AI Post Transformers

This episode explores whether large language models can estimate their chances of success before acting, update those estimates during a task, and use them to decide when to abstain from costly work. It explains why that matters for agentic systems in coding and software environments, where overconfidence can lead to wasted effort, unsafe actions, or expensive mistakes, and connects the paper to e...

This episode explores ChartNet, a 1.5 million-sample multimodal dataset designed to improve how vision-language models read and reason about charts. It explains why chart understanding is harder than OCR or captioning alone, because models must connect visual marks, axes, legends, numerical values, and language-based reasoning with high precision. The discussion places ChartNet in the context of e...

This episode explores whether language models can express uncertainty in natural language in a way that is actually calibrated and useful, rather than merely sounding cautious or confident. It focuses on the paper’s central distinction between token uncertainty and epistemic uncertainty, arguing that next-token probabilities are a poor proxy for whether a model truly knows an answer. The discussio...

This episode explores a 2026 paper proposing that multiple LLM agents should share transformer KV-cache state, not just text, so they can avoid repeatedly paying the prefill cost of rereading the same plans, critiques, and intermediate outputs. It explains the systems background behind prefix caching, vLLM’s PagedAttention, and SGLang, then focuses on why multi-agent workflows break the exact-pref...

This episode explores Paul Werbos’s 2004 review of reverse differentiation and argues that reverse-mode automatic differentiation, backpropagation, hand-coded adjoints, and adjoint circuits are largely the same core idea expressed in different technical communities. It explains the mechanics of automatic differentiation and reverse mode in clear terms, then traces how these methods diverged histor...

This episode explores how test-time compute should be allocated in large language models, using a recent study that compares parallel sampling, majority voting, shortest- and longest-trace selection, and beam-style search under a common evaluation setup. It explains the paper’s central argument that there is no single best inference-time strategy: some model families behave like short-horizon reas...

This episode explores a 2026 paper on making reasoning models faster and cheaper by shortening their chain-of-thought without sacrificing too much accuracy. It explains the paper’s core claim that short-context RL post-training can itself push models toward concise reasoning, while also unpacking the instability this creates and how the proposed Step-Level Advantage Selection method is meant to st...

This episode explores Apple’s Stochastic KV Routing approach for reducing transformer inference costs by letting some layers reuse key-value caches from earlier layers instead of storing separate caches at every depth. It explains why KV cache memory becomes a major bottleneck for long-context autoregressive decoding, and why depth-wise cache sharing is a different idea from token eviction or temp...

This episode explores how spiking neural networks differ from conventional deep networks by representing information as time-based spikes rather than continuous activations, and why that makes them both appealing and difficult to train. It breaks down the main research directions in the field, including local spike-timing-dependent plasticity, direct gradient-based training of spiking models, and...

This episode explores a survey of FPGA-based neural network accelerators for space applications, focusing on what the literature actually demonstrates rather than treating “space AI” as a single vague category. It explains why FPGAs are appealing for onboard inference, including tight control over hardware, energy efficiency, and the ability to support vision, autonomy, compression, navigation, an...

This episode explores World-R1, a post-training method for improving 3D consistency in text-to-video generation without redesigning the underlying model architecture. It explains how the approach combines reinforcement learning, pretrained 3D reconstruction critics, vision-language rewards, and camera-motion-focused prompt data to push generated videos toward more stable geometry under viewpoint c...

This episode explores the 1997 LSTM paper as a targeted fix for why standard recurrent neural networks failed on long-range dependencies: backpropagated error signals either vanish or explode when carried through many time steps. It explains how earlier training methods like BPTT and RTRL worked, why the underlying optimization problem was mathematically ill-conditioned, and how LSTM’s memory cell...

This episode explores Compressed Convolutional Attention (CCA), a new approach that pushes transformer attention itself into a shared compressed latent space instead of only compressing the KV-cache. It walks through the evolution from standard multi-head attention to MQA, GQA, and MLA, explaining why earlier efficiency methods mostly targeted decode-time memory while leaving much of the attention...

This episode explores AgenticQwen, a system for training small open-weight language models to handle industrial-scale tool use through repeated reinforcement learning and synthetic data generation. It explains the paper’s central idea of dual data flywheels: one that turns reasoning failures into harder verifiable tasks, and another that expands simple agent workflows into branching, tool-using tr...

This episode explores a 2025 DeepMind and University of Alberta preprint arguing that AI is reaching the limits of learning from human-generated data and that the next major advances will come from agents learning through interaction and feedback in environments. It explains the shift from static pretraining to grounded reinforcement learning, defining key ideas like long-term reward optimization,...

This episode explores Leo Breiman’s “Statistical Modeling: The Two Cultures” as a sharp argument about a core divide in data science: building explicit probabilistic models to explain the world versus training algorithms that simply predict well on new data. It traces how that split maps onto inference versus prediction, connects Breiman’s critique to earlier ideas from Tukey and Box, and shows ho...

This episode explores Moonshot AI’s KIMI K2.5, a 2026 multimodal model that aims to improve text reasoning, visual understanding, and agent-style task execution within a single open system. It explains the paper’s two main bets: training text and vision jointly from early stages rather than bolting vision on later, and using an external “agent swarm” orchestration layer to split wide-search tasks...

This episode explores ScoutAttention, a systems paper on speeding up long-context LLM inference by managing KV cache growth more intelligently instead of treating it as a pure memory-capacity problem. It explains why long prompts, retrieval-heavy inputs, and extended reasoning traces make decoding increasingly constrained by memory traffic, and why simply offloading cache data to CPU memory can st...

This episode explores DeepSeek-V4’s claim that million-token context windows may finally be practical for real-world use, not just benchmark demos. It explains how the model combines hybrid attention, mixture-of-experts routing, and manifold-constrained hyper-connections to reduce the usual memory and compute costs of long-context transformers while trying to preserve reasoning, coding, and agent...

This episode explores whether the Muon optimizer can truly scale from promising small experiments to frontier-style language model training, and what it would mean if its reported roughly 2x compute-efficiency gain over AdamW holds up. It examines how Muon differs from standard optimizer setups by applying momentum plus orthogonalization to matrix-shaped hidden-layer weights while keeping embeddin...

This episode explores a Princeton paper on whether multiple long-running, tool-using AI agent trajectories can be combined more effectively by an “aggregator agent” that selectively inspects the full traces, rather than by simple answer voting or compressed summaries. It explains why aggregation gets much harder for long-horizon agentic tasks like web research, navigation, and software repair, whe...

This episode explores TokenDance, a systems approach for serving many LLM-based agents more efficiently by collectively sharing transformer KV caches across synchronized conversation rounds. It explains why multi-agent workloads are fundamentally different from ordinary chat serving: agents persist across rounds, accumulate large KV caches, and often follow an “all-gather” pattern where each agent...

This episode explores RetrievalAttention, a 2024 paper that tries to make long-context LLM inference much cheaper by retrieving only the most relevant key-value cache entries during decoding instead of scanning the entire history every time. It explains why long-context serving is bottlenecked less by raw FLOPs than by memory traffic and KV-cache growth, citing concrete figures such as roughly 125...

This episode explores a 2023 paper on Deep Spiking Q-Networks, asking whether a directly trained spiking version of DQN can compete with earlier conversion-based spiking reinforcement learning methods on Atari while retaining the energy-efficiency promise of spiking neural networks. It explains the technical foundations behind spiking networks, including leaky integrate-and-fire neurons, surrogate...

This episode explores DreamerV3, a world-model reinforcement learning system that claims to use one main configuration across more than 150 tasks spanning Atari, ProcGen, DMLab, robot control, visual control, BSuite, and Minecraft. It explains how world models work—learning compact environment dynamics so an agent can train on imagined futures—and why that approach is appealing for sample efficien...