Glossary · Term

kernel

← all terms

Definition

Plain language

A small, central piece of code that does one critical job — on a graphics chip it's a fast math routine, and in a proof checker it's the trusted core that verifies every step.

As stated in the literature

On a GPU, a compute routine implementing a primitive operation (e.g., a CUDA kernel); writing fast custom kernels is a frequent optimization target for self-improving agents and is benchmarked by KernelBench. In a proof assistant like Lean, the kernel is instead the small, heavily-scrutinized trusted core that mechanically checks each proof step. Distinct from an OS kernel.

Also called: kernels, CUDA kernel, GPU kernel

Why it matters: Faster kernels can dramatically speed up training and inference, making them a prime target for agents that optimize their own performance.

For example, a custom kernel might multiply two matrices on a GPU far faster than the default code by squeezing the most out of the chip.

Heard on the show

“There's a language called Lean 4 where you write math as code, and a small, paranoid core program — the kernel — checks every logical step against a fixed set of rules.”
Episode 188 — A Coding Agent Found a Hole in a Peer-Reviewed STOC Proof for Five Dollars

Mentioned in 19 episodes

  1. 188
    A Coding Agent Found a Hole in a Peer-Reviewed STOC Proof for Five Dollars
  2. 177
    Why Raw Profiler Data Made an AI Worse at Writing GPU Code
  3. 139
    When Optimizing One GPU Kernel Quietly Breaks the Whole System
  4. 125
    AI Coding Agents Run a Marathon, and Fewer Than One in Three Finish
  5. 124
    A Cheap Model With the Blueprints Beats Expensive Models Working Blind
  6. 107
    How a Market of Crippled AI Agents Outscored One Unrestricted Model
  7. 101
    Treating Math Formalization Like a Codebase, and Where the Agents Cheat
  8. 091
    When Better Fine-Tuning Can't Help: A Geometric Impossibility in LLM Causal Reasoning
  9. 088
    Two Levers for Self-Improving AI: When Rewriting Code Isn't Enough
  10. 075
    Growing Code and Proof Together: Verified Systems in Ten Hours Instead of a Year
  11. 068
    The OS Trick That Makes Tree Search Practical for Coding Agents
  12. 065
    One Loop to Optimize Them All: A Universal API for LLM-Driven Discovery
  13. 062
    Treating Hallucinations as Exploits: A Gate-Based Architecture for Agent Safety
  14. 053
    An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script
  15. 046
    When the AI Optimizer Edits the Grade Book: Why Harnessing Evolution Needs a Wall
  16. 036
    Sparse Attention Was the Wrong Frame. Treat It as Geometry Instead.
  17. 033
    Echo: The Paper Arguing You Never Needed a KV Cache for Retrieval
  18. 027
    When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure
  19. 007
    Exploration Hacking: When Models Sabotage Their Own RL Training

Related concepts

Related terms