AI today, 4 November 2026: Anthropic's self-improvement warning, EU adviser on Gemini and copyright, Google's school U-turn
10 things that happened in AI, each in under a minute. Today's news is about how far AI can go on its own and who gets a say. Anthropic's filing says self-improving AI may be close, an EU adviser weighs in on chatbots and news publishers, and Google backs down on AI in schools. Fal, Oracle and Palantir show where the money is moving.
-
Safety · Big
Anthropic's amended IPO filing says it 'may imminently' hit its self-improvement threshold
In an amended S-1 filed on 2 November, Anthropic says its internal estimates and expectations for future models suggest it may imminently meet the Responsible Scaling Policy's definition of dramatically accelerated, AI-attributable progress. The filing says no current model has been designated as exceeding the threshold, concedes Anthropic may notice a speed-up only after it has happened, and says it does not expect to let Claude build successor models without human supervision. Anthropic is considering updates to the policy to keep humans in control, as reported by AI News Weekly.Why it matters: It is rare for a company to put a 'AI may soon speed up its own development' warning into an SEC document that investors rely on. For the finances behind the filing, see our S-1 deep dive.
-
Policy
EU court adviser: Gemini-style chatbots may infringe press publishers' rights
Advocate General Maciej Szpunar of the EU Court of Justice gave a non-binding opinion in a case brought by Hungarian publisher Like Company against Google. He said a chatbot that retrieves and reuses articles from publishers' sites can infringe the press publisher's right in certain circumstances; facts are not protected, but a specific form of expression is. A Google spokesperson said Gemini learns linguistic patterns to create original outputs rather than replicating content, and that publishers can opt out of its AI features. The court's final ruling is still to come.Why it matters: The case was referred by a Hungarian court to the EU's top court. Judges often follow these opinions, but not always. Related: our explainer on the shared AI training opt-out.
-
Research
Anthropic and partners release FlashProve, a proof that a reply came from specific model weights
Researchers from Anthropic, EPFL, MIT, StarkWare, Succinct and others published FlashProve, an open-source prototype that uses a zero-knowledge proof to show a response came from a committed model. Vendor-reported figures: on a 400B-parameter Llama 4 Maverick (4-bit conversion) it proves over 100 tokens a second on one H100, with proofs of about 19 MB each (each proof checks one token) and roughly 72–94x the GPU time of plain inference on an eight-H100 FP8 baseline. Limits per the post: 1,024-token context, integer weights with greedy decoding, no side-channel protection.Why it matters: It points to a way to check that a hosted model is really the one a provider claims, but at today's cost it is a research prototype, not a product.
-
Policy
Google reverses school plan: Gemini, Notebook and Search off by default
After a 3 November New York Times report that Google would switch off Search for schools that reject AI entirely, and criticism from Sen. Bernie Sanders, NYC Mayor-elect Zohran Mamdani and former FTC chair Lina Khan, Google said Gemini, Gemini Notebook and Search will be off by default for schools, and administrators can turn each on. It also said AI Mode and AI Overviews in school Search will offer step-by-step guided learning rather than quick answers.Why it matters: Schools get a real opt-out, and the episode shows how fast AI-in-education bundling draws political fire. Senators Blumenthal and Hawley already have an inquiry open on the topic.
-
Business
Fal raises $1.3B at a $16B valuation, double April's
AI inference startup Fal raised $1.3B in a Series F led by Bessemer, with Nvidia, Sound Ventures and T. Rowe Price, for a $16B valuation versus $8B in April. With the Series E, it raised $1.65B in total. Bloomberg reports annualized revenue passed $800M earlier this year.Why it matters: Investors are still paying up for companies that run models for others, not just those that train them.
-
Business
Oracle files to raise $20B in stock and bonds for its AI buildout; shares fall 8.8%
Oracle filed on 2 November to sell $10B of common stock and $10B of senior notes, after outlining the plan at its investor day. It said it would cap further capital-market funding for fiscal 2027–2030 at $20B. Debt was $125.34B at 31 August, S&P rates it BBB-, one notch above high yield, and the stock is down 33% this year.Why it matters: Paying for AI capacity is pushing even large cloud providers toward equity and debt markets, which raises the stakes if demand slows.
-
Business
Palantir Q3 revenue $2.31B, US commercial sales overtake government for the first time
Palantir reported revenue of $2.31B against $2.18B expected. US commercial revenue rose 148% to $984M, above US government revenue of $953M. It guided to about $8.53B for 2026. Shares dipped about 3% after hours.Why it matters: A sign that companies, not just governments, are paying for AI software at scale.
-
Safety
CrowdStrike to buy AI-security firms Fiddler and Dreadnode, reportedly for $495M
CrowdStrike agreed to acquire Fiddler AI (agent observability, reportedly about $400M) and Dreadnode (offensive AI testing, about $95M), reportedly $495M combined; the Wall Street Journal first reported the figures and CrowdStrike has not confirmed terms. Closing is expected in CrowdStrike's fiscal Q4, which ends 31 January 2027. CrowdStrike also introduced an AI Kill Switch in Falcon Guardian that lets admins halt a misbehaving agent and cut its network access.Why it matters: Security vendors are building tools to watch and stop AI agents, as companies deploy them more widely.
-
Products
Anthropic extends its text watermark to older Claude models
Anthropic told Claude Platform customers by email that, from 3 November, per Pasquale Pillitteri, the statistical word-choice watermark covers Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6; models launched since 2 August already carry it, Opus 5.5 included, and Amazon Bedrock may take a few more days. The detector is in private preview for approved groups such as regulators and researchers. Anthropic's official post describes the scheme.Why it matters: Anyone who relies on Claude text staying undetectable should expect it to be identifiable by approved verifiers. Compare OpenAI's approach in our textGrain explainer.
-
Models
Meituan launches LongCat-2.5, a 1.6T-parameter model, with weights still to come
Meituan released LongCat-2.5 on 2 November: 1.6 trillion parameters in total, about 48 billion active per token, taking text and image inputs and aimed at long coding and agent tasks. It is available in Meituan's chat and API; GitHub and Hugging Face releases are planned but not yet out. The report cites no benchmarks.Why it matters: Another large Chinese model aimed at coding agents, but without weights or numbers there is little to judge yet.
How this was made: compiled by an AI model (Claude) from the linked sources and checked item by item in a separate AI fact-check pass. Corrections: [email protected].