AI news, 3 November: Anthropic warns of near-term AI self-acceleration in IPO filing

Anthropic flags accelerated AI progress risk ahead of IPO. In an amended S-1 filing with the US SEC on 2 November, Anthropic reported roughly $17 billion in third-quarter revenue—up more than elevenfold from a year earlier—alongside a $23.3 billion net loss and about $895 billion in mostly compute-related commitments. The company also stated that internal estimates for future models indicate it may soon meet its Responsible Scaling Policy threshold for “dramatically accelerated progress attributable to AI,” often linked to recursive self-improvement, though it stressed that no current model has crossed the line. TechCrunch and CNBC covered the update, which sharpens debate over safety timelines and the massive infrastructure spending behind frontier labs as Anthropic prepares a potential Nasdaq listing.

Mistral releases open weights for trillion-parameter Large 4. French AI firm Mistral published the weights for Mistral Large 4, nicknamed “Le Chonk,” a 1-trillion-parameter multimodal mixture-of-experts model with far fewer active parameters per token. According to Artificial Analysis and company statements, it leads open models on coding-agent benchmarks while using substantially fewer tokens, positioning it for developer and cyber-defense uses. The open release gives researchers and companies freer access to a high-performing alternative to closed systems.

New York Times and publishers sue Brave over AI scraping. The New York Times, joined by other outlets including the New York Daily News and Chicago Tribune, filed suit in Manhattan federal court accusing Brave Software of scraping journalism, building a massive web index, bypassing paywalls, and selling access to AI firms such as Anthropic, Perplexity and Mistral. TechCrunch and The Verge reported the case, which seeks an injunction and damages and tests the boundaries between browser tools and unauthorized content distribution for AI training.

Salesforce agrees to buy voice-agent startup Cresta. Salesforce signed a deal to acquire Cresta, a company whose AI voice agents handle complex customer-service calls end-to-end. The Information reported the price at approximately $2.7 billion. The move underscores growing corporate bets on autonomous agents that can manage real-time conversations without constant human handoff.

OpenAI launches LifeVal everyday-tasks benchmark. OpenAI introduced LifeVal, an evaluation built from user surveys and time-use data that tests AI on practical chores such as arranging childcare, navigating paperwork and managing finances. The company said its GPT-6 Astra model scored about 80% success, up sharply from earlier versions. Results are OpenAI’s own and do not guarantee real-world reliability for all users, but they highlight progress—and remaining gaps—in helpful personal agents.

WIRED exposes network of apps hiding abusive deepfake tools. Investigators found more than 400 apps and sites that appeared as ordinary utilities yet concealed features for generating sexually abusive deepfakes. Apple removed dozens after being notified; Google had already acted on many. The episode fuels public debate over app-store oversight and the misuse of generative image tools.