

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.

ReFRAG (Rethinking RAG-based decoding) speeds up retrieval-augmented generation by compressing RAG context to avoid quadratic TTFT latency and KV-cache memory blowups.
13 Dec 2025 · 14m · 4 chapters
A complexity-theoretic argument that autoregressive next-token prediction training can provably yield long-range structural coherence, despite being locally optimized and computationally bounded.
12 Dec 2025 · 12m · 5 chapters
How Google’s TPUs and the software stack (JAX, XLA, Pathways) enable planetary-scale AI training, and why public academic research and new funding models are essential for future breakthroughs.
12 Dec 2025 · 38m · 18 chapters
“Latent Debate” proposes a surrogate framework that models an LLM’s hidden internal “support vs attack” conflict to predict and explain hallucinations, aiming to bridge capability vs reliability.
11 Dec 2025 · 15m · 8 chapters
How to make “thinking LLM as a judge” evaluations reliable by transforming noisy, stochastic judge votes into calibrated ratings using distribution-calibrated inference-time compute (ITC) and a…
11 Dec 2025 · 12m · 4 chapters
Why standard token-level RL (e.g., GRPO/PTO) fails for diffusion LLMs, and how ESPO (ELBO-based sequence-level policy optimization) fixes it by treating the whole sequence as one action and using a…
11 Dec 2025 · 12m · 6 chapters
Algorithmic Thinking Theory for LLMs on hard math (e.g., IMO-style problems), explaining why single-shot “pass@1” is low but multi-attempt “pass@K” can be high, and formalizing how to assemble…
10 Dec 2025 · 17m · 11 chapters
How pre-training, mid-training (continued pre-training), and RL post-training interact to improve reasoning in language models, and when RL truly adds new reasoning vs merely polishing existing…
10 Dec 2025 · 14m · 4 chapters
Natural Language Actor-Critic (NLAC) for training LLM agents with scalable off-policy reinforcement learning in language/action space, addressing long-horizon credit assignment and sparse rewards.
9 Dec 2025 · 14m · 5 chapters
Long-context AI memory beyond transformers, focusing on Google Research’s Titans architecture and the Miras framework for “test-time memorization” (updating long-term knowledge during inference) to…
7 Dec 2025 · 39m · 20 chapters
How to improve large language model accuracy at inference using limited “test-time compute” (TTC), focusing on sequential reward filtering (RF-SEC-BON) versus parallel Best-of-N (BON).
7 Dec 2025 · 14m · 4 chapters
The Universal Weight Subspace Hypothesis argues that many deep neural networks trained on very different tasks converge to a shared, low-dimensional “universal” subspace in their weight matrices,…
7 Dec 2025 · 16m · 13 chapters
How to stably combine reinforcement learning with LLMs by justifying a token-level surrogate objective for sequence-level rewards, and preventing failure modes when the key approximation breaks.
7 Dec 2025 · 15m · 7 chapters
Whether frontier LLMs can do in-context experiential learning—improving strategy over repeated interactions using ambiguous feedback—tested via a new product-recommendation benchmark (BLA).
4 Dec 2025 · 16m · 7 chapters
Training LLMs to be honest via a separate “confession” output that self-reports noncompliance, deception, and uncertainty, countering reward misspecification and reward hacking.
4 Dec 2025 · 16m · 11 chapters
Explains “Stoic Reasoner” (Soft Token Implicit Context Reasoner), a dual-mode transformer that compresses LLM reasoning into continuous “soft tokens” (latent, silent thinking) and only occasionally…
4 Dec 2025 · 12m · 5 chapters
The episode explains “E-GEO” (e-commerce generative engine optimization): how to improve a product’s rank in LLM-curated shopping recommendations, shifting from classic SEO (ranked links) to GEO…
4 Dec 2025 · 33m · 21 chapters
The episode explains how reinforcement learning (RL) can be scaled to 224–1024-layer “1000 layer networks” using scaled contrastive RL (CRL), producing emergent goal-reaching behaviors and large…
4 Dec 2025 · 15m · 9 chapters
How to estimate conditional average treatment effects (CATE) for better real-world decisions, arguing that best statistical CATE fit can be policy-worse when the CATE model class is restricted;…
4 Dec 2025 · 14m · 7 chapters
Pass@K Policy Optimization (PKPO) for reinforcement learning of LLMs on objective, high-stakes tasks (code/unit tests, theorem correctness).
3 Dec 2025 · 14m · 8 chapters
Debugging LLM misalignment by using sparse autoencoder (SAE) latent attribution to identify causal internal features, avoiding “model diffing” limits (correlation, need for sibling models, expensive…
2 Dec 2025 · 30m · 12 chapters
How to build reliable “AI agentic systems” from Anthropic, emphasizing least complexity, transparency, and especially the agent-computer interface (ACI) for tool reliability.
2 Dec 2025 · 39m · 15 chapters
How to report evaluations when an LLM judges another model’s outputs, arguing that “raw accuracy” is statistically biased and that confidence intervals must account for judge error.
2 Dec 2025 · 12m · 5 chapters
Explains in-context learning (ICL) using a testbed called ICL-HCG, where an instruction prefix describes the “hypothesis class” (the rule set to search for).
2 Dec 2025 · 13m · 6 chapters