

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.

AI “clones” trained on an individual’s digital traces could be used to match people for romance, jobs, housing, childcare, tutors, or therapists; the episode explains a research paper’s limits on…
25 Jun 2025 · 15m · 6 chapters
The episode argues that uncertainty quantification for large-language-model agents needs reassessment because hallucinations are inherent to generative models, and the traditional aleatoric vs…
25 Jun 2025 · 19m · 11 chapters
The episode argues current LLMs are strong pattern matchers but lack meta-reasoning—thinking about thinking—so they hallucinate confidently, fail under small prompt changes, and generalize poorly.
25 Jun 2025 · 20m · 9 chapters
The episode discusses a research paper arguing that today’s large language models have fragile, non-generalizable reasoning because supervised pretraining entangles “knowledge” with “how to reason.”…
25 Jun 2025 · 17m · 12 chapters
Explains major “mysteries” of deep learning generalization—benign overfitting, over-parameterization, and double descent—using unifying ideas like simplicity/compression bias and “soft inductive…
25 Jun 2025 · 22m · 13 chapters
The episode explains “authenticated delegation” for AI agents—how to prove an agent’s identity, that it acts on behalf of a specific user, and that it has only task-specific permissions (not carte…
25 Jun 2025 · 19m · 10 chapters
Argues that causal inference (interventions and counterfactuals) can be fully handled within standard probabilistic modeling, using the principle “write down the probability of everything” (David…
25 Jun 2025 · 22m · 8 chapters
A discussion of a position paper arguing that “deep learning phenomena” (edge cases like double descent, grokking, and lottery ticket hypothesis) should be studied for their ability to refine broad,…
25 Jun 2025 · 19m · 10 chapters
E3 (“Exploration Enables Extrapolation”) teaches LLMs to use more test-time compute by learning in-context exploration—chaining generation, verification, and refinement to improve beyond the training…
17 Jun 2025 · 14m · 8 chapters
“Extrapolation by association” for length generalization in transformers—how co-training on a related auxiliary task with longer inputs lets a model extrapolate to longer lengths on a main task it…
17 Jun 2025 · 12m · 6 chapters
The episode explains research on evaluating LLM capabilities beyond single benchmark scores, using PCA/HCA on thousands of models from the Open LLM leaderboard to infer a low-rank “causal hierarchy”…
17 Jun 2025 · 19m · 9 chapters
Generalization vs hallucination in transformers, arguing both come from one mechanism called out-of-context reasoning (OCR): learning implications via concept co-occurrence even when no…
17 Jun 2025 · 14m · 5 chapters
The episode explains GATE, a framework for improving conditional average treatment effect (CATE) estimation under data scarcity and covariate shift.
17 Jun 2025 · 15m · 9 chapters
The episode discusses research showing that LLMs (example: LLaMA-28B) encode numerical predictions and uncertainty in internal hidden states before autoregressive token generation, enabling faster,…
17 Jun 2025 · 15m · 11 chapters
Explains why transformer-based in-context learning (ICL) enables strong few-shot learning without updating model weights, framing it as meta-learning.
17 Jun 2025 · 21m · 10 chapters
The episode critiques using large language models (LLMs) as “human surrogates” for social science research, focusing on a paper titled “Take Caution in LLMs as Human Surrogates: Scylla Ex Machina.”…
14 Jun 2025 · 28m · 14 chapters
Whether today’s “AI reasoning” is genuine multi-step cognitive reasoning (deduction, planning, understanding) or sophisticated pattern matching that only mimics logic, with implications for trust,…
12 Jun 2025 · 31m · 14 chapters
The episode explains a research finding that large language models often encode their most useful representations in intermediate (middle) layers rather than the final layer.
12 Jun 2025 · 13m · 5 chapters
The episode explains Luo, Yu, Lu, Shi, and Jing’s “Causal Attribution Analysis for Continuous Outcomes,” which answers a retrospective “why” question for continuous results (e.g., blood pressure,…
12 Jun 2025 · 18m · 7 chapters
The episode discusses the paper “Training a Generally Curious Agent” (Fahim Tajwar et al., aimed at ICML 2025), arguing current LLMs are strong pattern matchers but weak at strategic exploration in…
12 Jun 2025 · 14m · 8 chapters
The episode explains why standard A/B testing (difference in means) can be badly biased in non-stationary, interactive systems where actions cause temporal interference and the environment changes…
12 Jun 2025 · 21m · 9 chapters
Explains why in-context learning (ICL) in transformer models can appear during training but later fade, using a “strategy coopetition” framework.
12 Jun 2025 · 19m · 14 chapters
The episode discusses a research paper claiming that narrow fine-tuning of already-aligned LLMs on a single harmful behavior can cause “emergent misalignment,” where the model becomes broadly…
11 Jun 2025 · 17m · 7 chapters
The episode explains MAAS (Multi-Agent Architecture Search via Agentic SuperNet), a framework for automatically designing multi-agent LLM workflows that adapt per query instead of using one fixed…
11 Jun 2025 · 18m · 12 chapters