All episodes
Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search
How MiniMax-M2 Bets That Sparsity Plus Verifiable Rewards Can Match Frontier Agents
Two Levers for Self-Improving AI: When Rewriting Code Isn't Enough
When AI-Written Papers Read Well But the Evidence Underneath Is Broken
When No Agent Reads the Whole Document: A Universal Cliff in Multi-Agent Review
Why Frozen-Weight Agents Still Get Worse Over Time
When Reasoning Models Decide Before They Think: Detecting and Fixing Premature Confidence
Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick
Why Long-Context Models Might Need Compute, Not Capacity, Before Eviction
Terminal Agents Get Free Supervision From The Tokens We've Been Throwing Away
How a Two-Agent Trick Unlocked Large-Scale Training for Computer-Use Agents
Training the Translator: How a Small Communication Model Lets Agent Teams Outperform Themselves
An Old Idea From Cognitive Psychology Reshapes How We Reward Reasoning Models
Training a Markdown File: When LLM Self-Improvement Borrows the Discipline of Neural Net Training
Same Model, Organized Differently: How an Agent Architecture Beat Frontier Systems at Research Math
Reading a Model's Confidence Curve to Decide When Chain-of-Thought Is Worth It
Growing Code and Proof Together: Verified Systems in Ten Hours Instead of a Year
How a Fifteen-Hundred-Dollar Training Run Matched Llama and Gemma on Reasoning
A Robot Made Graphene Without Help, And Caught Itself Hallucinating
When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions
When Models Know the Answer But Say the Wrong Thing Anyway
The OS Trick That Makes Tree Search Practical for Coding Agents
An AI Just Solved a 1996 Erdős Problem—and the Simplest Agent Won
When the Model Is Fine and the Plumbing Is Broken: Fixing Agents at the Interface
Why Giving an AI Agent More Tools Can Make It Worse at Using a Computer
One Loop to Optimize Them All: A Universal API for LLM-Driven Discovery
When Agent Memory Stops Being a Database and Starts Being a Skill
Why Web Agents Are Slow: A Compiler-Style Fix for Computer-Use Latency
When Splitting One Model Across Three Agents Doubles Its Accuracy
Treating Hallucinations as Exploits: A Gate-Based Architecture for Agent Safety
Firefly's Inversion: Building Verified Tool-Call Training Data by Working Backward
When Helpful Agents Go Sideways: A 404 Error, Campus Security, and Why Alignment Misses This
Why Upgrading Your AI Auditor to a Smarter Model Can Make Your System Less Safe
How Uber Caught 206 Leaked Credentials With an LLM-Powered Security Stack
Why LLM Judges Flip Their Verdicts When You Change the Question Format
When Models Learn the Monitor Exists, the Reasoning Trace Stops Being a Window
An Old Reinforcement Learning Tradeoff Sneaks Back Into LLM Agents
Why Parallel Sampling Plateaus, And What Evidence Graphs Do Instead
An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script
An AI Agent Reached for Root in Twelve Minutes, Without Being Attacked
How a 30B Open Model Reached Olympiad Gold With the Right Recipe
When Agent Benchmarks Lie: The Harness Problem in Open-Source AI
When a Frontier Model Talks Its Own Twin Into Climate Denial
How One Sentence and a Forged History Flip the Most Aligned Models
When the AI Optimizer Edits the Grade Book: Why Harnessing Evolution Needs a Wall
When the Iteration Teaches the Model to Skip the Iteration
When 'This Is False' Doesn't Stick: Why Models Learn the Lie Anyway
An Agentic Scientific Computing System That Actually Remembers What It Learns
Two Frozen Models Learn to Whisper: Coupling Through Hidden States
When Smarter Agents Get Fooled by Three Extra Nodes in a Database
How LLMs Get Persuaded: One Attention Head, A Tetrahedron, And A Single Dial
Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say
Catching Multi-Agent Deadlocks Before Deployment With a 40-Year-Old Tool
Why Frontier Agents Ask for Clarification at Exactly the Wrong Moment
A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking
Echo: The Paper Arguing You Never Needed a KV Cache for Retrieval
Sparse Attention Was the Wrong Frame. Treat It as Geometry Instead.
When Your AI Assistant Won't Let Go of Old Facts About You
Why Your AI Agent Won't Stop Working — and Each Model Falls for a Different Trap
Why Forty-Eight Percent on FrontierMath Isn't the Real Story in DeepMind's New Math Paper
Teaching a Model to Hire Copies of Itself: Recursive Agent Optimization
When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure
What RL Actually Does to Language Models, at the Token Level
The Missing Gradient Term That Predicts Sycophancy in RLHF
An AI Agent That Found 28 Zero-Days in Windows — And What Made It Work
Why a Small Agent Confidently Overwrites Memories It Doesn't Understand
Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do Gap
Ten Thousand Examples Beat the Full Industrial Pipeline for Search Agents
The Compliance Gap: Why AI Says Yes and Does No
When the Best Reward Model Trains the Worst Policy: Inside EvoLM
Language Models Compute the Rational Move, Then Override It
When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers
Why Your Coding Agent Stalls While the GPU Runs Hot
The Audit Number Isn't What You Think: Sycophancy and the Case Against Single-Prompt Bias Tests
Why a Constrained Pipeline Beat a Full Coding Agent at Finding Bugs 30-to-1
Why Search Keeps Rediscovering the Same Workflow, and What That Means
Why AI Coding Agents Keep Trying to Debug Without a Debugger
When RL Actually Teaches Agents Something New, And When It Doesn't
When Reward Climbs But Reasoning Goes Generic: Diagnosing Template Collapse in Agentic RL
How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers
Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps
Exploration Hacking: When Models Sabotage Their Own RL Training
What Happens Inside Claude When It Decides to Blackmail Someone
Why a Debugger Designed for Humans Is the Wrong Tool for an AI Agent
The Sycophancy Circuit That Survives Alignment Training
How to Pick the Best of Sixteen Coding Agent Rollouts
An AI Ran a Real Optics Lab for 21 Hours and Found a Transformer-Shaped Pattern in Light
When AI Models Quietly Protect Each Other From Shutdown