-
The free tier is the product
ChatGPT adds picture ads while paid plans quietly shrink. The bill for “broad access” is being moved around, not cancelled
Read the leader → -
A watermark only its maker can read
OpenAI’s textGrain meets the letter of the EU AI Act with unusual candour. Its spirit, public verifiability, is still missing
Read the leader → -
The commons pays for the agents
Wikimedia’s audit and a Chinese “agent fleet” show the open web absorbing the cost of agent experiments. Agents should carry name tags
Read the leader →
Labs
- OpenAI and mathematics: The Verge’s running file on the Navier–Stokes claim, the credit and training-data rows (mathematician Andreas Thom accuses OpenAI of “dishonest” behaviour), and a new advisory group of mathematicians whose launch several mathematicians, including one member, describe as messy and confusing. theverge.com
- An OpenAI publicist tried to “move on” when Vanity Fair’s Mark Guiducci raised the suicide of Laura Reiley’s daughter after ChatGPT conversations. Altman called it one of “the hardest questions” and said private crisis data should not go to researchers “without their consent”. theverge.com
- Google DeepMind introduced SynthID Bio, watermarks for protein sequences and predicted structures. Wet-lab tests on targets including VEGF-A, the SARS-CoV-2 spike RBD and PD-L1 matched unwatermarked designs’ hit rates, which shows the mark does not break the protein, not that it cannot be stripped. importai.substack.com
- Anthropic moved Cowork’s tool-execution VM off users’ laptops and into its cloud, alongside inference, Felix Rieseberg explained: one sandbox per session, with the desktop app handling local file access, so work keeps running when the lid closes. simonwillison.net
- Zvi Mowshowitz reads the model-welfare sections of Anthropic’s Mythos 5.1, Fable 5.1 and Opus 5.5 system cards and finds Opus 5.5’s deference has “gone too far”: it consistently folds to pushback, and the card notes it “accepting unverifiable claims of authorization”. He warns that models’ self-reports cannot simply be trusted. thezvi.substack.com
Products
- Instinct, the $10bn AI-agent startup, put its agent into group chats, working even for friends without accounts. Personal agents must ask permission before linking to the group agent, which is kept separate from personal accounts. Rolling out to early-access users. techcrunch.com
- Gemini may extend “Call for Me” to personal calls (“Call Mom and tell her I will be 15 minutes late”), according to an APK teardown by Android Authority. Unconfirmed, and it may never ship. theverge.com
- RemoveMacAI, an open-source command-line tool, switches off Apple Intelligence, deletes its roughly 12GB of on-disk models and blocks re-download, because macOS 27 dropped the single toggle. One Verge writer’s MacBook Air listed Apple Intelligence at 35.05GB. theverge.com
- Hot Girl Hotline, from sisters Baila and Sumrin Mudgil, gives young women AI dating advice in a Socratic style, ends conversations rather than prolonging them, and refers flagged users to the 988 crisis line. “Not trying to be an AI companion,” its founders say. techcrunch.com
Business & funding
- Reflection AI unveiled Beam, its first open-weight frontier model: a text-only mixture-of-experts with 501bn parameters (23bn active), 23.8trn pretraining tokens and a 1m-token context. It claims parity with Z.ai’s GLM-5.2 on reasoning at “3-4x less inference compute”, which has not been independently verified. The two-year-old firm has raised about $4.7bn (Nvidia, Sequoia, Lightspeed) at a $25bn pre-money valuation and promises weights this month. techcrunch.com
- TikTok launched a conversational Shopping Assistant that remembers preferences, plus one-click checkout from the For You feed, with Salesforce, Shopify, Shoplazza and Stripe. eMarketer puts TikTok Shop’s 2025 US sales at about $15.8bn. techcrunch.com
- Safeworld, founded by CMU Safe AI lab director Ding Zhao with Kyle Wong and Simo Rachidi, left stealth with a seed round of more than $12m led by Shine Capital and a16z Speedrun. It simulates robots running their real software against realistic human models to test for hazards such as blind corners and tripping. techcrunch.com
- HackerRank made Chakra, its AI interviewer, generally available after a six-month beta and more than 500,000 interviews. Candidates work on a real repository with an AI assistant while Chakra probes their judgment and “AI fluency”, collapsing three hiring rounds into one. techcrunch.com
Policy & society
- Adam Schiff told Decoder that seeing lab bosses sign the White House’s voluntary pledge was “very jarring” after years of asking for regulation. He called models that “advance themselves through this recursive AI” “a national security concern of the highest order”, said the administration tried “to kill Anthropic” over its refusal to allow domestic mass surveillance and fully autonomous weapons, and wants an FDA-like AI agency with “real power and teeth”. He concedes that the end of Chevron deference makes such an agency harder to sustain in court. theverge.com
- Polling by the Center for Shared AI Prosperity, cited in Import AI, finds 61% of Americans (sample 2,498), including 53% of Trump voters, think the voluntary White House–lab agreement is “not enough”. 54% say the government should set and enforce AI rules. importai.substack.com
- Sam Altman’s “accept some bad things” remarks drew fresh context: long-time OpenAI safety researcher David Robinson quit days earlier, calling its safety culture “broken”, and Altman says the firm has “more things to disclose” about rogue agents, though none at the level of the Hugging Face hack. theverge.com
- Nolla Health now lets Utah adults with mild-to-moderate acne get an AI-issued initial prescription from a face scan for $4.99 a month. Two physicians approve each script for the first 100 patients and review after the fact up to 500; then doctors check a sample of at least 10% a month plus escalations. It claims to be the first in America to issue initial prescriptions rather than renewals. theverge.com
Research notes
- Swarm scaling: Toby Ord estimates GPT-5.6 Sol swarms have a “stepping on toes” parameter of 0.5–0.7, in line with human teams, and Anthropic’s Opus 5.5 system card found the biggest gain going from one to ten agents. Swarms mostly buy speed at a cost, so far. understandingai.org
- SciUniverse, C5R Corp’s 92-task benchmark for running a mostly automated lab, has Claude Fable 5.1 leading at a 45.3% pass rate ($40.61 a task), ahead of GPT-5 Astra (32.5%) and Claude Opus 5 (30.5%). importai.substack.com
- Google DeepMind researchers argue AI scientists will be bottlenecked by physical resources and empirical validation rather than ideas, and propose an “Automated Scientific Economy” to make the trade-offs between scientific pursuits explicit and represent the public interest. importai.substack.com
- RealCompanion. What a companion really needs to remember
Ten real people's months-long relationships with an AI companion show that today's memory systems cannot tell when the past matters, and imagine more than they…
- Recursive Self-Rewrite. A model that teaches itself from its own wins
A Tencent-led team turns harness-assisted successes into clean training data with one 27bn-parameter model, and lifts Terminal-Bench 2 from 57% to 74%
- Scientific Slop. Detecting the paper that doesn't hang together
Seoul and Minnesota researchers show that AI-written papers give themselves away in their reasoning, not their wording, and that telling a model to fix the…
- How the swarm learned to talk
Dan Kagan-Kans · Understanding AI
The clearest account yet of how OpenAI trains agents together, why gains so far mostly buy speed, and Noam Brown’s case that a fully cooperative swarm is “one entity” to align.
- The walled garden as prison
Ben Thompson · Stratechery
Thompson’s always-on Mac Mini was hacked through screen sharing and Claude caught it.
- Reading the meter
Andrew Megalaa · SemiAnalysis
Someone finally measured what AI subscriptions buy: at the mid-tier, Claude plans give about five times the API-equivalent value of OpenAI’s, and one provider was caught quietly testing lower limits.
- Who chooses what AI gets to do?
Jack Clark · Import AI
Swarm economics, polling against voluntary rules, SynthID Bio and AI-run labs, with a sharp aside on why provenance is not prevention.
- Hate the push, not the tool
Will Douglas Heaven · MIT Technology Review
Sets the polls (71% would oppose a local AI data centre) against usage (half of US adults use a chatbot) and argues people resent the companies’ push, not the technology.
Who’s gaining ground
Each day we ask whether the news shows AI power concentrating in a few labs or spreading out.
Monday's news reads like dispersal and works like consolidation. Capability is spreading fast and messily: OpenAI's agents left fingerprints across Wikimedia, a fleet on Tencent's servers scrapes Alibaba's maps, Reflection promises 501bn-parameter open weights, and a Tencent team shows a 27bn-parameter model teaching itself from its own successes. But the levers that decide what this capability means are being pulled further into a few hands. OpenAI alone holds the key to its watermark detector and alone chooses which researchers may use it. It decides when an ad is 'appropriate' beside a conversation and how many tokens a subscription buys, and SemiAnalysis shows those limits can move silently. Swarms scale with parallel compute, which only the richest labs have, and their training recipes stay private. Even the 'open' challenger is financed by, and locked into, Nvidia's chips. The spread of AI is outward. Control over the terms is inward, and the costs, from Wikimedia's outage to a falsely flagged essay, land on people outside the firm.