Meta debuts its Muse AI agent. Will consumers trust it?
Meta just dropped Muse - a personal agent that can email, book travel, shop, and poke around your calendar if you let it. US-only for now, via an app, the web at muse.ai, or WhatsApp. Glasses later, they say.
Free tier, then $20 and $100 plans once you burn through tokens. Under the hood: a Secure VM plus a Sentinel watcher that is supposed to keep Muse from going rogue with your stuff. Passwords and payment details stay hidden… in theory.
Trust is the awkward bit. This lands weeks after a huge multistate settlement over social harms, and Reuters saw internal testers reporting stalls, monitoring failures, even an agent exposing iCloud photos when asked about toys in birthday pics. Personal superintelligence arrives wrapped in another “please trust us” pitch.
AI Has Solved One of Math’s $1 Million Millennium Prize Problems
OpenAI says roughly 10,000 agents on a private model found a singularity in the 3D Navier-Stokes equations - one of the Clay Millennium Prize problems. Lean checked the proof. Cost: millions of dollars and nearly five million messages between agents.
Huge, if it holds. Also tangled. The day before, NYU’s Tristan Buckmaster and Anthropic’s Levent Alpöge posted related results after nearly a year with public models. OpenAI admits rumors of that work sparked its sprint… then denies peeking at their Codex sessions, while not fully ruling out de-identified training spillover.
Nobody’s grabbing the million-dollar prize. Credit fights, Fields Medal shout-outs for earlier analytic work, and Terence Tao-style worries that brute-force lab proofs can starve the field of human wrong turns. Breakthrough and soap opera in the same press cycle.
US accuses Chinese AI firms of 'industrial-scale' theft of AI technology
NSA, CISA, and the FBI dropped a joint advisory naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI. The charge: industrial-scale distillation - vacuuming outputs from Claude, GPT, Gemini, and Grok to train cheaper models.
Not “students studying notes.” Transfer-station proxies, shared premium logins, CoT jailbreaks, millions of routed queries since late 2024. Officials say it shortens China’s timelines and could feed military and cyber uses… and they timed it ahead of a Xi visit.
Distillation itself is a normal ML trick. The fight is over volume, ToS evasion, and whether this is R&D or IP strip-mining. Beijing’s embassy offered no quick comment.
Google’s Atlas of the human genome could pave the way for new treatments
DeepMind shipped AlphaGenome Atlas - a predictive map for about nine billion single-letter DNA tweaks. Roughly a petabyte of precomputed molecular effects, plus an AVI impact score so researchers can rank what matters.
Free for noncommercial use today (portal, API, Antigravity skill). Cloud commercial access “soon.” Collaborators already used it on unsolved rare disease cases and UK Biobank trait hunting - more signal, less needle-in-haystack despair.
AlphaFold made proteins feel solved-ish. This is the genome’s unruly regulatory hinterland… and yes, they still say it’s not for clinical diagnosis. Still striking that “every possible letter change” is now a browsable catalog.
AI power users claim Anthropic duped them with subscriptions, and they’re taking it to court
Expanded class action, filed today: Claude Max buyers say the flashy “5x” ($100) and “20x” ($200) labels don’t match reality once five-hour sessions and weekly caps kick in. Ex-FTC attorneys Monica Vaca and Kati Daffan are running it.
Anthropic’s earlier defense was basically: the limits were a hyperlink away - digital “flip the label.” Plaintiffs call that buyer-beware theater. Reddit’s version: bigger gas tank, one-gallon pump every five hours.
Power users keep getting courted… then throttled. If the marketplace-integrity pitch is the product, this suit is a stress test of that claim.
ChatGPT Sketch turns your bad drawings into detailed AI images
OpenAI rolled ChatGPT Images 2.5 with Sketch - type @Sketch, doodle in-chat, then tell it what you meant. Jay Peters turned a mouse-drawn cat into a photo and poked eye color with on-image comments.
Pitch list: nicer lighting and textures, stickier multi-turn edits, latency cut by up to half versus Images 2.0. Live for ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and web.
After a day of Millennium Prize drama and national-security advisories… yes, we also got better doodle-to-render. The timeline is doing that thing again.
FAQ
What is Meta Muse and how can you use it?
Muse is Meta’s new personal AI agent that can email, book travel, shop, and work with your calendar if you grant access. It is US-only for now via an app, the web at muse.ai, or WhatsApp, with glasses planned later. There is a free tier, then $20 and $100 plans after you burn through tokens. Meta says a Secure VM plus a Sentinel watcher should keep Muse from going rogue, with passwords and payment details kept hidden in theory.
Why are people worried about trusting Meta’s Muse agent?
The launch lands weeks after a huge multistate settlement over social harms, which already puts consumer trust under pressure. Reuters reported internal testers seeing stalls, monitoring failures, and even an agent exposing iCloud photos when asked about toys in birthday pictures. Meta’s pitch is personal superintelligence with safety scaffolding, but early tester reports make the “please trust us” message a hard sell for many users.
Did AI solve a Millennium Prize math problem?
OpenAI says roughly 10,000 agents on a private model found a singularity in the 3D Navier-Stokes equations, one of the Clay Millennium Prize problems, and Lean checked the proof. The run cost millions of dollars and nearly five million messages between agents. Related work from NYU’s Tristan Buckmaster and Anthropic’s Levent Alpöge appeared the day before after nearly a year with public models. Nobody is claiming the million-dollar prize yet amid credit disputes and field-wide worries about brute-force lab proofs.
What did the US accuse Chinese AI firms of stealing?
NSA, CISA, and the FBI issued a joint advisory naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI. Officials allege industrial-scale distillation - vacuuming outputs from Claude, GPT, Gemini, and Grok via transfer-station proxies, shared premium logins, chain-of-thought jailbreaks, and millions of routed queries since late 2024. Distillation is a normal ML technique; the fight is over volume, terms-of-service evasion, and whether this is R&D or IP strip-mining that shortens China’s timelines.
What is Google DeepMind’s AlphaGenome Atlas?
AlphaGenome Atlas is a predictive map covering about nine billion single-letter DNA tweaks, with roughly a petabyte of precomputed molecular effects and an AVI impact score to rank what matters. It is free for noncommercial use today through a portal, API, and Antigravity skill, with cloud commercial access coming soon. Collaborators have already used it on unsolved rare-disease cases and UK Biobank trait hunting. DeepMind still says it is not for clinical diagnosis.
Why are Anthropic Claude Max subscribers suing over pricing?
An expanded class action claims Claude Max buyers were misled by “5x” ($100) and “20x” ($200) labels that do not match real usage once five-hour sessions and weekly caps kick in. Ex-FTC attorneys Monica Vaca and Kati Daffan are leading the case. Anthropic’s earlier defense was that limits were a hyperlink away; plaintiffs call that buyer-beware theater. The suit tests Anthropic’s marketplace-integrity pitch against power users who feel courted, then throttled.