

530 episodes
by Enoch H. Kang · English
Cut through the noise. We curate and break down the most important AI papers so you don’t have to.

Recurus architecture for long-horizon autonomous agents, addressing “omitted actions” where models claim completion without executing tool/API steps.
31 Aug 2026 · 21m · 12 chapters
TailSFT (Filtered Fine-Tuning) fixes the “SFT trap” where standard supervised fine-tuning collapses probability mass onto easy majority answers, hurting later reinforcement learning (RL) that needs…
30 Aug 2026 · 22m · 10 chapters
SPADE (Self-Play in Adaptive Synthetic Executable Environments) proposes self-play training where one LLM acts as both environment designer and reasoning agent, generating infinite executable Python…
29 Aug 2026 · 22m · 10 chapters
The episode argues that evaluating AI “skills” via static checks (format/LLM-as-judge/linters/security scanners) is unsafe because skills aren’t executed; instead it promotes ACS (Agentic Continuous…
27 Aug 2026 · 28m · 13 chapters
How Meta researchers improve offline evaluation of ranking/recommendation algorithms by predicting “impression share shift” across objective buckets (clicks, video views, conversions) before…
25 Aug 2026 · 22m · 16 chapters
QWM (Q-Learning with World Models) for faster, safer robotic learning by combining model-free Q-learning with short-horizon world-model lookahead to avoid compounding simulation bias.
23 Aug 2026 · 25m · 15 chapters
Conformal language modeling via posterior sampling, a MIT statistical method to reduce LLM hallucinations during text generation by conditioning token sampling on a calibrated “high-trust” region,…
20 Aug 2026 · 22m · 10 chapters
Explains why RLHF reward models trained with pairwise rankings (Bradley-Terry) fail under adversarial distribution shift, enabling reward hacking, and presents Bon Voyage, a method to train reward…
20 Aug 2026 · 22m · 10 chapters
Why AI “agent skills” (distilled procedural checklists) improve performance, and when they fail.
18 Aug 2026 · 20m · 12 chapters
Epistemic stability of AI “judges” (LLMs used for grading, moderation, and reward modeling) under silence, pressure, and persistence; argues that standard accuracy tests are misleading because models…
15 Aug 2026 · 20m · 11 chapters
The episode explains the “Data Manifold Oracle” (DMO), a training-free method to predict neural scaling laws (the “floor” error limit and the “slope” learning rate) from raw text using standard…
15 Aug 2026 · 22m · 12 chapters
A July 2026 research framework for making interpretable, bias-resistant discoveries from unstructured text using sparse autoencoders, high-dimensional multiple-hypothesis testing (KFWER), Gaussian…
11 Aug 2026 · 24m · 14 chapters
The “incentive collapse paradox” in AI-human workflows: if pay depends only on final accuracy, near-perfect AI makes workers rationally free-ride (exert less effort), potentially requiring infinite…
11 Aug 2026 · 21m · 11 chapters
Continual learning agents that learn over a lifetime without catastrophic forgetting, using modular memory (core model + working memory + long-term memory) and two regimes: external interaction and…
10 Aug 2026 · 27m · 14 chapters
Harness RL for meta-learning/self-improvement at test time by moving adaptation from updating model weights to revising the “harness” (instructions, memory, tool rules, verification).
8 Aug 2026 · 22m · 9 chapters
The episode argues that large language models often fail in strategic settings because they assume opponents are perfectly rational “Nash-type” optimizers, creating “Nash traps” where the…
7 Aug 2026 · 22m · 10 chapters
When LeJEPA (Joint Embedding Predictive Architecture) learns a reliable “world model,” i.e., linearly identifiable latent variables from pixel observations, and when that guarantee fails in real…
7 Aug 2026 · 23m · 17 chapters
The episode explains “role drift” in compound LLM systems trained with outcome-only reinforcement learning, where modules abandon their intended roles while still achieving high final accuracy.
3 Aug 2026 · 21m · 13 chapters
Whether online reinforcement learning fine-tuning needs a pre-trained Q-function (critic) and why naive offline Q pretraining can hurt.
1 Aug 2026 · 24m · 10 chapters
The episode argues that traditional keyword “blue link” search is dying and being replaced by AI-native delegated decision-making, where consumer agents interpret intent, evaluate options invisibly,…
29 Jul 2026 · 19m · 8 chapters
How to evaluate LLM outputs reliably without opaque “holistic” scores, using binary yes/no checklists (Binavol) to catch specific errors and enable automated self-correction.
28 Jul 2026 · 5m · 4 chapters
How reasoning emerges from the standard AI pipeline (pretraining + supervised fine-tuning + reinforcement learning), using chess and math as controlled testbeds; includes scaling laws, compute…
24 Jul 2026 · 22m · 12 chapters
Whether LLM “self-explanations” (the reasons they give alongside answers) are faithful to their actual internal decision logic, and whether explanations improve prediction of future behavior.
23 Jul 2026 · 15m · 9 chapters
How inference-time “parallel reasoning” works for language models using sequential Monte Carlo (SMC) with process reward models (PRMs), including math guarantees and why theory can fail on real…
19 Jul 2026 · 23m · 11 chapters