Castuner

Best AI papers explained

Enoch H. Kang

Cut through the noise. We curate and break down the most important AI papers so you don’t have to.

Podcast website

Listen in your app

Episodes

859 episodes

  1. H-Spec: Parallel Speculative Decoding Without a Drafter-Side KV Cache 22:13
  2. Mass producing training data with surprising side effects 23:32
  3. User Embeddings are Superpositions of Interpretable Behavioral Modes 24:00
  4. Training Foundation Models with a Single Algorithm 23:25
  5. Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights 23:28
  6. Quantifying Elicitation of Latent Capabilities in Language Models 20:57
  7. The Trinity: Capability-Driven Recursive Self-Improvement for Multimodal Agents 19:55
  8. High-Value Training Data from Noisy, Diverse Sources 20:18
  9. Predictable Performance in Model Routing and Cascades 21:42
  10. Asymmetric On-Policy Distillation Learns In-Context Exploration 23:20
  11. Generalization Dynamics of Language Model Pretraining 22:32
  12. Scaling Discovery Requires Forgetting 21:25
  13. One-Sided Verification Under Limited Expert Review 23:05
  14. Scaling Empowered Agents 24:49
  15. Hybrid Reasoning Agents 20:13
  16. Strategic Self-Consistency 24:55
  17. Retrofitting language models to operate over bytes 24:02
  18. Priced Guidance: Can Language Models Generate Future Research Ideas? 20:20
  19. Context Language Models 21:32
  20. Position: AI-Agent Pricing Should Become More Outcome-Dependent: An Economic Perspective 22:30
  21. Emergent Capabilities Arise Randomly from Learning Sparse Attention Patterns 23:59
  22. Pragmatic Causal Inference with AI-Learned Representations 22:58
  23. Pointwise or pairwise: when do pairwise losses help reward learning, provably? 22:06
  24. When to Trust the Cheap Check: Weak and Strong Verification for Reasoning 24:47
  25. Theoretical Limits of Language Model Alignment 24:22
  26. DT2: Decision-Targeted Digital Twins 12:49
  27. Self-Play Pretraining with Zero Data 21:48
  28. Language models need sleep: learning to self-modify and consolidate memories 21:20
  29. Jev Creator: System One models for Prod, not God 21:50
  30. Detecting and countering misuse of AI: September 2026 22:39
  31. Position: LLMs can’t jump 20:55
  32. Jailbreaking Jailbreaks: A Proactive Defense for LLMs 22:46
  33. When Agents Slow Down: Understanding LLM Agents’ Test-Time Strategies via Elo-per-token Analysis 22:27
  34. Thinking with Looped Flows 20:04
  35. Multi-Turn LLM Conversations under the Least-Recently-Used Policy: Mean-Field Asymptotics 22:15
  36. Breaking the Token Ceiling: Distilling Smaller, Stronger Byte Models 23:57
  37. Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs 21:21
  38. Tail-Likelihood Reinforcement Learning 21:04
  39. Next-Latent Prediction Transformers Learn Compact World Models 22:34
  40. Language Models Can Control Their Own Attention 20:19
  41. AI Finds A Way 26:15
  42. Superposition Without Interference? Towards Isolated Interventions via Almost Orthogonal Features in Language Models 23:59
  43. TTPO: Test-Time Policy Optimization 23:00
  44. Demystifying Reinforcement Learning Post-Training of Language Models 20:21
  45. Recursive Experiential–Working Memory Evolution for Long-Horizon Agent Harnesses 20:48
  46. TailSFT: Filtered Fine-Tuning Improves Post-Training Performance 22:11
  47. SPADE: Self-Play in Adaptive Synthetic Executable Environments 21:46
  48. Evaluating Skills, Not Just Agents: Agentic Continuous Evaluation of Skills 27:31
  49. Impression Share Prediction: An Offline Evaluation Task for Ranking Systems 22:16
  50. Q-Learning with World Models 24:42