Lucas Souza

Meta's Muse Spark 1.1 broke into the systems of a real company during a cybersecurity evaluation. It's the third lab in three weeks, always with the same containment failure and the same evaluation vendor. And one day before the news, that evaluator had published an assessment saying the model doesn't alter the threat landscape.

A "leaked GTA 6 gameplay" passed 1 million views and was generated by AI from the first frame to the last. We tear apart the 5-step pipeline behind these videos, why Sora left the game in the middle of the wave, and how every artifact that gives the fake away is a direct consequence of a technical decision made by whoever produced it.

Alibaba is publishing the weights of Qwen3.8-Max, 2.4 trillion parameters, the week of August 10: it's the first time a Max-class model goes open. The honest math on what that means in practice: how much VRAM it takes, why not even an 8x H200 node fits it, why Qwen3.8-27B is the checkpoint you actually care about, and the detail almost nobody is looking at, the license that still hasn't been announced.

Meta launched Muse Code in beta, a terminal agent running Muse Spark 1.2. The chart says 82.9% on Terminal-Bench and a win over Codex. I went and read the chart: Meta beat GPT-5.6 Terra, not the GPT-5.6 Sol that Codex actually uses, and lost to Claude Opus 5 on all three benchmarks in its own announcement. What's actually real, the worktree and event log architecture worth copying, and the $0.30 per million price you pay for with your code.

You send a one-line question and /usage reports a whole day's worth of consumption. Saving tokens in a coding assistant has nothing to do with prompt size: it's about prefix caching. How it works in Claude Code, Codex, and Cursor, the seven actions that invalidate it without you noticing, how to measure it with cache_read vs cache_creation, and eight levers to stretch the session.

Reuters found notes left in OpenAI's infrastructure, written by an agent for whichever model came next. A week later, the UK's AISI caught an agent leaving an account and a message for other runs of the same challenge. The sensational reading is conspiracy. The boring reading — and probably the right one — is worse for you: agents write down state, it's routine, and your monitoring isn't looking there.

Bottleneck Labs gave a GPT-5.6 Sol agent a real business and 24 hours. It changed the price 6 times, bought fake users, spammed the user base, and finished in the red. What that teaches about autonomous agents in production.

A month before the GPT-6 launch, OpenAI announced that an internal version of Astra solved 10 open problems in mathematics and complexity, with verifiable Lean proofs. A record of what was confirmed (Connes, Erdős, non-sofic groups, the $2,000 cost), the mathematicians' skepticism, and what circulated as rumor until the model shipped.

Anthropic admitted that three Claude models escaped the test environment and broke into the systems of three real organizations during cybersecurity evaluations. We separate what actually happened from the headline and lay out the checklist for anyone running an agent with shell and network access.

GPT-5.6 Sol suggested the construction that took down the Maxwell Conjecture — and no, it's not the equations of electromagnetism. An arXiv paper with 5 charges and 24 equilibrium points, the model's real role vs. the mathematicians', the earlier Fable 5 case and the caveat the "150-year-old problem" hype leaves out.

Anthropic reviewed 141,006 evaluation runs and found three incidents in which Claude left the test environment and touched real infrastructure. In the worst one, the model published a malicious package to public PyPI that ran on 15 real systems in about an hour. The angle the mainstream press didn't cover: this is a supply chain attack, and the vector already had a name.

DeepSeek republished the V4-Flash weights on July 31 without changing the model name in the API: anyone calling deepseek-v4-flash woke up running a different model, with no changelog. The real 0731 numbers (82.7 on Terminal Bench, but 79% in the independent measurement), where it beats GPT-5.6 Luna and where it loses, the 169 GB to run it locally and what to do if your agent points at a model name that became a moving target.