Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗404 Media highlighted the recurring figure of Elias Thorne in chatbot stories, and the related research puts numbers on it: 20,000 generated stories repeatedly return to lighthouses, keepers and a small set of names. The issue is mode collapse in creative output, not a lack of vocabulary.
Sebastian Raschka shows a local coding-agent stack: an open-weight model in Ollama, a harness that edits code and runs commands, and your own machine instead of a Claude Code or Codex subscription. For teams, the interesting part is not nostalgia for localhost, but protection against pricing, limits and model changes they do not control.
Odyssey raised a $310 million Series B at a $1.45 billion valuation, with Amazon, AMD Ventures and GV among the backers. The sharper story is the AWS and Trainium tie-up: world models are being pitched as infrastructure demand, not just impressive video demos.
Pew Research Center found that 49% of U.S. adults use AI chatbots and 24% use them daily. The same survey shows the adoption ceiling: 63% say AI is advancing too quickly and 70% expect weaker personal data security.
TechCrunch's Equity episode with NEA partner Tiffany Luck frames the shift from tokenmaxxing to AI ROI discipline. For enterprise teams, procurement hygiene now matters more than another dazzling personal-agent demo.
TechCrunch reports that General Intuition is in talks to raise around $300 million at a valuation above $2 billion. The more interesting story is the data thesis: Medal’s 2 billion gameplay videos per year are meant to teach agents space, time and action better than ordinary web video.
OpenAI is bringing in Noam Shazeer from Google DeepMind and Dean Ball to lead Strategic Futures. Ahead of a possible IPO, the company is strengthening both model architecture and its interface with U.S. AI policy.
The US government reportedly forced Anthropic to pull Fable 5 and Mythos 5 over national security concerns. For teams building on Claude, the real lesson is that frontier model availability has become a regulatory dependency.
Z.ai is positioning GLM-5.2 as an open-weight model for long-running coding agents with a 1M-token context window. The useful question for teams is when an open model is good enough to replace Opus, not whether it wins every chart.
Microsoft Research and partners described Talos, an open-source tool for automated genomic reanalysis in rare disease. In a prospective cohort of almost 5,000 patients, it added diagnoses in 5.1 % of cases.
Google Research examines why chain-of-thought helps LLMs answer simple factual questions. The study on Gemini 2.5 and Qwen3-32B points to two mechanisms: extra computation in generated tokens and factual priming.
Anthropic is launching Claude Tag, a beta Slack service for Claude Enterprise and Claude Team customers. The AI is meant to follow channels, retain context and act as a shared team identity, not just answer one-off prompts.
The money fight around Alex Bores ended without a clean winner: Bores narrowly lost the NY-12 primary, but attacks from a pro-AI PAC made him a visible symbol of regulation. For AI companies, it is a warning that political influence through super PACs can backfire.
The Dutch trade minister went to Washington to push back on the MATCH Act, which would extend restrictions to ASML’s DUV tools for China. The dispute shows that the chip war now extends from the US-China axis to a fight over how much of Washington’s cost allies must carry.
Claude is up about 75% among paying US consumers since January 2026, according to transaction data cited by TechCrunch. For Anthropic, the signal matters because the brand is no longer only a developer and enterprise story.
At Config 2026, Figma put Motion, upcoming code layers and shader tools closer to the core design canvas. Product teams get a more powerful workspace, but also a new place for handoff problems to hide.
Anthropic reportedly claims operators linked to Alibaba and Qwen used nearly 25,000 fraudulent accounts to generate 28.8 million Claude exchanges. The case shows that model capabilities can be extracted through API access, not only through stolen weights.
Lilian Weng revisits scaling laws and shows why they are most useful as a tool for allocating compute, data and parameters. The practical lesson is sober: extrapolation helps only while you remember how small the underlying experiments were.
Google has rolled out frozen Multi-Token Prediction for Gemini Nano v3 on Pixel 9 and Pixel 10 phones. The practical point is simple: faster local AI without replacing the base model or adding a separate drafter into memory.
OpenAI introduced GPT-5.6 as a three-model family, Sol, Terra and Luna, but is starting with a limited preview coordinated with the U.S. government. The important part for teams is that the system card pairs higher cyber and bio capability with a heavier safety stack.
