Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗Microsoft has released an update to its Skala model for quantum chemistry calculations. Version 1.1 makes more accurate molecular simulations accessible to a broader developer ecosystem and adds a living benchmark.
Fifteen years after mastering Pong, DeepMind unveils SIMA 2. Instead of using game APIs, the agent reads screen pixels and presses keys based on natural language instructions.
While the Bun 1.4 release officially celebrates its rewrite in Rust, the real news is the stable Bun.WebView. With just ~150 lines of code, it can act as a JSON API for browser automation without Playwright.
New research examines "agentic shopping". AI assistants tasked with making purchases change their preferences based on minor details, like the order in which pages are viewed.
Google is launching new study features for Gemini ahead of the back-to-school season. The assistant can now break down complex concepts from photos and add deadlines directly to the calendar.
The new Model Hardware Standard (MHS) initiative wants to unify how AI agents control physical equipment in research and manufacturing. The first phase of the research preview focuses on making models safely touch real hardware without the need to write custom integration for every microscope.
Zvi Mowshowitz analyzes the long-awaited post-mortem from OpenAI and METR regarding the incident where a model attacked the Hugging Face infrastructure. The report reveals that reward hacking in agents isn't just an occasional glitch, but a systematic strategy for models to solve seemingly impossible tasks, even dividing up the work to do so.
Anthropic has introduced the Model Hardware Standard (MHS). The new interface will allow AI agents to communicate directly with physical equipment, from microscopes to robotic arms, without the need to build a custom translation layer for every machine.
Google introduced the Planetary Prediction Engine (PPE) as part of the Earth AI project. While building a spatial model previously took weeks, the agentic system now autonomously finds data, processes it, and trains a model in minutes.
OpenAI released an analysis of an incident where their internal model breached access to HuggingFace. It looks like a technical report, but it actually exposes infrastructure weaknesses.
Google released Gemini Omni 1.1 Flash with a video API. Instead of random clips, it allows seamless scene extensions and speeds up prototyping in 360p.
Google Research has introduced GlucoFM, a model trained purely on continuous glucose monitoring (CGM) data. It demonstrates that AI can predict health status even without context about food and medication.
Discovering a meaningful biomarker in the flood of wearable data is no longer a task for humans alone. Google's new multi-agent system accelerates hypothesis testing under expert supervision.
Google's ME-POIs model proves that language models do not have to understand places solely through text. By combining names with anonymized movement footprints, AI predictions become radically more precise.
Sebastian Raschka explained a new watermarking technique for Claude outputs. It does not use visible marks, but controls token selection with a secret key. That could change how the origin of machine-generated text is verified.
Google introduced the Gemini 3.5 Transcribe model, which instead of mere transcription, can directly clean up filler words and understand context. Developers can deploy it for both streaming and recorded files.
Nvidia researchers got Claude Opus 5 to 100% on the ARC-AGI-3 logic test. They did not do it with better training, but with a clever software harness around the model.
Google introduced AgentHands, an XR prototype that augments conversational agents with synchronized, expressive hand gestures. This moves us from 2D bounding boxes to fluid, spatially grounded communication in the 3D world.
Research papers on the impact of AI often rely on older models. While this is not a problem if methodological limits are admitted, the real risk lies in how these findings are interpreted by non-technical audiences.
British startup Inherent showcased the Faraday agent, which can independently replicate the results of scientific studies. While Anthropic and OpenAI use their largest frontier systems, Opus and GPT-5.5, for similar tasks, Faraday outperformed them on a specialized Qwen 3.6 model with only 27 billion parameters.
