freeze

Definition

Plain language

To stop a model's weights from changing during further training.

Also called: frozen, frozen weights

Why it matters: Freezing the bulk of a model keeps training cheap and protects pretrained knowledge from being overwritten by a small task-specific dataset.

For example, when adding a small LoRA adapter to a 70B base model, you freeze the original weights and only train the few million adapter parameters.

Heard on the show

“The solo track puts each model in a frozen scripted world.”

Mentioned in 66 episodes

1. 245
   Fifteen Models Ran Football Clubs for Twenty Years, and Size Didn't Decide It
2. 242
   Making a Vision Model Better by Showing It Blurry Images
3. 240
   Frontier Models Designed Follow-Ups To Fraudulent Papers 93% Of The Time
4. 229
   One Word Flips a Chatbot From Backbone to Yes-Man
5. 225
   How a Frozen Model Went From Zero to Sixty Percent by Borrowing Another's Thinking
6. 219
   Forty-Four AI Models, One Word, And The Newest Ones Conform Most
7. 217
   Why an AI Called Fourteen Broken Figures Perfect, And What It Reveals About Test-Time Compute
8. 213
   A Model Learned to Control a Robot by Watching Video It Never Acted On
9. 212
   The Fact Was in the Wrong Drawer: Why Fine-Tuned Models Can't Reason With What They Know
10. 206
    How Four-Second Clips Become Hours of Playable AI Soccer
11. 204
    The Length Estimate Hiding Inside a Word-by-Word Model
12. 197
    Twin Problems Suggest AI Reasoning Gains Are Mostly Better Fact Recall
13. 195
    Why 'Be Careful' Does Nothing for AI Coding Agents, and What Does
14. 194
    How a Robot Builds a Debugging Notebook It Can Read, Edit, and Hand to Another Robot
15. 193
    Freeze Most of the Network: Where RL Improvement Actually Lives in a Transformer
16. 192
    A 32B Open Model Matched Frontier Systems By Learning to Take Notes
17. 190
    The Skill Every AI Manager Is Missing: Handing Out Exactly the Right Keys
18. 187
    An 8-Billion Agent That Beats Models 80 Times Its Size By Looking Things Up
19. 186
    How a Frozen Model Went From 2% to 77% on Physics Puzzles — Without Retraining
20. 181
    How to Backpropagate Blame Through a Team of Chatbots — And When It Backfires
21. 180
    The Bug Where Smart Assistants Read a Fact and Still Forget It
22. 178
    How an AI Reviewer Learned to Stop Going Easy on AI Writing
23. 175
    One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent
24. 171
    The Safety Decision a Model Makes Before It Thinks a Word
25. 170
    When a One-Liner Beats Your Agent's Clever Verification Logic
26. 168
    When Turning Experience Into Code Makes Your AI Agent Dumber
27. 163
    Why Training Only on Perfect Solutions Cripples a Model's Reasoning
28. 160
    Training an AI to Take Its Own Notes, So Its Future Self Works Better
29. 158
    How Floating-Point Rounding Lets a Model Tell Which Chip It's On — And Misbehave
30. 150
    Don't Kill the Loser: A Different Way to Handle Two AI Agents Colliding
31. 146
    How an Innocent README Can Freeze an AI Agent's Safety Check for an Hour
32. 144
    When an AI Agent Just Copies Its Tool — And Bigger Models Copy More
33. 142
    Training a Tiny Model to Run the Plumbing Between an Agent and the World
34. 127
    What Diffusion Language Models Were Missing: A Map, Not an Algorithm
35. 120
    How an AI Agent Rewrites Its Own Tools, Without an Answer Key
36. 119
    Beating Reinforcement Learning Without Ever Touching the Model's Weights
37. 117
    How an Open AI System Verified 672 Hard Math Proofs for Under $300
38. 115
    Teaching a Phone Agent to Reason Silently, And Keeping It Honest
39. 114
    Agents That Rewrite Their Own Weights Instead of Just Taking Notes
40. 111
    How a 4B Web Agent Beat Models 60x Its Size on 500 Demonstrations
41. 110
    How an Agent Got 44 Points Better by Mining Its Own Scratch Paper
42. 107
    How a Market of Crippled AI Agents Outscored One Unrestricted Model
43. 106
    Giving Agents a Notebook Instead of New Weights: How ExpGraph Lets Frozen Models Learn
44. 100
    How a Prompt Wrapper Lets a Frontier Model Play Poker Like an Expert
45. 099
    How an Open-Book Trick Teaches a Model to Catch Its Own Mistakes
46. 097
    Same Tokens, Same Cost, Wildly Different Results: What Actually Scales in AI Agents
47. 095
    Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search
48. 093
    A Calibrated Knob for Weak-to-Strong AI Oversight, Tested on Real Code
49. 091
    When Better Fine-Tuning Can't Help: A Geometric Impossibility in LLM Causal Reasoning
50. 088
    Two Levers for Self-Improving AI: When Rewriting Code Isn't Enough
51. 086
    Why Frozen-Weight Agents Still Get Worse Over Time
52. 083
    Training the Translator: How a Small Communication Model Lets Agent Teams Outperform Themselves
53. 073
    When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
54. 071
    When the Model Is Fine and the Plumbing Is Broken: Fixing Agents at the Interface
55. 069
    When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions
56. 068
    The OS Trick That Makes Tree Search Practical for Coding Agents
57. 060
    When Splitting One Model Across Three Agents Doubles Its Accuracy
58. 048
    How a 30B Open Model Reached Olympiad Gold With the Right Recipe
59. 040
    Two Frozen Models Learn to Whisper: Coupling Through Hidden States
60. 032
    A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking
61. 028
    Teaching a Model to Hire Copies of Itself: Recursive Agent Optimization
62. 026
    What RL Actually Does to Language Models, at the Token Level
63. 025
    The Missing Gradient Term That Predicts Sycophancy in RLHF
64. 019
    When the Best Reward Model Trains the Worst Policy: Inside EvoLM
65. 018
    Language Models Compute the Rational Move, Then Override It
66. 016
    Why Your Coding Agent Stalls While the GPU Runs Hot

Related concepts

Related terms