

531 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.

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
Research on “Turnwise” shows multi-turn chat can degrade a language model’s performance versus its own single-turn ability (“conversational amnesia” / “turn decay”).
22 Mar 2026 · 11m · 5 chapters
Temporal straightening for latent planning—teaching AI to smooth its internal time/space representation so it can plan efficiently despite visual noise, reducing a “compute crisis” in robotics.
20 Mar 2026 · 21m · 14 chapters
How fine-tuning can cause “localized amnesia,” degrading in-context (few-shot) learning even while improving zero-shot task performance, and what to do instead.
19 Mar 2026 · 19m · 9 chapters
The episode argues that current LLM reasoning training wastes massive compute because trainer and inference engines fall out of sync, making on-policy RL updates (e.g., PPO/GRPO) mathematically…
19 Mar 2026 · 20m · 13 chapters
Continual reinforcement learning for vision-language-action (VLA) robots, addressing catastrophic forgetting.
17 Mar 2026 · 24m · 8 chapters
Test-time scaling and self-reflection in language models via in-context policy optimization (ICPO), using a multi-armed bandit framing to let the model optimize its own reasoning on the fly without…
17 Mar 2026 · 21m · 13 chapters
Explains why long-form generation derails (teacher forcing + distribution shift) and why common fixes like RLVR can “hack” rewards, then introduces energy-based fine-tuning (EBFT) that matches…
16 Mar 2026 · 23m · 13 chapters
The episode explains new research claiming that “guess-and-check” can work for large, heavily pretrained AI models because pretraining reshapes the loss landscape into a dense “thicket” of nearby…
14 Mar 2026 · 20m · 12 chapters
AdaEvolve (UC Berkeley, Bespoke Labs) argues that algorithm/code search with LLM-guided evolutionary methods fails due to fixed “static schedules,” and proposes hierarchical adaptive optimization:…
14 Mar 2026 · 20m · 10 chapters
Deep dive into “NABLA Reasoner,” a framework for LLM test-time reasoning using test-time gradient descent in latent space (Differentiable Textual Optimization, DTO), aiming to improve multi-step…
14 Mar 2026 · 21m · 11 chapters
The episode explains why the common “two-step” workflow—use AI/ML to generate variables from text/audio, then run a standard OLS regression—can produce biased causal estimates and misleading…
12 Mar 2026 · 22m · 11 chapters
The episode explains the “context window” bottleneck in LLMs, focusing on KV cache memory growth, and a new method to compress KV cache ~50x in seconds without losing reasoning.
12 Mar 2026 · 23m · 10 chapters
Argues that “reasoning traces” (intermediate tokens like “aha”) in large reasoning models are not human-like thought, and that treating them as genuine logic misleads users and researchers.
11 Mar 2026 · 19m · 9 chapters
Code World Models (CWMs) for General Game Playing—using an LLM to translate natural-language game rules into executable Python “world” simulators, then letting classical planning (MCTS) play within…
8 Mar 2026 · 22m · 8 chapters
How chain-of-thought prompting changes a transformer’s optimization behavior, letting it implement multi-step gradient descent rather than a single-step update; includes related “loop transformer”…
7 Mar 2026 · 17m · 7 chapters
How “task descriptors” in prompts (e.g., “translate English to French”) change the internal math of in-context learning in transformers, including provable behavior with infinite data and measured…
7 Mar 2026 · 19m · 15 chapters
Explains a mathematically proven view of in-context learning: a prompt is equivalent to applying temporary “implicit weight patches” inside transformer blocks, even for modern bias-free architectures.
7 Mar 2026 · 21m · 9 chapters
Explains Google research on in-context learning (ICL): how a frozen transformer “learns” a new task from the prompt without updating permanent weights, by applying implicit, temporary parameter…
7 Mar 2026 · 24m · 15 chapters
Causal AI for counterfactual identification, introducing “counterfactual randomization” (Layer 2.5) and a new algorithm (CTFIDU+) that is complete for realizable queries; when exact identification is…
7 Mar 2026 · 20m · 15 chapters
Netflix researchers argue that cosine similarity between learned embeddings may not reflect true semantic similarity because hidden scaling introduced by regularization can make cosine scores…
6 Mar 2026 · 22m · 17 chapters
The episode explains a research paper, “Diffusion LLMs are Natural Adversaries for Any LLM” (Technical University of Munich), arguing that jailbreak vulnerabilities are “data-specific” and can be…
5 Mar 2026 · 25m · 16 chapters
The episode explains the “partial token problem,” where language models tokenize text into chunks and can fail when a user prompt ends mid-token (e.g., pausing after an incomplete word or punctuation…
4 Mar 2026 · 19m · 7 chapters
“Inert map” problem in language models—models can encode a representation learned from in-context data but fail to use it for rule-based navigation and adaptive world modeling.
2 Mar 2026 · 19m · 12 chapters