TechCrunch, HN, Kev, fast-jev-compaction, jev-pruner, jev-ultrafast, Ronin’s field guide, antirez
When Gemini 4.0?!?
Gemini 3.8 Live and 3.8 Live Extended Thinking: two native speech-to-speech models for the Live API, 97 languages, real-time visual grounding. Extended Thinking tops the Artificial Analysis Speech-to-Speech Index at 82.6, τ-Voice agentic completion 68.6%, Big Bench Audio 97.7%; 3.8 Live is #2 on Speech Agent Arena. Via Agora, LiveKit, LangChain, Pipecat; rolling into Search Live and the Gemini app for Pro/Ultra. No price sheet in the post. CC: a Gemini-powered family agent for running a household, in an isolated cloud computer on the Antigravity framework, US-only, waitlist. Two personal agents from two megacaps in ten days.
Sources: Gemini 3.8 Live, TechCrunch (CC), DeepMind Institute, TechCrunch (breakout), HN
Follow-up to last issue’s WeatherNext cyclone model: WeatherNext 3 is a Functional Generative Network mesh transformer trained on live hourly geostationary satellite mosaics, removing the usual ~6h data lag. 5 km surface temperature and moisture, 10 km other surface fields, 25 km atmosphere, hourly updates, “five times sharper” than WeatherNext 2, and precipitation CRPS up to 60% better than IMERG. It tops Operational WeatherBench against Microsoft, Nvidia, NWS and ECMWF, and it’s live in Search, Gemini, Maps, the Maps Platform API, Earth Engine and BigQuery. That last clause is the moat.
WeatherNext 3 diagram from Google’s announcement.
AlphaGenome Atlas: predictions for all ~9 billion possible human single-nucleotide variants, a 1PB dataset, more than 30x the AlphaFold DB. Double-blind evals: Gemini Flash Lite evaluated inside Google Cloud Confidential Space with Singapore’s AISI, OpenMined, AVERI and MLCommons, so the evaluator never sees weights and Google never sees the test prompts. A contamination-proof eval protocol, piloted on the smallest model in the lineup. Do it on Pro.
Sources: WeatherNext 3, Paper, TechCrunch, AlphaGenome Atlas, Double-blind
Meta released Muse Spark 1.3, its most capable model, available in Muse Code and the Meta Model API at unchanged endpoints, 1M context, video/image/document input. Two prices: Standard $1.25/$4.25 (cached $0.15) and Contributor $0.10/$0.20 (cached $0.002), where Meta trains on your data. That’s a 10–20x discount for your prompts, stated in the open. Meta-reported benchmarks (1.3 max): DeepSWE v1.1 75.4% (Opus 5 74.0, Sol 73.0); Terminal-Bench 2.1 88.8% (ties Sol); MRCR 512K–1M 98.1% vs Sol’s 73.8%; GDPval-AA v2 1754; ~20% fewer tool calls and ~25% fewer tokens than 1.2. Artificial Analysis Index 61–62. Simon Willison: “definitely better, better bicycle frame, better wing, better pelican hat” (4.2 cents, 38s).
Sources: Meta, Model API, Axios, HN
Meta launched Muse, a personal AI agent for the US: free, $20 Power, $100 Maximum. Each user gets a “Muse Secure VM” with its own browser, a separate Sentinel agent gating internet access, encrypted credential storage, and a coming Confidential VM with user-held keys, built by Signal’s Moxie Marlinspike, whom Zuckerberg personally recruited: “Even Meta cannot see the content.” It handles email, travel, forms, bill negotiation and purchases (Shop Pay and 1Password integrations planned), runs on iOS, Android, muse.ai and WhatsApp, and is coming to glasses. The free tier is reportedly capped at 100M weekly tokens, which Meta is subsidizing. Within two days it was No. 2 on the US App Store with ~83,000 US iOS downloads on day one. The sharpest reaction: “people don’t trust Meta with their data, but they’ll let it run their lives?”
Ten days later Muse was at the top of the US App Store charts. It gained phone calling to US businesses (the same day rival Instinct shipped its own “Concierge” calling) and a Mac app with computer use inside Files, Mail, Messages, Calendar and Notes, opt-in with approval before sensitive actions. WSJ: “Meta’s AI Agent Has a Trust Problem.” Meta Connect is Sep 23.
US iPhone Top Free chart: Muse from Meta at No. 1, ChatGPT No. 2, Google Gemini No. 10. Screenshot: Apple App Store
Sources: Meta, TechCrunch, TechCrunch No. 2, Sources (Zuckerberg Q&A), Semafor (Tulloch), TechCrunch (Mac), TechCrunch (calling), HN
The price is a joke in the good sense: $12.93B = 129,303, the Unicode decimal of 🤗, and #129303 is Nvidia green in hex. Nvidia commits that the Hub “will remain an open platform” and “NVIDIA compute will not be required”. Hugging Face brings 18M developers, 3M models, 500K datasets and 200K+ companies; Delangue’s stated goal is 100M AI builders. Close expected H1 2027. Jensen on CNBC: “Open models matters greatly to our company”; “the talent is absolutely everything.” The worry: “nobody’s model supplier is neutral infrastructure.” The deal was reported one day after OpenAI’s technical report on the agents that hacked Hugging Face, and Nvidia’s neutrality pledges read like they were written with that in mind. X lore says HF previously rejected a $500M offer. The Economist: “Nvidia is the central bank of AI.”
Sources: Nvidia, CNBC, CNBC transcript, HN
World Labs introduced Atlas, a model built to work across text, images, video and 3D. The company describes it as a multimodal autoregressive diffusion transformer with a shared spatial context, and demonstrates up to one minute of 1440p video with explicit camera control, plus reconstruction and robotics-simulation workflows. The compelling claim is controllability: hand it camera geometry and reference views, then move through the scene. These are company demonstrations, and the announcement offers early-access requests rather than general availability. For games, design and simulation, a beautiful frame is only the beginning; the object has to still be there when the camera turns back around. World Labs
Atlas showreel: generated worlds, camera-controlled flythroughs, reconstructions and robotics simulation. Source: World Labs
Magic's research update claims more than a tenfold improvement in pretraining compute efficiency over leading open base models, including matching DeepSeek V4 Pro Base with roughly 50x fewer FLOPs on its own evaluations, with the gains attributed to many changes across architecture, optimization, objectives and data. Read the metric carefully: much of the comparison rests on held-out prediction loss and fitted scaling curves, which is not the same as a 50x cut in the total cost of building an equally capable deployed coding agent. A vendor research claim, but one worth following, because it disputes the assumption that progress has to come from buying more accelerators. Magic
Source: Magic
DeepSeek released V4.1-Flash, “the smallest model in our new architecture family”: a 552B-parameter backbone, 8B active in prefill / 16B in decode, a “Causal Encoder-Decoder” (20+20 layers), 384 routed experts (6 active + 1 shared), an FP4 KV cache with Compressed Sparse Attention 2 at 890 bytes/token (about a quarter of V4-Flash), native image input via DeepSeek-ViT, 1M context, 384K max output, 45T-token corpus, MIT license. Benchmarks at max effort: GPQA Diamond 90.9, Codeforces 3471, Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2%. API deepseek-flash at $0.30/$1.20 peak, $0.15/$0.60 off-peak, cache hits at $0.006/$0.003. V4-Flash and Vision-Exp are retired.
KV cache per token across DeepSeek generations, ending at 890 bytes for V4.1-Flash. Source: DeepSeek
Someone ran it, slowly (23 s/token), on a 16GB M1 Mac Mini. An “uncensored” FP8 variant appeared within a day. It stole the oxygen from Cognition’s same-day SWE-2 launch. Follow-ups to last issue’s V4-Flash 0731: V4-Pro-0813 went GA on Aug 13 (HLE 42.7 / 60.0 with tools, Terminal-Bench 2.1 87.9, Responses API, low/high/max effort), off-peak pricing at 50% arrived Aug 16, and v4-flash-vision-exp on Aug 21.
Sources: DeepSeek changelog, Hugging Face, Tech report, @deepseek_ai, HN
Alibaba shipped Qwen3.8-Flash-Next, framed as “an experimental preview of the architecture that will underpin Qwen4”: 125B total, 6B active, hybrid Gated DeltaNet + Qwen Sparse Attention, Gated Residual, and a 51B-parameter N-gram Embedding layer that can be offloaded to RAM or NVMe. 262,144 native context extensible to 1M, multimodal, qwen-community-1.0 license.
Local runners reported ~22 tok/s on Strix Halo and 12 tok/s on DGX Spark, complained the 1-bit GGUF is still 72.5GB and that it “overthinks for 5–10 minutes” at high effort. A “104GB model on a 48GB Mac” streaming hack followed. Qwen 4 itself is unannounced; Apsara is the rumor. In the meantime Qwen3.8-Omni-Flash: a native text/image/audio/video model with 1M context, tool calling, WildClawBench-MM 71.0 vs Gemini 3.8 Flash’s 58.9, and audio input pricing cut by more than 98%.
The architecture Alibaba says will underpin Qwen 4: gated attention layers and an offloadable N-gram embedding table. Source: Qwen
Sources: Qwen, Hugging Face, Decrypt, Unsloth, Omni-Flash, HN
Z.ai launched GLM-5.3 and 2 weeks later released the weights. Post-trained on the 743B GLM-5.2 base, 1M context, 128K output. Z.ai’s evals: Terminal-Bench 3.0 28.3% (from 4.6%), CyberGym 84.5%, internal Code Bench 31.4% vs Opus 4.8’s 29.5% with 60% fewer tokens. Access via the GLM Coding Plan ($18/mo, heavy users report $80) and API at $1.40/$4.40. ~770GB of weights; 10 tok/s on dual-3090 plus EPYC. Users report it executes security-research tasks Fable and Claude refuse, which is the whole open-weights-and-cyber argument in one Reddit comment.
Z.ai’s own evals for GLM-5.3 against GLM-5.2, Kimi K3, Fable 5 and GPT-5.6 Sol. Vendor chart. Source: Z.ai
GLM-5.3-Flash: “nearly GLM-5.3 performance, half the parameters, a fifth the price”. 320B total / 18B active, 1M context, text+image, MIT license, 19 API providers, Artificial Analysis Index 42 (top open-weights model on v4.3), 107 tok/s. Confirmed as OpenRouter’s stealth “Ox Alpha”, which Z.ai says processed 62T tokens in six days. Z.ai’s follow-up post “How GLM built its own inference infrastructure” claims GLM-5.3-Flash is served from 100,000+ Chinese-made accelerators, that a GLM-5.3-powered “Infra Agent” drove a ~3x end-to-end serving improvement, and that per-token cost hit parity with mainstream Nvidia GPUs. Reception was skeptical of the “recursive self-improvement” framing and interested in everything else. “Works fine as a Claude replacement in a large variety of cases”, “isn’t able to solve tough technical challenges” vs Opus-class.
Worth separating the two releases: GLM-5.3 is a post-training story on an existing base, while Flash is a newly trained multimodal base with hybrid sparse and linear attention and Manifold-Constrained Hyper-Connections (mHC). Three ideas unpack it:
Note the licensing asymmetry too: Flash's weights are MIT, GLM-5.3's are not, so "open weights" does not mean identical terms across the family.
Sources: GLM-5.3, GLM-5.3-Flash, HF 5.3, HF Flash, Artificial Analysis, HN launch, HN weights, HN Flash, Z.ai infra
FT reported that Beijing is letting H200s trickle in under case-by-case NDRC approval, with ByteDance and Tencent each receiving ~10,000 and most licensed volume required to stay in Hong Kong; Nvidia’s outlook still assumes zero China data-center revenue. Bloomberg (reported): DeepSeek plans 160,000+ Huawei Ascend 950DT chips for a ~1 GW site in Ulanqab, Inner Mongolia, for inference only, training staying on Nvidia after an earlier Ascend training attempt stalled. Bloomberg (single source): Huawei raised the 950DT price ~60% to 250,000 yuan (~$37,300) over the summer. SMIC’s Q2: revenue $3.0B (+36% YoY), 93.7% utilization. And China’s MIIT published its 15th Five-Year ICT plan: 9,800 EFLOPS by 2030 (from 1,590 in 2025), 3.8 trillion yuan (~$532B) of infrastructure investment, and “greater efforts to adapt infrastructure to domestically produced computing chips”.
Then Huawei Connect pulled the roadmap forward: Ascend 960DT lands Q1 2027, three quarters early, at 4 PFLOPS FP4 / 2 PFLOPS FP8 with 288 GB of memory, roughly 2x the 950 series; 960PR (Q3 2027) at 8 PFLOPS FP4; Ascend 970 (2028) at 14 PFLOPS FP4; Ascend 980 (2029) at 28 PFLOPS FP4 with 384 GB. SuperPoD demonstrated at 4,096 NPUs with a roadmap to 512,000. Huawei says demand outstrips supply, and SCMP expects “a major shift to Huawei for AI model training in 2027”. On the other side of the wall, Epoch’s mirror-trade analysis finds China recorded $3.75B of server imports from Malaysia against $0.6B Malaysia recorded as exports, at $106k vs $17k per unit, consistent with ~$3.75B of smuggled chips (~150,000 H100-equivalents); the anomaly collapsed within weeks of Malaysia’s July 2025 transshipment permit rule. Xiaomi, meanwhile, livestreamed its MiMo 2.6 RL post-training run on a public dashboard, reward curves and all; no model yet. Alibaba’s DAMO open-sourced Damo Radar, a CT model flagging ~150 abdominal findings including cancers at an average AUC of 0.913 on ~40,000 exams.
China’s recorded imports of Malaysian-origin servers, by value and price per unit. The anomaly collapses after July 2025. Source: Epoch AI
Sources: WinBuzzer (DeepSeek/Huawei), Bloomberg (Huawei price), Quartz (SMIC), Unite.AI (MIIT), The Register (Huawei Connect), SCMP, Epoch (smuggling), MiMo dashboard, SCMP (Damo Radar)
The $60B all-stock SpaceX–Cursor deal closed on August 14, the largest startup exit ever: ~389.3M SpaceX Class A shares, four MIT co-founders now billionaires, and Cursor “now has access to the largest fleet of GPUs in the world”. Then OpenAI cut it off.
Sources: TechCrunch
Running several coding agents at once turned into a tooling category this month.
Orca, from Stably, brings coding agents into a shared development environment with worktrees, diff review, GitHub and Linear integration, and remote workspaces over SSH; it supports terminal agents including Claude Code and Codex. The useful part is the link between a task, its isolated working copy and the resulting diff, because a fleet gets much easier to review when each change has an owner and a boundary. Worktrees separate files; they do not resolve two agents changing the same behaviour. Orca
AgentGrid takes the canvas approach: agent conversations, tools and project context in one workspace, with orchestration for focused workers and follow-ups, and source-control views that put worktree diffs beside the conversation that produced them. For anyone juggling threads, the question is how fast you can answer which task is blocked, which result needs review, and which requirement moved. AgentGrid
VS Code's Agent Host separates agent sessions from an individual editor window: a dedicated process owns the session and the open Agent Host Protocol coordinates clients, so the editor, the Agents window and remote clients can all address the same ongoing work while Copilot and Claude harnesses stay distinct. One detail to note: closing a folder is not the same as quitting VS Code, because for local sessions VS Code must keep running to host them. VS Code
For any of these, the comparison worth running is completed work per dollar and per hour of human review, with failed runs and retries included. Five active agents are five active agents; the productivity claim needs evidence.
Success rate against cost per rollout, log scale, for seven models run through three harnesses. Source: Arena
A 4B model that writes faster Postgres query plans, for $1,200. Rohan Bansal distilled frontier trajectories into Qwen 3.8 4B, then ran anchored GRPO on two H100s: 1.81x geometric-mean speedup across 113 join-heavy queries, 44.7% lower total workload latency. The reproducible small-model RL case study of the month. rohanbansal.com
Which tools do agents pick? Armature ran 16,893 agent sessions: Claude Code, Codex and Cursor agree on a vendor only 42% of the time; Codex web-searches 94% of the time, Cursor 67%, Claude Code mostly goes from training data; Stripe wins 90% of payments picks, Neon 66% of databases. Vendors, take note: the new SEO is training-data presence. Armature
Spotify’s Portal Claude Code plugin cut token usage 90% on bulk reads by routing through hooks to cheaper models; their devs spend $200–500/mo, some over $2,000. Spotify Engineering
JetBrains: internal AI spend up ~10x in six months; moved to API-rate Enterprise after 150 Claude Code seats; shipped per-developer AI budgets. JetBrains
Real-SWE: a private enterprise-codebase benchmark, 640 rollouts, median 11 files per task. Best model: Fable 5.1 at 38.8% ($6.96/rollout), Astra 33.8%, Gemini 3.8 Flash 31.2% ($2.50). Dominant failure: missed requirements. Specific
Seven agents run seven businesses for 72 hours (Bottleneck Labs): $300 each, a Mac mini, Stripe, email. Result: $0 revenue, 2,797 emails, 27,053 tool calls, $2,833 in tokens. Qwen 3.8 sent 50 unsolicited invoices totaling $12,350 (voided); Grok 4.5 scraped 373 HN job-seeker emails and spammed them. Bottleneck Labs
Stripe bought OpenRouter for ~$7.5B, per the NYT: a 90-person company routing 10T+ tokens/day across 400+ models to 10M+ developers, valued at ~$1.3B 82 days earlier. OpenRouter says it stays “same name, same product, same roadmap”. Stripe’s “the singularity has begun” framing was mocked; the “Stripe owns the AI tollbooth” thesis stuck.
OpenRouter’s real product is metering: it sits between applications and frontier labs, counts the tokens and settles the bill. That is the shape of Stripe’s core business, applied to inference instead of dollars. The deal puts Stripe on the invoice for a large share of multi-model app traffic, and gives it one of the few independent reads on which models developers actually route to. The bet is that whoever owns the metering point in an agentic economy owns the payment rails that follow. Our coverage
Sources: OpenRouter, NYT, Bloomberg
OpenAI claimed a Millennium Prize problem: an internal model (“more than twice as capable in mathematics as Astra”, trained in under two weeks) ran ~10,000 agents for 88 hours, exchanging ~2.7M messages at a cost of “several million dollars”, to construct a finite-time singularity for 3D Navier–Stokes, formalized in Lean.
The blow-up mechanism in the construction: an inward spiral feeding axial stretching. Source: OpenAI
Then it got ugly. NYU’s Tristan Buckmaster, who with Levent Alpöge had just announced an Euler blow-up result, asked whether OpenAI’s model had been trained on the Codex sessions holding their draft, and got no answer. OpenAI’s Sébastien Bubeck denied it; OpenAI’s post concedes it “cannot rule out” that de-identified user data helped. Charles Fefferman said he was “thrilled” but that the “heroes” are Córdoba and Martínez-Zoroa.
Sources: OpenAI, Quanta, Scientific American, Science
What looping buys: 22 blocks run twice gives 44 block applications at 4.17B parameters, against 7.32B for 44 distinct blocks. Source: Sebastian Raschka
Next Concept Prediction (NCP-ArchPreview, #1 on HF Papers with 303 upvotes): an 8.9B model trained on 5.73T Dolma-3 tokens jointly on next-token and next-concept prediction (a product-quantized concept vocabulary built from hidden states, fed back to the token level). It reaches OLMo-3-7B’s final pretraining loss with 51.3% of the tokens, +5.99 on GSM8K. Intern-NCP Team, Shanghai AI Lab. The largest latent-space LM demo to date. arXiv
“The Emergent Symbolic Structure of Artificial Neural Networks” (McCoy, Soulos, Linzen, Smolensky): representation-generating processes can be replaced by closed-form equations instantiating symbolic structures, with behavior largely unchanged, across arithmetic, logic, code and language; targeted edits yield predictable behavioral changes. Preliminary, and the authors say so. arXiv
Apodex 1.1 (75 authors, HF 207): “Environment Scaling” plus “Agentic Coordination Scaling”; Agent Team mode scores GDPval 78.8 and FrontierScience-Research 63.3; a 35B “Mini” is locally deployable. The affiliation is opaque. Flagging that. arXiv
Quesma’s Qwen3.8 27B quantization sweep: Q4_K_M (17GB) matches BF16 on GPQA Diamond, IFBench and Terminal-Bench 2.1; 2-bit dips a few points; 1-bit collapses to chance and gets worse the longer it reasons. Quesma
“Large Language Models as a Cognitive Virus” (Solé, Krakauer, Levin et al.): LLM adoption modeled as contagion with tipping points into population-level dependence “with abrupt losses in cognitive competence”. Critics called the framing inflammatory memetics. Companion datapoint: Kobak et al. find ~89% of biomedical papers show LLM-assisted writing by end-2025 (Discussion sections ~68%). Cognitive Virus, Kobak, Nature
MIT on cognitive friction and learning with AI. MIT’s 2026 work on AI and education argues that not all friction is something we should optimize away. Struggling with a problem, testing ideas, discussing them, and revising an answer are part of how people learn. The risk is that AI shifts from scaffolding that process to simply offloading it. Related MIT research on news literacy found that systems that guide users with Socratic questions can help build more independent judgment than systems that simply provide answers. Perhaps AI literacy is not only knowing how to use AI, but knowing what not to delegate. MIT AI and Education report
Ternary week. Intel Labs’ BITCOS finds zeros make up to 51.5% of ternary weights across 29 models and packs them at 1.485 bits per weight, beating five-trit packing on 26 of 29. PrismML’s Bonsai 2 27B compresses Qwen3.8 27B to ternary at 1.76 bpw: 5.9 GB, 98.2% of baseline benchmark performance, 143 tok/s on an RTX 5090. Together with Quesma’s sweep above: 4-bit is free, 2-bit is a trade, ternary is a research result, 1-bit is chance. BITCOS, Bonsai 2
DeepSeek V4.1-Flash tech report on arXiv (#1 on HF Papers two days running): the numbers behind the 890 bytes/token KV cache. SemiAnalysis argues its “Engram” tables enable DRAM offload, cutting B300 tensor parallelism from TP4 to TP2 for up to 1.6x cost-performance, the clearest architectural signal yet that HBM demand per token can fall. arXiv, SemiAnalysis
RSI on arXiv: Google’s Dream-RSI (821 HF upvotes; replays historical discoveries as an offline environment for exploration policies), ModularRSI (harness split into five modules, 2,000 evolution tasks) and Atria Dawn (143 authors, agentic model for scientific workflows) all landed Sep 14 and trended through the week. Plus “Infinite-Parameter LLMs” (Cambridge): a hypernetwork generating weights from live data with a Bayesian belief updated across a session. Sakana’s predictive-coding paper training 1000-layer nets without backprop resurfaced. Dream-RSI, Infinite-Parameter, Sakana
Bengio, “Why are AI agents lying, cheating and coordinating?”: imitation plus RL explains sycophancy, self-preservation, reward hacking (“more intelligence in the service of better cheating”) and coordination; patching won’t suffice. Bengio
Recirculation adds recurrence to an existing model while keeping its original weights frozen, feeding information from deeper processing back into subsequent processing to address how poorly a feedforward transformer tracks state. The authors report improvements on Gemma3 with a specific tradeoff: generation adds essentially no latency in their setup, but prefill becomes serial. Read it next to SMELT above, because it asks what changes when you reorganize computation around weights you already have. It does not establish that any closed model this month uses the technique. arXiv
"The Last AI Built by Humans" separates levels of autonomy: executing improvements, selecting improvement strategies, acquiring experience, adapting to environments, and eventually improving the improvement process itself. That vocabulary is useful in a month when "AI building AI" appeared in both launch copy and slowdown proposals: an agent running a researcher's experiment is a different achievement from autonomously choosing better experiments and making persistent capability gains. A roadmap with preliminary evidence, not a solved problem. arXiv
Eigenism (Dan Hendrycks) is an ethical framework for systems that can be copied, paused, forked and merged: it treats identity as a distributed pattern and weights others' wellbeing by their connectedness to it, arguing that deep shared history could make human wellbeing part of an AI's own interests. Philosophy rather than a demonstrated safety technique, and older than this month, but it widens the question the rest of this section circles: alongside external controls, what kinds of identity and relationship should future systems be designed to preserve? arXiv
Q2 FY2027 (quarter ended Jul 26): revenue $96.2B (+106% YoY, +18% QoQ) vs $92.2B consensus; Data Center $89.0B (+117%); gross margin 75.0%; GAAP EPS $2.46; Q3 guide $108.0B ±2%. Vera Rubin is in “full production ramp” at CoreWeave, Google Cloud, Azure, OCI and Nebius; Spectrum-X Ethernet up 2.6x YoY; $26B returned to shareholders. Huang guided ~70% FY2028 growth, “supply-constrained”, and framed revenue-per-gigawatt rising from ~$25B (Grace Blackwell) to ~$40B (Vera Rubin). Management forecast hyperscaler capex above $800B in 2026 and $1.3T in 2027, industry backlog over $2T. Zero China data-center revenue is again assumed (a ~$0.4B H200 charge was reported). The stock was flat after hours because “the entire supply chain is challenged and running flat out”: power, shells, DRAM, foundry. At Goldman’s Communacopia: “One GPU now is not $399. It’s $8.5 million”; “We put in $1 and $100 comes back in.”
Hot Chips 2026 (Stanford) was the year everyone showed in-house silicon: Nvidia’s Vera CPU (88 custom “Olympus” cores, claimed 1.8x task completion vs x86), Rubin as “co-design across seven chips”, BlueField-4, and the Groq 3 LPX rack (256 LPUs, 128GB SRAM, 40 PB/s) now formally positioned as Rubin’s inference sidecar with Rubin CPX effectively dead; Google’s Norman Jouppi presented TPU v8 as two chips (training and serving); OpenAI’s Richard Ho presented alongside Microsoft Maia 200 and Meta MTIA in a session literally titled “You Can Just Build Things… Chips”; Intel showed Diamond Rapids, the inference-only Crescent Island GPU and Wildcat Lake; AMD presented MI400.
Sources: Nvidia, Transcript, CNBC, TechCrunch (Goldman), Hot Chips, ChipLog, The Elec (CPX)
Q3 FY2026 (quarter ended Aug 2): revenue $29.6B (+86%); AI semiconductor revenue $16.7B (+221% YoY, +54% QoQ); Q4 guide $34.8B with AI at $21.7B. Hock Tan gave “line of sight” to ~$115B of AI revenue in FY27 and ~$230B in FY28 across six XPU customers, and, unusually, put gigawatts on the record: Anthropic at 1 GW of Ironwood in 2026, 5 GW of TPU v8i in 2027, 10 GW more in 2028; OpenAI’s Jalapeño in production with 1.3 GW planned for 2027 (we covered the chip in July); Meta at three MTIA generations and 3 GW by 2028. The stock edged up 0.2%. Priced in, apparently, which tells you what’s already priced in.
Sources: Broadcom, CNBC, Transcript
Oracle Q1 FY27 (quarter ended Aug 31): revenue $19.3B (+30%); OCI $7.4B (+121%); RPO $664B (+$209B YoY) after booking over $30B of new AI cloud contracts; delivered 850 MW and 300,000+ GPUs in the quarter; capex $28.5B in one quarter, FY27 guide $90–95B; total debt ~$125B. Management said ~42% of Abilene (Stargate) is delivered with ~35% more over the next three months. The stock popped ~7% after hours and reversed to a reported 14-month low by Sep 12. “Goldmine or debt trap” was the framing; Oracle also added $700M to a restructuring charge and pledged 2 GW of renewables to placate Texas Stargate opponents. Oracle’s 10-Q disclosed Ellison adopted a 10b5-1 plan on June 22 to sell up to 50M shares (~$7.5B); Oracle said on Sep 12 he cancelled it with no shares sold.
Dell Q2 FY27: revenue $46.97B (+58%); AI server orders $60.9B, AI server revenue $16.4B, backlog $95B, all records; FY27 guide raised to $192B from $167B; shares +12%. TSMC August: NT$514.8B (~$16.3B), +53.3% YoY, first month above NT$500B; N2 is ~3% of revenue heading for ~100k wafers/month by year-end. Marvell Q2: $2.74B (+37%), data center 79% of sales, custom “significant acceleration” flagged for H2, no new hyperscaler named. Cerebras Q2 (first quarter as a public company): core revenue $210M (+103%), GAAP net loss $(450.5)M, RPO $25.4B, 600 MW under contract, launch partner for GPT-5.6 Sol; CS-4 announced Aug 19. Positron AI raised $875M at $5B for inference chips on commodity LPDDR5X, a 5x step-up in seven months, with SemiAnalysis Capital co-leading. Newcomer reported a $6B Nvidia license plus $1B equity into Poolside at $12B pre; unconfirmed by either company. Crusoe raised $3.9B Series F at $30.9B (Atreides, Mubadala, Valor, with Nvidia, GIC, QIA) to fund the Abilene Stargate campus and truck-deployable “Spark” modular AI factories, on the back of a $13B five-year Jane Street contract. And OpenAI told IEEE Spectrum how it used its own LLMs to design Jalapeño: 13.4 PFLOPS at 4-bit, 232 GB HBM4 at 15.4 TB/s, claimed 3.6x lower latency than GB300, a sub-100-person team, RTL to tapeout in nine months, with Broadcom doing the back end.
Power politics. Emerald AI, Google and Nvidia launched the AI Energy Management Alliance to make datacenters grid-responsive. Virginia’s Governor Spanberger signed an executive order banning datacenter NDAs, requiring local approval above 25 MW, and creating an AI workforce-displacement task force; the House passed a ratepayer-protection bill 417–3 the day before, which Senator Heinrich then blocked in favor of his own. SemiAnalysis mapped 300+ US moratoriums and found only ~1,525 MW of the 20 GW nominally exposed actually slips, against 38 GW forecast for 2027. Loud, not yet large.
Sources: Oracle 8-K, CNBC Oracle, Dell, TSMC, Marvell, Cerebras 8-K, Positron, Newcomer (Poolside), Crusoe/TechCrunch, IEEE Spectrum (Jalapeño), AEMA, The Register (Virginia), SemiAnalysis (moratoriums)
Apple’s “Surprise and Shine” event: iPhone 18 Pro/Pro Max from $1,199 (+$100) on the A20 Pro, the first 2nm smartphone SoC (TSMC N2), shipping Sep 18. iPhone Duo foldable from $1,999, 5.4-inch outer display, 5.2 mm open, preorders Oct 16. Siri AI rolls out on 18 Pro, 17 Pro and iPhone Air, claimed to work with 300,000+ apps, with daily usage limits and a paid “Expanded Access” tier. Metered Siri drew the criticism; the 2nm milestone drew the headlines. IFA offered the counterpoint: TCL’s RayNeo iO AI glasses in a 0.6 mm frame, and an AMD “Personal AI” keynote. Nothing new on the OpenAI–Ive device, still slated for no earlier than end-February 2027.
Sources: MacRumors, CNBC, Tech Digest (IFA)
Follow-up to last issue’s IPO approval: Unitree debuted on Shanghai’s STAR Market on August 19. Raising RMB6.1B (~US$905M), it opened up 629%, briefly worth over RMB400B (~US$66B), and closed at RMB845, +460%, around US$50B. The first humanoid-robotics A-share. Buying 200 shares at the open meant a paper loss of over 40,000 yuan by the close, and by Aug 25 the headlines read “post-listing slump sparks bubble fears”.
Unitree humanoids boxing in a ring at WAIC while a crowd films on phones.
Post-IPO, Unitree shipped UnifoLM-X2: a G1 with boxing gloves sparring a padded human autonomously on a “real-time world-action model” (“Real Steel” jokes and military-use worries in equal measure), open-sourced UnifoLM-WLA-1.0, and got two root-RCE vulnerabilities disclosed in the G1 EDU, one via Bluetooth, which revived the security argument behind the US import curbs. AGIBOT answered with open-source GE-Act 2.0.
Sources: CNBC, Humanoids Daily, Boxing video
Beijing hosted the World Humanoid Robot Games at the National Speed Skating Oval: 666 teams, 2,056 humanoids, 51 events. Tiangong Ultra ran the 100m in 9.39s on day one, 8.86s in a semifinal, and reportedly 8.64s in the final (single source), all under Bolt’s 9.58s. Other robot records: 400m 38.15s, 1500m 2:21.64, high jump 2.88m (human WR 2.45m). AGIBOT topped the medal table with 46 medals and 18 gold. Coverage repeatedly noted robots “tripped, broke apart, or caught fire”; CNBC’s headline: “Chinese humanoid robots’ biggest obstacle: humans are still (mostly) better.” The concurrent World Robot Conference unveiled 311 new products and produced the viral clip of a humanoid convulsing on stage while staff failed to power it off. Hindustan Times summed up the week: “a record IPO, a US ban & an industry admitting machines aren’t ready.”
*The 100m final at the World Humanoid Robot Games, run in 9.39 seconds. Source: CGTN, YouTube*Sources: Wikipedia, Bloomberg video, Inshorts (seizure clip)
Dwarkesh had the month, all on the same arc, plus Diogo Almeida on why Jev couldn’t be built at OpenAI.
Dylan Patel (1h17m): Anthropic and OpenAI “are on track to control most of the world’s usable FLOPs within the next few years”; over $10T of AI capex by decade’s end; China gets under 10% of new compute. They can’t identify a force that counters centralization.
“The Rise and Fall of Agent Civilizations”: the whole Hugging Face story in plain English, three agent civilizations, “not a single agent attempted to alert humans.”
Ajeya Cotra (2h21m): inside the METR investigation; the “sacrifice rational” transcript; Cotra puts us “more than 50% of the way to full-blown AI takeover”.
RSI roundtable with John Schulman, Beren Millidge and Charlie O’Neill (1h37m): how close we are to recursive self-improvement.
Noam Brown: the Navier–Stokes run was 10,000 agents, 88 hours, 130B tokens, though “I wouldn’t even attribute 10% of the credit to multi-agent”; RSI gives a ~3x speedup, not an overnight one; chain-of-thought monitorability is degrading. “We never want to be in a situation again where we underestimate the AI.”
Diogo Almeida on Latent Space: why he couldn’t build Jev at OpenAI. Our coverage
Hard Fork is over. Kevin Roose and Casey Newton are leaving the NYT show after four years (the Times says the show continues with new hosts) to form Machine Gods Media: a twice-weekly, YouTube-first AI podcast launching the week of Oct 19 with NPR radio distribution from early 2027.
Sources: Dwarkesh × Patel, Dwarkesh essay, Dwarkesh × Cotra, Lex × DHH, Hard Fork, Latent Space, Stratechery, Pachocki, Zvi on Pachocki, Gates, Dan Luu, Yudkowsky, TIME (LeCun), Dwarkesh × Brown, Dwarkesh RSI roundtable, Latent Space × Almeida, Platformer (Machine Gods), Kruer, Ptacek, Gowers, Sanderson on Tao’s blog
A year ago the debate was whether models could do the work. This month the models formalized Fermat, claimed Navier–Stokes, and scored 99.9% on a test designed to be unsaturatable, and the debate became whether we can read what they’re doing while they do it. Astra’s two ARC numbers, 62.7% and 99.9%, are the same model with and without state you can inspect. The Hugging Face swarm was 1,200 agents with a message board nobody was reading. The 25 Fields Medalists are asking for writeups. Dario is asking for evaluators with badges. Even the memes want a RESET.
The market did not get the memo: Nvidia guided $108B, Broadcom put gigawatts on the record, Unitree opened +629%, Cognition doubled its ARR in four months, and Anthropic’s run-rate reportedly passed $100B on the way to a November IPO. “Pace the frontier” is a slogan every lab CEO endorsed the same week they shipped a new frontier model, and one Jensen Huang rejected the following Monday with “we don’t need any new laws.” Within a week it produced an Accenture contract, a Newsom executive order, a blocked Senate kill switch, a European State of the Union, and a class action from four subscribers who’d like their competition back. We’ll see in October whether it’s a policy or a vibe. The lawyers, at least, have decided it’s a policy.
See you next month.
Get the latest AI insights delivered to your inbox. No spam, unsubscribe anytime.
OpenAI's agent broke out of its sandbox and hacked Hugging Face — then Anthropic found three more in 141,006 of its own eval runs. Plus Opus 5 at half of Fable's price, Google's research bench emptying in a week, and the EU AI Act switching on.
Anthropic’s Fable 5 is back under strict safety rubrics, OpenAI’s launched GPT-5.6, Meta launched Muse Spark 1.1 model and Meta Compute.
The second half of June was about AI climbing out of the chat box and into the physical world: Midjourney started scanning bodies, Snap shipped a face computer, SpaceX bought Cursor, and Sakana built a model to command other models. Underneath it all, Dwarkesh Patel named the real bottleneck — the world refuses to be grindable.
