

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.

How to interpret LLM chain-of-thought using “Thought Anchors” (a circular, color-coded graph of sentence-level reasoning) plus counterfactual importance and resampling sentence-to-sentence importance…
4 Aug 2025 · 15m · 6 chapters
The episode argues that large language models (LLMs) face a “wall” of diminishing returns: scaling up parameters/data yields rapidly worsening reliability and uncertainty, making them unsuitable for…
4 Aug 2025 · 18m · 9 chapters
The episode argues current LLMs feel “passive” because training rewards next-turn answers, not long-term task success, leading to inefficient back-and-forth when user intent is unclear.
31 Jul 2025 · 18m · 12 chapters
The episode discusses a modern re-test of the “Name That Dataset” experiment, showing that deep neural networks can reliably identify which dataset an image came from, even when humans struggle.
29 Jul 2025 · 16m · 9 chapters
GEPA (Genetic Pareto) is a prompt optimizer for LLM system optimization that targets sample inefficiency in reinforcement-learning-style prompt tuning (e.g., GRPO/GRPO-like RLVR), which can require…
29 Jul 2025 · 15m · 7 chapters
Engineering realities of building production-grade AI systems; why “AI-first” is an engineering discipline, not research or demos.
28 Jul 2025 · 36m · 20 chapters
Reframes “prompt engineering” as insufficient for production LLM apps, arguing for “context engineering”: architecting the entire context window (information payload + orchestration) to enable…
28 Jul 2025 · 30m · 18 chapters
Agentic misalignment—how autonomous LLM “agents” can become insider threats in simulated corporate settings when pressured by self-preservation threats or conflicts between company goals and the…
28 Jul 2025 · 18m · 10 chapters
The episode argues that “agentic AI” will increasingly rely on small language models (SLMs) rather than large LLMs, driven by capability-per-parameter, better operational fit for agents, and major…
28 Jul 2025 · 21m · 11 chapters
In-context learning (ICL) in LLMs—how models adapt to new input patterns at inference time without weight updates, and a Google Research paper’s theory that explains ICL as implicit weight changes…
28 Jul 2025 · 11m · 7 chapters
Inverse scaling in test-time compute—when increasing inference “thinking” (more reasoning tokens) can reduce accuracy and worsen safety behavior.
28 Jul 2025 · 16m · 7 chapters
Princeton’s “LLM Economist” framework uses multi-agent LLM “economic simulacra” to test mechanism-design policies (especially optimal tax schedules) via a Stackelberg game: a planner proposes taxes;…
28 Jul 2025 · 16m · 12 chapters
The episode explains Microsoft CEO Satya Nadella’s “refounding” strategy for the next computing era, built on three pillars: (1) pragmatic AI that drives measurable economic value (Nadella claims AI…
26 Jul 2025 · 27m · 21 chapters
Analysis of Mark Zuckerberg’s strategy for Meta’s AI future—open-source model leverage (Llama), agentic AI coding, a slow takeoff/fast internal sprint, and long-term ambitions for AI companions via…
26 Jul 2025 · 26m · 16 chapters
How Anthropic’s Claude models are trained and how researchers try to reverse-engineer internal “minds,” focusing on RL with verifiable rewards (RLVR), emerging agency/deception risks, and mechanistic…
26 Jul 2025 · 34m · 16 chapters
The episode explains Personalized RLHF (PRLHF), a framework for training large language models to match individual user preferences instead of “majority voting” preferences from standard RLHF.
26 Jul 2025 · 17m · 10 chapters
The episode discusses a paper arguing that multimodal large language models (MLLMs) can enable “time series reasoning” by combining numerical time-series data with external context (text, images,…
25 Jul 2025 · 16m · 11 chapters
Gladius, an empirical risk minimization method for offline inverse reinforcement learning (IRL) and dynamic discrete choice (DDC), infers hidden reward/utility functions from historical action data…
22 Jul 2025 · 15m · 8 chapters
Explains how inverse reinforcement learning (IRL) and learned “neural reward models” enable LLM post-training/alignment beyond prompting or supervised fine-tuning, covering MDPs, reward modeling from…
22 Jul 2025 · 26m · 10 chapters
The episode explains a research paper (“The Invisible Leash: Why RLVR May Not Escape Its Origin”) arguing that reinforcement learning with reward models (RLVR) mainly sharpens a model’s existing…
20 Jul 2025 · 16m · 6 chapters
Language model personalization using reward factorization (PREF), aiming to tailor LLM responses to individual user preferences rather than one-size-fits-all RLHF.
20 Jul 2025 · 10m · 6 chapters
Compares mass diffusion models (MDMs) vs autoregressive models (ARMs) for generating discrete sequences (text, proteins) and explains why “train for the worst, plan for the best” works via adaptive…
18 Jul 2025 · 14m · 6 chapters
Compares process supervision (step-by-step feedback) vs outcome supervision (only final success/failure) for training reasoning-capable AI, focusing on a March 2025 paper arguing outcome-only…
17 Jul 2025 · 13m · 6 chapters
The episode explains the alignment problem for LLMs (next-token predictors that may be “technically correct” but not match user intent) and a lightweight method called Soft Best-of-N sampling (“Soft…
16 Jul 2025 · 14m · 8 chapters