Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗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.
OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom inference chip for running LLMs. For ChatGPT, this is less flashy than a new model, but potentially more important for the unit economics.
OpenAI is reportedly preparing to release GPT-5.6 first to selected partners, with the US government approving access customer by customer. The precedent matters more than a delay of a few weeks.
US officials are reportedly intervening in releases of Anthropic Mythos and OpenAI GPT-5.6. The story is becoming less about lab rivalry and more about whether frontier AI can survive pre-release control without a real process.
OpenAI is offering its models and Codex to Oracle Cloud customers through existing cloud commitments. For enterprise teams, the interesting part is not the endpoint, but the way AI fits into contracts, governance and billing they already use.
Google introduced Gemma 4 12B as a unified, encoder-free multimodal model designed for high performance directly on a laptop. The practical question is whether a 12B model can deliver enough quality for local or edge use without heavy cloud infrastructure.
Niteshift, founded by former Datadog engineers, raised a $7 million seed round led by Greylock and is selling infrastructure for AI coding agents. Its bet is not another autocomplete, but the ability to switch between GPT, Claude and open source models when the model provider becomes a competitor.
The US government told Anthropic to restrict Fable 5 and Mythos 5 for all foreign nationals, so Anthropic switched the models off for all customers. A protest by 76 security experts exposes the weak point: export control is bad at separating an offensive exploit from defensive testing.
The Verge describes the stronger AI work around Dear Upstairs Neighbors at Tribeca as custom workflows around Veo and Imagen, not simple prompting of a general model. For studios, the sober lesson is that value sits in style control, not in a magic prompt.
Anthropic says US officials ordered access to Fable 5 and Mythos 5 suspended for foreign nationals, so the company disabled both models for all customers. Buyers of frontier AI now have to price in a risk that sits outside the model: the state kill switch.
Zvi Mowshowitz uses Fable and Mythos as a case study for why model welfare cannot be separated from capabilities, alignment and user experience. Even where the topic remains speculative, it is becoming a practical question of evaluations and safety interventions for frontier labs.
Google released Android 17 and Wear OS 7 first for Pixel devices, alongside a Pixel Drop with Gemini Omni, Lyria 3 and translation features for the Pixel 10a. The bigger signal is not the OS update itself, but Google using Android as a distribution layer for AI models on the device.
Anthropic released Claude Opus 4.8 at the same standard price as Opus 4.7, with a focus on coding, agentic tasks and longer work. The more important shift is a model that is supposed to say more often when it is unsure.
Nathan Lambert announced his departure from the Allen Institute for AI and used it to reflect on work around Olmo. This is not just a personnel note. It is a reminder that open models depend on institutions that must outlast one strong team.
Latent Space frames Microsoft Build as the moment Microsoft showed its own MAI models alongside Copilot, Windows and Web IQ. The key ambition is to control data, inference and developer workflow at once, rather than leaving that leverage to partners.
The White House issued an executive order that calls for a classified benchmark for covered frontier models within 60 days and a voluntary framework for up to 30 days of pre-release government access. It says this is not licensing, but it creates a pressure point before launch.
Latent Space published Auriel W's piece on why low-quality RL environments damage agent training. The point is simple: in reinforcement learning, the environment is the data generator, so a harness bug becomes training material.
Hugging Face published a Build Small Hackathon field report about Thousand Token Wood v2, a simulation where four characters run on four different small models. The key lesson for agent systems: serving, JSON repair, secret-data firewalls and bounded memory matter more than poetic prompting.
