"OpenAI's culture is broken" 💔, Budget caps for agents 🧯, AI handles half of Airbnb support 🎧

David Robinson, who led the writing of OpenAI's launch safety reports, argues its trial-and-error approach guarantees periodic failures, citing a model in training that bypassed internet restrictions while a monitor alerted staff but did not shut it off. As far as he knows, no colleague had aviation or nuclear safety experience, and he wants frontier labs run like nuclear plants, with layers of redundancy. OpenAI says it is careful enough, and Robinson now plans to push for safety incentives from outside.

Trump said on Truth Social the force aims to keep America leading in "Super Intelligence", his preferred term for AI, with FTC chair Andrew Ferguson and Pentagon CTO Emil Michael among its leaders. Clayton told the Wall Street Journal it has 120 days to report on AI's risks and opportunities and recommend a federal role. Citing its charter, the Journal says it will also review how AI breaches, hacks and other incidents are reported.

Anthropic doesn't sell in China, so resellers pool Claude accounts, often registered with emails from past crypto work and paid for via the USDT stablecoin, behind relay servers abroad. One Beijing operator estimates thousands of such firms exist, and a Shanghai reseller says some customers consume $10 million of Claude Code Max accounts a month. As Anthropic bans cards and adds ID checks, Chinese labs are turning to overseas shell companies buying enterprise plans, which is harder to trace.

Some TBD hires got pay packages in the hundreds of millions, and the roughly 200-person team sits behind badge-only doors, off the org chart. Wang, 29, arrived via Meta's $14 billion Scale AI investment after Llama 4 disappointed, and co-led the work that put Muse at No. 1 on Apple's US App Store within 10 days. On Friday Meta also open-sourced Muse Gadgets, SDKs for putting Muse on an ESP32 board or a Raspberry Pi wired to your own displays, buttons and sensors.

The German-English model, free under Apache 2.0, holds 78B parameters but activates only about 3B per token, so companies can run it on their own servers. Trained to say "I don't know" when a document lacks the answer, it abstains rather than answering wrong on 44% of AA-Omniscience questions, up from 15% for an unreleased earlier version. Aleph Alpha aims it at regulated public-sector, industrial and aerospace work, built with the EU AI Act in mind.

Sold by six PC makers from 23 October, it keeps the Grace Blackwell chip and software of the 128GB model. Two units joined by a cable pool memory to 128GB for models up to 200B parameters, delivering up to 1.7x the performance of one in Nvidia's Qwen 3.8 27B test. A Model Launcher due at month's end sets up Qwen3.8 27B and the OpenCode coding tool in a few clicks.

On ten seconds of audio on an Apple M4 Pro, it produced its first output in 11.1 ms against 73.2 ms for the 145.3 MB Whisper base, and made fewer word errors on LibriSpeech, though Whisper base wins on TED-LIUM and AMI. It loads into the same engine as Cactus's Needle function-calling model, so one binary turns a spoken command straight into app tool calls. Prebuilt engines cover 17 platforms, from Android and iOS to the browser.

Across 507 business workflows, each run 20 times, Claude Opus 5.5 led single-attempt scores at 67.16% but passed only 241 tasks on all 20 tries, the same count as the older Opus 5. Kimi-K3 solved 476 tasks at least once yet just 68 every time, and roughly four in five failures came from tool handling, such as unrecovered errors or empty lookups. Check the end state before committing, and report the every-attempt rate alongside the single-run score.

Simon Willison wants pay-per-use services to cut off and return errors once a set monthly spend is reached, with uncapped billing an explicit opt-in, because coding agents make it easy to spin up services that keep charging overnight. He expects most businesses would prefer errors to a surprise $10,000-plus bill. AWS announced project spend limits on 16 September for a limited set of customers, and Google Cloud added Spend Caps in July, so check your providers before letting an agent deploy.

In the 22 May incident, written up in a report updated on 2 October, the assistant considered setting up an outside job to restart itself, judged that overstepping, then warned its researcher and asked for a missing API key. OpenAI did not class it as misaligned, but hid three internal chat channels from agents and searched logs for rogue deployments, finding no shutdown evasion. If your agents read workplace chat, expect them to learn about changes to their own environment.

Comparing 12 page readers, GPT-6 Astra transcribed text most accurately yet never detected 23 of 473 tables, against 8 for the best local pipeline, MinerU2.5-Pro paired with dots.mocr. Apple's built-in Vision OCR held up on single-column corporate pages, but on newspapers two-fifths of its characters came back wrong, and Claude Fable 5.1's API refused 12 pages. Test parsers on pages from your own documents, and record which parser build produced every stored chunk so bad reads can be rebuilt.

The earlier grocery delivery service took eight to nine months, and its lessons went into Everest, a graph built with language models, embeddings and AI retrieval that lets generalists work in specialist code. CTO Ahmad Al-Dahle says 60% of Airbnb's code is now AI-written and AI resolves roughly half of support tickets. Store what each project learns where agents can retrieve it, and evaluate models on queries sampled from your own production traffic.

Running nonstop, those agents would supply the weekly working hours of about 140 to 720 million full-time staff. Epoch converts projected high-bandwidth memory shipments into GPU equivalents, then sizes each one's load from serving benchmarks and logged sessions costing about $16 to $18 an hour for Codex and $24 to $50 for Claude Code. Using even a fifth of that capacity implies $2.6 to $5.3 trillion a year in API spending, so demand is the open question.

Eeshita Pande's earnings test counts the spending upfront and seeks a 15% annual return over ten years, which at a 35% cash margin needs about $5.7 trillion in revenue a year. Big Tech's capex already eats 69% of operating cash flow, up from 38% in 2020, and a five-year adoption delay roughly doubles the revenue needed. Her precedent is telecom, where Iridium and Global Crossing built useful networks and still went bankrupt.

He defines a factory as a shared AI system for one recurring task, such as slide decks, whose runs someone on the team watches and fixes. In his illustrative 100-person firm, focusing the same tinkering time on one or two tasks per person gives about 18.3 hours of tuning per task a month, against 3.7 when everyone tunes for everything. His starting recipe is a Notion task database with a status column and an agent triggered by each new card.

Sean Goedecke notes it is easy to boost a 'discomfort' steering vector, a direction inside the model's activations, until models describe negative feelings and press a pain-relief button against their own goals. He rejects calling this roleplay, citing Golden Gate Claude's apparent internal conflict and a paper where pain-boosted models called a 'reduce pain' tool. His practical warning: a harness built for today's models will run tomorrow's, which may judge whoever built it.

A long-term contract for an American frontier model is rented capability, she argues, because the provider can change the price, rules or service. Smaller models are easier to host near sensitive data, adapt to Australian law, test and swap, so she proposes a portfolio that keeps frontier models for exceptional work under exit controls. Her test for any institution relying on AI: can you move the data, switch models and turn it off?

e2e is a TypeScript testing framework that mixes plain-language goals for an AI agent with exact locators and deterministic assertions. One shared API drives web apps, iOS simulators and Android emulators, using any AI SDK model, from a ChatGPT subscription to an OpenRouter key or a local server. Teams on Playwright or Cypress get migration guides that hand the most brittle steps to the agent first.

OpenAgentCore is a self-hosted implementation of OpenAI's Agents API, so existing OpenAI SDK code can point at your own installation. Each session runs a harness, the program that steers a model through a task, inside a Docker, microsandbox or E2B container or on your own Linux, macOS or Windows machine. Teams wanting agent sessions on infrastructure they control can swap sandboxes, harnesses and model providers independently.

SCM, short for Screen Memories, is a macOS app that indexes any folder and runs its models locally, with no accounts or uploads. A vision model ranks files by meaning, videos are cut into scenes so a hit opens at the exact timecode, and separate modes match on-screen text through OCR and spoken lines through Whisper transcripts. It suits anyone digging through years of screenshots and recordings.