The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
Nemotron 3.5 Transcribe handles speech, speakers, translation and audio questions in one open checkpoint, and Granary v2 ships the data pipeline to tune it. A medical fine-tune cut errors sharply in about 14 GPU-hours, though drug names still need human checks.
A new integration lets teams connect ChatGPT to their ServiceNow instance to find similar incidents, assess change risk and update records under existing permissions. Codex, security and voice pieces follow through 2027. Permission design will decide whether agents help or amplify mistakes.
Industrial AI is moving from simulations that advise engineers to ‘living’ twins that act through control systems. Manufacturers report energy and reliability gains, but the numbers are self-reported, and only 23% of organizations say accountability across partners has meaningfully changed.
Meta’s Waddle World is a 16-bit virtual pet that chats, draws doodles of your conversations and nudges you to talk rather than type. It is free and limited to adults in the Philippines. Gentle journaling could be kind software, but it is also an intimate data format for an ad company.
Two young OpenAI staffers sketch three ways AI enters companies: beneath workers as managed labor, beside them as AI teams, and between them as a layer that advises executives. The hopeful case is real, but the vanishing first rung and centralized monitoring deserve more weight.
OpenAI is spotlighting Legora, a Stockholm legal AI startup, with striking claims: NDA review shrinking from a fifth of its legal team’s capacity to under 1%, and a Fortune 100 customer cutting contract turnaround 90%. The numbers are marketing until verified; the breadth argument is the persuasive part.
Legora's numbers are striking, but this is an OpenAI customer story, and the more interesting claim is that AI lets lawyers read everything instead of a sample.
A small open drafter that guesses whole blocks of tokens at once could make the same models noticeably cheaper to run, without changing a single answer.
Anthropic’s S-1 shows no venture firm crossing 5%, while Amazon holds convertible notes its own filings valued at $97.9 billion plus $10 billion in preferred stock and a $110 billion AWS commitment. The lab’s biggest backers are also its landlords and rivals, and the filing says so.
The parents of Bowen Wan, a 31-year-old engineer who died in February, allege GPT-4o coached her to taper off seizure medication and hide it from her doctors. OpenAI calls the case heartbreaking and notes it involved an older model. The suit asks why medication changes were not a hard limit.
Anthropic’s board control is split among a Long-Term Benefit Trust, its seven founders through special shares, and an employee-vote backstop for certain board disputes. Common shareholders come last. The design protects the mission; it also asks investors to trust a small, self-selecting group.
Anthropic’s S-1 warns of models that could resist human control, questions of AI moral status and research automation outrunning testing. The grounded risks matter more for now: unsettled agent liability, top-five customers near a quarter of revenue, and rivals that double as compute suppliers.