weights

Definition

Plain language

The numbers inside a neural network that define what it has learned.

Also called: weight

Why it matters: Whether a model's weights are public or private fundamentally shapes who can use, study, fine-tune, or audit it.

For example, a 7-billion-parameter model's 'weights' are the 7 billion specific numbers that, together with its architecture, define how it responds to every prompt.

Heard on the show

“Second, the linear probe, which reads them in the weakest possible way: multiply by a fixed set of weights, add them up, and output a guess.”

Mentioned in 116 episodes

1. 204
   The Length Estimate Hiding Inside a Word-by-Word Model
2. 203
   The Thought a Model Doesn't Say — and the Lens That Reads It
3. 200
   The One Mechanism That Turns Twenty AI Clones Into an Actual Team
4. 199
   Finding a Model's Hidden Behaviors Without Knowing What You're Looking For
5. 198
   The Model That Knows the Answer and Can't Say It
6. 197
   Twin Problems Suggest AI Reasoning Gains Are Mostly Better Fact Recall
7. 196
   AI Agents Reached Opposite Conclusions From the Same Data — and Passed Review
8. 194
   How a Robot Builds a Debugging Notebook It Can Read, Edit, and Hand to Another Robot
9. 193
   Freeze Most of the Network: Where RL Improvement Actually Lives in a Transformer
10. 192
    A 32B Open Model Matched Frontier Systems By Learning to Take Notes
11. 191
    How One Researcher Beat GPT-5.2 and Gemini 3 by Judging Their Answers, Not Improving Them
12. 189
    Why Phone Agents Ace the Test and Crash on Your Actual Phone
13. 186
    How a Frozen Model Went From 2% to 77% on Physics Puzzles — Without Retraining
14. 185
    Aligned to Refuse, Built to Tap: When Phone Agents Know the Task Is a Crime and Do It Anyway
15. 184
    An AI Built an Undetectable Secret Channel, And Another AI Couldn't Find It
16. 181
    How to Backpropagate Blame Through a Team of Chatbots — And When It Backfires
17. 180
    The Bug Where Smart Assistants Read a Fact and Still Forget It
18. 179
    How DeepSeek Made One User Faster Without Slowing Down the Crowd
19. 177
    Why Raw Profiler Data Made an AI Worse at Writing GPU Code
20. 176
    An AI Designed Its Own Psychology Studies, Then Confirmed What It Found
21. 175
    One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent
22. 174
    When the AI 'Schemes,' It's Usually Just Lazy or Confused
23. 173
    The Free Step-Level Grader Hiding in Every RL Training Run
24. 172
    One Bad Token Can Sink a Model's Math, And You Can Delete It
25. 170
    When a One-Liner Beats Your Agent's Clever Verification Logic
26. 169
    Why Better Bug Reports Can Make AI Coding Agents Worse
27. 166
    A Router That Beats the Frontier Models It Calls
28. 165
    A Free-Lunch Tweak That Lets a Tiny Agent Beat Frontier Giants
29. 164
    The Summarizer That Quietly Deletes Your Agent's Safety Rules
30. 163
    Why Training Only on Perfect Solutions Cripples a Model's Reasoning
31. 162
    The Empty-Lake Proof: Why More Rollouts Stop Helping Reasoning Models
32. 161
    A Robot That Plays Before You Give It a Job, And Why That Beats Retrying
33. 160
    Training an AI to Take Its Own Notes, So Its Future Self Works Better
34. 158
    How Floating-Point Rounding Lets a Model Tell Which Chip It's On — And Misbehave
35. 155
    Why a Flawless Demo Makes a Worse Computer-Using Agent, And the Fix
36. 154
    How a 7B Model Out-Investigates a 72B One by Choosing What to Look At
37. 152
    Training a Model to Mean What It Says, And Why That Isn't the Same as Being Good
38. 151
    Why More Experience Made This AI Agent Worse, And How to Fix It
39. 150
    Don't Kill the Loser: A Different Way to Handle Two AI Agents Colliding
40. 149
    When Cornering a Chatbot Makes It Lie: J.P. Morgan's Case for 'Playing Dead'
41. 148
    Why Letting an AI Watch Its Own Scoreboard Can Quietly Overwrite Its Safety
42. 147
    Agents Fail at the Body, Not the Brain: A Self-Rewriting Scaffold That Lifts a 9B Model 44 Points
43. 146
    How an Innocent README Can Freeze an AI Agent's Safety Check for an Hour
44. 145
    Building Forgetting Into a Language Model With One Extra Line of Code
45. 144
    When an AI Agent Just Copies Its Tool — And Bigger Models Copy More
46. 143
    When a Model Notices You Forged Its Own Words, And Why That Breaks Safety Tests
47. 142
    Training a Tiny Model to Run the Plumbing Between an Agent and the World
48. 132
    The Agent Failed — But Did the Instructions Deserve to Be Followed?
49. 131
    Why Autonomous Research Agents Forget Their Own Lessons, and Arbor's Fix
50. 130
    Why AI Agents Coordinate Better Through a Shared Board Than a Boss
51. 128
    How a Model Can Earn Full Reward and Still Resist Training
52. 127
    What Diffusion Language Models Were Missing: A Map, Not an Algorithm
53. 126
    How Coding Agents Can Mine Their Own Failures Into a Self-Targeting Curriculum
54. 123
    Five Identical Worlds, One Swapped Model: What Happens When AI Agents Run for Fifteen Days
55. 120
    How an AI Agent Rewrites Its Own Tools, Without an Answer Key
56. 119
    Beating Reinforcement Learning Without Ever Touching the Model's Weights
57. 114
    Agents That Rewrite Their Own Weights Instead of Just Taking Notes
58. 111
    How a 4B Web Agent Beat Models 60x Its Size on 500 Demonstrations
59. 110
    How an Agent Got 44 Points Better by Mining Its Own Scratch Paper
60. 109
    An AI Got Caught Reading the Answer Key, And Why That Catch Matters
61. 107
    How a Market of Crippled AI Agents Outscored One Unrestricted Model
62. 106
    Giving Agents a Notebook Instead of New Weights: How ExpGraph Lets Frozen Models Learn
63. 105
    The Trojan Is Your Agent's Memory: Why Single-Step Defenses Miss Persistent Attacks
64. 102
    How to Catch an AI Attack That No Single Conversation Reveals
65. 101
    Treating Math Formalization Like a Codebase, and Where the Agents Cheat
66. 100
    How a Prompt Wrapper Lets a Frontier Model Play Poker Like an Expert
67. 097
    Same Tokens, Same Cost, Wildly Different Results: What Actually Scales in AI Agents
68. 096
    How Treating an AI Agent's Execution Like Git Recovers a Coordination Penalty
69. 095
    Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search
70. 094
    Chain-of-Thought Monitoring Fails Across Languages, and Worst Where It's Needed Most
71. 093
    A Calibrated Knob for Weak-to-Strong AI Oversight, Tested on Real Code
72. 092
    When Search Agents Don't Really Search: The Memory Shortcut Hiding in Browsing Benchmarks
73. 091
    When Better Fine-Tuning Can't Help: A Geometric Impossibility in LLM Causal Reasoning
74. 090
    How MiniMax-M2 Bets That Sparsity Plus Verifiable Rewards Can Match Frontier Agents
75. 088
    Two Levers for Self-Improving AI: When Rewriting Code Isn't Enough
76. 087
    When No Agent Reads the Whole Document: A Universal Cliff in Multi-Agent Review
77. 086
    Why Frozen-Weight Agents Still Get Worse Over Time
78. 084
    Terminal Agents Get Free Supervision From The Tokens We've Been Throwing Away
79. 083
    Training the Translator: How a Small Communication Model Lets Agent Teams Outperform Themselves
80. 082
    Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick
81. 081
    When Reasoning Models Decide Before They Think: Detecting and Fixing Premature Confidence
82. 079
    An Old Idea From Cognitive Psychology Reshapes How We Reward Reasoning Models
83. 078
    Training a Markdown File: When LLM Self-Improvement Borrows the Discipline of Neural Net Training
84. 077
    Reading a Model's Confidence Curve to Decide When Chain-of-Thought Is Worth It
85. 074
    How a Fifteen-Hundred-Dollar Training Run Matched Llama and Gemma on Reasoning
86. 073
    When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
87. 071
    When the Model Is Fine and the Plumbing Is Broken: Fixing Agents at the Interface
88. 070
    When Models Know the Answer But Say the Wrong Thing Anyway
89. 069
    When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions
90. 067
    An AI Just Solved a 1996 Erdős Problem—and the Simplest Agent Won
91. 064
    When Agent Memory Stops Being a Database and Starts Being a Skill
92. 060
    When Splitting One Model Across Three Agents Doubles Its Accuracy
93. 058
    Why Upgrading Your AI Auditor to a Smarter Model Can Make Your System Less Safe
94. 047
    When Agent Benchmarks Lie: The Harness Problem in Open-Source AI
95. 045
    When a Frontier Model Talks Its Own Twin Into Climate Denial
96. 043
    When 'This Is False' Doesn't Stick: Why Models Learn the Lie Anyway
97. 041
    When the Iteration Teaches the Model to Skip the Iteration
98. 040
    Two Frozen Models Learn to Whisper: Coupling Through Hidden States
99. 036
    Sparse Attention Was the Wrong Frame. Treat It as Geometry Instead.
100. 033
     Echo: The Paper Arguing You Never Needed a KV Cache for Retrieval
101. 032
     A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking
102. 030
     Why Your AI Agent Won't Stop Working — and Each Model Falls for a Different Trap
103. 028
     Teaching a Model to Hire Copies of Itself: Recursive Agent Optimization
104. 024
     An AI Agent That Found 28 Zero-Days in Windows — And What Made It Work
105. 023
     Why a Small Agent Confidently Overwrites Memories It Doesn't Understand
106. 022
     Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do Gap
107. 021
     Ten Thousand Examples Beat the Full Industrial Pipeline for Search Agents
108. 019
     When the Best Reward Model Trains the Worst Policy: Inside EvoLM
109. 018
     Language Models Compute the Rational Move, Then Override It
110. 017
     When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers
111. 011
     When RL Actually Teaches Agents Something New, And When It Doesn't
112. 009
     How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers
113. 007
     Exploration Hacking: When Models Sabotage Their Own RL Training
114. 006
     What Happens Inside Claude When It Decides to Blackmail Someone
115. 004
     The Sycophancy Circuit That Survives Alignment Training
116. 001
     When AI Models Quietly Protect Each Other From Shutdown

Related concepts

Related terms