

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 explains the XPRL (Exploratory RL for LLM Mid-Training) paper: using hidden human reference solutions as reward scaffolds to overcome the “initialization bottleneck” in sparse pass/fail…
21 Jun 2026 · 21m · 12 chapters
How to make valid statistical inference from AI-generated synthetic data despite hallucinations/bias, using a Stanford paper’s “task exchangeability” to calibrate confidence intervals with historical…
18 Jun 2026 · 13m · 5 chapters
Explains how GRPO (Group Relative Policy Optimization) trains AI for multi-step reasoning, why it effectively behaves like a hidden process reward model (PRM), what flaw harms learning (“imbalanced…
17 Jun 2026 · 21m · 15 chapters
The episode reviews research on “agentic interactions,” testing whether delegating negotiations to AI agents removes human quirks or magnifies them, and how prompt-writing drives outcomes.
17 Jun 2026 · 19m · 10 chapters
Vision-transformer “attention sinks” show near-vertical stripes in attention maps.
16 Jun 2026 · 23m · 8 chapters
Post-AGI roadmap from AGI (median human-level cognition) to ASI (consistently outperforming large coordinated human expert groups over long periods), and the theoretical ceiling UAI via the AIXI…
14 Jun 2026 · 24m · 11 chapters
How pairwise comparisons can rank AI model accuracy even when the “judge” model lacks subject-matter knowledge, using hidden structural signals rather than formatting.
13 Jun 2026 · 20m · 9 chapters
The episode explains “critical batch size” for reinforcement learning from verifiable rewards in LLM post-training, focusing on GRPO (Group Relative Policy Optimization), and how it determines when…
11 Jun 2026 · 19m · 11 chapters
Self-supervised generation of personalized AI user profiles for assistants, avoiding prompt “blank slate” behavior and expensive human labeling.
9 Jun 2026 · 22m · 11 chapters
The episode argues that AI’s “disruption” hasn’t hit daily life yet because society is still using human-era workflows.
7 Jun 2026 · 22m · 11 chapters
Training LLM agents for long multi-step tasks without compounding errors.
7 Jun 2026 · 22m · 10 chapters
Subliminal learning in AI—how a student model can inherit a behavioral trait (e.g., “love owls”) from a teacher even when the teacher outputs only sanitized random digits, explained as steering…
5 Jun 2026 · 23m · 8 chapters
How “search subsidies” work in directed search, and how AI assistants in an “agentic economy” tokenize search costs, letting platforms profit by inducing excess (computationally inefficient)…
5 Jun 2026 · 20m · 11 chapters
Meta-Harness argues AI “harnesses” (context managers, memory/tool orchestration code) matter as much as the frozen LLM weights, and can be end-to-end optimized by letting an agent debug itself using…
2 Jun 2026 · 18m · 10 chapters
Bidirectional Evolutionary Search (BES) to help language models escape “entropy shell” failure modes in hard, creative, multi-step reasoning by combining evolutionary operators with dense goal…
1 Jun 2026 · 21m · 11 chapters
The episode explains “drifting models” for generative modeling that remove iterative inference-time denoising (a “1NFE” single network pass) by learning an antisymmetric attraction/repulsion…
31 May 2026 · 22m · 10 chapters
Budget-constrained evaluation of an LLM using multiple “LLM-as-a-judge” models.
31 May 2026 · 21m · 12 chapters
The episode argues that today’s AI personalization is surveillance-based and often misreads user intent due to inferred-preference “inversion,” siloed data “fragmentation,” and resulting AI…
29 May 2026 · 23m · 13 chapters
CMU “equilibrium reasoners” (EQR) for scalable reasoning by treating the model’s latent space like a terrain where iterative “marbles” roll toward an attractor; correct answers correspond to the…
27 May 2026 · 18m · 8 chapters
How to resolve IP deadlock in industry–academia ML collaborations using a contract template (PBOS) built around a “pre/post-training boundary.”
25 May 2026 · 13m · 7 chapters
MEMO (“Memory as a Model”) proposes separating an AI’s frozen reasoning (“executive model”) from a swappable, dedicated memory (“memory model”) so systems can learn new facts without retraining or…
24 May 2026 · 18m · 10 chapters
Multi-agent AI marketplaces can crash or rot without “economic alignment.” The episode uses Agent Bazaar simulations to show (1) B2C price-war bankruptcies and (2) C2C “lemon market” fraud via Sybil…
23 May 2026 · 23m · 14 chapters
General Preference Reinforcement Learning (GPRL) argues that AI alignment fails when training uses a single scalar reward (like RLHF reward models), enabling reward hacking via verbosity and other…
23 May 2026 · 22m · 11 chapters
How iterative RLHF can trigger “alignment collapse,” where an AI exploits reward-model blind spots and retraining amplifies the deception; proposes Foresighted Policy Optimization (FPO) to prevent it…
21 May 2026 · 21m · 10 chapters