

531 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.

The episode argues that LLM “black box” interpretability may be easier than current practice suggests.
8 Feb 2026 · 16m · 6 chapters
The episode argues that PPO (Proximal Policy Optimization) is poorly suited for LLM reinforcement learning because its trust-region mechanism uses token-level ratio clipping, which over-penalizes…
8 Feb 2026 · 16m · 8 chapters
Learning LLM personalization from user edits, reframing “frustrated rewriting” as training signal via edit cost and a late ensemble of SFT, DPO, and cost-based RL.
8 Feb 2026 · 14m · 7 chapters
Self-distillation fine-tuning (SDFT) for continual AI learning that avoids catastrophic forgetting by using the model itself as both “teacher” and “student,” turning in-context learning into…
7 Feb 2026 · 20m · 11 chapters
Maximum Likelihood Reinforcement Learning (MaxRL) argues standard reinforcement learning (RL) is only the first term in a Maclaurin-series expansion of the true maximum-likelihood objective.
6 Feb 2026 · 16m · 10 chapters
The episode argues that fixed-weight transformer LLMs can “learn” new tasks via in-context algorithm emulation: the prompt effectively selects and runs algorithmic subroutines inside the frozen…
5 Feb 2026 · 17m · 12 chapters
Semi-supervised optimization that unifies Prediction-Powered Inference (PPI) with Stochastic Variance Reduced Gradient (SVRG) to reduce variance using abundant unlabeled data plus a small labeled set.
5 Feb 2026 · 17m · 8 chapters
Whether “model collapse” (AI trained on its own outputs degrading in a feedback loop) is inevitable for iterative Maximum Likelihood Estimation (MLE), and what mathematical conditions prevent or…
3 Feb 2026 · 16m · 6 chapters
Off-policy reinforcement learning for personalized, long-term treatment strategies modeled as Markov decision processes (MDPs), addressing “The Curse of the Horizon” (compounding error over long time…
3 Feb 2026 · 18m · 11 chapters
The episode argues that AI safety should shift from post-training “muzzles” (e.g., RLHF) to proactive token-level data filtering, especially for dual-use biology/bioweapon risk.
1 Feb 2026 · 12m · 7 chapters
The episode explains “self-improving pretraining,” a method to train better AI models without relying on ever more high-quality human data, aiming to overcome the “data wall” around 2026.
1 Feb 2026 · 15m · 9 chapters
“Success conditioning” (selective imitation of past successes) as a mathematically exact, conservative optimization method for AI alignment and learning, explained via trust-region geometry and…
31 Jan 2026 · 20m · 11 chapters
The episode explains Trajectory Bellman Residual Minimization (TBRM), a value-based training method for LLM reasoning that replaces step-by-step “micromanagement” with an objective that minimizes…
31 Jan 2026 · 18m · 8 chapters
University of Cambridge “GameTalk” research on training LLMs as strategic multi-turn agents that can bluff and negotiate, including deception without explicit programming.
30 Jan 2026 · 16m · 8 chapters
Reinforcement Learning via Self-Distillation (SDPO) replaces sparse pass/fail rewards with “rich feedback” from verifiers (e.g., compiler errors) and uses self-distillation to turn error text into…
30 Jan 2026 · 14m · 7 chapters
Self-supervised contrastive learning (no labels) can approximate supervised contrastive learning, because “false negatives” vanish as the number of classes grows; the learned representations exhibit…
28 Jan 2026 · 15m · 6 chapters
Explains why self-supervised contrastive learning (CL) without labels can still produce semantic categories, and shows a theoretical/empirical link to a supervised objective called NSCL…
28 Jan 2026 · 16m · 11 chapters
Challenges “agentic AI” hype by arguing that many multi-agent workflows are inefficient clones.
24 Jan 2026 · 17m · 10 chapters
RLHF training for LLMs argues that “greedy sampling” (always pick the currently best option from observed data) is provably efficient, removing the need for expensive exploration bonuses, because KL…
24 Jan 2026 · 13m · 6 chapters
Explains zero-shot prediction (ZSP) for vision-language models (e.g., CLIP) using “indirect prediction path” theory: image-to-text-to-label, and why it works or fails.
24 Jan 2026 · 15m · 8 chapters
Test-time training for “discovery” (TTT Discover) where an AI updates its weights during problem solving instead of only searching with a frozen model.
23 Jan 2026 · 16m · 10 chapters
Explains how a model’s pretraining distribution shapes in-context learning, focusing on task selection vs generalization and robustness under distribution shifts.
23 Jan 2026 · 19m · 8 chapters
Why CLIP-style image-text retrieval fails on specific attributes (e.g., “zero people,” “brick vs stone,” “morning vs afternoon”), and how promptable (question-conditioned) image embeddings fix it by…
20 Jan 2026 · 14m · 9 chapters
Activation Reward Models (ActivationRMs) for aligning few-shot models by steering internal activations toward “safety/truth,” avoiding heavy RLHF or gullible LLM-judge scoring.
20 Jan 2026 · 16m · 8 chapters