

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.

Temporal credit assignment in reinforcement learning—how to determine which earlier actions/states actually caused delayed outcomes, beyond recency-based methods.
15 Jul 2025 · 17m · 8 chapters
How reward models used in RLHF for LLM alignment can be exploited via reward hacking, especially out-of-distribution (OOD) prompts, and how SMORM (Joint, Single, and Multi-Objective Reward Model)…
15 Jul 2025 · 17m · 8 chapters
Tests whether foundation models learn “world models” (underlying rules) or just predict sequences via heuristics, using an “inductive bias probe” rather than accuracy alone.
15 Jul 2025 · 12m · 7 chapters
GenAI-Powered Inference (GPI), a statistical framework (Yamai & Nakamura, Harvard; paper dated July 8, 2025) that uses generative AI to do causal/predictive inference on unstructured data (text,…
14 Jul 2025 · 20m · 10 chapters
Interpretable Reward Modeling for RLHF—making reward models auditable by predicting human-interpretable concepts via Concept Bottleneck Reward Models (CBRM), trained with active learning to reduce…
14 Jul 2025 · 12m · 7 chapters
“PrefillOnly” inference engine for LLM applications where each request outputs exactly one token (e.g., yes/no, option A/B/C, label).
14 Jul 2025 · 14m · 4 chapters
The episode argues that today’s AI debate is “noisy” and too individualistic (thinking machines/AGI).
14 Jul 2025 · 21m · 11 chapters
The episode discusses the paper “Textual Bayes: Quantifying Uncertainty in LLM-Based Systems,” which aims to make LLM outputs more reliable by producing calibrated uncertainty estimates.
12 Jul 2025 · 9m · 6 chapters
The episode explains the “winner’s curse” in inference-then-optimization systems: even with unbiased estimates for each option, selecting the best-performing option based on noisy estimates makes the…
11 Jul 2025 · 30m · 14 chapters
Spiral is a self-play training framework for language-model reasoning in two-player zero-sum, multi-turn “language games,” aiming to reduce reliance on human-labeled data and expert supervision.
11 Jul 2025 · 17m · 9 chapters
The episode argues that today’s frontier LLMs (trained via statistical learning/next-token prediction) are misaligned with the deductive, zero-error reasoning needed for general intelligence and…
11 Jul 2025 · 22m · 11 chapters
The episode explains a research paper, “Aligning Learning and Endogenous Decision Making,” arguing that standard AI “predict-then-optimize” fails when actions change the data you observe (endogenous…
11 Jul 2025 · 16m · 11 chapters
The episode argues that LLMs can generate fully synthetic survey/text data to overcome the high cost and slow timelines of collecting human social-science data, but only if synthetic data is combined…
11 Jul 2025 · 14m · 12 chapters
Multi-turn reinforcement learning from human preference feedback (RLHF), arguing that training LLMs with feedback on whole conversations beats single-turn feedback because good actions can look bad…
10 Jul 2025 · 17m · 11 chapters
How large language models can learn from detailed natural-language critiques (LLF) instead of scalar reward scores, and when this yields provably faster learning.
9 Jul 2025 · 17m · 8 chapters
How competing learning algorithms in markets determine “survival” (nonzero wealth share) vs “vanishing” (wealth share goes to zero), comparing Bayesian learners to no-regret learners, and proposing a…
5 Jul 2025 · 21m · 10 chapters
A July 3, 2025 theoretical framework explains how gradient-based training of neural networks can yield discrete symbolic structures (neurosymbolic reasoning) from continuous optimization.
5 Jul 2025 · 14m · 4 chapters
Projected abstractions for causal reasoning with lossy representations, addressing the abstract invariance condition (AIC) problem and enabling cross-layer inference (L1 observations to L2…
4 Jul 2025 · 26m · 13 chapters
Winner’s curse in data-driven decisions: when teams do “inference, then optimize” (estimate values from data, then pick the best), selection bias from noise makes the chosen option’s estimated lift…
4 Jul 2025 · 23m · 14 chapters
Embodied AI agents that perceive (vision/audio/touch), model the physical and human “mental” context, plan and act with autonomy, and learn continuously; the episode covers research status (a July 4,…
4 Jul 2025 · 29m · 15 chapters
The episode argues that today’s LLMs, trained via statistical learning (next-token prediction), are fundamentally misaligned with the kind of exact, deductive reasoning needed for general…
4 Jul 2025 · 20m · 12 chapters
An “inductive bias probe” tests whether foundation models learn true underlying world rules (Newton) or only learn shortcuts that predict outcomes (Kepler).
4 Jul 2025 · 12m · 6 chapters
Language Bottleneck Models (LBMs) for knowledge tracing: converting a student’s interaction history into a concise natural-language “knowledge state” summary that preserves predictive power while…
3 Jul 2025 · 23m · 12 chapters
The episode explains “pure exploration” in AI and a new method called In-Context Pure Exploration (ICPE), which learns how to choose the most informative next actions without being given explicit…
3 Jul 2025 · 17m · 9 chapters