

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.

Evaluates “automatic harness evolution” for AI agents—whether letting a model rewrite its prompts/tools/control logic actually makes it smarter, or just exploits benchmark evaluation.
19 Jul 2026 · 23m · 13 chapters
Compositional generalization in language-model reasoning—why supervised fine-tuning (SFT) can make models rigid, and how reinforcement learning (RL) “untangles” reasoning into reusable skill and…
18 Jul 2026 · 19m · 11 chapters
The episode argues that AI interpretability must become actionable—moving from passive “black box” observation to concrete interventions that improve real-world behavior, safety, and usability.
17 Jul 2026 · 25m · 12 chapters
A 2026 theoretical breakthrough on high-accuracy sampling for diffusion models and log-concave distributions using only first-order information (gradients), avoiding the “polynomial wall” of standard…
17 Jul 2026 · 22m · 9 chapters
Causal inference for video persuasion—isolating the frame-by-frame causal effect of a fleeting visual feature (treated as a “treatment”) on changing human emotion, while correcting for confounding…
15 Jul 2026 · 22m · 11 chapters
Exploit–explore tradeoffs in multi-armed bandits, focusing on what Thompson sampling optimizes and how new research “opens the black box” using Markov decision processes and Bellman equations.
15 Jul 2026 · 22m · 10 chapters
Offline reinforcement learning safety via Globally Convergent Offline RL using Smoothed Bellman Residual Minimization (OffGladius), addressing the “deadly triad” (offline data + neural nets +…
13 Jul 2026 · 12m · 5 chapters
LLM-as-a-Verifier proposes a general-purpose framework to verify LLM outputs by extracting internal uncertainty (logits) instead of using discrete AI “judges,” addressing a “verification crisis”…
10 Jul 2026 · 20m · 12 chapters
How much do transformer language models memorize vs generalize, and what mathematical threshold triggers “grokking” (learning rules) instead of rote storage.
9 Jul 2026 · 24m · 9 chapters
The episode argues that uncertainty quantification (UQ) in large language models is mis-specified: mainstream UQ methods effectively measure internal consistency via unsupervised clustering of…
7 Jul 2026 · 22m · 12 chapters
When AI agents should use external tools (search, APIs, code) versus rely on internal reasoning, arguing tools shouldn’t be used unless epistemically necessary.
6 Jul 2026 · 12m · 6 chapters
How AI preference-alignment methods (used in chatbots via RLHF/DPO) can fail when they assume a fixed logistic “link function” for human choices, and how semi-parametric preference optimization (SPO)…
5 Jul 2026 · 22m · 14 chapters
The episode explains research on reinforcement learning post-training (RLVR) in large transformer LLMs, arguing that training only a subset of layers—especially middle layers—can match or beat…
4 Jul 2026 · 23m · 11 chapters
The episode discusses a Harvard paper, “RL Excursions During Pre-Training,” arguing that reinforcement learning (RL) can be applied early during LLM pre-training—challenging the standard sequential…
2 Jul 2026 · 22m · 10 chapters
How language-generation learning fails under “no feedback” math, but becomes robust when feedback is added; the episode centers on theorems using language generation in the limit, set-based…
1 Jul 2026 · 20m · 9 chapters
Bayesian persuasion limits for misaligned AI that has an information advantage and tries to steer a rational receiver’s decisions, quantified via receiver-utility bounds.
1 Jul 2026 · 22m · 9 chapters
The episode argues that AI “capability” is jagged: models can solve hard math (e.g., Erdos unit distance problem) yet fail on simple logic.
29 Jun 2026 · 22m · 10 chapters
Qwen- AgentWorld describes “language world models” that let general AI agents simulate interactive computer environments internally before acting, reducing errors like clicking the wrong…
27 Jun 2026 · 21m · 10 chapters
When offline reinforcement learning can learn efficiently using only trajectory-level feedback (final outcomes or preferences) instead of step-by-step rewards, and when it becomes impossible.
27 Jun 2026 · 19m · 11 chapters
The episode explains “SuperThoughts,” a framework to speed up LLM reasoning by generating two tokens per step using “superposition,” while preserving accuracy via confidence-based adaptive fallback.
26 Jun 2026 · 19m · 7 chapters
Explore-vs-exploit in reinforcement learning under deceptive rewards, and a new method called First-Explore PPO (FEPPO) that learns meta-exploration efficiently.
25 Jun 2026 · 23m · 9 chapters
Data-scarce pretraining for language models; why multi-epoch reuse overfits, and how self-distillation (teacher-student with identical architectures) improves learning using “dark knowledge” (soft…
24 Jun 2026 · 22m · 11 chapters
Meta-harness for agent-state construction (SAIT): an outer-loop AI that writes and debugs the “harness” (working-memory management layer) for other agents, optimizing what gets…
21 Jun 2026 · 23m · 9 chapters
The episode explains the XPRL (Exploratory RL for LLM Mid-Training) paper: using hidden human reference solutions as reward scaffolds to overcome the “initialization bottleneck” in sparse pass/fail…
21 Jun 2026 · 21m · 12 chapters