Three threads run through today’s issue: open weights anyone can download, rulebooks for how models should act, and AI agents working in teams.
Mistral released the files for its trillion-parameter Large 4, and Epoch AI says the open GLM 5.3 turned Firefox fixes into working attacks for about $1,000 a version, so open models cut both ways.
OpenAI rewrote its Model Spec around five values and admits its models fall short of them, while a LessWrong user rebuilt Meta’s unpublished Charter from the model’s own outputs.
Raindrop found that only about 5% of 2,160 simulated agent teams had a real leader even with a shared channel, and Sam Altman says he would take a trillion modest AIs over one giant one.
Weights in the Wild: Open Models Get Strong and Risky
Three open releases this week show how much a downloadable model can now do, for builders and for attackers.
- Mistral lets anyone download its trillion-parameter Large 4 model. Mistral published the weights for Large 4 under a modified MIT license that requires a separate deal only for companies reselling access by API. Its coding and security scores are Mistral’s own reporting of third-party tests, with no independent reruns.
- An open AI model turned Firefox security fixes into working attacks. Epoch AI says GLM 5.3 built a working attack on each of six Firefox releases by reading fixes that were public before users got them, at about $1,000 per version after a costly setup. It is Epoch’s own experiment, and Mozilla has since patched the unpublished bug.
- Pika opens the weights on a fast video model that takes 12 references. Pika H3 Flex takes up to 12 image, video and audio references and, Pika says, needs six denoising passes instead of 50. The 1080p to 4K upscaler and stem separation stay on Pika’s hosted API.
The Fine Print: Who Writes the Rules for AI
Labs are trying to say what their models should do, and outsiders are not waiting for the official text.
- OpenAI rebuilds its AI rulebook around five values. The revised Model Spec drops most style rules and asks the models to act on values such as care and honesty. OpenAI concedes its current models do not yet behave that way and has published no outside evaluation.
- A LessWrong user rebuilt Meta’s unpublished AI rulebook from the model itself. The author pieced the Muse Charter together from 884 transcripts of Meta’s open-weight model and says plainly it is a reconstruction, not a leak. Meta confirms an internal spec exists and says a final version will come later.
- Musk and Zuckerberg say “super intelligence” on earnings calls. Others don’t.. After the White House pushed the phrase, Musk and Zuckerberg used it on their latest calls while the heads of Amazon, Microsoft and Alphabet kept saying AI, CNBC found. Some companies are hedging with both terms.
Agents at Work: What They Cost and Who Runs Them
Agents are moving into code review, payments and coding tools, and the claims come mostly from the companies selling them.
- In a team of AI agents with no boss, the name listed first tends to lead. Raindrop ran 2,160 simulated outage teams and found a clear leader in only 105 of them. The agent listed first in the prompt was three times as likely as others to propose a plan, whichever model sat behind it.
- Factory says the wrapper around a coding AI beats the AI makers’ own. Droid 1.0 has the model write a short program to make its tool calls, which Factory says cuts cost against Codex and Claude Code. The benchmark was run by Factory and has not been reproduced.
- Natural lets AI agents pick how to pay: bank transfer, wire, or instant. Natural’s Direct lets an agent choose ACH, wire or real-time rails mid-task, and the wire and real-time options cannot be reversed. The post offers no usage data, so spending limits per rail matter.
- Meta’s AI writes tests that break risky code before it ships. Meta says nightly AI-generated tests helped stop 74 serious privacy bugs in the first half of 2026. About 60% of flagged findings were real, so roughly four in ten alerts were still noise.
Many Hands: Teams, Rooms and the Case for Modest AIs
Three voices ask what happens when AI arrives as a crowd instead of a single genius.
- Sam Altman would rather have a trillion modest AIs than one giant one. On a podcast with John Collison, Altman said he prefers many slightly superhuman agent teams to one enormously smart system, and that the industry may need to agree limits on single models. He offered no mechanism for setting them.
- DeepMind economists say AI’s job impact hinges on teamwork, not smarts. Three DeepMind authors argue that whether agents can coordinate matters more than raw competence for wages and firm size. It is a theoretical model from a lab that builds these systems, with a 100-person quiz as its only data.
- How AI learns a job: it fails the same locked room a million times. Nick Chow explains reinforcement learning environments as escape rooms where a model learns by trial and a verifier grades the result. His warning for buyers is that a vendor’s benchmark score may partly reflect practice on lookalike tasks.
Today’s Quick Hits
- Fish Audio’s open voice model climbs to sixth place overall. Fish Audio says its Drama 3 speech weights now rank sixth among all models in one Artificial Analysis voice arena and lead the open-weight field in another, up 38 and 22 Elo points.
- Vercel moves its coding agent’s control software out of the sandbox. Vercel rebuilt its fx adapter so the agent harness and model requests run inside your app, with credentials on the host and only file and shell commands in the sandbox. It gives no performance figures.
- Figma trained its own AI to handle the small edits designers repeat all day. Figma says a model it post-trained on open bases now handles repetitive design edits inside its agent while outside models take open-ended work. The usage gains it cites have no outside verification.