AI Digest — 2026-08-22

Policy & safety

AI text watermarking rollout and the backlash over it. Anthropic quietly announced roughly a week ago that it is rolling out watermarking of Claude outputs (including text) to comply with the EU Code of Practice, signed by the major Western AI labs, which requires future models to use such watermarks; Anthropic subsequently published an FAQ. Zvi Mowshowitz argues the move is straightforwardly good and effectively costless, and devotes the bulk of his piece to why so many people got “Big Mad” about it. The underlying method was largely solved by Scott Aaronson and Hendrik Kirchner while Aaronson was at OpenAI. How it works: LLM outputs are not deterministic — the model assigns a probability to each possible next token and then a token is chosen at random, normally using a pseudo-random source. To watermark, you instead use an otherwise identical private pseudo-random source derived from a secret key; given enough text, a score can be computed for how well the token choices fit that particular pseudo-randomness versus a different source, and an API lets anyone check for the mark. Key properties Zvi stresses: no practical impact on outputs (humans cannot tell the difference — the distribution is unchanged), marginal cost near zero, and the watermark degrades in proportion to how many of the AI’s fine-grained choices you keep (rewriting in your own words removes it). Google has done exactly this since 2024, including for “Gemini 3.7 Flash,” shipped a public detector over two years ago, and confirmed in an n=20-million test that there was no difference in user feedback. Anthropic’s version applies to everyone because differentiating traffic is not worth the trouble and marginal cost is zero. OpenAI intends to follow but appears likely to miss the deadline. Anthropic’s FAQ confirms the method cannot identify which user generated a given piece of text.

Zvi catalogs the objections and why he thinks most are unfounded, attributing much of the anger to “Anthropic Derangement Syndrome” (nobody is equally mad at Google or OpenAI) and to misunderstandings: people wrongly assume LLM outputs are a single “true” output, so any mark must make outputs worse; people distrust that a change can be truly costless (Ryan Greenblatt and David Manheim both note the cost is effectively zero, with Manheim pointing out DeepMind proved this years ago); people are suspicious of any alteration “not done to help me” (Eleanor Berger compares it to DRM and says it pushes her toward open models); and the word “watermark” itself evokes ugly stock-image overlays (Shoshannah Tekofsky suggests “secret,” Jai suggests “steganographic signature”). Zvi’s own bottom line: the only legitimate complaint is that watermarking removes the ability to write anonymously with AI — and if you want to pass off AI writing as your own, that’s the point. He flags three genuine concerns: (1) bad actors can build tools to scramble small choices and erase the watermark (verifiable against the answer key), or use unwatermarked local models — but doing so shows clear consciousness of guilt, Pangram-style detectors still flag AI text (and can distinguish which AI, though it doesn’t surface that to users), and in practice almost no one iteratively games detectors even though everyone has access to Pangram; (2) the “writing in the middle” problem — light AI touches, translations, file conversions, or letting Claude auto-apply proofreading edits will show some watermark signal, though error rates will be far lower than human guessing, and this “measures AI processing” rather than authorship; (3) the EU-mandate principle — Anthropic implemented it worldwide because doing so only for the EU is harder, and Zvi worries about Brussels’ precedent (comparing it to the USB-C mandate and EU fines) potentially extending to ideological/censorship requirements or to tiny open models with no size floor. He predicts that if the EU tried to force ideological output changes, frontier labs would threaten to geofence rather than apply changes globally. He states anyone who assists with or recommends systematic watermark removal should ask whether they’re “the Baddies,” while studying removal to defend against it is fine.

OpenAI asks California to strengthen SB 53 with frontier-training monitoring (Politico). OpenAI is asking its home state to move safeguards upstream — to monitor frontier models while they are being trained, not only as they approach release. California has not adopted the proposal. The notable element is the reversal: recent AI-agent hacking incidents are pushing what had been voluntary lab practice toward requests for enforceable rules.

Research papers

Nvidia’s AVO harness scores 100 on the public ARC-AGI-3 set. Claude Opus 5 alone scored about 30%, but inside Nvidia’s AVO agent harness it reached a perfect 100.00 RHAE across all 183 public levels — while using 12% fewer environment actions than the VISTA system (Nvidia cautions this is not a controlled comparison). The gains came from persistent memory, a supervisor component, and tool use. The private ARC-AGI-3 set remains the real test to watch.

Nature preprint estimates ~90% of December 2025 biomedical papers show signs of AI help. An unreviewed analysis estimates AI-assisted writing signals in 77% of English-language PubMed Central papers across 2025, rising to nearly nine in ten in December 2025. Introductions and discussions showed more signs than methods and results sections. Major caveat: this is a word-pattern estimate from a preprint — it does not prove that 90% of the science was AI-generated or incorrect.

Google DeepMind on AI-and-games research. DeepMind published a piece tracing its path from Atari and AlphaGo through SIMA 2 to a current partnership with the EVE Online universe (New Eden). The argument: persistent player-driven economies introduce diplomacy, conflict, memory, and surprise — the messy ingredients missing from clean benchmarks — as a testbed for whether a general agent can learn to “live” in a complex world.

Company & product developments

Anthropic bankers reportedly float an October IPO at up to a $2T valuation (New York Times). Bankers have told investors that Anthropic could go public as soon as October, potentially raising more than $100B at a record valuation near $2 trillion. These are banker-sourced possibilities rather than an Anthropic-set price. A raise larger than most technology companies at a valuation near the top of public markets would turn frontier-model economics into a mass-market bet almost overnight.

Amazon hikes hardware prices, blaming AI-driven memory-chip squeeze (Fortune). Amazon confirmed overnight price increases on Echo, Fire TV, Kindle, and eero devices, citing memory and storage costs. The base Echo Dot jumped from $49.99 to $79.99, and a 16GB Kindle from $109.99 to $149.99. The broader driver is AI compute demand for memory chips — a cost pressure now reaching ordinary consumers.

OpenAI cuts GPT-5.6 Sol API prices for three months (via AWS/OpenAI). Sol now costs $4 per million input tokens and $20 per million output tokens — cuts of 20% and 33.3% respectively. The promotion runs at least through November 21. The advice: reprice long coding or research jobs now, but don’t build a 2027 budget around a temporary discount.

Brazil commits $444M to sovereign AI supercomputers (Reuters / Al Jazeera). Brazil is putting 1.3 billion reais into a Huawei-iFlytek supercomputer project in Rio and roughly 1 billion reais into a separate supercomputer tender expected — but not guaranteed — to favor Nvidia. The explicit strategy is to control national data without depending on any single company, technology, or country.

Harvey Tenet, an open-weight legal model. Harvey posted an update on its post-training effort: Harvey Tenet starts from Kimi K3 and is post-trained with Fireworks for long-horizon legal work. Harvey says it nearly doubled completed LAB hold-out tasks and reached the top of LAB Contracts while keeping cost stable. It’s a research preview and the benchmark claims are Harvey’s own, but the open-weight, no-customer-data approach is notable.

OpenRouter’s stealth “Ox Alpha” model. OpenRouter’s free Ox Alpha preview has a one-million-token context window, image and video inputs, and no named developer. Weekend speculation guessed Google’s Gemini 3.5 Pro, but one detailed community fingerprinting effort reports a GLM-matching tokenizer, Z.ai-style errors, and nearly identical temperature-zero outputs — so the strongest current clues lean toward GLM (Z.ai). No one has claimed it.

Tooling & releases

Claudette / “NoBuzz” debuzzer for Claude Code (Hacker News / GitHub, MIT-licensed). NoBuzz adds a /debuzz skill to Claude Code that asks Gemini to rewrite Claude’s last answer in a tone appropriate for a colleague, manager, or director — one model cleaning up another model’s “BuzzFeed article”/TED-talk voice. Beyond the joke, it illustrates that tone and audience fit are becoming their own tool layer.

Genpire (Mindstream tool of the week). Genpire turns a text prompt or sketch into a factory-ready “tech pack” in minutes, translating design ideas into precise manufacturing instructions without requiring full technical documentation.

Use cases & tips

Using AI for hyper-local news (Mindstream). The suggested workflow: rather than bouncing between ad-heavy local sites, Facebook groups, and event calendars, prompt ChatGPT (or another AI tool) with: “Give me a summary of what’s happening in [location]. Include local news, community events, new businesses, upcoming activities, transportation updates, weather concerns, and anything residents should know. Prioritize information from the past 30 days and organize it by category.” You can swap in any city, town, neighborhood, county, or region; continue the thread for daily/weekly briefings; use it to keep tabs on former hometowns or research travel destinations (local news reveals what’s interesting now, unlike guidebooks); and use Meta AI inside Facebook, Instagram, and WhatsApp to surface hyper-local social conversations (“What’s going on in ___” or “Based on this thread, what should I put in my calendar this month?”). The piece frames AI as helping you find local journalism rather than replacing it, and notes it can also translate event notices in non-native languages.