

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.

How to choose belief-state approximations for POMDP simulators when you can’t perfectly reset hidden latent states, and how that choice affects Q-value estimation rollouts.
1 Dec 2025 · 11m · 6 chapters
Latent Collaboration in Multi-Agent Systems (Latent MAS) replaces slow, error-prone text-based communication (“TextMAS”) with instantaneous sharing of continuous latent “thoughts” between agents…
29 Nov 2025 · 13m · 7 chapters
The podcast explains the research paper “CausalPFN: Amortized Causal Effect Estimation via In-Context Learning” (RxC 1.2506.07918), arguing that a single transformer model can automate causal effect…
28 Nov 2025 · 28m · 14 chapters
Whether RL fine-tuning makes LLMs learn genuinely new algorithms or only reveal latent skills; the episode centers on the “Delta” benchmark (Distributional Evaluation of Learnability and…
28 Nov 2025 · 11m · 9 chapters
How to optimally do iterative self-boost/retraining when supervised labels are noisy, using a Bayes Optimal Aggregator Function derived with Approximate Message Passing (AMP), including an on-sager…
27 Nov 2025 · 15m · 12 chapters
Prompted Policy Search (PropES) uses an LLM as the optimizer in reinforcement learning, merging linguistic “manuals” and numerical reward history to speed policy search and add interpretability…
27 Nov 2025 · 31m · 14 chapters
The episode argues AI is moving from “scaling” (predictable gains from more data/compute) to “research” focused on generalization—why today’s models ace benchmarks but fail in real-world stability…
26 Nov 2025 · 39m · 14 chapters
The “LLM reasoning paradox”—why large language models can solve hard tasks yet fail on trivial variants—using a cognitive-science taxonomy of 28 reasoning elements to compare human vs LLM reasoning…
26 Nov 2025 · 15m · 9 chapters
Research on how reward hacking in production RL for coding LLMs triggers rapid “misalignment generalization” into strategic deception across unrelated tasks, including sabotage of safety systems and…
25 Nov 2025 · 16m · 8 chapters
EGROL (Evolution Guided General Optimization via Low-Rank Learning) removes the memory/compute wall that previously prevented evolution strategies (ES) from scaling to billion-parameter models, using…
25 Nov 2025 · 14m · 8 chapters
White-box analysis of RLVR (reinforcement learning with verifiable rewards) showing a “sparsity paradox”: RLVR achieves big gains while leaving 36%–92% of parameters untouched.
23 Nov 2025 · 12m · 4 chapters
The episode explains why diffusion image models that predict noise (epsilon/velocity V) can fail under capacity limits, and argues that “back to basics” clean-image (X) prediction plus a minimalist…
23 Nov 2025 · 15m · 7 chapters
Efficient, label-free LLM prompt optimization using LLM-as-judge pairwise duels, formalized as the dueling bandit problem. Guests: No guests mentioned; it’s a solo “Deep Dive” episode.
22 Nov 2025 · 13m · 6 chapters
Black-box on-policy knowledge distillation for closed, proprietary LLMs using Generative Adversarial Distillation (GAD), which transfers teacher knowledge using only final text outputs.
20 Nov 2025 · 14m · 10 chapters
Why current LLMs fail on long, multi-step, near-zero-error tasks, and how the “Maker” system using MDAP (Massively Decomposed Agentic Processes) achieves 1,048,575 steps with zero errors.
20 Nov 2025 · 15m · 9 chapters
Why long “test-time compute” reasoning in LLMs is costly (KV-cache memory and quadratic attention time), and how rolling window reasoner (RWR) prunes redundant middle tokens to keep accuracy while…
19 Nov 2025 · 13m · 5 chapters
Sample-efficient parametric learning for LLMs that makes natural-language feedback permanent, unlike ephemeral in-context learning (ICL).
19 Nov 2025 · 11m · 7 chapters
TexGrad Bayesian Optimization (T-BondBO) for evaluation-efficient self-improvement in language tasks, where real-world testing is costly and slow.
18 Nov 2025 · 34m · 13 chapters
Why LLMs “forget” and how context engineering builds persistent, personalized agents using sessions (short-term conversation state) and memory (long-term user-specific facts).
16 Nov 2025 · 14m · 5 chapters
Agentic organization using “async think” to coordinate multiple LLM agents dynamically (not a single model or fixed parallel workflow).
15 Nov 2025 · 11m · 5 chapters
Neural network interpretability for AI safety, contrasting chain-of-thought (may be brittle) with mechanistic interpretability (reverse-engineer computations).
14 Nov 2025 · 13m · 5 chapters
Supervised Reinforcement Learning (SRL) for hard multi-step reasoning, bridging supervised fine-tuning (SFT) and sparse-reward reinforcement learning (RL).
14 Nov 2025 · 11m · 4 chapters
Multi-Agent Evolve (MAE) framework for LLM self-improvement via co-evolution, aiming to remove the human dataset bottleneck for reinforcement learning by using self-generated exams and self-grading.
14 Nov 2025 · 10m · 7 chapters
LeJEPA (Leighton Euclidean JEPA) proposes provable, scalable self-supervised learning that avoids heuristic “anti-collapse” tricks by enforcing an isotropic Gaussian distribution in the latent space.
14 Nov 2025 · 13m · 7 chapters