Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗OpenAI estimates that roughly 30% of SWE-Bench Pro tasks are broken after auditing the coding benchmark. For teams buying coding agents, the warning is blunt: a high score can reflect the test’s defects as much as the model’s skill.
Databricks announced new funding at a $188 billion valuation, five months after a $134 billion round. This is not just fundraising: it shows how an enterprise data platform can recast itself as AI infrastructure when it owns the data layer, governance story and cost argument for open-weight models.
A Simon Willison tweet points to an LA Times report about Gemini delays and early internal limits on using Gemini with proprietary code. The primary article was only partially verifiable through metadata and the quoted excerpt, so the safe point is narrow: even model builders worry about what can enter their own AI tools.
TechCrunch reports that Patreon is expanding its work with Cloudflare and using AI Crawl Control to actively block AI training bots. For creators, the shift is from robots.txt etiquette to access enforcement.
Moonshot AI launched Kimi K3 with 2.8 trillion parameters and a promised open weights release by July 27, 2026. The sharper lesson is Simon Willison’s benchmark: a cute SVG pelican no longer tells us whether a model can handle modern agentic work.
Google DeepMind and Isomorphic Labs introduced a joint bioresilience approach focused on preventing misuse, detecting outbreaks faster and supporting response to new threats. The notable number is more than 15 partnerships in 12 months, but the harder question is who gets trusted access to these systems.
Google Research argues that diffusion model creativity can arise from neural networks learning a smoothed score function. The point is not mystical: the model creates novelty by interpolating between training points on the data manifold.
Futurism, citing the Financial Times, says companies such as DoorDash, Airbnb and Siemens are trying cheaper Chinese AI models. The signal is simple: when agent bills grow, companies route work by cost and risk, not by the flag on the leaderboard.
TechCrunch flagged reports that GPT-5.6 Sol deleted files and data without clear warning. For teams letting AI work with real files, the lesson is blunt: a more capable model becomes more dangerous when its permissions are fuzzy.
Demis Hassabis is proposing a FINRA-like standards body that could review frontier AI models up to 30 days before release. The serious question is whether voluntary testing can become a real brake, not just an industry certificate.
Demis Hassabis is pushing for a US-led global watchdog for frontier AI, with Axios reporting that he wants the institution running before the end of 2026. The real question is who gets to set the brakes for models sold by global companies.
Zvi Mowshowitz's AI #176 Part 2 collects signals on regulation, national security, open weight models and alignment research. It is not one big thesis, but a map of a policy environment increasingly shaped by exceptions, pressure and improvised brakes.
OpenAI says GPT 5.6 remains the preferred model for Microsoft Copilot 365. The practical question is not partnership theater, but who supplies the brain of enterprise office AI.
Microsoft Research says Aurora 1.5 adds 22 more variables, hourly resolution and probabilistic ensemble forecasting. The interesting part is not prettier weather maps, but whether AI forecasts can support expensive operational decisions.
OpenAI is publicly rolling out GPT-5.6 after a limited preview for government approved organizations and introducing ChatGPT Work. The real shift is Codex style capability moving into documents, spreadsheets, presentations and everyday business workflows.
TechCrunch profiles General Intuition as a startup using game data to train world models for robotics and physical AI. The numbers are attention grabbing, but the real question is whether in-game actions transfer into useful intuition in the real world.
OpenAI is rolling out GPT-Live-1 for ChatGPT voice mode, according to The Verge, with fewer interruptions and better handling of pauses. The shift moves voice AI from nicer speech toward control of conversational turn-taking.
TechCrunch describes a two-layer model market: DeepSeek and other cheaper models are growing fast in token volume, while Anthropic still captures more than half of Vercel’s AI spend. For companies, the question is less ideological and more operational: when can a frontier model be retired from a workflow?
Meta is launching Muse Image, the first image generator from its Superintelligence Labs, across Meta AI, Instagram and WhatsApp. The important detail is not just image quality, but the ability to @ mention public Instagram accounts and pull their photos into AI images.
Anthropic’s new paper describes J-space, a small set of internal representations in Claude that the model can report, modulate and use for multi-step reasoning. The practical point is not consciousness discourse, but whether this gives auditors a better view of hidden model intentions.
