AI Post Transformers

This episode explores "Post-Training Science for Supervised Fine-Tuning," a study from Baseten researchers that treats SFT hyperparameter choices — learning rate, batch size, LoRA rank, epochs, and optimizer — as empirical questions rather than inherited folklore. Using controlled, one-variable-at-a-time sweeps across Qwen3 and Llama models ranging from 0.6 billion to 235 billion parameters (inclu...

This episode examines the "Reversal Curse," a 2023 finding from researchers at Vanderbilt, UK AI Safety Institute, Apollo Research, NYU, Sussex, and Oxford showing that autoregressive language models trained on "A is B" statements fail to infer "B is A," even though the two are logically equivalent. Using the example of Valentina Tereshkova, the discussion shows how a model finetuned on "Tereshkov...

This episode examines a solo-authored arXiv paper by Charles O'Neill (Baseten), "Can a Language Model Learn Facts Continually in Its Weights?", which asks whether facts written into a model's weights via LoRA adapters remain usable after dozens or even a hundred subsequent training updates, rather than simply measuring whether accuracy holds up. The discussion traces the theoretical lineage behind...

This episode traces the "Knowing-Using Gap" — the puzzling phenomenon where fine-tuning a language model on a new fact produces instant, perfect recall of that fact in isolation, yet the model fails when asked to actually reason with it in multi-step tasks. Drawing on a July 2026 arXiv paper from HKUST researchers, the discussion covers how the authors use a novel "self-patching" technique — build...

This episode explores Patchscopes, a unifying framework from Ghandeharioun et al. (Google Research and Tel Aviv University, ICML 2024) for inspecting hidden representations of language models. Rather than decoding internal states through narrow tools like probing classifiers, logit lens, tuned lens, or activation patching, Patchscopes patches a hidden representation from a source prompt directly i...

This episode explores Tensor Cache, a memory architecture for transformer inference that addresses the tradeoff between unbounded KV cache growth and the "amnesia" problem of sliding-window attention. Rather than deleting evicted tokens, the approach folds them into a fixed-size associative memory matrix using an outer-product mechanism rooted in Schmidhuber's 1992 fast-weight memory concept, comb...

This episode explores MemGPT, a UC Berkeley paper proposing to manage large language model context windows the way operating systems manage virtual memory, paging information in and out rather than trying to cram everything into a fixed prompt window. The hosts unpack why this matters: self-attention's quadratic cost caps context size, and even models with huge windows suffer from the "lost in the...

This episode traces the evolution from classic knowledge distillation to on-policy distillation and finally to Lightning OPD, a technique from NVIDIA researchers for post-training large reasoning models. The hosts unpack why on-policy distillation offers denser training signal than RLVR's sparse end-of-trace rewards, but has historically required an expensive live teacher model running alongside t...

This episode examines SAC, a disaggregated KV-cache system built to serve sparse-attention LLMs like DeepSeek-V3.2 by pairing prefill and decode instances with a CXL-based memory pool instead of RDMA. The discussion traces how sparse attention mechanisms — specifically DeepSeek Sparse Attention's Lightning Indexer, which selects only the top-2048 relevant KV entries per layer — break the assumptio...

This episode explores a paper introducing Gimbal, a serving system for Mixture-of-Experts LLMs that coordinates two scheduling decisions previously handled separately: which backend engine a request enters through and where individual experts are physically placed across GPUs. The hosts walk through why request-count-based load balancing fails for MoE models — a 200-token and a 2,000-token request...

This episode examines "Rethinking KV Cache Eviction via a Unified Information-Theoretic Objective," a paper that attempts to derive an optimal approach to KV cache eviction from first principles rather than stacking heuristics. The discussion covers the two dominant camps in cache eviction—attention-pattern-based methods like SnapKV and H2O versus structure-aware methods like KeyDiff and Knorm—and...

This episode examines a black-box audit method for probing what large language models associate with a given name, applied across eight models including GPT-4o, GPT-5, Grok-3, and several locally-run open models. The researchers built WikiMem-style completion probing that works without token probabilities, testing models against 100 famous public figures and 100 invented synthetic names to isolate...

This episode closes a three-part arc on "Understanding and Steering the Cognitive Behaviors of Reasoning Models at Test-Time," examining how researchers identify and manipulate the specific attention heads responsible for triggering non-linear reasoning detours in language models. The discussion covers the technical pipeline: segmenting chains of thought at delimiter tokens, fitting per-head linea...

This episode explores "Thought Anchors: Which LLM Reasoning Steps Matter?" by Paul C. Bogdan and Uzay Macar, with senior authors Neel Nanda and Arthur Conmy, examining which specific sentences in a chain-of-thought trace carry disproportionate causal weight over a model's final answer. The discussion covers why standard mechanistic interpretability tools, built for single forward passes, break dow...

This episode explores DeltaProduct, a linear-RNN architecture that improves state-tracking by generalizing DeltaNet's single Householder reflection into a product of multiple reflections per token. The discussion traces the theoretical foundation: transformers and diagonal linear RNNs like Mamba face a proven complexity-class ceiling (TC0 vs NC1) that prevents them from tracking permutation-group...

This episode explores OPSDL (On-Policy Self-Distillation for Long-Context Language Models), a technique out of Baidu that tackles the gap between a model's advertised context window and how much of it the model can actually reason over faithfully. Rather than training on a separate reward model or human-labeled preferences, OPSDL has the same model supervise itself: a version reading a short, evid...

This episode explores a technique called "Still," which compresses transformer key-value caches into a fixed-size representation in a single forward pass. The discussion walks through why KV caches balloon with long-context agents, the two-axis taxonomy of compression methods (selection versus synthesis, per-context versus amortized), and how prior work only amortized selection while synthesis rem...

This episode explores why identical reinforcement learning recipes produce wildly different reasoning ability in similarly-sized language models, examining "Cognitive Behaviors that Enable Self-Improving Reasoners" from Stanford and SynthLabs. Training Qwen-2.5-3B and Llama-3.2-3B on the number-puzzle game Countdown with identical PPO settings, Qwen jumps to roughly 60% accuracy while Llama platea...

This episode explores Parcae, a paper on scaling laws for stable looped language models, where instead of stacking distinct transformer layers, a single block is applied repeatedly through the residual stream — echoing Universal Transformers and ALBERT's weight-tying but tackling the training instability that has historically plagued the approach. The discussion centers on reframing looped inferen...

This episode explores how speculative decoding — the standard trick for speeding up autoregressive LLM inference by having a cheap draft model propose tokens for batch verification — breaks down when applied to Mixture-of-Experts models. The discussion traces why decoding is memory-bandwidth bound rather than compute bound, then shows how MoE routing decouples a draft token's acceptance probabilit...

This episode explores how inference-time scaling breaks down once large language models shift from short chat responses to long chain-of-thought reasoning, drawing on Micron and Argonne National Laboratory's research spanning models from 8 billion to 671 billion parameters. It explains the divide between compute-bound prefill and bandwidth-bound decode phases, and how reasoning traces exceeding te...

This episode explores STEEL, a sparsity-aware fused attention design for running long-sequence prefill inference efficiently on AMD's XDNA neural processing unit. The discussion contrasts spatial-dataflow NPU architectures, where compute tiles are explicitly scheduled with no dynamic cache management, against GPU SIMT execution, and explains how the causal attention mask creates load imbalance tha...

This episode examines HyperOffload, a graph-driven scheduling system from Shanghai Jiao Tong University and Huawei that shifts LLM memory offload and prefetch decisions from a reactive runtime into compile-time graph scheduling for terabyte-scale "SuperNode" hardware. The hosts scrutinize the paper's headline 26% peak memory reduction, arguing it's largely definitional since it comes from offloadi...

This episode examines SoundnessBench, a new benchmark testing whether frontier LLMs can judge the underlying soundness of a research proposal before any experiments are run, rather than just executing and scoring completed work like prior agent benchmarks (MLE-Bench, PaperBench, InnovatorBench). Built from 1,099 ICLR proposals labeled with reviewers' soundness sub-scores rather than acceptance out...

This episode closes out a trilogy on self-improving coding agents by examining the Huxley-Gödel Machine, which formalizes AI self-improvement as a tree-search problem and challenges the field's default assumption that an agent's raw benchmark score is the right signal for choosing which agent to build on next. The discussion traces the lineage from Schmidhuber's 2003 Gödel Machine — a proof-based...