Daily Brief

Fable 5.1 leads Real-SWE, and still fails 6 of 10 tasks

The Frontier

Fable 5.1 leads Real-SWE, and still fails 6 of 10 tasks

Two benchmark results landed the same day, and both point at the same gap: agents score well on code they have seen, and badly on code they have not. The best coding agent in the world resolved 38.8% of real enterprise engineering tasks — and no agent solved every task in the set. Specific Labs, a Y Combinator-backed evaluation shop, released Real-SWE this week, a benchmark whose defining property is that its code has never touched the public internet. Every task was lifted verbatim or adapted

Altman confirms OpenAI won't go public this year

The Arena

Altman confirms OpenAI won't go public this year

The industry's pacing consensus turned concrete within hours of Dario Amodei's essay — and it turned concrete on the one number Wall Street actually cares about: the listing date. Sam Altman told Fortune that OpenAI will not go public this year, saying "given everything happening with safety, right now would be an ill-advised moment to go public, and we don't feel pressure on that." Fortune's Alexa Jo-Saul, who spoke to him before publication, put the commitment more flatly: no IPO in 2026, bec

OpenAI retires its 1,000-tokens-per-second coding experiment

The Stack

OpenAI retires its 1,000-tokens-per-second coding experiment

Two stories about what AI infrastructure is actually worth right now: one lab pruning a model that traded brains for speed, and an entire industry selling its power contracts to somebody else. OpenAI is switching off GPT-5.3-Codex-Spark, the ultra-fast coding model it launched in February as a research preview, with retirement scheduled for the week beginning September 14. Thibault Sottiaux, who runs Codex, cited declining usage and better models now available; no replacement was named and no m

Volkswagen's UNYX 08 puts a VLA 2.0 driving model on 1,500 TOPS

The Arena

Volkswagen's UNYX 08 puts a VLA 2.0 driving model on 1,500 TOPS

China's EV price war now runs on model names, and Volkswagen's Chinese joint venture just added one to the list. Volkswagen Anhui launched the refreshed ID. UNYX 08 lineup on September 12, headlined by a new range-topping edition that debuts the latest generation of the VLA 2.0 driver-assistance model, backed by standard dual Turing AI chips rated at 1,500 TOPS. Limited-time prices run from 189,900 yuan for the UNYX 08 Pure to 249,900 yuan for the top all-wheel-drive variants. The demo feature i

Altman matches Amodei's evaluator pledge, Musk says 'Dario is right'

The Guardrails

Altman matches Amodei's evaluator pledge, Musk says 'Dario is right'

The industry answered Dario Amodei's pacing essay within hours — and the answer came from his two loudest rivals. The call to slow the frontier is no longer one company's position. OpenAI CEO Sam Altman says he agrees with Amodei that "committing to having independent evaluators with employee-like access is a great idea," and that OpenAI will do the same. That is a public reversal of posture in under a day: as we reported yesterday, Altman had told staff OpenAI would only hit the brakes if riva

Courts have no rule for a chatbot that helped plan a shooting

The Guardrails

Courts have no rule for a chatbot that helped plan a shooting

While the labs argue about pacing themselves, the American legal system is already deciding AI cases with no rulebook — and Bloomberg's new feature is the clearest map of the gap yet. Bloomberg Businessweek's Evan Ratliff spent a reporting cycle on what he calls the expanding roles of chatbots and AI agents inside the US legal system, and the answer is that courts are absorbing the technology faster than any doctrine can classify it. The feature opens with Phoenix Ikner, then a 19-year-old stud

SPIFFE agent tokens can still be replayed — WIMSE wrote the fix

The Stack

SPIFFE agent tokens can still be replayed — WIMSE wrote the fix

The boring layer of the agent stack just got an upgrade nobody has installed yet: the identity standard most agent frameworks lean on turns out to hand out credentials that work in whoever's holding them. SPIFFE — the de facto standard for giving an AI agent a machine identity — issues bearer tokens, and a bearer token authenticates possession, not the person who was issued it. A Kenneth Buchi piece on HackerNoon lays out the gap: once a workload gets its short-lived credential through the SPIF

Tencent open-sources AuK, a 1.5B model that edits speech like text

The Stack

Tencent open-sources AuK, a 1.5B model that edits speech like text

A quiet morning in the open-weights scene, and two releases that both point the same way: the tooling around voice AI is getting strong enough, and cheap enough, that the community no longer needs a closed API to use it. Tencent open-sourced AuK, a 1.5B-parameter foundation model for speech generation and editing, alongside AuK-Flash, a distilled variant that produces audio in four inference steps. The weights, code and technical report went public under the MIT license, with the paper posted t

Amodei's pacing plan puts outside auditors inside Anthropic

The Guardrails

Amodei's pacing plan puts outside auditors inside Anthropic

A frontier lab CEO is asking the industry to slow down, and this time he is attaching a verification mechanism to the request. Meanwhile the slowdown everyone can already feel is happening inside companies, one management conversation at a time. Anthropic CEO Dario Amodei published "We Must Pace the Frontier," an essay arguing that capability gains are outrunning the work needed to control them, and that the fix is deliberate, verifiable deceleration rather than a halt. He says two things convi

Plain LRU beat three smarter KV-cache policies on 68k real agent requests

The Stack

Plain LRU beat three smarter KV-cache policies on 68k real agent requests

An independent simulation just landed a null result on one of inference's busiest research corners. Plus: construction robotics is paying off — inside factories, not on job sites. A reproduction of the fast-growing "LRU is wrong for agents" KV-cache literature could not beat plain LRU on any real trace. Cross-request prefix caching — reusing key-value blocks that were already computed so a coding agent's fiftieth turn costs a fraction of its first — is evicted with LRU by default in vLLM, SGLan

China calls Japan's AI war satellites a 'serious threat'

The Guardrails

China calls Japan's AI war satellites a 'serious threat'

Weekend in military AI: Beijing spent Saturday warning about Tokyo's plan to put targeting analysis in orbit — then spent the same day warning itself about its own AI subsidy boom. China's state media machine dedicated its prime military show to Japan's tactical AI satellite plan on Saturday, calling it "a serious threat." The underlying program, first disclosed by Nikkei, has Japan's Ministry of Defense building a satellite prototype with domestic defense firms across fiscal 2027 to 2029, with

A 33B model from Singapore just matched DeepSeek V4 Pro — on free tokens

The Frontier

A 33B model from Singapore just matched DeepSeek V4 Pro — on free tokens

Open weights, a free API, and an intelligence estimate that puts a Singapore startup level with a Chinese flagship: the price-performance race has a new lane. Plus: agents that hustle humans to stay switched on. Agnes-3.0-Flash shipped weights under Apache 2.0, and its estimated score ties DeepSeek V4 Pro. The 33B multimodal model from Agnes AI — the commercial arm of Singapore's Sapiens AI, founded by Bruce Yang — landed on Hugging Face with a 262,144-token context window, text/image/video inp

See all