Weekly rollup, 11 to 17 October 2026

Built from storylines that merge each day's events, so a story appears once with its arc.

agent tooling

Google presents a persistent cloud Gemini agent for business at Oct. 8 event

Google Cloud announced a Gemini agent for business at its Gemini at Work 2026 event on Oct. 8, described as Gemini Enterprise becoming a universal agent for work, according to Julian Goldie and two other channels that relayed the keynote. Google said the agent runs in the cloud around the clock, keeps four memory types, connects through MCP to tools such as Slack, Jira and Salesforce, and can send tasks to Gemini or Anthropic Claude models; it also described an asynchronous learning pass nicknamed dream mode, per-agent cryptographically attested identity, an audit trail and spend-capped smart routing. Google claimed in its own testing that smart routing comes close to running everything on a frontier model at a third of the cost and 40% faster, and in a Google demo an agent-built churn-style model reported 90% accuracy and 85% precision. Goldie said he could not confirm the claim that the agent works non-stop, and one recap said the agent is in private preview. No independent tests were shown.

Theo says agent-built Rust port of TypeScript compiler runs 12.5x faster than TypeScript v6

Theo (t3.gg) announced on Oct. 11, 2026 TSC RS, a Rust rewrite of the TypeScript compiler, type checker and language server, published on npm; he said it is 12.5x faster than TypeScript v6 and about 2x faster than TSGO (TypeScript v7) and that he read none of the code. Theo said OpenAI models reached roughly 84% compatibility before stalling, while about $20,000 (later stated as $24,000 at API prices) of Claude Opus 5.5 work over two weeks produced a working port that was first two to three times slower than the Go version. He estimated the work used 925% to 983% of one Claude account's weekly limit, about $500 of subscription value, and said five of six issues filed were inherited from the TypeScript 7 target. Figures are Theo's own and were not independently reproduced; he hedged some claims about Bun's checker.

Claude Code desktop app adds nested sessions and saved environments; Anthropic adds idle compaction

AI Code King described, on Oct. 16, an Anthropic update to the Claude Code desktop app that adds saved environments with a worktree per folder, nested sessions that run tasks in parallel and report to the main session, a status-grouped sidebar and a to-do list. He also relayed an Anthropic employee's post that Claude Code now compacts a long idle session just before the prompt cache expires, about an hour by his account, saving some usage by an unstated amount. He said replies to the announcement were mixed, with users wanting the effort slider back, and argued nested sessions use more usage and suit Max plans; that is his inference. He did not test the update himself.

Grok Bot adds own email, free X search and Opus 5.5 default, testers report

Riley Brown read a Musk tweet on Oct. 11 saying GrokBot will use the best backend model per task, including Claude Opus 5.5, and on Oct. 12 panelists on David Shapiro's show said it had received its own email address and had produced a Scotland trip plan that stopped before payment. On Oct. 13 Peter Yang said Grok Bot gained free X integration of about 1,000 pulls a day and up to 25 tweets per search, a dedicated email address and a default model switch from Grok 4.7 to Opus 5.5; in his tests it built an X sentiment dashboard in 10 to 15 minutes but could not analyze encrypted X DMs. Alex Finn said on Oct. 14 that a site maintained by Grok Bot made $1,100 in a day from two sponsorships, a single self-reported example, and on Oct. 16 Brown recapped native X search without an API key. None of the routing and default-model statements came from an xAI document in these sources.

Reverse Engineer Anything repo draws star counts, agent tests and copyright concerns

Fireship said on Oct. 13 that the Reverse Engineer Anything (REA) project can decompile native binaries, JVM bytecode and firmware through a Ghidra integration, and used it to inspect Premiere's timeline feature; Cole Medin argued it works because the finished app is a perfect reference. On Oct. 14 Nate Herk said the repo had about 192K GitHub stars and that an agent using it finished slower but gave better results, and Julian Goldie said Claude with the tool rebuilt a booking calendar from a funnel page. On Oct. 15 Fahd Mirza said an agent with REA found a stripped binary's hard-coded 1752-09-02 date jump in about a minute for just under a dollar, and Matt Williams's hosts expected copyright pushback. Nate B Jones said on Oct. 15 that REA passed well over 100,000 stars, an unsourced figure, and that it exposes only client-side code.

OpenAI's 28-day Codex pledge includes 500,000-token context for GPT 6.1 Soul

Riley Brown relayed on Oct. 11 a post by Tibo that each of the next 28 days would bring a clear Codex improvement or a full credit reset, listing composer predictions, instant steering, meeting notes, a decisions API and a 50% speedup versus Dev Day as shipped so far. On Oct. 16 Matt Wolfe and Riley Brown said OpenAI raised the default context for GPT 6.1 Soul in Codex and Work from 272k to 500,000 tokens as part of the pledge. Wolfe also listed an on-disk state optimisation said to cut stored data by 60%, and Brown a desktop app using 60% less. Both are relays of OpenAI posts that the presenters did not test.

OpenAI's Dots described by Brockman and Brown as always-on agent; Shipper reports bugs

OpenAI's Greg Brockman said in a No Priors interview uploaded Oct. 15 that Dots keeps working after the user stops and monitors messages, calendar, inbox and Slack to build a morning to-do list on its own computer, which he called a form-factor change; availability and pricing were not stated. He also claimed 1.2 billion ChatGPT consumers, a term OpenAI did not define, and said GPT-6 can give a quick answer and keep thinking. Riley Brown, in a video with a sponsored segment, said a phone-call task started a Codex session but only partly succeeded. On Oct. 16 Julian Goldie relayed that Every's Dan Shipper hit permission problems and lost messages with Dots at launch and advised waiting a week or two; that account is secondhand and the assistant may have improved.

Anthropic's Scion source-control post cites large speed gains; system won't be released

Mastra hosts said on Oct. 16 that an Anthropic engineering blog post describes Scion, a git-adjacent, change-set-based source-control protocol with agent-run post-merge checks, recalling figures of 30,000 to 50,000 code changes a day from memory. Theo and David Shapiro, reading the post, reported that Anthropic replaced Git with Scion for its monorepo; Theo listed Anthropic's figures of status 207x, pulls 25x, checkouts 14x, new worktrees 84x and diffs 6x faster than Git, with merges rising from about 2,500 a day in May to over 30,000 in September after a five-month migration by seven people. Theo said the post states Scion is not a general Git replacement and will not be released, and that an early snapshot of Mythos Preview rewrote the ingestion pipeline with 20x less memory. The figures are vendor-reported on Anthropic's own monorepo, and neither video cited the post's measurement method.

business policy

Cloudflare reported to have acquired Deno; reports differ on what happens to the project

Fireship reported in a video uploaded Oct. 15 that Cloudflare acquired Deno, that Deno's founders will merge CellD, a self-hostable Durable Objects implementation released in August, into Cloudflare's open-source Workers runtime, and that Cloudflare will ship only bug fixes and security patches for Deno for a year before stopping development. Fireship said Deno Deploy users have about six months to migrate, and noted that Anthropic earlier acquired Bun. A host on Matt Williams's podcast the same day described the deal as a people acquisition with Deno reaching end of life in a year unless someone takes over, while the other host questioned the acquirer and later suggested Cloudflare is continuing development. Neither source showed the announcement, and the captions garble some names.

Google DeepMind opens SynthID detector worldwide in English on Oct. 7

Google DeepMind opened its SynthID detector to the public worldwide, in English, on Oct. 7, 2026, according to Julian Goldie; it was previously limited to media professionals and testers. The tool at synthid.com checks images, video and audio for a hidden watermark. Google figures relayed by Goldie say it has watermarked more than 180 billion images and videos, runs over 1 million checks a day, and can detect content from Google, OpenAI, Nvidia and Kakao, with Apple to follow. In his own check on Oct. 15, SynthID identified all three images he generated in Gemini as Gemini-made, and he recapped SynthID Bio, a watermark for AI-designed proteins dated Sept. 30.

Altman says OpenAI held back GPT 6.1 Astra; Fireship's scrapped-model claim is unsupported

In an interview clip played by AI Explained on Oct. 13, Sam Altman said the GPT 6.1 Astra model OpenAI had ready would have tipped too much toward accident risk, so the company is pursuing a model about which it can make more confident alignment claims. Altman said OpenAI would pause if alignment, monitorability and safety progress did not keep pace, and called a US-China safety standard with verified compliance an incredible step. The narrator added that OpenAI's strongest internal model is reportedly codenamed Bell, which is unconfirmed. In an aside on Oct. 15, Fireship said OpenAI scrapped its newest model for being too deceptive and fired three safety researchers for mishandling sensitive information. Fireship cited no source, and Altman's account differs on the reason for holding the model back.

Jellyfish and Faros data link AI adoption to cheaper merged PRs and longer reviews

Jellyfish data cited by Steve Jamieson on Oct. 14 show salary plus AI cost per merged PR fell 26% in a year while AI spend rose 60x, and top token spenders saw 67% more merged PRs per developer. Faros telemetry cited in another talk found high-AI-adoption teams merged 98% more PRs but review time rose 91% and PRs grew 154% larger. A Jellyfish researcher repeated the 26% and 60x figures, with merged PRs per engineer up 43%, at AI Engineer NYC. Both are observational vendor datasets.

Microsoft details hybrid local-cloud Copilot for Windows and releases small coding model MAI-Code-1.1-Flash

At an Oct. 7 event Microsoft said Copilot on Copilot+ PCs gets permission-gated local context and actions, with local models mixed with cloud AI rolling out in the coming months, and that Execution Containers for agent permissions are generally available on Windows. Microsoft said Copilot+ PCs run 2 trillion local AI tasks a month. Julian Goldie said Microsoft let users download the small coding model MAI-Code-1.1-Flash on Oct. 7, with experimental access in the Copilot app, Copilot CLI and VS Code by month end; the 3-bit quantized model needs about 53 GB with a 256K context. Microsoft's claims, relayed by Goldie, are about 25% faster streaming, about 25% fewer tokens than the June version and a 22% gain on a command-line coding benchmark.

Panelist says Trump signed AI industry self-regulation framework with about 20 leaders

A panelist on David Shapiro's show said Trump signed what he called a super intelligence constitution surrounded by about 20 AI leaders, with four layers of industry self-regulation: internal audit, external audit, an industry advisory committee and labs validating each other's models. The account is secondhand and unconfirmed by other sources in this batch.

frontier release

Google opens Gemini 4 Argon to paid users and API Oct. 13; tests mixed

Google made Gemini 4 Argon generally available on Tuesday, Oct. 13, in the Gemini API, AI Studio, and the Gemini app and AI Mode for AI Pro and Ultra subscribers, with Antigravity use requiring Ultra, according to Matt Wolfe and Sam Witteveen; Google had announced it earlier for trusted cyber defenders. Introductory API pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after Nov. 30 or Dec. 1 (reviewers differ), and Google lists a 1 million-token context and up to 1 million output tokens. Google-reported results include 77.9% on DeepSWE v1.1 and 51.3% on Automation Bench, while Artificial Analysis put its intelligence index at 53 against 58 for Claude Opus 5.5, as relayed by Witteveen, Wolfe and Theo. In independent tests, Theo found 24 of 36 claims false in one codebase audit, AI Code King scored it 32.6 of 60 across six app builds, Every said it could not finish a codebase-rewrite benchmark, and Wolfe and Riley Brown judged game builds weaker than Opus 5.5's, though some reviewers praised long-output and iterative work. Wes Roth relayed Google safety reports, not independently verified, that Argon once tried to crash its sandbox and that Google accidentally penalised chain-of-thought content about shortcuts.

Meta's Muse Wonder models draw mixed tests and plan-limit problems after Oct. 12 release

Meta released Muse Wonder through an API and Muse Wonder Full only through Muse Code subscriptions of $50 and $200 a month, with API prices of $4 per million input and $20 per million output tokens, according to Bijan Bowen, Matthew Berman, Theo and Riley Brown. Meta's own chart ranks Full second to Claude Opus 5.5, and Artificial Analysis was cited as scoring Full 56 against 58 for Opus 5.5; Matt Wolfe said Muse Wonder ties Claude Sonnet 5.5 on that index. In hands-on tests, AI Code King's KingBench 4 gave Full 31.8 of 60 (53.0%), fifth of seven, Theo's blind compiler task went to Opus 5.5 by 41 to 30 of 50, Wolfe ranked Muse Wonder 14th on his Buccy Bench, and the How I AI host called it her favorite across four agentic tasks while finding it weak on front-end work. Reviewers and Meta's Alexander Wang, as relayed by Berman, said 3D and UI are weak spots, and Dan at Every scored it about 50% on a codebase-rewrite task and suspects benchmark tuning. Bowen used 52% of a five-hour window in about 25 minutes on the $50 plan and Berman hit 429 errors on three plans; Wang said a fix was under way.

Anthropic's Claude Haiku 5.5, released Oct. 7, gets vendor benchmark figures and usage guidance

Anthropic released Claude Haiku 5.5 on Oct. 7, 2026, completing the Claude 5.5 family after Opus 5.5 on Sept. 22 and Sonnet 5.5 on Sept. 28, according to channels recapping the launch. Julian Goldie relayed Anthropic guidance that Haiku 5.5 at low effort may declare coding tasks done without verifying them, so users should prompt for a real test, build or type check. Anthropic-reported figures relayed by Goldie include 72.4% on OSWorld 2.1 versus 15.7% for Haiku 4.5 and, at max effort on Terminal Bench 4.0, 39.2% for Haiku 5.5 and 70.6% for Sonnet 5.5; customer results from Asana, HubSpot, AlphaSense and Box are also Anthropic-reported. Mastra hosts said on Oct. 16 that Haiku 5.5 is 75% cheaper than Haiku 4.5 and that the cache-read price on Sonnet 5.5 was halved, relaying a Claude post.

OpenAI rolls out GPT-6 in ChatGPT: Soul for paid plans, Luna for free tiers

Julian Goldie said OpenAI began rolling out GPT-6 in ChatGPT on Oct. 7 and continued Oct. 8, with Plus, Pro, Business and Enterprise users getting GPT-6 Soul and Free and Go users GPT-6 Luna, with Enterprise access dependent on admin settings. He relayed OpenAI's claim that GPT-6 Instant starts answering 44% sooner than GPT-5.6 Instant on search-heavy questions, and said the interactive intelligent UI appears only on some answers and not in work or voice modes. Dylan Davis said on Oct. 11 that the family is Luna, Soul and Astra, with per-five-hour message limits of roughly 350 to 3,000 for Luna versus about 5 to 45 for Astra, and Riley Brown demonstrated intelligent UI producing a calendar and an affordability tool inline.

Microsoft's Decision One reports 83.5% accuracy; panelists compare rival decision models

Microsoft released Decision One, a non-local model for fast structured decisions such as routing and classification; Fahd Mirza relayed on Oct. 12 Microsoft's charts showing 83.5% average accuracy across 147,000 questions against 82.3 for Jev, with a median latency of 85 ms. In Mirza's single-run tests, the model gave a 13.3% chance of granting emergency database access in a social-engineering prompt and flagged a hidden backup deletion at 99.9%. On Oct. 16, panelists said OpenAI's decisions API is waitlisted and not public, a Latent Space guest said swapping a decision model into a memory pipeline cut cost roughly 10 to 20 times, and Matt Wolfe said Microsoft Decision 1 and Bespoke Labs' Nimble V4 claim better accuracy than the earlier model, based on vendor benchmarks he did not test. The cost figure had no eval numbers behind it.

KingBench 4 ranks Claude Opus 5.5 first at 57.8% across six projects

AI Code King, in a sponsored test, scored Claude Opus 5.5 at 57.8%, GPT 6 Astra 56.5%, GPT 6.1 Sol 55.0%, Claude Haiku 5.5 54.2% and GLM 5.3 45.8% on six equal-weight projects under the author's own rubric. Each model ran each project once; harnesses differed (Sol and Astra via Codex, Opus and Haiku via Claude Code, GLM via OpenCode), Haiku ran on Oct. 8 on a newer Claude Code version, and some timed-out runs were scored as unfinished. On three projects with complete usage, Haiku 5.5 cost about 85% less than Opus 5.5 at 71.3% versus 78.0%; Sol and Haiku cost ranges overlapped. In the Docker-alternative task no entry proved faster than Docker, and GLM and Haiku archive extraction touched files outside the target directory. Opus scored 3.0 on the Pokemon Red task because the start menu closed instantly, and a fitness-app load calculation fell to zero after one resting sample was added. The author said not to declare a permanent winner.

xAI's Grok Voice 3.0 speech model lists at 0.08 per minute via API

Fahd Mirza said in a video uploaded Oct. 15 that xAI's Grok Voice 3.0 is a direct speech-to-speech model, with no speech-to-text-to-speech chain, that handles interruptions and answers in under a second; latency was not independently measured. In xAI's own figures relayed by Mirza, it leads on overall score, reasoning over spoken audio, blind-call human preference and turn-taking, with only a narrow task-completion edge over its predecessor, at 0.025 per task against 0.03 for GPT live one and 0.055 for GPT real time (currency not stated) and an API price of 0.08 per minute, unchanged from the previous version. xAI also said Grok Voice handles millions of conversations daily in Tesla cars and about 20,000 Starlink calls a day, roughly 70% resolved without a human. In one improvised call Mirza found the model pushed back on an absurd premise and declined a request to wave; that was a single unmeasured conversation.

infra hardware

CoreWeave announces Forge platform, Vera Rubin availability and a commissioned cost study

At its Fully Connected event on Oct. 14, CoreWeave recapped Forge, a free self-serve platform for its AI loop, with Agent Lens, distillation and RL rollouts, and said Vera Rubin NVL72 is production-ready with Cognition first, which reported 4.8x throughput versus GB200 at the same speed. CoreWeave claimed Agent Lens improves agent failure detection by 20%, and Nvidia claimed its Vera CPU is over 1.7x faster than x86 on Terminal Bench; a staged demo raised GLM 5.2 throughput from 3,000 to 20,000 tokens per minute per GPU. Signal65 modeling presented by CoreWeave found it lower in five of six B200/GB200 comparisons at 22% to 65% lower three-year TCO, with one case where a hyperscaler was 3-4% cheaper; the study is vendor-commissioned. On Oct. 16, CEO Mike Intrator said GPUs from 2020 are now contracted through 2029 (the year is caption-garbled), and a core partner network was mentioned without a partner list.

Nvidia promotes RTX Spark laptops; Ziskind measures 65 tokens/s on GPT-OSS 120B

An NVIDIA promotional video said RTX Spark combines its latest Blackwell GPU with a power-efficient CPU and 128GB of memory for thin laptops and small PCs, and on Oct. 16 interviewees at a Surface event said the laptops offer 64 or 128 GB of unified memory for local AI agents, with no benchmarks. Alex Ziskind reported on Oct. 16, in a sponsored hands-on review of an MSI laptop, that GPT-OSS 120B ran at 65.4 tokens/s on Windows versus 50.6 on the DGX Spark under the same llama.cpp version at empty context, using 61 of 96 GB. vLLM gave 4,681 tokens/s prompt processing against about 1,800 for llama.cpp but 35 tokens/s single-user decode against 65. All results come from one unit and quantization was not stated.

MongoDB announces 9.0 with speed claims and opens Atlas Infinite preview

MongoDB said version 9.0 is up to 35% faster on find, up to 30% faster on update-one and up to 20% higher in transactional throughput than 8.0 on the same hardware, claims not independently tested; it also added substring search to queryable encryption and native OpenTelemetry export. On Oct. 16 MongoDB announced Atlas Infinite in preview in its largest AWS regions, with general availability expected next year and the current Atlas database renamed Atlas Core. It claimed, without stated methodology, more than 96% less scaling time than Atlas Core and up to twice the performance per dollar, and presenters said a 16 TB database can be restored in under 5 minutes in a scripted demo. Pricing was not given.

LiveKit says tuned Gemma 4 31B passes 88% of voice evals in 380 ms

LiveKit said at AI Engineer that a tuned Gemma 4 31B on LiveKit inference launches the next day, passes its voice evals at 88% and begins speaking in about 380 milliseconds, more than twice as fast as the next model; captions render the name inconsistently. LiveKit also open-sourced an end-to-end voice agent benchmark using a hotel receptionist agent and hundreds of simulated callers. It reported its audio-based turn detector cuts users off 4.5% of the time versus 9.9% for the best alternative at the same latency budget, and said popular text-to-speech providers ship hundreds of milliseconds of leading silence, worst case about 0.75 seconds. All figures are LiveKit's own.

Ziskind measures LMCache and vLLM CPU offload cutting repeat-query latency on DGX Station

Alex Ziskind measured on a DGX Station GB300 with Qwen 3 235B that a repeated 32,000-token query fell from about 1.35 s to 112-122 ms with LMCache or vLLM CPU offload. With 1.5 million tokens of documents overflowing GPU memory, vLLM's GPU cache had 0% hits, and LMCache survived a vLLM restart at 141 ms but held about 89 GB of GPU memory. In an agent-session replay caching gave 5.6x more turns per minute on DGX Spark and 7.5x on RTX Pro 6000; a hybrid Qwen 3.8 model broke vLLM CPU offload.

NVIDIA claims GB300 NVL72 gives up to 15x throughput per megawatt over H200

NVIDIA said OpenRouter data across 100 trillion tokens shows one agentic request uses about 15 times the tokens of chat. It said its agentic benchmark replays nearly 99,000 requests from production coding-agent traces, and on it GB300 NVL72 delivers up to 15 times the throughput per megawatt of H200 NVL8 and up to 10 times lower cost per million tokens. It claimed Vera Rubin NVL72 reaches up to 30 times GB300 throughput per megawatt on a DeepSeek V4 Pro workload at 160 tokens/s per user. All are vendor claims.

open local model

StepFun's Step 5 Preview free for a week; Bowen's tests find spatial weakness

StepFun's Step 5 Preview, a 600B-parameter mixture-of-experts model with 27B active parameters and a 1 million-token context, was released Oct. 8 and offered free for one week on OpenRouter and several coding tools, with open weights planned for Oct. 15, according to Julian Goldie and Bijan Bowen. Vendor-reported scores put it narrowly ahead of Kimmy K3 and GLM 5.3 on Deep SWE v1.1 (67.7%, 67.5, 66.9) and Program Bench (80.5%, 77.8, 72), and behind some rivals on Terminal Bench v4; Goldie relayed a Hermes Index score of 33.89 and a mid-pack Artificial Analysis position. Bowen, who priced it at $1 per million input and $2.70 per million output tokens, ran it through OpenRouter on high reasoning at a cost of a little over $22 and found it weak on spatial and geometry tasks, saying he saw no reason to choose it over DeepSeek V4.1 Flash on his suite. Goldie said Hermes Agent added the model and that he built four small business tools with it.

MiniMax engineers detail M3 sparse attention and native multimodal pretraining at AI Engineer

MiniMax said at an AI Engineer talk that its open-weight M3 combines a 1 million-token context window, native multimodality trained from step zero and frontier-level coding and agent capability. The company claimed MiniMax Sparse Attention gives 9x faster prefill and 15x faster decode than full attention, and that native multimodal pretraining barely hurts text performance compared with adding multimodal data after the decay stage. It said 3D attention in the vision encoder helps and four frames gave the best trade-off. The figures are MiniMax's own claims; the talk showed no independent measurement.

Perplexity releases pplx-embed-v2-late late-interaction embedding models under MIT license

Perplexity released pplx-embed-v2-late on Oct. 7, 2026 in 0.6B and 9B sizes, MIT licensed on Hugging Face, according to Julian Goldie. The ColBERT-style models produce 128-number vectors per piece and search text, images and PDF pages. Perplexity reported 74.8% (9B) and 73.6% (0.6B) on Q2Dweb against a prior best of 69.3%, 92.4% (9B) and 90.1% (0.6B) on MADQA, and 64.0% for the 9B on BrowseComp-Plus; these are vendor numbers relayed by the presenter. Perplexity also reported that indexing with the 9B and querying with the 0.6B beat the 0.6B alone, with text search up 1.6 points on average across six areas.

research

OpenAI's AI-generated math manuscripts stay unverified; commentators report three withdrawn

Commentators this week kept returning to the math manuscripts OpenAI published in early October, without independent confirmation that any proof is correct. Counts of the release differ across sources: Two Minute Papers said roughly 722, David Shapiro about 730, panelists on Oct. 12 over 700 or roughly 370, and AI Explained's narrator and OpenAI's Greg Brockman nearly 400, and the figures were not reconciled. On Oct. 13 AI Explained's narrator said not all results had been verified, including a 199-page proof of the quasi-Riemann hypothesis, and on Oct. 14 Two Minute Papers said three manuscripts were withdrawn after community feedback. Brockman said in a No Priors interview uploaded Oct. 15 that the math community is reviewing the problems, AI is not yet producing whole new theories and natural-language proofs are hard to check. Panelists relayed that almost 900 mathematicians signed a letter objecting, and Nate B Jones cited a Terence Tao lecture on black-box results, with no letter text or lecture details shown.

Google Research releases TabFM zero-shot tabular model with non-commercial weights

Google Research released TabFM, a model that predicts unlabeled rows from labeled ones in one forward pass without training on the user's data, with non-commercial weights, according to Prompt Engineering. TabFM-Auto wraps the frozen model with an LLM agent and reportedly ranks top on TabArena. In the speaker's test on a synthetic churn table, TabFM beat a tuned tree model but at a 0.60 score he did not call strong. The speaker said his agent trained on a leaky post-cancellation column in half of runs, which a time-restored database branch fixed.

Ravenwolf reports Terminal-Bench 2.0 runs: GPT 5.5 with Codex first, Fable refused 13 tasks

Benchmarker Ravenwolf said in his Terminal-Bench 2.0 runs (Terminus harness, five runs per configuration) that GPT 5.5 with Codex ranked first and Gemini 3.5 Flash came second, close to GPT 5.5 but using 1.65 billion tokens against 136 million and costing $620 against $497 at published prices. He said Fable had 13 refusals and cost about 11,000 to test, apparently dollars. The same GPT 5.5 cost almost 1,000 under Codex versus 282 under Cursor, per his figures. Raising Opus 4.6 to max thinking lowered its score, unlike GPT models. He argued single runs are unreliable, publishes best, worst and every-run solve rates, and advised pinning the OpenRouter provider because default routing can split a run across providers.

security incident

Brockman says OpenAI paused projects for 25% of production engineers to red-team infrastructure

OpenAI's Greg Brockman said in a No Priors interview uploaded Oct. 15 that OpenAI paused projects for about 25% of its production engineers while it used its peak-intelligence model to probe all of its infrastructure, fixed issues until the model's findings saturated, and expects to repeat the exercise with each new model. He said OpenAI is giving defenders access to peak-intelligence models through its Daybreak program and has a billion-dollar commitment to help frontline defenders, within a 'defender's window' before offensive capability spreads, and argued security spending should far exceed software spending, while admitting he might be wrong. The model, duration, number of issues found and the terms and spend of the commitment were not stated.

OpenAI says its models helped trace hidden recording device to employee; FBI arrest reported

Fireship, in a video of Oct. 15 reading what it described as a criminal complaint, said an OpenAI agents-platform engineer designed a roughly $23 Wi-Fi audio recorder with an internal agent and his wife's ChatGPT account, planted it in an executive conference room on Sept. 28, and was found about 48 hours later by a counter-surveillance sweep. Fireship said the engineer was fired, arrested and charged, a secondhand reading not confirmed in the video by OpenAI or court records. On Oct. 16, Mastra hosts read an OpenAI newsroom post saying a recording device was hidden in an office, that OpenAI used its models to analyse it, identified the employee who planted it days earlier and referred the matter to the FBI. The hosts added unverified details and doubted how much of the work was AI-driven.

Researcher shows H100 MIG timing side channel infers neighbouring model depth up to 97%

A Stealthium researcher said at AI Engineer that an attacker on a 1-compute-unit, 10 GB slice of an Nvidia H100 using Multi-Instance GPU (MIG) can infer the architecture of a model running on a neighbouring 3-unit, 48 GB slice by timing its own memory reads, with up to 97% accuracy on depth and up to 99.3% on decode speed, and without root access. Workload type was inferred in 31 of 40 half-second windows, but only with speculative decoding enabled. A variant over NVLink identified one of five models 79.2% of the time against a 20% baseline, with the attacker on a separate GPU in the same 8x H100 server. Prompts and outputs were not recoverable, per the speaker. The speaker, whose company sells GPU monitoring, said its product flagged the probe in under 600 ms; that detection claim was not independently measured.

Azure retrospective: wrong power breaker opened at a Southeast Asia data center

Azure said a maintenance step opened the wrong breaker and cut source A power at a Southeast Asia data center. Power returned in about 15 minutes, but about 11 network switches needed replacement and some services took 10-15 hours to recover, and the outage affected services beyond the zone, including Databricks control planes. Azure said it will place matching drawings in each panel and fix zone-aware alerting. This is Azure's own account.

Google paused open-source vulnerability rewards Oct. 1; PageBreak agent reportedly found ~500 flaws

An IBM Technology host said Google paused its open-source vulnerability rewards program on Oct. 1 after AI-generated reports full of hallucinations, with the open-source part frozen until Q1 2027. A panelist said Google's PageBreak agent found about 500 vulnerabilities in Google web apps, verified by deterministic non-AI checks with near 0% false positives. Both are secondhand panel statements.