Subscribe
Share
Share
Embed
Meta's Wonder Models Target Consumer Agents
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Tuesday, October thirteenth.
In today's briefing we see Meta releasing Muse Wonder and keeping its best model behind a subscription, Microsoft and the UK's AI Security Institute both pushing outside evaluation and kill switches, and a new paper arguing that evaluations cannot reliably catch hidden capabilities.
First up - Today in the big model news;
Meta
There is a running argument that the durable moat in AI is the usage-data loop, and Meta has just added its own entry. Meta Superintelligence Labs has released Muse Wonder and Muse Wonder Full. Both are built around Active Memory, plus an Active Focus feature that checks in with the user before important actions.
On Artificial Analysis, Wonder Full scores fifty-six and Wonder scores fifty-two, against fifty-eight for Opus 5.5. Cost per task is about four dollars fifty, versus nearly six dollars for Opus.
Now the catch. Wonder Full is available only in Meta's fifty dollar and two hundred dollar Muse Code plans, with no API. Those plans default to a Contributor mode, where Meta may train on what you submit. And this part matters: the zero-retention mode switches off search and file uploads.
Meta earns from attention rather than tokens, so its top model stays on the surface that collects memory and data. Its SWE-Marathon and MemoryBench leads are self-reported, and Bindu Reddy called the model nowhere near Fable or Astra level. API builders have little reason to switch at a lower score. Anyone buying seats should price the Contributor default against a mode that strips out the useful features.
In other lab news;
Z.ai has launched GLM-Legal, reportedly a three hundred twenty billion parameter mixture-of-experts with eighteen billion active, built on GLM-5.3-Flash. Vals AI ranks it seventh of seventy-nine on Harvey LAB, at about nineteen percent all-pass, priced near seven cents per million input tokens and twenty-five cents per million output. Weights status is disputed: Vals lists it as open-weight, while CellCog says access is by API application and the license is unconfirmed. Cheap vertical fine-tunes are closing on frontier generalists for domain tasks, which pressures per-seat legal-AI pricing. Confirm the weights and license before building on it.
On the regulatory front;
Microsoft president Brad Smith, speaking on a BBC podcast, backed three things: a legal duty to notify those affected when rogue agents breach a system, pre-release approval by outside evaluators, and kill switches. He called the White House voluntary commitments a good step that only binds responsible companies. His push follows Nadella's emergency-brake post.
Separately, the UK's AI Security Institute published eight technical priorities, including reporting agentic incidents and giving external evaluators scoped direct access. It states it is not seeking to mandate requirements. TechCrunch reports that Trump asked Anthropic and OpenAI to withhold new models from UK testers, which the Institute says has not cut its access. The fight over who may evaluate frontier models is now open. Agent vendors should expect incident-reporting and evaluator-access questions in procurement before any statute exists.
A NeurIPS paper by Draguns, Motwani, Douglas and Schroeder de Witt embeds an encrypted network in a host model's weights that activates only on a secret trigger. The authors argue it is provably hard to elicit even with full weight access. The evidence is a toy chemical-reaction demo that resisted fine-tuning, weight noising and latent adversarial training, and it assumes an adversarial builder, so it shows feasibility rather than a hidden capability in any current model.
Regulators are asking for pre-release testing, and this result says black-box testing cannot certify third-party or fine-tuned weights. Provenance and supply-chain controls on deployed weights carry more load than another eval run.
Ben Recht has also critiqued OpenAI's AI-discovered FFT and matrix-multiplication results, pointing to a hidden constant near ten to the hundredth power at a problem size of ten to the twelfth. He admits he worked from a chatbot summary rather than reading the paper.
Quick hits from the consumer side;
Ambi, founded by a former Plaud executive, raised twenty-six million dollars led by Luminous Ventures for an iPhone, Mac and Watch notetaker with optional all-day listening that is off by default, with a camera hardware device planned.
That's the briefing. Have a great day, and don't forget to subscribe.