AI Briefing
TLDR AI news
The latest AI and developer news from TLDR AI, with short summaries.
Visit website →Jury Dismisses All Claims in Musk's Lawsuit Against OpenAI CEO Sam Altman
A California jury dismissed all claims in Musk's lawsuit against Sam Altman.
Generalization Dynamics of LM Pretraining
During pretraining, LMs repeatedly oscillated between pattern imitation and generalization.
HRM-Text (GitHub repository)
HRM-Text has released a method for pretraining a 1B model for about $1000.
Political censorship revealed inside an LLM's weights
Qwen3.5-9B's political censorship was found to be a small circuit that decomposes into 3 axes.
Qwen3.7 Preview Launches on Arena
Qwen3.7 Preview launched on Arena's Text and Vision categories, ranking 13th and 16th respectively.
NVIDIA Delivers First Vera CPUs for Agentic AI to Major AI Labs
NVIDIA has delivered its first Vera CPUs for agentic AI to major AI labs and OCI.
Skills launched for web, iOS, and Android
Grok Skills has launched on web, iOS, and Android, supporting document automation and personalized workflows.
Fine-tuning NVIDIA Cosmos Predict 2.5 with LoRA/DoRA for Robot Video Generation (9 min read)
Fine-tuning NVIDIA Cosmos Predict 2.5 with LoRA and DoRA achieves efficient robot video generation in a single-GPU environment.
Experimenting Better with LLM Evals: The Funnel Is Not a Fork in the Road
LLM evals should be used as a pre-experiment validation funnel, not as a replacement for A/B testing.
Model Half-Life
Analyzing the release cadence of major AI models showed that deployment speed accelerating past the 'half-life' framing was the more accurate finding.
Headroom GitHub Repository
Headroom is a local-first compression layer that reduces AI agent input by 60-95% while being able to restore the original.
Gemini App Adds 'Extended' Reasoning Level and Third-Party App Integrations
The Gemini app is expanding its reasoning level options and third-party app integrations.
Agent Evaluation: A Detailed Guide
It summarizes how to design agent evaluation harnesses that go beyond static benchmarks.
Cursor Releases Composer 2.5
Cursor has released Composer 2.5, a coding agent built with reinforcement learning, synthetic data, and new distributed training techniques.
Apple Silicon Costs More Than OpenRouter
OpenRouter costs about 1/3 as much as Apple Silicon and is 2x faster, based on comparable models.
Token Economics: The 62.5-Minute Rule for Claude Caches
It's advantageous to keep a Claude cache alive if you plan to write to it again within 62.5 minutes.
Winners and Losers of the AI Gold Rush
As the AI boom builds wealth for a few, anxiety and layoffs are growing for the many.
Notes on Pretraining Parallelization and Failed Training Runs
It identified broken causality and numerical bias as the core causes of pretraining failures.
Codex to enable control of other desktop devices via Computer Use
OpenAI is expanding Codex's Computer Use to remote desktop control.
DeepSeek-V4-Flash, why LLM steering is interesting again
The emergence of DeepSeek-V4-Flash is bringing renewed attention to LLM steering, a technique for manipulating model activations.