TwIL-8B is open-sourced with 8.8 billion parameters and reaches 93% of frontier performance on coding benchmarks and 92% on instruction following. The post shares Andrej Karpathy’s five-point thesis on small LLMs and claims a 1B-parameter core can capture most intelligence.
Key points
TwIL-8B has 8.8 billion parameters.
It achieves 93% of frontier performance on coding benchmarks.
It achieves 92% of frontier performance on instruction-following benchmarks.
Karpathy’s thesis lists five arguments that a ~1B-parameter core can contain the cognitive core of LLMs.
Sources
Orbit Wars winner Isaiah Pressman top spot with 15B steps
Orbit Wars winner Isaiah Pressman took the top spot with reinforcement learning that simulated 15 billion steps. He discusses predictability risks in a bonus episode of Beyond the Leaderboard.
Key points
Isaiah Pressman took the top spot in Orbit Wars.
The reinforcement learning simulation ran 15 billion steps.
Sources
MaxxBench 89 DeepSWE tasks with small instruction changes
MaxxBench introduces 89 DeepSWE tasks with small changes to the instructions or starting code. Models clustered around 74% on DeepSWE show a 23 pp spread.
Key points
Tasks: 89 DeepSWE tasks with small changes to the instructions or starting code
Models: clustered around 74% on DeepSWE end up 23 pp apart
Sources
TristanTube – 47K videos, TV mode added
TristanTube launches with 47,000 curated videos and a TV mode for relaxed watching. The tool was built in three minutes using Claude Code. A separate HackerTube hack filtered 43 million comments to 6,454 recommended videos for a two-hour viewing session.
Key points
47,000 videos included, plus TV mode for laid-back watching.
Claude filtered 43 million comments to 6,454 videos, yielding ~2 hours of content.
Sources
- Original post
- My YouTube tier list, six years later - Tristan Hume - Linked page
- karpathy.ai - Linked in the post
Inkling 2 Small free limited-time access on OpenCode
Inkling 2 Small is now available for free on OpenCode for a limited time to collect usage data. The model ranks 16th in last week’s OpenCode usage with 496B tokens processed.
Key points
Free limited-time access announced for Inkling 2 Small on OpenCode.
Ranked 16th in last week’s OpenCode usage with 496B tokens processed.
Sources
- Original post
- Inkling 2 Small Usage, Cost & Rank - Linked page
Social Humanoid framework deployed on Unitree G1
Social Humanoid is a framework for socially responsive human–humanoid interaction, deployed on a Unitree G1. It combines high-level social reasoning with low-level control in a hierarchical system.
Key points
Framework: Social Humanoid
Platform: Unitree G1
Architecture: hierarchical system merges high-level social reasoning with low-level control
