loss

Definition

Plain language

A number that measures how wrong a model's outputs are, which training tries to make smaller.

Why it matters: Loss is the dial the optimizer actually moves, so understanding what's in it is essential to understanding what a model is learning.

For example, a language model's cross-entropy loss of 2.3 on a held-out batch means it's assigning, on average, that much surprise to each true next token.

Heard on the show

“… Among the patches the search returns, one hacks at less than half the baseline rate with no loss in real task performance — while the top raw performer actually hacks a bit more, which is why the …”

Mentioned in 39 episodes

1. 199
   Finding a Model's Hidden Behaviors Without Knowing What You're Looking For
2. 198
   The Model That Knows the Answer and Can't Say It
3. 197
   Twin Problems Suggest AI Reasoning Gains Are Mostly Better Fact Recall
4. 181
   How to Backpropagate Blame Through a Team of Chatbots — And When It Backfires
5. 173
   The Free Step-Level Grader Hiding in Every RL Training Run
6. 167
   How Teaching an AI to Predict, Not Act, Made It a Better Actor
7. 165
   A Free-Lunch Tweak That Lets a Tiny Agent Beat Frontier Giants
8. 163
   Why Training Only on Perfect Solutions Cripples a Model's Reasoning
9. 153
   Catching a Lie From the Inside, When the Words Look Completely Honest
10. 148
    Why Letting an AI Watch Its Own Scoreboard Can Quietly Overwrite Its Safety
11. 141
    How Two Tokens Reopened a Reasoning Method the Field Had Given Up On
12. 139
    When Optimizing One GPU Kernel Quietly Breaks the Whole System
13. 133
    How MiniMax Turned a Reward-Hacking Disaster Into Olympiad Gold
14. 131
    Why Autonomous Research Agents Forget Their Own Lessons, and Arbor's Fix
15. 128
    How a Model Can Earn Full Reward and Still Resist Training
16. 127
    What Diffusion Language Models Were Missing: A Map, Not an Algorithm
17. 115
    Teaching a Phone Agent to Reason Silently, And Keeping It Honest
18. 108
    The Reasoning Cliff: Why Thinking Longer Makes Models Worse at Exact Step-by-Step Tasks
19. 106
    Giving Agents a Notebook Instead of New Weights: How ExpGraph Lets Frozen Models Learn
20. 098
    Finding Millions of Readable Concepts Inside a Real, Deployed AI Model
21. 095
    Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search
22. 093
    A Calibrated Knob for Weak-to-Strong AI Oversight, Tested on Real Code
23. 087
    When No Agent Reads the Whole Document: A Universal Cliff in Multi-Agent Review
24. 084
    Terminal Agents Get Free Supervision From The Tokens We've Been Throwing Away
25. 080
    How a Two-Agent Trick Unlocked Large-Scale Training for Computer-Use Agents
26. 077
    Reading a Model's Confidence Curve to Decide When Chain-of-Thought Is Worth It
27. 074
    How a Fifteen-Hundred-Dollar Training Run Matched Llama and Gemma on Reasoning
28. 070
    When Models Know the Answer But Say the Wrong Thing Anyway
29. 060
    When Splitting One Model Across Three Agents Doubles Its Accuracy
30. 053
    An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script
31. 043
    When 'This Is False' Doesn't Stick: Why Models Learn the Lie Anyway
32. 041
    When the Iteration Teaches the Model to Skip the Iteration
33. 040
    Two Frozen Models Learn to Whisper: Coupling Through Hidden States
34. 038
    How LLMs Get Persuaded: One Attention Head, A Tetrahedron, And A Single Dial
35. 032
    A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking
36. 026
    What RL Actually Does to Language Models, at the Token Level
37. 025
    The Missing Gradient Term That Predicts Sycophancy in RLHF
38. 009
    How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers
39. 001
    When AI Models Quietly Protect Each Other From Shutdown

Related concepts

Related terms