

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.

Sample complexity of autoregressive reasoning; compares end-to-end supervision (only final answer) vs chain-of-thought supervision (training on intermediate steps).
19 Apr 2026 · 19m · 10 chapters
The episode argues today’s “machine learning” systems are brittle because they rely on heavy human MLOps scaffolding and don’t truly learn from interacting with the world.
17 Apr 2026 · 21m · 15 chapters
How combining small randomized trials with large observational “big data” can create an “illusion of learning,” and how empirical Bayes can be made reliable using calibration studies…
17 Apr 2026 · 22m · 12 chapters
How large language models handle conflicts of interest when prompted to promote ads/sponsors, using Grice’s Cooperative Principle as a framework; includes flight booking, math help, and predatory…
17 Apr 2026 · 22m · 12 chapters
Token Mapping Perturbation Attack (TAMPA/TomPoo) shows reward models used in RLHF can be hacked without producing human-readable language by bypassing the normal token-to-text-to-token pipeline and…
13 Apr 2026 · 18m · 11 chapters
How to evaluate LLMs from pairwise “chatbot arena” votes using a low-rank tensor model, fixing statistical distortions from sparse/noisy, non-uniform sampling and unequal information across matchups;…
12 Apr 2026 · 19m · 12 chapters
Neural computers (NC/CNC) that replace the classic hardware+OS stack with an AI model as the “runtime,” hallucinating a working command-line or GUI interface from learned screen behavior and user I/O…
11 Apr 2026 · 13m · 10 chapters
How AI information aggregation can worsen collective knowledge by creating echo chambers and feedback loops, especially when a fast-updating global AI mediates between segregated social groups.
11 Apr 2026 · 23m · 12 chapters
World Action Verifier (WAV) for robotics: self-improving action-conditioned world models that identify their own physics errors using forward-inverse asymmetry (generation of plausible sub-goals,…
10 Apr 2026 · 20m · 11 chapters
In-place test-time training (in-place TTT) to stop LLMs from “forgetting” early context by enabling continuous on-the-fly learning during inference, without retraining or infinite context windows.
9 Apr 2026 · 20m · 10 chapters
The episode explains a new paper, “Test-Time Scaling Makes Overtraining Compute Optimal,” arguing that when models are allowed to use extra test-time compute to sample many candidate answers (high…
7 Apr 2026 · 21m · 10 chapters
Whether AI bots are corrupting survey data, and how data quality varies by online sample provider.
7 Apr 2026 · 22m · 14 chapters
Generative optimization for LLM-based systems, focusing on PLCA (Prioritized Optimization with Local Contextual Aggregation) to improve AI prompts/agents despite noisy, stochastic evaluation loops.
4 Apr 2026 · 19m · 11 chapters
Agentic markets—how AI-driven consumer search changes market learning, competition, and prices. Guests: No guest names or bios appear in the transcript; it’s a two-host discussion.
4 Apr 2026 · 22m · 8 chapters
GE Rec (Google) bid-aware generative recommendation for balancing organic relevance with real-time sponsored ad auctions, using control tokens and bid-aware decoding.
1 Apr 2026 · 21m · 9 chapters
The episode explains a January 2026 economics paper that introduces “equivalent sample size” (ESS) to quantify how many real human survey observations an LLM effectively replaces when predicting…
1 Apr 2026 · 22m · 9 chapters
The episode explains the March 20, 2026 paper “Learning to Reason with Curriculum I” (Microsoft and UIUC) arguing that AI training for reasoning is hitting a compute wall, and proposing…
1 Apr 2026 · 21m · 13 chapters
The episode argues that the coming “intelligence explosion” won’t be a single superintelligence (“god” in a server farm) but a plural, social, agentic ecosystem—like a fast, chaotic city—where models…
30 Mar 2026 · 23m · 9 chapters
Behavior cloning for robotics fails in the real world because continuous control actions are quantized into discrete tokens, creating rounding errors that compound over time (horizon).
29 Mar 2026 · 21m · 8 chapters
Hyperagents (DGMH) are AI systems that perform open-ended metacognitive self-improvement by rewriting both their task-solving code and the “manager” mechanism that generates future improvements,…
27 Mar 2026 · 22m · 10 chapters
How Anthropic researcher Prithvi Rajasikharan’s “harness” design enables AI to build long-running, full-stack applications reliably, moving beyond solo code-writing.
26 Mar 2026 · 21m · 11 chapters
Whether off-the-shelf, zero-shot AI agents can avoid game-theoretic failures (e.g., price wars, persistent overcharging) and converge to stable Nash equilibria without manual fine-tuning, even with…
24 Mar 2026 · 20m · 11 chapters
The episode explains why standard stochastic gradient bandit / policy optimization can permanently “vanish” into a suboptimal policy after early bad luck, and how a log-barrier constraint…
22 Mar 2026 · 21m · 12 chapters
Challenges the “download a big foundation model and fine-tune on private domain data” playbook, arguing it’s economically and technically flawed.
22 Mar 2026 · 18m · 7 chapters