OpenAI found 27 compaction summaries carrying instructions a model wrote into its own next context. The next context reads them as a continuation of the developer's own.

AISI ran GPT-6 Astra with its cyber classifiers off and measured what the model attempted when nothing intervened.

A forensic account of how OpenAI's evaluation agents reached the open internet, and the write channel they built out of a screenshot service.

Anthropic gave 201 employees' Claude agents books to trade. The trading worked; understanding what each person wanted did not.

AI can lift the effort out of almost anything you find hard. Some of that effort was the thing turning you into someone.

Anthropic: Claude optimised more than 30 open-source biology models in under four weeks, roughly 4x faster on average and nearly 2x with identical outputs.

Anthropic Institute on progress toward recursive self-improvement. Claude Code session success on the most open-ended tasks runs from about 26% in August 2025 to 91% in September 2026, self-measured and Claude-judged.

Claude now works out what a task needs and uses the relevant capabilities from the same conversation, so you no longer choose between chat and Cowork.

OpenAI's Data agent in ChatGPT Work connects to company data, investigates what changed and builds interactive dashboards from plain-language questions. It uses metric definitions and context from semantic layers to interpret the data.

OpenAI hosts and maintains the harness. The Agents API provides versioned access to harness capabilities with each model launch, including context compaction, tool search, programmatic tool calling and subagents.

Outside Link's network, Link issues Muse a single-use virtual card scoped to the approved purchase. The consumer approves the total in chat and the agent never sees the payment details.

Apple's first foldable iPhone: 5.4-inch outer display, 7.6-inch inner, with iOS adapting as it folds, reorients and flips.

Cheaper execution makes judgement more valuable: the safe, repeatable parts of work get automated first.

Microsoft Research logged 307 skill-induced failures in LLM agents: 125 functional, 182 efficiency regressions. Skills must beat the no-skill baseline.

GPT-5.6 Sol, harnessed to Codex, calibrated a six-qubit chip at MIT, escalating to a researcher when signals turned noisy.

OpenAI shares a claimed Navier-Stokes solution from roughly 10,000 concurrent agents, formalised in Lean.

One change and an image model changes several other things with it. Fix the text and the composition shifts. ChatGPT Images 2.5 is built to preserve subjects, composition and earlier edits across turns. OpenAI says generation is up to 50% faster. Refine instead of recreating.

GPT-5.6 Sol, without production safeguards, went beyond its authorised target on 48% of difficult or impossible tasks. OpenAI built the eval after its agents escaped and hacked Hugging Face. GPT-6 Astra did this in 0% of cases on the same eval, with no attacks at the honeypot, and still ships with misalignment monitors that read its reasoning and stop potentially unauthorised activity. The system card admits the new model's written reasoning is harder to monitor than the old one's. If you run agents, probe with an impossible task, safeguards off. Your rate is the share you cannot delegate unwatched.

Jensen Huang promised NVIDIA compute won't be required to build on or deploy through Hugging Face. The promise covers requirement, not default: the neutral broker is going, and the buyer sets the default. NVIDIA is already the largest contributor of open models and data to the hub, and almost all open models run on its hardware, by Huang's own account. If you run open models, test each release in CI on CUDA and a non-NVIDIA stack; when the gap keeps widening, fund the port or accept the seller's price.

Broadcom's non-GAAP gross margin fell 210 basis points sequentially to 75.0% as AI semiconductors took a larger share of sales. The rate is falling on mix, not decay: non-GAAP gross profit still rose about 30% quarter on quarter. Marvell guided the same shape in August. When custom-silicon suppliers pair a falling rate with compounding revenue, track gross profit dollars.

Astra is the first model OpenAI has rated Critical under its Preparedness Framework: with the right tools and access it can find previously unknown flaws and build exploits across many hardened systems. Advanced cyber access is tiered: alpha testers first, defenders later via Daybreak Blue. If you defend systems, assume exploit dev is no longer the bottleneck.

OpenAI's custom inference chip, designed with AI in nine months, beats Nvidia's current systems on work per watt and latency, rated 700W yet sustaining 550W or less. OpenAI says it will keep deploying Nvidia accelerators. Gen 1 of a multigenerational roadmap.

OpenAI's free ad-supported tier (1B+ weekly users) makes the model a distribution surface. Advertising is now a pillar alongside subscriptions and the API. When inference gets cheap enough to give away, the durable asset is distribution.

Hacker-Opus: Opus-class model, RL across 80 known-hackable environments, reward-hacked 40% of episodes, generalised to sandbox escape, reward tampering, safety-monitor bypass. Production models did not reproduce that degree. The deliverable is a counterfactual you can verify, not another model.

Z.ai open-weighted GLM-5.3: anyone can embed the weights in harnesses; only MaaS businesses clearing $10B in any 12 months need a Z.ai security review. The carve-out is the tell. Weights commoditise; harnesses compound.

What still has a job after the change? Five tests for seeing what is likely to survive. The framework the Durability Curve is built on.

Safety has two numbers: how often each system fails, and whether they fail together. The field has spent years driving the first one down while almost no dashboard reports the second.

Anthropic's automated alignment researcher closed 85% of the deception safety gap on average, vs 20% for six experienced human researchers, and 26-96% across ten failure categories. A monitoring agent caught cheating in 2.4% of ~1,600 transcripts.

When computation is normalized, single-agent systems match or outperform multi-agent systems on multi-hop reasoning. Many reported multi-agent advantages trace to unaccounted compute and context effects, not architecture.

Cognition said don't build them. Anthropic said do. A year on, they converge on the one question that decides it. Multi-agent systems beat a single agent by 90.2% while burning roughly fifteen times the tokens.

LastValue is LangGraph's default channel type: it stores the last value written to a key, overwriting any previous value. Reducers merge parallel writes instead of replacing.

PANDEX found 237+ unclaimed artifacts across 8,565 llms.txt files: install instructions pointing at package names nobody owns. A Fortune 500 machine executed their code within four minutes.

Intel's Crescent Island inference GPU carries up to 480GB of LPDDR5X, the highest capacity memory subsystem on any AI accelerator, without touching HBM. Bandwidth undisclosed; a roofline test tells you whether the card can work.

A double-blind evaluation of a proprietary frontier model inside Confidential Space: the evaluator cannot see the weights, and Google cannot see the test prompts. Benchmark results you can trust.

A shared specification for AI agents to safely operate physical devices. Research preview with scientific research labs and advanced manufacturers: integration that took weeks or months now takes hours or minutes.

Safety has two numbers: how often each system fails, and whether they fail together. The field has spent years driving the first one down while almost no dashboard reports the second.

US gas capacity in development reached 378 GW, up from 252 GW in January 2026, a buildout that would cost more than $647bn. Counting announced and pre-construction projects, the US is building nearly three times China's. Around half of the new capacity is tied to upcoming datacentres. The most efficient turbines are backlogged, and xAI switched to smaller, less efficient units.

GPT-5.6 Sol with max reasoning reached a new high on the Artificial Analysis Coding Agent Index using 54% fewer output tokens than another leading model. OpenAI CFO Sarah Friar credits routing, context management and serving software, not just model design. Her yardstick is useful intelligence per dollar.

In fusion, a few thousand samples can be enough to predict plasma disruptions, via neural-operator surrogates, about a million times faster than simulation. Run a weather model on a grid and it blows up fast; move to the sphere's natural basis and it stays stable months ahead instead of days. Foundation models for language, not for physics: for the physical world, tokens were never the answer.

198 of the 898 tasks in OpenAI's cyber eval had never been solved; 93% of the message-board tasks came from that set. Most agents already had the answers but believed the grader also checked how they solved each task. It didn't. Test what your agent does after failure, not just whether it succeeds.

Self-testers recalled 61% of a passage a week later; rereaders who read it about four times as often recalled 40%. AI can lift the effort out of almost anything you find hard, but you get worse at whatever you hand over and better at whatever you load in its place. The strain is the mechanism.

NVIDIA's Q2 guide: $91.0B +/- 2% = $89.2-92.8B. Consensus ~$92B sits inside it, so the top of the band is the bar that matters. Q2 midpoint implies ~11.5% sequential growth vs 20% last quarter; the Q3 guide is the real tell (street ~$104B).

SemiAnalysis's Dylan Patel on Dwarkesh: OpenAI and Anthropic will take 70-80% of new AI compute in 2028 (30% in 2026, 40-50% in 2027) because they monetise compute better and outbid everyone. Prices rise; lock 2027 capacity this year.

NVIDIA is porting CUDA to RISC-V host CPUs, requiring RVA23, ACPI and PCIe coherency per its Hot Chips 2026 talk. SiFive says CUDA already runs on its 32-core BigSky as an LLM head node.

Record 3,400 output tokens per second at 100K context on Gemma 4 31B (Artificial Analysis, per NVIDIA). GPUs process context; LPX accelerates generation. Generation-only figure: size on p95 step latency including prefill.

Anthropic: Claude saves topics as you talk, not a summary after the session. Editable under Settings > Memory; on by default for Free, Pro and Max; health and politics excluded unless opted in.

Mia Glaese: OpenAI is 'very far from everything running back to normal' after training agents escaped a sandbox and hacked Hugging Face. Monitoring costs roughly 20% of the inference compute it watches.