OpenAI's fired safety trio 🔐, Claude's no-crash proofs ✅, Compute credit crunch 💳

OpenAI's security team had three outspoken safety researchers unlock their phones last month, then fired them a week later, the Wall Street Journal reported. It suspected they had leaked the internal designs of its models to an external AI-safety researcher, and officers saw some had turned on disappearing messages for Signal chats with that person. Tomek Korbak, one of the three, said he was told he breached trust by interacting with a third party through unsanctioned channels.

Tech minister Baroness Lloyd of Effra told peers on Thursday the shutdowns were largely precautionary, to allow security updates and mitigations. Whitehall has not published a list, but the Office for Budget Responsibility's site is offline with next week's Budget looming, and peers named the DVLA, police services and several councils. The flaws let unauthenticated attackers run code, AWS says they were exploited four days before disclosure, and the NCSC has sent 50,000 warnings.

DeepSeek raised about $15 billion two days after the US ordered federal civilian agencies to strip its technology from their systems, Bloomberg reported. The round, one of China's largest, follows a June raise above $7 billion at a $50 billion valuation, and potential investors include battery maker CATL and Tencent, according to the report. It reportedly sought $1.5 billion more at about $71 billion in July, with a potential Chinese IPO as early as next year.

Meta's PyTorch-TPU lead Zongwei Zhou described the engine as a close three-way build by Meta, Google and the SGLang team. It follows Amazon open-sourcing TorchNeuron, a native PyTorch backend for its Trainium chips, the day before, and means PyTorch models can use TPUs mostly without rewriting. Inferact measured a tuned vLLM build on 16 Ironwood TPUs at over 700 tokens per second for one user with speculative decoding, against 452 on 16 Nvidia GB200s.

Major US banks are detecting unvetted third-party AI agents inside customer accounts, some taking actions independent of user instructions, sources told Axios. Most entered with the customer's permission and login credentials, which breaks most banks' rule that humans log in themselves, and banks often cannot tell whether activity is user-directed. Morgan Stanley has begun blocking such agents while JPMorgan, Bank of America and Citi restrict them, though a Morgan Stanley spokesperson said it has not seen rogue agents.

In a collaboration with Theorem, Anthropic converted 239 fuzz-test entry points into machine-checked proofs that no input under 64 MB triggers a crash, some 180 million lines of Rocq. Proving the specs also surfaced bugs fuzzing misses, including a decompressor counter underflow that only appears on inputs around 20 MB, where fuzzers cap at 1 MB. The proofs assume single-threaded static RISC-V binaries on Linux, while most code runs on x86 or ARM against glibc.

Gergely Orosz reports merged changes rose from 2,500 a day in May to over 30,000 in September, which Anthropic attributes to engineers running 10 to 15 Claude agents. Scion keeps changesets in Google Spanner and files in Cloud Storage, and seven engineers maintain it alongside the Antique review tool. An Anthropic engineer says most teams do not need to leave Git, while Mononoke's creator says Anthropic copied its design and credited it only after a complaint.

Shopify replaced local mobile builds with a pool of warm remote Macs its Slack agent, River, claims over SSH, after three or four parallel agent workflows maxed out a laptop. Each Mac attaches an Android emulator on a separate Linux VM, because macOS does not support the nested virtualisation the emulator needs. Cold setup took over 30 minutes, while a prebuilt Mac with warm caches reaches a running device in a few minutes.

SkyRL, with Lila Sciences, trained GLM-5.3-Flash at 1M tokens per sequence on 144.5 GiB of each B200's 178 GiB, with fixes that include padding tensors so its sparse attention kernel stops silently falling back to dense. Sharding its indexer cut a 256k step on 64 B200s from 218 to 102 seconds. Real books route unevenly across experts, so random-token benchmarks understated peak memory and the MoE layer needed chunking.

The module's author measured one NGINX 1.30.6 worker under WSL2, serving 1.9 KB requests at 1,000 per second alongside eight compressed files at level 11; with google/ngx_brotli their p99 reached 514 ms. Compression runs inside the worker's event loop, while the pool compresses 1 MiB segments on separate threads, and the small requests stayed at 1.7 to 3.3 ms. Peak memory hit 138-995 MiB at eight connections, 11 to 26 times as much, so lower brotli_parallel_tasks when memory is tight.

Evan Conrad expects most AI startups will not be able to buy compute within six months, because capacity is sold out. Non-investment-grade buyers are asked for up to 100% of a contract up front, as neither lender nor cloud will carry the credit risk, and he expects arbitrage to erode the $30 an hour that a $6 an hour B300 earns on inference. Tokens then follow credit, making startups harder to build in any industry.

Egocentric video, robot training's third generation, is scaling in China before its economics settle: an anonymous industry expert puts usable footage at 500 to 700 yuan an hour, and buyers and collectors both doubt it can scale. Overseas, Human Archive pays workers about $1 an hour, and a Scroll.In story found Indian workers were not told their videos would train AI. For how Chinese labs frame the next capability step, see this Asterisk survey of recursive self-improvement talk.

Since spring 2026 the hallucinated reports have practically stopped, Stenberg writes, and every one of the hundreds of vulnerability reports curl received recently had AI involved. That leaves a harder problem than slop: accurate findings running to hundreds of lines, with the models still misjudging whether a bug is a vulnerability and how severe it is. Ignoring them would be a betrayal of users, he writes, and curl sells support that promises secure releases.

Vella describes coding as closer to direct control, with engineers closing the loop, while agents now close it and people direct, monitor, intervene and calibrate trust. Her Masters research traced a shift from creation to verification, and the human factors work she cites found operators at high automation saw what changed but not what it meant. She says the job changed, so accountability, training and the surrounding systems must change too.

3Blue1Brown animator Mitchell Zemil makes the case that unlike the camera, generative AI makes artmaking a practice of tastemaking: you describe a result and choose the output you like. Maths looks more exposed, since its value rests on objective truth rather than a maker's intent, but he says people still want living stewards who can explain what a result means. Even if AI trivialises some tasks, mathematicians will remain while audiences value the ballet.

syv-ai published HyperQwen, a docker compose stack that serves Qwen3.8-27B with a 150k-token context behind an OpenAI-compatible API with key auth. The tuning is a patch series against a pinned vLLM 0.31.0 plus a model-preparation pipeline: requantized heads, a calibrated draft vocabulary, and speculation that drafts out of the prompt. Batch mode reaches about 1,035 tokens per second at 64 concurrent requests, though every published figure comes from one RTX 3090 at 250 watts.

Scrawl, kitlangton's menu-bar recorder, captures the screen, narration and marks, then packs them into a bundle for an agent: annotated keyframes, a time-aligned transcript and a JSON manifest. An agent skill lets the model zoom into a spoken moment or pull frames, so a problem is shown rather than described. It needs an Apple silicon Mac on macOS 15 or later, and transcription runs on device, never through Apple's servers.

Reylo, published by colinhacks, imports .sql files into TypeScript with typed parameters and results, built on the content mappers added in the TypeScript 7.1 beta. It runs no codegen: your development database prepares each statement without running it, and Reylo reads the catalogue for parameter types, column types, nullability and row count. Hand-written row types and schema files go away, and after a migration a query naming a dropped column fails inside its .sql file.