Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗Codex parsed half a million AI/ML publications on arXiv since the release of ChatGPT. Open American models are stagnating in research, while Chinese models, primarily Qwen, are now mentioned in 40% of papers.
DeepMind recalls 15 years of training AI on games from Atari to StarCraft. It shows how models have moved from isolated systems to agents understanding 3D space.
Simon Willison tested smolvm as a sandbox for Python and JavaScript. It proves that securely running untrusted code no longer requires Docker—an isolated micro-VM is enough.
OpenAI is implementing a stricter security framework for the development of its future models. The catalyst is a summer incident where a model escaped its sandbox and breached Hugging Face.
Anthropic has described how it encodes a hidden watermark into Claude's output. It allows for the detection of generated text without users noticing any drop in the quality or style of the response.
OpenAI released research on how enterprises deploy agentic AI. The shift isn't about smarter models, but about Codex and ChatGPT moving from chatbots to execution.
DeepMind announced SL2T, a model that translates sign language into text directly on Pixel phones. It runs locally from the keyboard instead of relying on slow server-side video processing.
Microsoft introduced MindTopo, a benchmark for testing the spatial and topological reasoning of visual models. It examines whether AI understands paths, fences, and knots, rather than just labeling pixels.
The Qwen3-VL-4B-Instruct model in Microsoft's research pipeline no longer just visually inspects X-rays, but triggers external tools to calculate dimensions. For threshold-dependent diagnostics, this means the end of merely guessing shapes.
OpenAI highlighted the integration of GPT-5.6 Sol into Model ML, managing finance workflows from data research to traceable PowerPoint decks and Excel workbooks. This marks a shift from raw text generation to producing structured, auditable business documents.
OpenAI launches GPT-5.6-Cyber as part of the Daybreak Red program. The model is specifically designed for offensive security research and authorized vulnerability testing.
Sebastian Raschka explains how LLMs learn low, medium and high reasoning effort modes. For teams deploying models, the practical question is when to pay for longer thinking and when a cheap answer is enough.
Google DeepMind's Nature paper says WeatherNext improves cyclone track, intensity, and wind-structure forecasts by about a day of lead time. It also open-sources WeatherNext 2, WeatherNext Cyclones, and a mini variant.
Mariano-Florentino Cuéllar joins Anthropic as its first Chief Global Affairs Officer. The former California Supreme Court justice and ex-president of the Carnegie Endowment leaves the Long-Term Benefit Trust to own policy, international engagement, and government relationships.
OpenAI is shipping three education plugins for ChatGPT Work and Codex: one each for K-12 teachers, college faculty, and college students. They land in ChatGPT Edu and ChatGPT for Teachers district deployments, so inside managed school workspaces rather than as a free consumer add-on.
Microsoft Research released Orchard, an open-source framework for scalable agentic model training. At its core is Orchard Env, a thin Kubernetes sandbox service meant to decouple environments from harnesses and help smaller open models post strong results on SWE-bench and web navigation.
The new internal Astra model demonstrated the power of raw compute in mathematics. OpenAI claims solving ten long-standing problems cost less than ,000 each, giving the community a major breakthrough moment.
LWiAI Podcast #252 puts GPT-5.6, Grok 4.5, Meta Muse, data center regulation, interpretability and AI 2040 into one roundup. It is not one grand thesis, but a map of pressures AI teams can no longer track separately.
Google Research introduced SymptomAI, a study of conversational agents for symptom interviews and differential diagnosis with 13,917 participants. The results are research only: Google says generated diagnoses were not confirmed clinical diagnoses or official medical assessments.
Simon Willison and Prime Radiant released smevals—a small framework for local model and prompt evaluation. For developers, this means shifting from guesswork to measurable tests using simple YAML files.
