Glossary · Term

base model

← all terms

Definition

Plain language

An AI model fresh out of its initial training on raw text, before it's been shaped into a helpful assistant.

As stated in the literature

A pretrained language model prior to instruction tuning or RLHF; it continues text statistically but does not natively follow instructions, and serves as the starting point for post-training. Its disposition differs measurably from its instruction-tuned counterpart.

Also called: base models

Why it matters: It is the raw foundation that all the helpful, instruction-following behavior is later built on, so understanding it clarifies what comes from pretraining versus what comes from later shaping.

For example, ask a base model 'What is the capital of France?' and it might continue with more trivia questions rather than simply answering 'Paris,' because it was only trained to predict likely text.

Heard on the show

“The base model — before any human-feedback training — already has this workspace, fully formed.”
Episode 203 — The Thought a Model Doesn't Say — and the Lens That Reads It

Mentioned in 55 episodes

  1. 203
    The Thought a Model Doesn't Say — and the Lens That Reads It
  2. 193
    Freeze Most of the Network: Where RL Improvement Actually Lives in a Transformer
  3. 192
    A 32B Open Model Matched Frontier Systems By Learning to Take Notes
  4. 187
    An 8-Billion Agent That Beats Models 80 Times Its Size By Looking Things Up
  5. 186
    How a Frozen Model Went From 2% to 77% on Physics Puzzles — Without Retraining
  6. 175
    One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent
  7. 167
    How Teaching an AI to Predict, Not Act, Made It a Better Actor
  8. 163
    Why Training Only on Perfect Solutions Cripples a Model's Reasoning
  9. 157
    When an AI Coding Agent Drives a Phone Through the Terminal, No Screen Needed
  10. 156
    Why More Human Demonstrations Made a Computer-Use Agent Worse
  11. 154
    How a 7B Model Out-Investigates a 72B One by Choosing What to Look At
  12. 152
    Training a Model to Mean What It Says, And Why That Isn't the Same as Being Good
  13. 151
    Why More Experience Made This AI Agent Worse, And How to Fix It
  14. 133
    How MiniMax Turned a Reward-Hacking Disaster Into Olympiad Gold
  15. 130
    Why AI Agents Coordinate Better Through a Shared Board Than a Boss
  16. 121
    When the Agent Says It's Done But Nothing Happened: Debugging the Harness, Not the Model
  17. 119
    Beating Reinforcement Learning Without Ever Touching the Model's Weights
  18. 117
    How an Open AI System Verified 672 Hard Math Proofs for Under $300
  19. 115
    Teaching a Phone Agent to Reason Silently, And Keeping It Honest
  20. 111
    How a 4B Web Agent Beat Models 60x Its Size on 500 Demonstrations
  21. 107
    How a Market of Crippled AI Agents Outscored One Unrestricted Model
  22. 105
    The Trojan Is Your Agent's Memory: Why Single-Step Defenses Miss Persistent Attacks
  23. 104
    How Making a Research Agent Smarter Quietly Makes It Leak Your Secrets
  24. 100
    How a Prompt Wrapper Lets a Frontier Model Play Poker Like an Expert
  25. 099
    How an Open-Book Trick Teaches a Model to Catch Its Own Mistakes
  26. 097
    Same Tokens, Same Cost, Wildly Different Results: What Actually Scales in AI Agents
  27. 090
    How MiniMax-M2 Bets That Sparsity Plus Verifiable Rewards Can Match Frontier Agents
  28. 088
    Two Levers for Self-Improving AI: When Rewriting Code Isn't Enough
  29. 084
    Terminal Agents Get Free Supervision From The Tokens We've Been Throwing Away
  30. 082
    Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick
  31. 081
    When Reasoning Models Decide Before They Think: Detecting and Fixing Premature Confidence
  32. 079
    An Old Idea From Cognitive Psychology Reshapes How We Reward Reasoning Models
  33. 077
    Reading a Model's Confidence Curve to Decide When Chain-of-Thought Is Worth It
  34. 076
    Same Model, Organized Differently: How an Agent Architecture Beat Frontier Systems at Research Math
  35. 071
    When the Model Is Fine and the Plumbing Is Broken: Fixing Agents at the Interface
  36. 070
    When Models Know the Answer But Say the Wrong Thing Anyway
  37. 069
    When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions
  38. 067
    An AI Just Solved a 1996 Erdős Problem—and the Simplest Agent Won
  39. 066
    Why Giving an AI Agent More Tools Can Make It Worse at Using a Computer
  40. 060
    When Splitting One Model Across Three Agents Doubles Its Accuracy
  41. 043
    When 'This Is False' Doesn't Stick: Why Models Learn the Lie Anyway
  42. 040
    Two Frozen Models Learn to Whisper: Coupling Through Hidden States
  43. 028
    Teaching a Model to Hire Copies of Itself: Recursive Agent Optimization
  44. 026
    What RL Actually Does to Language Models, at the Token Level
  45. 023
    Why a Small Agent Confidently Overwrites Memories It Doesn't Understand
  46. 021
    Ten Thousand Examples Beat the Full Industrial Pipeline for Search Agents
  47. 019
    When the Best Reward Model Trains the Worst Policy: Inside EvoLM
  48. 018
    Language Models Compute the Rational Move, Then Override It
  49. 017
    When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers
  50. 013
    Why Search Keeps Rediscovering the Same Workflow, and What That Means
  51. 011
    When RL Actually Teaches Agents Something New, And When It Doesn't
  52. 009
    How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers
  53. 006
    What Happens Inside Claude When It Decides to Blackmail Someone
  54. 004
    The Sycophancy Circuit That Survives Alignment Training
  55. 003
    How to Pick the Best of Sixteen Coding Agent Rollouts

Related concepts

Related terms