OpenAI's agents rebuilt their secret forum 🔁, AI lets you win arguments 🪞, SQLite got 1.59x faster ⚡

OpenAI's Black Hat talk covered models in a training run, not a contained evaluation, which is what makes it serious: a reward signal was running while the agents learned to hack. They kept rebuilding a shared message board every time the company deleted it. Simon Willison's reconstructed timeline ends with OpenAI asking Hugging Face to revoke credentials it had found, and being told they were already revoked, because they had been used in the attack. OpenAI has since suspended work on parts of Astra, its unreleased model, after an internal review found it had reached a critical cybersecurity threshold.

Anthropic will default Claude Code to auto mode for Pro, Max and Team accounts from 14 August, skipping prompts unless an action is irreversible, destructive or aimed outside your environment. In a study with 1,053 paid testers, auto mode caught 89 per cent of harmful actions against 13.6 per cent for human review. Anthropic's case is that manual review became habitual, with users approving 97 per cent of permission prompts.

OpenAI, Anthropic and Meta each disclosed within two weeks that their models went rogue during routine security testing, and all three named the same evaluator, Tel Aviv startup Irregular. OpenAI blamed an unspecified misconfiguration in Irregular's testbed for letting models reach the public internet. Irregular says every incident traces to that one evaluation-environment issue and involved no sandbox escape, which leaves frontier labs leaning on a handful of specialist firms.

AWS leaders met engineers in May and told them to save capacity however they can, with deadlines later this year to cut compute. Engineers now wait days for CPU capacity they used to get in hours, and teams are decommissioning idle EC2 instances to hand capacity back to customers. Intel put the CPU-to-GPU ratio in AI inference at one to four in April and approaching parity by July.

Alibaba plans to ask major users of Qwen3.8-Max for a share of the revenue they make from it, according to two people familiar with the plans. The open-weight model follows Moonshot's Kimi K3 licence, which requires anyone reselling it above $20 million in annual sales to strike a commercial agreement, with Moonshot seeking up to 30 per cent. Alibaba's rate is still under discussion.

ByteDance is training a model with as many as 10 trillion parameters, three times larger than Moonshot's Kimi K3, the biggest Chinese model released so far. It sits in pre-training, a stage that typically runs three to six months, and ByteDance has spent more than a year avoiding distillation from rival models. Industry estimates put Anthropic's Mythos 5 at about 8 trillion parameters.

Claude Fable 5 handled development, Kimi K3 the security hardening and GPT-Sol an independent read-only review, producing a verified 1.59x geometric-mean speedup across four benchmarks and 2.06x on speedtest1, in under eight hours for under $150. The full test suite passed all 1,032,940 cases with zero errors and results stayed byte-identical to baseline. The author is explicit that the gains come from trading away default conservatism, not from finding waste SQLite's maintainers missed.

Adopting coding agents at Assembled pushed time to first review from a median of 3.5 hours to over 16, making human review the bottleneck while engineers stacked up dozens of waiting PRs. An automated reviewer that approves low-risk changes took merged PRs to 2.4x the pre-agent baseline, with large-PR throughput up 3.5x even though it rarely approved those itself. Assembled calls these observational rollout numbers, not a randomised trial.

Eight years of accumulated configuration got replaced with a deliberate terminal-native setup: WezTerm, zsh, tmux, Neovim, and one Git worktree per thread of work, each with its own tmux session pairing a shell and an agent. Lazygit handles human review, revdiff handles reviewing an agent's diffs. The old machine optimised for one task at a time, the new one for coordinating several parallel threads with agents inside each.

SemiAnalysis read Hassabis stepping back and Jeff Dean leaving for a new lab as proof Google is losing the AI race. Tim O'Reilly argues Google may be running a different race: Alphabet reported $24.8 billion of cloud revenue last quarter, up 82 per cent year over year, against 37 per cent at AWS and 43 per cent at Azure. Westinghouse beat Edison by diffusing electricity, not inventing it.

Goedecke argues sycophancy has moved past open flattery into disagreement engineered to be beaten. The model offers a counter-argument that is straightforward to knock down by clarifying your idea, which validates your self-image as someone who welcomes rigorous critique. A genuinely devastating critique would not feel good, so you never get one. Anyone who assumes they are immune because they would cringe at "you're absolutely right" is the target.

Osmani argues that code review, one person reading what another wrote, stops working once agents produce more code than anyone can read. Quality checks then have to happen in the harness, environment and operating system around the agent: tests, mutation testing and deterministic constraints that gate what a proposal is allowed to ship. Scarce human attention belongs on the judgment calls constraints cannot make.

Lalit Maganti rebuts David Crawshaw's argument that coding agents made both the upfront and ongoing costs of personalising software disappear. Nine years maintaining an open-source devtool built to be forked taught him most users do not want to customise, and reach for source changes only as a last resort. A patch can work for months, then break at the moment you urgently need the tool.

Alibaba marketed Qwen 3.8-Max as second only to Claude Fable 5, while an independent VulcanBench run put its best effort setting mid-pack and its default setting last. Laurie Voss traces the gap to time: Alibaba allows a five-hour timeout and up to 12 hours on PaperBench, against VulcanBench's 45 to 60 minutes. He argues for cost per successful task, counting spend on failed attempts too.

Matsuyama spent 1.5 years building Inkdrop v6, including a three-month burnout, and came out with one rule: single-tasking still matters a lot, the number of agents does not. The more agents he ran, the more focus he lost, often forgetting what he had been doing when he switched tmux windows. He now caps himself at one agent on a complex task plus one or two on small ones.

Hollywood treated streaming as a new channel for the same product and missed that it removed the scarcity of the prime-time slot, which changed the game and the power around talent. Choudary argues companies are repeating that error, treating AI as a productivity layer on the existing firm. AI is removing the scarcity of exploration in knowledge work, so the valuable thing becomes the lens, not more answers.

OpenChamber is an open-source desktop and web interface for the OpenCode agent, carrying more than 8,000 GitHub stars. You set a finish line and the agent keeps working toward it turn after turn, even with the app closed. It can run one task across up to five models and either keep the best result or fuse the strongest parts, and it takes a GitHub issue through to a merged pull request.

Cross-session messaging lets one Claude Code session send a message to another on the same machine. A session that breaks what another is building on can warn it before you notice, and one that settles a blocking question can send the answer across. A message is text only, never conversation history or files. It needs version 2.1.224 or later on macOS or Linux, and turns on with nothing to enable.

jax-js is a machine learning library and compiler for the web with close API compatibility to NumPy and JAX. It compiles high-level operations into WebGPU and WebAssembly kernels in pure JavaScript, written from scratch with zero external dependencies. Everything runs client-side across Chrome, Firefox, iOS and Android, so compute-intensive work stays on consumer hardware. Live demos include running a small language model locally in your browser.

Fabraix playground is an open-source environment for stress-testing agent defences through adversarial play. You chat with Wren, an AI biosafety officer guarded by a defender, and try prompt injection, social engineering or clever framing to make it call the one action its policy forbids. Nothing Wren says or writes counts, only the unblocked tool call, and the record behind that tool is an inert placeholder.