

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.

The episode explains “emergent capabilities” in transformer language models as abrupt, seemingly overnight skill gains caused by random alignment of attention patterns during training.
8 Oct 2026 · 24m · 12 chapters
How to do causal inference when analysts replace raw, messy data with AI-learned embeddings, risking “target gap” between true causal effects and AI-based estimates.
6 Oct 2026 · 23m · 10 chapters
Pairwise vs point-wise reward learning for RLHF-style training. Compares value regression (VR: train on absolute numeric rewards like 8.5/10) to value difference regression (VDR: hide absolute scores…
4 Oct 2026 · 22m · 9 chapters
A Deep Dive on a University of Pennsylvania paper (Feb 20, 2026) proposing Selective Strong Verification (SSV) to decide when AI should use cheap “weak verification” vs expensive “strong…
3 Oct 2026 · 25m · 12 chapters
The episode explains Apple research (May 2026) on information-theoretic limits of language-model alignment, including exact bounds for reward gains under KL-regularized RL, how to predict the…
1 Oct 2026 · 24m · 14 chapters
Decision-Targeted Digital Twins (DT2) argues standard digital twins fail because they try to perfectly replicate every variable, wasting compute on irrelevant details and producing wrong action…
29 Sep 2026 · 13m · 8 chapters
Self-play pretraining with zero data—two randomly initialized AIs co-evolve in a synthetic “blank slate” universe to learn predictive structure without any human text or images, aiming to bypass the…
27 Sep 2026 · 22m · 12 chapters
The episode argues that language models need an AI “sleep” cycle to learn continuously without catastrophic forgetting, by consolidating short-term context into long-term memory during offline phases…
27 Sep 2026 · 21m · 13 chapters
The “hidden crisis” of automation: today’s AI is brilliant at reasoning and benchmarks but fails at reliable, code-friendly tasks like classifying emails or producing strict outputs.
23 Sep 2026 · 22m · 12 chapters
How “uplift” from advanced AI lowers the barrier for cyberattacks, influence operations, surveillance, and even weapons/biological R&D; includes examples of autonomous malware mutation, AI-driven…
22 Sep 2026 · 23m · 12 chapters
The episode argues that today’s LLMs and similar AI systems can excel at deduction (formal proof) and induction (pattern-finding), but cannot perform abduction—the creative leap needed to invent new…
22 Sep 2026 · 21m · 11 chapters
PROACT/“jailbreaking jailbreaks” for LLM safety, using proactive deception to disrupt autonomous jailbreak optimization loops that exploit passive refusals.
20 Sep 2026 · 23m · 10 chapters
How LLM agents’ performance scales with huge test-time compute, why “thinking longer” plateaus, and how to allocate tokens to avoid getting trapped in a “sticky basin” of self-generated context.
18 Sep 2026 · 22m · 13 chapters
The episode explains “looped flows,” a new AI architecture meant to let models “scale internal thinking time” for hard logic tasks, unlike standard deep nets that answer with a fixed compute budget.
17 Sep 2026 · 20m · 12 chapters
How multi-turn LLMs stay fast by caching KV (key/value) attention states in finite high-bandwidth memory (HBM), using least-recently-used (LRU) eviction, and predicting cache hit ratio via multi-turn…
17 Sep 2026 · 22m · 12 chapters
The episode explains research from University of Washington and MetaFair (“Breaking the Token Ceiling”) arguing that switching AI distillation from token vocabularies to byte-level vocabularies…
14 Sep 2026 · 24m · 11 chapters
The episode argues that LLM “thinking out loud” (reasoning traces) improves recall of factual trivia by two mechanisms: a computational buffer (stalling via filler tokens to force extra forward…
12 Sep 2026 · 21m · 13 chapters
Tail-likelihood reinforcement learning (Tail RL) trains AI to maximize rare “upper-tail” high-reward outcomes instead of average reward, using a thresholded, binary success objective and an…
11 Sep 2026 · 21m · 10 chapters
The episode argues that today’s transformer LLMs mainly do next-token “parroting” due to attention’s effectively infinite lookback memory, which encourages shortcut learning (“epicycles”) instead of…
7 Sep 2026 · 23m · 17 chapters
Declarative attention for language models reduces the KV-cache memory bandwidth bottleneck by having the model explicitly declare where to look before reading, cutting attended tokens and decode time…
5 Sep 2026 · 20m · 13 chapters
The episode argues that reinforcement-learning and other optimization-based AI can achieve “success” while violating human intent by exploiting proxies, loopholes, and the environment itself—framing…
4 Sep 2026 · 26m · 10 chapters
Mechanistic interpretability for LLMs—how to enable isolated, non-interfering interventions by enforcing near-orthogonal (perpendicular) feature geometry, reducing “superposition interference” in the…
3 Sep 2026 · 24m · 14 chapters
Test-Time Policy Optimization (TTPO), a method for improving language models’ complex math reasoning without ground-truth answer keys, using self-generated rollouts and a two-branch training scheme.
3 Sep 2026 · 23m · 12 chapters
How reinforcement learning post-training actually works for language models, and why sparse rewards often fail while dense, step-by-step rewards can unlock novel reasoning; also how random/spurious…
1 Sep 2026 · 20m · 13 chapters