904 summarised stories about AI Safety, each linking back to the original source. Browse all topics →
TechCrunch 2 months ago 24 ● 7 sources
Amazon is buying rare books, cutting their spines, and scanning them to provide text for AI model training, based on tracking of a book sent to its Las Vegas facility. The report says the shipment arrived at facility VGT3 in Las Vegas. This shifts rare-book handling toward bulk digitization for AI training, rather than preservation or resale.
Simon Willison’s Weblog 2 months ago 24 ● 7 sources
404 Media tracked a shipment of rare books from a dealer to an Amazon AI training facility after suspected AI-scan bulk orders. The bookseller received an order of around 1,000 books in July. The reporting links the shipment to a Las Vegas Amazon site (VGT3) where workers said books are destructively scanned for AI training, strengthening evidence of such practices.
404 Media 2 months ago 6
An expert witness hired by 3M used ChatGPT to draft parts of his expert report in a Houston explosion lawsuit tied to the 2020 Watson Grinding disaster. The record shows 3M paid about $90,000 for the analysis at a rate of $475 per hour, and the prompts sought to “show how 3M is 0% at fault.” The case exposes that ChatGPT prompts and conversations can be discoverable in litigation, and this changed the trial focus as attorneys used the AI use to challenge the report’s credibility.
404 Media 2 months ago 6
GetReal cofounder Hany Farid discussed deepfake detection and the evolution of AI-generated images with a podcast host. PhotoDNA, his perceptual hashing algorithm, is said to be used by every serious internet platform in the world. The conversation shifts from how well people detect AI images now toward how reality perception and defenses may change next.
IEEE Spectrum 2 months ago 23
Axiom Math used its AI system AxiomProver to automatically verify the proof of the 246 theorem about primes for the first time. The verified result states there are infinitely many prime pairs that differ by 246. It raises the prospect of reusable formal proof components to help formally verify AI-generated computer code for correctness, despite limitations of verification systems.
MIT Technology Review 2 months ago 43
A 10-year-old neurodivergent boy in New York uses an AI-powered social robot, Moxie, as a practice companion, but Moxie’s behavior and planned purpose for therapy fall short before she is due to die. Moxie originally sold for $1,499 plus a $40 monthly subscription before Embodied later lowered the price to $800. The situation underlines both the promise of embodied, robot-assisted interaction for autism skills and the risks when these devices fail to work reliably for vulnerable children.
Trending Topics 2 months ago 28 ● 4 sources
Anthropic described how Claude’s text watermark works and addressed backlash from subscribers after its existence was publicly noted on a support page. The watermark relies on a secret key over word-by-word synonym choices, and Anthropic says it adds no hidden characters or tokens. Rollout continues worldwide under the EU Code of Practice for transparency, with an (as-yet unavailable) detection API and potential use for filtering Anthropic output from future training data.
Fortune 8 ● 3 sources
Dario Amodei said AI is facing a crisis of trust in which people suspect companies and governments will “screw them over,” even as he pushed back on claims that he drives public doom. He cited his earlier 50% white-collar job wipeout prediction as context while saying he now describes AI as a multiplier of output rather than a destroyer of jobs. As a result, Anthropic is expected to focus more on delivering concrete results in biology and medicine while avoiding empty promises and emphasizing honest risk discussion.
The daily briefing
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.
TLDRocket
22 October 2026
The biggest AI story today is the next step in OpenAI’s distribution strategy: it’s bringing its most advanced large language models to Google Cloud through Alphabet’s Model Garden service. OpenAI says Astra, its flagship GPT-6 Astra model, will be paired with smaller GPT-6.1 Sol and GPT-6 Luna variants, priced at 5x and 100x less per token than Astra. That matters because token cost is where pilots either scale or stall, and Model Garden is Google’s attempt to turn model selection into something closer to procurement than experimentation.
Read the full briefing →