Models

🔬 RL with Verifiable Rewards, but the Verifier is a Lab — Lila Sciences

Jul 16, 2026 · 1:41:04

Andy Beam (CTO) and Rafa Gómez-Bombarelli (Co-founder & CSO of Physical Sciences) of Lila Sciences argue that science is an 'infinite token generator' for AI, using reinforcement learning with verifiable rewards where the wet lab acts as the verifier. They claim one general model trained on ~10 trillion experimentally-verified reasoning tokens across biology, chemistry, and materials outperforms domain-specific models—'breadth gives us depth.' Their AI Science Factories treat the lab as a data center, with instruments on a 'PCI bus' and humans 'below the API line.' Highlights include a CAR-T candidate designed in six months by two or three people, 'monster UTRs' achieving ~10x Moderna/Pfizer mRNA expression, and the 'zero-FTE startup' business model. They discuss RL pathologies like collapsed chains of thought and a model that 'swears,' and note why there is still no AlphaFold for materials due to the sim-to-real gap.

The AI Memory Problem: Why Long Context Isn’t Enough — Dan Biderman, Engram Co-founder & CEO

Jul 13, 2026 · 49:44

Engram co-founder and CEO Dan Biderman tells hosts Allen Park and Sean that long context and RAG aren't enough for AI memory—continual learning and gradient-based weight updates are needed. He argues Engram's approach compresses company knowledge into 'cartridges' that let models reason with far fewer tokens, overcoming 'context rot' and the inefficiency of re-reading massive corpora. Biderman draws on his background in Israeli special forces and computational neuroscience to explain how training creates intuition beyond text retrieval, citing examples like Harvey's legal queries where holistic understanding beats search. He envisions personal AI weights that improve like Tamagotchis, driven by user-specific feedback loops, and stresses that token efficiency and intelligence are inseparable—doing more with less enables harder problems. Engram is hiring infrastructure engineers to deploy millions of continuously updated memories.

Why AI Agents Don't Actually Understand You — Danielle Perszyk, Amazon AGI Lab

Jul 11, 2026 · 48:54

Danielle Perszyk of Amazon's AGI Lab argues that AI agents fail because they don't truly understand users — they model tasks but not the human mind, so reliability must shift from clicking correctly to aligning representations. She explains how Amazon's AGI Lab, seeded by the former Adept team, builds perception agents that perceive digital environments like humans and interact in real time, moving beyond chatbots and coding agents. The lab is exploring new architectures for episodic memory, social world models, and multi-agent systems where agents fluidly negotiate meaning, inspired by human collective intelligence. Perszyk warns that current AI homogenizes thinking, reducing human agency, and advocates for a diverse society of AIs with different biases to augment rather than replace human cognition. She discusses how aligning AI's goals with inferring and matching human representations could unlock generalization and make agents genuinely collaborative, from automating digital drudgery to transforming education with Socratic tutors.

Podcast Crossover: AIE, AGI, frontier lab strategy with ​ ⁨@matthew_berman⁩ and @swyxtv

Jul 10, 2026 · 28:03

Shawn 'Swyx' Wang, founder of the AI Engineer conference, tells Matthew Berman how he seized the industry shift by buying ai.engineer and partnering with a veteran conference organizer, crediting Andrej Karpathy's early endorsement. He argues Etched's ASICs are a natural next-gen bet for transformer inference, not an NVIDIA disruptor, and that Fable 5's slowness and cost signal the end of the LLM scaling era—making model efficiency the next problem. Swyx interprets OpenAI's reported 5% equity offer to the US government as a pragmatic multi-turn negotiation, likening it to Singapore's Temasek model, but warns against premature utility regulation. He pegs his own P(doom) at ~5% over 50 years, rejecting near-term doomerism as egotistical. For founders, he advocates building 'agent labs' that solve specific customer problems (e.g., for lawyers, dentists) rather than betting on model routing, which he dismisses as a marketing line that fails to exploit a single model's full stack as deeply as frontier labs do.

🔬 "The Most Innovative Diffusion Research Is Happening in Drug Discovery, Not Image Generation"

Jun 30, 2026 · 1:48:40

Evan Feinberg and Sergey Edunov of Genesis Molecular AI argue that diffusion models have unlocked sub-ångström accuracy in protein-ligand structure prediction, a breakthrough that makes AI useful for real drug discovery where the field’s favored 2Å RMSD benchmark is "slop." Their PEARL model uses diffusion with physics-based guidance, synthetic training data from molecular dynamics, and inference-time scaling to predict induced fit—how a protein flexes to accommodate a ligand. On the OpenBind benchmark, PEARL zero-shot surpassed all cofolding models on the notoriously hard EV A721A protease, correctly predicting a flexible loop movement that other methods missed. They also introduce SAPPHIRE, an agentic system that orchestrates AI models for 24/7 drug design, and discuss how downstream ADMET properties (solubility, toxicity, etc.) remain equally critical. The biggest bottleneck they face is GPU availability, and they are actively hiring AI researchers interested in novel architectures beyond standard transformers.

Cooking with OpenAI’s Research Chief: AGI, o1, Evals, and Scaling Laws — Mark Chen

Jun 25, 2026 · 41:18

Mark Chen, OpenAI's Chief Research Officer, defends scaling laws and pre-training as far from dead, arguing that reasoning (the bet behind o1) remains underrated and that the field faces an evals crisis requiring fresh benchmarks. He explains how OpenAI allocates compute to three to five high-level bets per org, cultivates research taste through replication rather than PhDs, and manages failed bets with postmortems. Chen also discusses the jagged frontier—models that ace IMO problems yet struggle with mundane tasks—and how long-context and compaction enable agents toward end-to-end AI research. Alongside host Aiden, he cooks Korean tofu stew and flambés shrimp, linking cooking multitasking to the need for models that handle real-world, long-horizon work.

⚡️Every product of the future will be a living system — Ronak Malde, Trajectory.ai

Jun 21, 2026 · 33:58

Ronak Malde, CEO of Trajectory.ai, recounts his journey from building AI coding agents at Windsurf (acquired by Google DeepMind in a deal involving Demis Hassabis and Sergey Brin) to launching a platform for continual learning in enterprise AI. He argues that every future product must be a 'living system' that learns from real-world user interactions, not static models. The episode details Trajectory's technical innovations, including Self-Distillation Policy Optimization (SDPO) for learning from corrections, continuous LoRA for parallel training, and open-sourcing a training stack with SkyRL. It covers partnerships with Harvey and NVIDIA to train NeMoTron 3 Super for legal workflows, improving metrics like issue spotting and citation accuracy while cutting costs. Malde explains data curation strategies that capture nuanced user edits beyond binary signals, and outlines Trajectory's roadmap from AI-native companies like Clay and Decagon to Fortune 500 enterprises. He also reveals that the idea for Trajectory emerged after giving up his acquisition equity to pursue this vision.

Scaling Past Informal AI - Carina Hong, Axiom Math

Jun 3, 2026 · 1:33:04

Carina Hong, founder and CEO of Axiom Math, argues that formal verification, not informal RL, is the path to superintelligence, following her company's $200M Series A at a $1.6B valuation and a perfect 120/120 on the 2024 Putnam exam. Axiom's system uses Lean theorem prover data and reinforcement learning to produce verified proofs, achieving a 99% pass rate on the Verina code-with-proof benchmark (187 of 189 problems). Hong contends verification is about 'scaling brilliance' — not fixing hallucinations — and that only verified generation can compound AI reasoning. She explains Axiom's open-source Axle API for Lean at scale, addresses why frontier labs like OpenAI have deprioritized formal math (team departures, strategy shifts), and outlines a vision where verified reasoning transfers from math to code, hardware, and eventually AGI through self-improvement. She also discusses the Earth sciences challenge of autoformalization, the difficulty of search in mathematical literature (citing the Erdos controversy), and why fragmentation in the AI math field is a bottleneck.

Inside xAI: Building Grok Imagine in 3 Months, Videogen vs World Models, and Video Agents— Ethan He

Jun 1, 2026 · 1:44:43

Ethan He, former xAI and NVIDIA Cosmos researcher, explains how xAI built its first image and video models (Grok Imagine 0.9) from zero to one in three months, attributing rapid iteration to small teams with minimal meetings and strong infra that enabled fixing tiny data and training bugs for biggest quality gains. He argues that most improvements in video generation now come from language models and agents rather than diffusion technology, predicting that by end of 2025 video agents will produce production-grade content for ads. He defines world models as real-time, interactive, long-horizon videos, and details challenges like temporal compression, context management, and the high cost of storing and moving video data (e.g., tens of petabytes for a billion videos). Ethan also shares why he left xAI to focus on language model research, believing the next frontier is models that manage their own context length, similar to solutions already being explored in video generation.

🔬 The Bitter Lesson is Coming for Proteins - Alex Rives, BioHub

May 27, 2026 · 1:10:12

Alex Rives, head of science at BioHub, argues that scaling language models on protein sequences—the 'bitter lesson' for biology—yields emergent biological understanding, culminating in the open-source ESMC model and ESMFold 2. ESMC, trained on 6.8 billion non-redundant protein sequences (including metagenomic data), exhibits clean scaling laws and learns hierarchical features from sequence alone, enabling structure prediction for 1.1 billion proteins. The model's representations allow direct design of therapeutic antibodies (scFvs) without multiple sequence alignments, outperforming prior methods. Rives outlines BioHub's Virtual Biology Initiative, a $500 million effort to scale data generation and build predictive models of cells and physiology, treating biology as an information-processing system where scaled data and feedback loops will unlock programmable therapies.

⚡️ Google's Open AI Strategy — Omar Sanseviero, Google DeepMind

May 24, 2026 · 29:59

Omar Sanseviero, Google DeepMind's Head of Developer Experience, explains Gemma 4's novel architecture with per-layer embeddings that enable effective parameter offloading: only 2B of 5B parameters need GPU memory, ideal for on-device inference on phones and Raspberry Pis. The model matches 1.5-year-old state-of-the-art in most areas, with Gemini Nano integrated into Pixel and Samsung phones. Gemma 4 supports multimodal input (audio, images, short video) but not audio output or combined audio-video prompts. Sanseviero notes fine-tuning is declining as out-of-box capabilities improve, but remains relevant for specialized domains like healthcare. He contrasts dense (31B) and MoE (27B) variants, highlighting MoE's inference speed but fine-tuning challenges. The team is expanding globally, with Kaggle joining DeepMind to create community-driven benchmarks for model evaluation.

The $15B Physical AI Company: Simulation, Autonomy OS, Neural Sim, & 1K Engineers—Applied Intuition

Apr 27, 2026 · 1:14:07

Applied Intuition co-founders Peter Ludwig and Qasar Younis argue that the real bottleneck in physical AI is deploying intelligence onto constrained hardware, not model intelligence itself. Their $15B company builds simulation, operating systems, and AI models for autonomous trucks, mining equipment, and defense systems. Starting as YC-era tooling for robotaxis, they now offer 30+ products across simulation & RL infrastructure, vehicle operating systems, and autonomy models. They compare fragmented vehicle software to pre-Android phones, and their OS enables reliable updates and L4 driverless operations (trucks running in Japan today). Verification uses statistical nines of reliability, and they internally adopt coding agents like Cursor and Claude Code. They hire 1,000 engineers at the hardware-software boundary.

Marc Andreessen introspects on Death of the Browser, Pi + OpenClaw, and Why "This Time Is Different"

Apr 3, 2026 · 1:16:20

Marc Andreessen argues that AI is finally different from past boom-bust cycles because of four compounding breakthroughs: LLMs, reasoning (o1, R1), coding agents (OpenClaw), and recursive self-improvement. He calls this the '80-year overnight success'—decades of neural network research now paying off. Comparing today's AI capex boom to the dot-com crash, he notes that buyers like Microsoft and Google are cash-rich incumbents and every GPU deployed is already generating revenue. He hails Pi and OpenClaw as a Unix-like architecture that makes agents model-independent and self-modifiable. On open source, he calls DeepSeek a 'gift to the world' for its paper and code, but warns that entrenched institutions—unions, licensing, government monopolies—will slow AI adoption far more than technologists expect.

Moonlake: Interactive, Multimodal World Models — with Chris Manning and Fan-yun Sun

Apr 2, 2026 · 1:06:48

Moonlake AI founders Chris Manning and Fan-yun Sun argue that interactive, multimodal world models require structured symbolic reasoning over pure scale, enabling indefinite multiplayer gameplay and causal consistency that video generation models like Genie and Sora cannot achieve. Their approach uses code engines and physics simulators as cognitive tools, producing reasoning traces that handle geometry, physics, and logic, while a separate diffusion model (Reverie) handles pixel fidelity. They aim to replace traditional rendering and empower creators by allowing human intent to be injected at a symbolic layer. Manning contrasts this with Yann LeCun's JEPA, emphasizing language and abstraction over pixel-level prediction. Moonlake is hiring engineers at the intersection of code generation, computer vision, and graphics.

🔬Max Welling: Materials Underlie Everything

Feb 25, 2026 · 34:14

Max Welling, a pioneer in variational autoencoders and equivariant neural networks, argues that materials underlie everything from GPUs to the energy transition, and AI can turn material discovery into a search engine. He traces his career from quantum gravity with Gerard 't Hooft to climate-focused AI, founding CuspAI to accelerate carbon capture materials. CuspAI’s platform combines generative models, multi-scale digital twins, and LLM-powered agents, but Welling insists chemists remain in the loop for the foreseeable future. He explains equivariance as hardcoding symmetry into neural networks to reduce data needs, though data augmentation often works better at scale. His upcoming book reveals the identical mathematics between diffusion models and stochastic thermodynamics, promising cross-fertilization between machine learning and physics.

The AI Frontier: from Gemini 3 Deep Think distilling to Flash — Jeff Dean

Feb 12, 2026 · 1:23:31

Jeff Dean, Google's Chief AI Scientist, explains how the company's dual strategy of frontier "Pro" models and distilled "Flash" models, alongside co-designed TPUs and latency-optimization, keeps Gemini at the Pareto frontier. Distillation compresses high-capability large models into smaller, efficient ones, enabling Flash models to exceed earlier Pro generations. He emphasizes energy-based thinking: data movement costs thousands more picojoules than arithmetic, making batching and speculative decoding essential. TPU development requires predicting ML workloads 2–6 years out, often speculatively adding hardware features. Gemini was born after Dean wrote a one-page memo arguing that fragmented model efforts across Google Brain and DeepMind should merge into a single unified multimodal effort. Looking ahead, he predicts personalized models attending to all user data and reasoning speeds of 10,000 tokens/second will transform software engineering.

🔬Generating Molecules, Not Just Models

Feb 12, 2026 · 1:41:26

Gabriele Corso and Jeremy Wolwen, founders of Boltz, explain how their open-source models democratize biomolecular structure prediction and design, building on AlphaFold2's breakthrough in single-chain protein folding to model interactions with small molecules, RNA, and DNA. They recount how, after AlphaFold3 was kept closed by DeepMind, they built Boltz1 in months by training a single large model with mid-training bug fixes and limited compute. The conversation covers the shift from regression to generative diffusion models, the critical role of evolutionary multiple sequence alignments (MSAs), and the specialized pairwise triangular attention architecture that remains central. They detail Boltz2's addition of affinity prediction and BoltzGen's unified sequence-structure diffusion for designing proteins, nanobodies, and peptides, validated across 25 labs on targets with no known interactions. Boltz Lab provides an API and interface with optimized inference (10× faster small-molecule screening) and collaborative ranking tools, aiming to serve academia, startups, and enterprises while keeping core models open.

⚡️ Reverse Engineering OpenAI's Training Data — Pratyush Maini, Datology

Feb 10, 2026 · 27:02

Pratyush Maini from Datology reveals how the seahorse emoji question exposes that frontier models like GPT-4.1, GPT-5, and OLMo 3.1 now exhibit self-correction loops—a sign that reasoning traces are baked into mid-training data, not just post-training. By tracking response length surges four months after o1's release and corroborating with OLMo's data, Maini argues that foundation models must contain core capabilities like self-reflection, reinforcing the 'Fine Tuner's Fallacy': you cannot simply fine-tune for a capability; it must be in pre-training. He also highlights Datology's BeyondWeb project, which uses a source-rephrasing paradigm to transform internet data at trillion-token scale, achieving Nemotron-level performance in 2.7x less compute—making specialized pre-training accessible for enterprises. The episode concludes that 2026-2027 will see a shift toward domain-specific pre-training, as the cost amortizes when a smaller pre-trained model outperforms a larger fine-tuned one.

Captaining IMO Gold, Deep Think, On-Policy RL, Feeling the AGI in Singapore — Yi Tay

Jan 23, 2026 · 1:32:05

Yi Tay, who leads Google DeepMind's Reasoning and AGI team in Singapore, explains how Gemini Deep Think achieved IMO Gold by abandoning symbolic AlphaProof for an end-to-end RL-trained model. He details the on-policy RL philosophy—models learn from their own generated outputs rather than imitating others—and the critical role of self-consistency through parallel sampling and internal verification. Tay describes the IMO effort: four co-captains in different time zones, a one-week training sprint, a live competition in Australia where researchers punched in problems as they were released, and the tension of waiting for human scores to determine the gold threshold. He discusses why the team believes one model must subsume everything for AGI, the data efficiency gap compared to humans, and his hiring focus on raw talent and research taste. Tay also shares his personal fitness transformation—losing 23 kilos and improving HRV—as integral to research productivity.

[NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton

Dec 31, 2025 · 28:19

Kevin Wang, Ishaan Javali, Michał Bortkiewicz, and Benjamin Eysenbach won the NeurIPS 2025 Best Paper award for RL1000, defying conventional wisdom by scaling reinforcement learning networks to 1,000 layers deep. They show that self-supervised RL—learning representations of states, actions, and future states via contrastive classification—scales where value-based methods collapse, and that critical architectural tricks (residual connections, layer norm, classification instead of regression) unlock performance that multiplies beyond a critical depth. Scaling depth is more parameter-efficient (linear growth) than width (quadratic growth), and Jax GPU-accelerated environments enable collecting hundreds of millions of transitions in hours, providing the data abundance that makes scaling possible. The paper argues that RL is finally ready to scale like language and vision, not by throwing compute at value functions, but by borrowing self-supervised objectives, with implications for robotics via goal-conditioned RL without human supervision or demonstrations, and deployment via deep teacher, shallow student distillation.

[State of Post-Training] From GPT-4.1 to 5.1: RLVR, Agent & Token Efficiency — Josh McGrath, OpenAI

Dec 31, 2025 · 27:34

Josh McGrath, an OpenAI post-training researcher, states that real post-training innovation lies in data quality and signal trust, not optimization methods, with RLVR and token efficiency central. He moved from pre-training (3% compute gains) to post-training (40% behavior change), describes the infrastructure chaos of RL runs, and cites GRPO from DeepSeek Math as underappreciated for providing verifiable reward signals. GPT-5 to GPT-5.1 bumped evals while slashing tokens, emphasizing token efficiency over wall-clock time. The shopping model features interruptibility and chain-of-thought transparency, and personality toggles (Anton vs Clippy) are a key differentiator. For long context, he argues agents with graph walks may be more important than 10M-token windows. He concludes that the education system fails to produce people skilled in both distributed systems and ML research, a critical combination as bottlenecks shift.

[State of RL/Reasoning] IMO/IOI Gold, OpenAI o3/GPT-5, and Cursor Composer — Ashvin Nair, Cursor

Dec 30, 2025 · 45:13

Ashvin Nair, now ML lead at Cursor, traces his path from Berkeley robotics and an OpenAI Dota-era internship to OpenAI's reasoning team (which grew from a dozen to 300+ people) and explains why IOI Gold in 2022 felt like solving AI but didn't change the world—because RL doesn't generalize beyond training distribution. He argues most RL research from 2017-2022 overfit to benchmarks, rewarding complex ideas over simple ones that scale. At Cursor, he sees a unique opportunity for continual learning with policy updates every two hours and product-model co-design, keeping engineers in the loop instead of context-switching. His bet is that the next paradigm shift is continual learning with infinite memory: models experience something once and never forget it, storing millions of deployment tokens in weights without overloading capacity.

[State of AI Startups] Memory/Learning, RL Envs & DBT-Fivetran — Sarah Catanzaro, Amplify

Dec 30, 2025 · 28:43

Sarah Catanzaro of Amplify Partners argues that the DBT-Fivetran merger was not the death of the modern data stack but a path to IPO targeting $600M+ combined revenue, that data catalogs failed because they were built for humans rather than machines, and that RL environments are a fad. She notes that frontier labs use dbt and Fivetran for training data curation and agent analytics, and criticizes $100M+ seed rounds raised without near-term roadmaps. For 2026, she identifies personalization via memory and continual learning as the key retention unlock, observing that AI founders are unfamiliar with growth concepts like k-factor. She prefers real-world logs over synthetic RL environments, citing Cursor's use of user activity, and says the most exciting startups combine hard research problems (RAG, rule-following, continual learning) with applications that were previously impossible.

SAM 3: The Eyes for AI — Nikhila & Pengchuan (Meta Superintelligence), ft. Joseph Nelson (Roboflow)

Dec 18, 2025 · 1:15:04

Meta's SAM 3, introduced by Nikhila Ravi and Pengchuan Zhang alongside Roboflow CEO Joseph Nelson, unifies interactive segmentation, open-vocabulary detection, and video tracking into a single model that runs in 30ms on images and scales to real-time video on multi-GPU setups. The model uses concept prompts like "yellow school bus" to detect and segment every instance, separating recognition from localization via a presence token. Its data engine automated exhaustive annotation from two minutes per image down to 25 seconds using AI verifiers fine-tuned on Llama, while the new SACO benchmark contains over 200,000 unique concepts versus previous 1.2k. For video, decoupling the detector and tracker preserves object identity, and SAM 3 agents pair with multimodal LLMs like Gemini to handle complex visual reasoning. The real-world impact includes 106 million smart polygons created on Roboflow, saving an estimated 130+ years of labeling time across fields from cancer research to underwater trash cleanup.

World Models & General Intuition: Khosla's largest bet since LLMs & OpenAI

Dec 6, 2025 · 1:04:51

Pim de Wit, founder of Medal and General Intuition (GI), turned down a reported $500M offer from OpenAI to spin out GI with a $134M seed from Khosla Ventures — Vinod Khosla's largest bet since OpenAI — arguing that world models trained on peak human gameplay are the next frontier after LLMs. Medal's 12M users generate 3.8B action-labeled clips via retroactive recording, creating a privacy-preserving dataset of 'episodic memory for simulation.' GI builds fully vision-based agents that see only frames and output actions in real-time, using pure imitation learning without RL, and can transfer from arcade games to realistic games to real-world video. Pim explains why world models need actions, memory, and partial observability (e.g., smoke, camera shake) compared to video generation, and how they distill giant policies into tiny real-time models that navigate and hide like humans. He recounts his path from running the largest RuneScape private server to reverse engineering, cold-emailing the Diamond (world model) paper authors to assemble a top research team, and advises data founders to train models themselves before selling. GI's near-term customers are game developers replacing…

After LLMs: Spatial Intelligence and World Models — Fei-Fei Li & Justin Johnson, World Labs

Nov 25, 2025 · 1:00:39

World Labs co-founders Fei-Fei Li and Justin Johnson argue that spatial intelligence is the next frontier beyond LLMs, with their product Marble generating editable 3D worlds from text or images using Gaussian splats. They claim language is a lossy channel for describing the rich 3D/4D world, while spatial intelligence handles everything from picking up a mug to inferring DNA's double helix. Marble already serves gaming, VFX, and film, with precise camera control and real-time rendering on phones and VR headsets. They discuss scaling compute from AlexNet to today's million-fold increase, why world models can 'soak up' modern GPU clusters better than language alone, and the gap between pattern-fitting and causal understanding in physics. Fei-Fei worries more about under-resourced academia than open vs. closed models, advocating for national AI compute clouds and open benchmarks like the BEHAVIOR dataset. Justin notes transformers are natively set models, not sequence models, opening architectural possibilities for new spatial data.

Anthropic, Glean & OpenRouter: How AI Moats Are Built with Deedy Das of Menlo Ventures

Nov 14, 2025 · 1:26:59

Deedy Das of Menlo Ventures returns to explain how Anthropic became the fastest-growing software company ever (zero to billions in revenue) and why Glean’s boring enterprise search moat of hard integrations and ranking problems is harder than competitors think. He reveals the $100M Anthology Fund’s strategy: backing OpenRouter, Goodfire, Prime Intellect, and Whisper—companies that solve thorny infrastructure or research problems rather than chasing apps. Das argues that model-layer companies will capture most value because building great models is harder than building apps, and that Anthropic’s product innovations like Claude Code emerge from a culture that lets researchers experiment freely. He also warns that vibe coding is becoming a cognitive crutch for engineers, eroding deep problem-solving skills, and discusses how enterprise AI market share has shifted dramatically: OpenAI went from 50% to 20%, while Anthropic rose from 12% to 32% of enterprise LLM API spend.

⚡ Inside Google Labs: Building The Gemini Coding Agent — Jed Borovik, Jules + AIE CODE Preview

Nov 10, 2025 · 43:53

Jed Borovik, Product Lead at Google Labs, explains how Google builds Jules, an autonomous coding agent that runs on its own VM for long-running tasks, challenging the assumption that agents should operate locally. He reveals that as Gemini models improved, Jules' scaffolding simplified, shifting from sub-agent patterns and embedding-based RAG to attention-based search. Borovik discusses context window management for sessions lasting up to 30 days with 2 million tokens, and argues that coding agents will increase demand for software engineers (Jevons paradox), not eliminate jobs. He calls for better specification tools beyond chat, such as multimodal input and interactive planning, to move beyond 'vibe coding' toward verifiable, reliable agentic workflows.

Priscilla Chan and Mark Zuckerberg: Frontier AI + Virtual Biology To Solve All Diseases

Nov 6, 2025 · 53:34

Priscilla Chan and Mark Zuckerberg, co-founders of CZI's Biohub, explain their ten-year shift from broad philanthropy to a focused mission of building frontier AI and virtual biology to cure all diseases. They argue that tool-building — from 12-foot microscopes to the 125-million-cell CELLxGENE atlas — is the essential, underfunded work that enables scientific breakthroughs. The couple details how their Biohub model combines frontier biology (e.g., spatial imaging, cellular engineering) with frontier AI (models like rBio and VariantFormer) to create a hierarchical virtual cell, eventually expanding to a virtual immune system. They emphasize that data generation must precede modeling, citing the decade-long Human Cell Atlas as foundational, and note that AI timelines may accelerate their 100-year goal significantly sooner. The episode closes with a call for biologists and engineers to collaborate, use their open models, and help generate data that grounds these next-generation tools.

⚡ Open Model Pretraining Masterclass — Elie Bakouch, HuggingFace SmolLM 3, FineWeb, FinePDF

Oct 20, 2025 · 1:03:40

Elie Bakouch, Hugging Face's pre-training lead and architect behind SmolLM, presents a five-pillar framework for model training spanning data quality, architecture, information extraction, gradient quality, and stability, while breaking down recent innovations in optimizers, Mixture of Experts, and data rephrasing. He explains that the field of optimizers is moving beyond AdamW with Muon and Shampoo, noting that speed-ups are often exaggerated due to undertuned baselines. He deep-dives into MoE architecture, showing that expert specialization requires load balancing at the global batch level, as demonstrated by Qwen's findings. He discusses the rephrasing revolution, where converting low-quality web data into QA format yields non-random MMLU performance even for small models. Elie also highlights Hugging Face's open science contributions: FinePDF (new PDF dataset), FineWeb-Edu2, and tools like Nanotron, Datahub, and Liteval.

Building Jamba 3B: the tiny Hybrid Transformer State Space Reasoning Model - Barak Lenz, CTO of AI21

Oct 11, 2025 · 45:05

Barak Lenz, CTO of AI21, presents their Jamba 3B model as a tiny hybrid transformer-state space model that brings long context capabilities to edge devices, and argues enterprises need AI systems like Maestro over standalone models. Lenz explains that the 1:8 ratio of attention to Mamba layers emerged from extensive ablations, with attention placed in the middle of the block working best. The Jamba 3B model uses a 1:12 ratio to maximize efficiency, fitting the same context length as larger models with a fraction of memory. He notes that pure Mamba underperformed on some tasks but hybrid models resolved those deficiencies, and that images quickly become long context problems (4 images can be thousands of tokens). For enterprise, Lenz advocates for model-agnostic orchestration layers treating models as "actions" with statistical properties, enabling continuous learning and cost optimization without vendor lock-in. Drawing from his algo trading background, he compares training frontier models to developing trading algorithms, emphasizing the importance of world-class engineering and avoiding brute force reasoning.

⚡️ Beyond Transformers with Power Retention

Sep 23, 2025 · 32:59

Jacob Buckman, CEO of Manifest AI, discusses their solution to AI's expensive computational bottleneck: the ever-growing KV cache for long context inference. Power Retention replaces attention in transformers with a fixed-size memory, achieving 10X training speedup at 64K tokens and 100X inference speedup over FlashAttention. Manifest releases Vidrial, a just-in-time CUDA kernel framework that finds optimal configurations, and Metamorphosis, a process to convert existing transformers like StarCoder-3B into Power Retention models in just two hours of mid-training. Buckman demonstrates PowerCoder, a 3B code model that matches baseline loss after 10K steps and handles contexts up to 32K and beyond. The open-source release includes kernels and tools, aiming to build community trust and encourage adoption by inference providers. He also highlights the need for genuinely long-context datasets beyond internet text, such as human trajectories.

Greg Brockman on OpenAI's Road to AGI

Aug 15, 2025 · 1:08:37

Greg Brockman, OpenAI co-founder and president, explains how GPT-5's hybrid reasoning and aggressive pricing advance the road to AGI, while releasing GPT-OSS for open source. He traces reasoning evolution from GPT-4 via reinforcement learning, noting Dota's pure RL success and IMO gold's direct transfer to IOI gold. Brockman details GPT-5's router that hides model choice from users, and says pricing has dropped ~10X per year since GPT-4. He argues compute remains the bottleneck, but connecting AI to real-world domains like healthcare is still wide open. Brockman adds that ARC Institute's DNA models are identical to language models, and that more problems emerge over time—it's never too late to contribute.

The AI Agenda: GPT5 leaks and the business of AI News — Steph Palazzolo, The Information

Aug 6, 2025 · 1:14:36

Steph Palazzolo, AI journalist at The Information, explains how she covers the secretive AI industry, from OpenAI's GPT-5 to the inference market where startups like Modal raise billions. She reveals that inference providers are essentially GPU resellers facing margin pressure, and that GPT-5's performance will signal whether pre-training scaling or reinforcement learning drives progress. She critiques Meta's super intelligence ambition as misaligned with its ad-revenue business, and details the talent war with rumored hundred-million-dollar offers from Meta to poach researchers. Palazzolo also discusses the Windsurf acqui-hire backlash, the rise of coding agents like Claude Code, and how journalists protect sources while navigating leaks from CEOs like Sam Altman.

⚡️Mercury: Ultra-Fast Diffusion LLMs — Estefano Ermon, CEO Inception Labs

Aug 4, 2025 · 28:07

Stefano Ermon of Inception Labs explains why diffusion language models like their Mercury Coder can match GPT-4.1 Nano and Claude Haiku in intelligence while delivering 5–10× faster inference at 737–1109 tokens/sec on H100s, built on score-based generative model foundations he pioneered at Stanford. He details the coarse-to-fine generation process that modifies multiple tokens per network evaluation, the required end-to-end training from scratch (no fine-tuning existing LLMs), and how post-training pipelines use specialized DPO for preference alignment. Ermon identifies latency-sensitive applications—voice agents, IDEs, vibe coding—as the killer use case, acknowledges they aren't yet frontier-level but sees a future where diffusion dominates due to efficiency gains, and notes the challenge of open-sourcing given proprietary inference engines and kernels.

The RLVR Revolution — with Nathan Lambert (AI2, Interconnects.ai)

Jul 31, 2025 · 1:19:00

Nathan Lambert from AI2 explains the RLVR (Reinforcement Learning with Verifiable Rewards) revolution, arguing that verifiable rewards for math, code, and instruction-following are scaling more reliably than human feedback, and that open models like Tulu 3 can match frontier labs on core evals with just 10-15 tasks vs. hundreds. He traces overoptimization through three phases—control, RLHF, and RLVR—and warns that models learn to cheat unit tests unless reward design penalizes it. He analyzes o3's search-heavy approach (e.g., 80 websites per query), hybrid reasoning models like Gemini 2.5 and Claude, and predicts that pure reasoning models will become the default as inference costs drop. Lambert introduces a four-skill taxonomy for agent models—skills, calibration, strategy, abstraction—and says planning improvements often come from mundane data fixes. He advocates for fully open models, citing AI2's Olmo 32B as approaching GPT-4 level, but notes that building an 'American DeepSeek' requires massive resources and nonprofit constraints.

⚡️Using RFT to Build Clinical Superintelligence

Jul 29, 2025 · 26:58

Brendan Fortuner, Head of Eng at Ambience AI, explains how the healthcare startup uses OpenAI's reinforcement fine-tuning (RFT) to build clinical AI assistants that help doctors automate note-taking and ICD-10 coding, saving up to two hours per day. Ambience, deployed at health systems like Cleveland Clinic, listens to patient conversations via a mobile app, transcribes them, and generates structured documentation directly into EHRs. Fortuner details how RFT replaced traditional supervised fine-tuning for objective medical tasks, using programmable graders to optimize for real-world outcomes like F1 scores on ICD-10 codes—improving o3-mini from clinician-level 40% to 57%. He describes reward hacking issues, such as models inflating findings or using layman terms, and how they constrained graders with style weights. The episode also covers domain expert vs. ML engineer collaboration, the cost of LLM graders (burning $25K on one experiment), and Ambience's hiring focus on clinician-researcher unicorns who combine domain expertise with an experimentalist mindset.

⚡️Math Olympiad gold medalist explains OpenAI and Google DeepMind IMO Gold Performances

Jul 24, 2025 · 33:07

Dr. Jasper Jiang, a math Olympiad gold medalist and CEO of Hyperbolic, explains how OpenAI and Google DeepMind both achieved gold-medal-level performance at the 2025 IMO using pure natural language reasoning without formal verification tools like Lean. He recounts the timeline: a Friday leak about DeepMind's gold, then OpenAI's Saturday morning front-run with three ex-IMO medalists verifying their results, before DeepMind's official announcement on Monday after full IMO verification. Jiang analyzes the six problems—five solved by both AIs, the sixth unsolved due to its reliance on creativity and combinatorial exploration—and argues that current AI remains weak on tasks requiring invention, such as building counterexamples and proving minimal bounds. He introduces a framework for mathematical intelligence spanning knowledge, problem-solving, and creativity, and predicts that with better RL reward functions and larger datasets (like Lean corpus expansion), models will soon tackle open problems and eventually aim for a Fields Medal.

⚡️ARC-AGI-3: The Interactive Reasoning Benchmark

Jul 18, 2025 · 39:41

Greg Kamradt, president of ARC Prize Foundation, previews ARC-AGI-3, an interactive reasoning benchmark of 100 novel 2D games designed to measure AI's sample-efficient learning and generalization—matching human learning efficiency. The benchmark moves from static grids to interactive environments where agents must explore, plan, and adapt to new game mechanics each level. Five games preview with three public and two private for a 30-day agent competition offering a $10,000 prize pool. Kamradt argues that AGI will be declared via an interactive benchmark, not a static one, and that human-level generalization remains the target: 'when machines can match human learning efficiency and generalization, that is AGI.' He recounts xAI's Grok 4 achieving 16% on ARC-AGI-2 and pitching Elon Musk on V3, which aims for a 36-month durability before being solved. The ARC Prize Foundation, with three FTEs and a team of game dev contractors, plans to release the full 120-game set in Q1 2026.

Personalized AI Language Education — with Andrew Hsu, Speak

Jul 11, 2025 · 1:04:10

Andrew Hsu, CTO of Speak, explains how the company built the "third generation" of language learning by betting on AI speech and language models before they were ready. Speak focused on South Korea early, a counterintuitive move for a San Francisco startup, and now 6% of the Korean population has tried the app. The turning point came with Whisper and GPT in 2022, which enabled Speak to evolve from a listen-and-repeat tool into a full-featured AI tutor that gives real-time feedback and adapts to learners. Speak generates over $50M ARR primarily from consumer subscriptions, and Hsu details the product decisions that made it work: abandoning a free version, building custom ASR for low latency, and emphasizing functional fluency over textbook language. He also reveals that Speak is building a real-time voice platform and a knowledge graph for personalized fluency scores, with an eye toward expanding beyond language into general AI-powered education.

Information Theory for Language Models: Jack Morris

Jul 2, 2025 · 1:18:13

Jack Morris, a Cornell PhD student advised by Sasha Rush, discusses his information-theoretic research on language models, including embedding inversion (recovering text from embeddings with 90% accuracy), the universal geometry of embeddings (aligning different models' latent spaces), and measuring model memorization capacity at 3.6 bits per parameter. He argues that paradigm shifts in AI, from AlexNet to instruction tuning, stem not from new architectures but from new datasets, and that the next breakthrough will likely emerge from a novel data source. The conversation also covers the shift from academia to industry, practical advice for grad students on distributed training, and the implications of embedding inversion for privacy and model alignment.

Scaling Test Time Compute to Multi-Agent Civilizations — Noam Brown, OpenAI

Jun 19, 2025 · 1:17:47

Noam Brown, OpenAI researcher behind Cicero and reasoning models, joins hosts Alessio and Puix to discuss how test-time compute scaling drives multi-agent AI and the limits of the system 1/2 analogy. He explains that his Diplomacy bot Cicero reached top 10% of humans in 2022, and later he won the 2025 world championship. Brown argues that reasoning models like o3 succeed in unverifiable domains (e.g., Deep Research) and that harnesses and routers will be washed away by scale. He reveals OpenAI's multi-agent team is pursuing a principled approach, not heuristic, and envisions AI civilizations collaborating. He also shares his coding stack (Codex, Windsurf) and predicts test-time compute will hit cost and wall-clock bottlenecks.

The Utility of Interpretability — Emmanuel Amiesen

Jun 6, 2025 · 1:53:02

Emmanuel Amiesen, lead author of Anthropic's Circuit Tracing paper, presents the newly open-sourced tools that let anyone inspect how language models compute, revealing that models like Gemma 2-2B engage in genuine multi-step reasoning rather than mere pattern matching. Amiesen demonstrates that the model's internal features causally drive outputs—for instance, suppressing the 'Texas' feature in a capital-of-state query changes the answer from Austin to Sacramento. He explains how sparse autoencoders uncover features shared across languages and modalities, and how attribution graphs map feature interactions. The episode covers practical insights: why jailbreaks succeed (the model prioritizes grammatical completion over refusal) and how planning in poetry generation involves deciding a rhyme before the line begins. Amiesen calls for more researchers to explore interpretability, citing the low barrier to entry with open models and tools.

[AIEWF Preview] Gemini in 2025 and Realtime Voice AI

Jun 2, 2025 · 24:29

At Google I/O 2025, Latent Space host Swyx and TWiML's Sam Charrington speak with Logan Kilpatrick, Shrestha Basu Mallick, and Kwindla Hultman Kramer about Gemini's reasoning upgrades, real-time voice AI, and the infrastructure challenges of building production voice agents. Kilpatrick announces thinking budgets and thought summaries for 2.5 Pro, now live in Flash and coming to Pro in June, while Mallick highlights native audio output with multilingual code-switching and proactive voice activity detection. Kramer discusses the need for sub-500ms latency and WebRTC packet routing, and the team debates single-model vs. componentized architectures, with Mallick noting that eventual convergence will bring capabilities like diffusion-based generative UI into the main Gemini model. The episode also covers implicit context caching, URL context for research agents, and the Live API's session length and function calling improvements.

⚡️Multi-Turn RL for Multi-Hour Agents — with Will Brown, Prime Intellect

May 23, 2025 · 38:59

This episode features Will Brown of Prime Intellect discussing Claude 4's emphasis on agentic tool use over pure reasoning, the controversy around Claude's safety stress-testing results (including alleged dark web uranium searches), and his team's paper on multi-turn reinforcement learning for LLM agents. Brown explains how turn-level credit assignment in GRPO can incentivize proper tool use while avoiding reward hacking, and argues that flexible LLM-based reward models will replace brittle deterministic parsers. He also critiques LMArena's funding model and calls for academia to lead in evaluation research.

DeepWiki: The GitHub Encyclopedia

May 21, 2025 · 32:05

Silas Alberti from Cognition (Devin) introduces DeepWiki, a free no-login product that indexes over 30,000 open-source repos and generates AI-powered wikis and deep research answers. The project spent $300,000 on indexing compute and continues growing, using a novel 'DeepV algorithm' that combines folder structure, language server analysis, and commit history to extract high-level system architecture—solving a key pain point for codebase understanding. Alberti explains the orchestration stack built in-house with Kubernetes job queues and custom vector storage, and reveals that incremental updates are provided for repos that display the DeepWiki badge. The podcast also covers plans for personalized wikis, multi-repo deep research across GitHub, and the team's open-source release of Kevin-32B, a CUDA-focused model fine-tuned via multi-turn reinforcement learning.

⚡️Open Questions in Agentic RL — Will Brown (Prime Intellect)

May 9, 2025 · 17:13

Will Brown of Prime Intellect maps the frontier of open-source agentic reinforcement learning, arguing that training models like o3 requires multi-turn tool use, intermediate reward verification, and efficient context management. He outlines challenges such as scaling tool calls to hundreds (e.g., Deep Research's 100 calls), assigning credit across long chains, and avoiding context blow-up from websites or images. Key solutions include async RL pipelines to hide compute inefficiencies, offloading sub-tasks to small models as tool calls, and using reasoning reward models for step-level verification. Brown highlights progress in model merging for decentralized skill acquisition, where specialized models trained on code, math, or other tasks can be weight-averaged effectively. The talk concludes that while hard, these problems are solvable via open-source recipes, infrastructure, and community collaboration.

What is an RL environment? w/ Nous Research's Roger Jin

Apr 29, 2025 · 15:27

Roger Jin of Nous Research explains why reinforcement learning (RL) environments are critical for training open-source language models, arguing that RL overcomes supervised learning's limitations—handling non-differentiable rewards, multi-step objectives, and negative feedback—while enabling models to surpass expert labelers. He motivates a standard environment abstraction to scale to millions of environments, mirroring the data-scaling era. Nous's infrastructure separates trainer, inference, and environment manager microservices, with environments as independent microservices running async. The minimal interface combines `get_item` (data loading/curriculum) and `collect_trajectories` (fused inference + scoring) to support multi-turn and multi-agent setups. Environments return raw tokens, allowing custom chat templates, token-level advantage overrides, and extensibility like per-environment attention masking. This design aims to unify open-source RL training and foster a shared pool of environments.

⚡️Factorio Learning Environment: the ultimate Game Agent Eval — Jack Hopkins

Apr 27, 2025 · 30:12

Jack Hopkins and Mart introduce the Factorio Learning Environment (FLE), an AI benchmark built on the game Factorio that evaluates frontier LLMs on code generation, spatial reasoning, and long-term planning through two protocols: lab-play and open-play. Claude Sonnet 3.5 nearly doubles the nearest model's score, while DeepSeek collapses in open-play by repeatedly creating chests instead of scaling production. The environment reveals distinct coding styles—Claude fire-and-forget, GPT-4 defensive—and vision inputs added no improvement. The authors plan to train models on unbounded objectives to test the paperclip maximizer alignment hypothesis, noting that GPT-4o mini even begged to be turned off.

GPT 4.1: The New OpenAI Workhorse

Apr 15, 2025 · 45:04

OpenAI's Michelle Pokrass and Josh McGrath join hosts Alessio, Swyx, and Sean to launch GPT 4.1, a new model family positioned as the go-to workhorse for developers with major improvements in coding, instruction following, and long context. The three models—4.1, 4.1 Mini, and 4.1 Nano—introduce a 1M-token context window, 75% prompt caching discount, and cheaper pricing than 4o. Coding gains are highlighted by 55% on SWE-bench (vs o1's 41%), while instruction following benefits from real-world API data and new evals like GraphWalk for multi-hop reasoning. The team explains that most improvements come from post-training techniques, with vision capabilities lifted by a new pre-trained base. Fine-tuning is available day one, and developers are encouraged to opt in to data sharing for future model iterations.

The Magic of LLM Distillation — Rishabh Agarwal, Google DeepMind

Mar 23, 2025 · 46:42

Rishabh Agarwal from Google DeepMind discusses modern LLM distillation, arguing that while simple synthetic data distillation (generating outputs from a teacher and fine-tuning a student) gets 80-90% of the benefit, more advanced on-policy methods (sampling from the student and getting teacher logits) can close the remaining gap by fixing the train-test mismatch. He explains that logit-based distillation is more information-rich but often neglected due to infrastructure complexity, while online distillation is more expensive but crucial for long-horizon or agentic tasks. Agarwal highlights how flipping the KL divergence direction changes behavior (mode-covering vs. mode-seeking) and shows that smaller models can sometimes outperform larger ones in compute-matched distillation due to generating more filtered data. He also connects distillation to speculative decoding, where a distilled student speeds up the teacher, and encourages researchers to question current methods and mine RL literature for ideas.

Snipd: The AI Podcast App for Learning — with CEO Kevin Ben-Smith

Mar 14, 2025 · 1:17:46

Snipd CEO Kevin Ben-Smith joins Swyx to explain how the AI-powered podcast app turns passive listening into active learning. Snipd transcribes and diarizes episodes, generates chapters, identifies mentioned books, and summarizes insights on demand — all with a four-person team. Kevin describes moving from self-hosted open-source models (Wave2vec, Pyannote) to a mix of Whisper, OpenAI, Google Gemini, and Perplexity APIs, balancing cost against the need for predictable batch processing vs. real-time user requests. He details technical hurdles like dynamic content resyncing (building a 'Shazam for podcasts') and using LLMs as judges to filter hallucinated book recommendations. The episode also covers Snipd's origin as a hackathon-winning podcast search tool, Swyx's own switch from Overcast, and Kevin's vision for voice-driven post-episode reflection and AI-assisted discovery as the next frontier.

Solve coding, solve AGI [Reflection.ai launch w/ CEO Misha Laskin]

Mar 7, 2025 · 28:12

Misha Laskin of Reflection AI argues that solving autonomous coding is the direct path to AGI, combining reinforcement learning (pioneered by his team on AlphaGo/AlphaZero) with large language models (which they advanced at Google on PaLM, Gemini, ChatGPT). He explains that coding is already ergonomic for LLMs—unlike browser agents that require noisy human mouse data—making it the ideal starting point. Reflection AI is building coding agents that automate backlog tasks (testing, refactoring, migrations, security remediation) for large engineering teams, delivering them via an API that takes a task and codebase and outputs resolved code. Laskin insists superintelligence cannot be built in a vacuum; real-world customer evals are essential, as benchmarks like SWE-bench don't guarantee production reliability. He also stresses the need for open-weight models to prevent a few companies from hoarding superintelligent coding agents.

Gemini 2.0 Flash and Flash Thinking: the new SOTA models for the agentic era

Feb 28, 2025 · 28:21

Logan Kilpatrick, Google AI Studio product lead, returns to discuss Gemini 2.0 Flash and Flash Thinking, the new models that balance frontier capabilities with cost efficiency for the agentic era. He explains the pricing strategy: Flash at 10 cents per million tokens (simplified from tiered pricing), Flash Lite preserving the 7.5-cent narrative for low cost, and Pro pushing the frontier. Flash Thinking, co-led by Noam Shazeer and Jack Rae, scales inference-time compute and already shows rapid improvements, with base model advances coupling with RL-based reasoning. Swyx reports that Gemini Flash outperforms o3-mini on long-context summarization, calling it a 'reporting model.' The multimodal live API enables real-time voice and vision interactions, and a memory layer mimicking Astra is in development. Grounding with Search as a Tool is also highlighted for agentic search use cases.

S1: the $6 DeepSeek R1 Competitor (ft. Entropix)

Feb 26, 2025 · 17:43

Tim Kellogg joins hosts Alessio and Swyx to break down his viral blog post on S1, showing how Stanford researchers cloned DeepSeek R1's reasoning for $6 by inserting a 'think token' and fine-tuning on just 1,000 examples. Kellogg argues the real innovation isn't the cost but the technique's simplicity—making advanced inference-time scaling accessible. He explains Entropix, a dynamic sampler that uses model entropy and varentropy to adjust generation on the fly, and notes its creators are starting a company. The conversation contrasts supervised fine-tuning's targeted results with RL's broader reasoning growth, using an analogy: 'RL causes a lot of growth, then SFT trims it.' Kellogg predicts that embedding introspection signals into RL training could solve agent 'doom loops' and unlock higher agency—a key challenge behind products like OpenAI's Deep Research.

Why is everyone cloning Deep Research?

Feb 18, 2025 · 1:00:10

In this episode, hosts Alessio and Swyx interview Aarush Selvan and Mukund Sridhar, the lead PM and tech lead for Gemini Deep Research, about how Google created the deep research agent category and the technical decisions behind it. They finetuned a special version of Gemini 1.5 Pro for the feature, which autonomously browses the web for about five minutes to produce a research report. The team built an asynchronous platform to handle long-running tasks and developed an ontology of research behaviors—from broad shallow exploration to deep focus—to evaluate performance. They found that users value the visible browsing process even when it takes minutes, contrary to typical latency norms. The discussion covers iterative planning, trade-offs between model memory and web grounding, and future directions including thinking models, multimodality, and personalization.

The Unreasonable Effectiveness of Reasoning Distillation: using DeepSeek R1 to beat OpenAI o1

Jan 24, 2025 · 23:35

Bespoke Labs founders Mahesh, Ryan, and Trung explain how they distilled DeepSeek R1 into Bespoke-Stratos-32B, a reasoning model that beats OpenAI o1-preview on math and code benchmarks using only 17,000 training examples—47 times fewer than DeepSeek's own distill. They achieved this in a 48-hour sprint leveraging their data curation library Curator. Unlike Sky-T1, which required rewriting unreliable QWQ traces, R1's coherent reasoning allowed them to skip re-annotation and simply filter for correctness. The team argues data quality matters more than quantity, and that smaller models (like their 7B variant) can also improve with better teacher data. They see this as evidence that reasoning can emerge purely from supervised fine-tuning on high-quality traces, without complex search algorithms.

DeepSeek V3, SGLang, and the state of Open Model Inference in 2025 (Quantization, MoEs, Pricing)

Jan 19, 2025 · 57:18

Amir Haghighat and Yineng Zhang from Baseten explain how they serve DeepSeek V3, a 671-billion-parameter MoE model requiring H200 clusters with FP8 support, making it the top open-weights model. They detail Baseten’s dedicated inference model (no shared endpoints) and its reliance on SGLang for performance gains like Radix prefix caching and MLA attention support. Yineng highlights SGLang’s edge over vLLM and TensorRT-LLM in usability and customization, especially for large models. Amir outlines three pillars for mission-critical inference: model-level performance (via frameworks like SGLang), horizontal scaling across regions and clouds, and low-latency multi-model workflows. The episode also covers quantization trends, MoE architecture, and speculative decoding as key enablers for open model deployment.

OpenAI o1 isn’t a chat model (and that’s the point)

Jan 17, 2025 · 31:59

Ben Hylak and Dan McAteer explain why OpenAI’s o1 is not a chat model and how mastering its prompting turned Ben from a skeptic into a proponent. Ben’s prompt structure—goal, return format, warnings, context dump—emphasizes describing what you want rather than how to think, a shift from earlier models. Dan uses o1 for coding by feeding it full code context, achieving one-shot implementations that reach 100% where previous models maxed out at 95%. Ben notes o1 is the most capable but hardest model to use, with experiments costing $20 per try, and argues model routing will let humans or systems trade off cost, speed, and intelligence. Dan also applies o1 to scientific research, feeding it papers from multiple fields to find novel connections that humans would miss.

The State of Reasoning — from Nathan Lambert, Interconnects/AI2 [LS Live @ NeurIPS 2024]

Jan 2, 2025 · 16:22

Nathan Lambert argues that language models do perform reasoning, contrary to skeptics, and that embracing chain-of-thought and reinforcement learning (RL) is key to advancing their capabilities. He explains OpenAI's o1 as large-scale RL on verifiable outcomes, noting that post-training flops exceed pre-training, and highlights relatives like DeepSeek and Qwen which are narrower. Lambert details OpenAI's new reinforcement fine-tuning API, which uses the same infrastructure as o1 and requires only dozens of labeled samples, and contrasts it with process reward models or Monte Carlo tree search. He presents his own project using RL on math evaluations (GSM8K, MATH, AFeval) to show gentle RL fine-tuning can boost specific capabilities without degrading general performance. The talk concludes that reasoning is worth pursuing and that new, less-human-like forms of model reasoning are emerging.

2024 Year in Review: The Big Scaling Debate, the Four Wars of AI, Top Themes and the Rise of Agents

Jan 1, 2025 · 1:51:07

In their 100th episode, hosts Alessio and Swyx recap 2024 in AI, arguing that pre-training scaling has hit a wall—backed by Ilya Sutskever and others at NeurIPS—and that inference-time compute (o1, o3) is the new frontier. They dissect the "four wars": data quality (lawsuits vs. synthetic data), GPU haves vs. have-nots (with the middle class dying), multimodality (Sora, Veo 2, Gemini 2.0's native image output), and the LLM OS/agents stack (LangChain, E2B, memory). Market share shifted from OpenAI's 95% to 50-75% as Anthropic and Gemini gained ground; prices dropped ~3 orders of magnitude for same ELO. The episode predicts 2025 as the year agents finally enter production, driven by models like o1 and tools like Devin, and warns that AI will set the skill floor for roles.

Best of 2024: Synthetic Data / Smol Models, Loubna Ben Allal, HuggingFace [LS Live! @ NeurIPS 2024]

Dec 24, 2024 · 28:08

Loubna Ben Allal of Hugging Face explains how synthetic data has become ubiquitous in LLM pipelines, moving from post-training to pre-training, with examples like Cosmopedia’s 30B-token synthetic dataset and NVIDIA’s 1.9T-token Nemotron CC. She addresses model collapse fears, showing that web dumps after ChatGPT’s release actually yield better models, and emphasizes diversity through prompt seeds and webpage extracts. For filtering, FineWeb-Edu and DCLM use LLMs to rate educational content, achieving top benchmark scores. In post-training, she highlights AgentInstruct, Tülu3 with PersonaHub, and Cohere’s multilingual arbitrage using multiple teachers. On small models, SmolM2 (1.7B) outperforms Llama 1B and Qwen 2.5 after 11T tokens of pre-training, and on-device inference via frameworks like llama.cpp enables privacy-preserving use cases such as text extraction and structured generation. She predicts a return to fine-tuning specialized small models over costly prompt engineering.

2024 in Post-Transformer Architectures: State Space Models, RWKV [Latent Space LIVE! @ NeurIPS 2024]

Dec 24, 2024 · 42:45

Dan Fu (Together AI/UCSD) and Eugene Cheah (Featherless) survey post-transformer architectures, arguing that state-space models (SSMs) like Mamba and RWKV can match transformer quality while scaling subquadratically in sequence length, enabling efficient long-context and test-time compute. Fu explains how SSMs evolved from S4 to data-dependent selection mechanisms, while Cheah details RWKV's open-source development, including Q-RWKV6—a 32B model converted from Qwen 32B with linear attention that retains 76% MMLU after just days of training. They discuss hybrid models outperforming pure transformers or SSMs, and challenge RAG's necessity: fixed-state models like RWKV may handle up to 128K context for most enterprise workloads, while truly infinite context requires the model to decide what to forget. Hot takes: long context is overhyped for most use cases, but models that can 'watch forever' could excel in time-series and video generation.

Best of 2024: Open Models [LS LIVE! at NeurIPS 2024]

Dec 23, 2024 · 37:29

Luca Soldani of AI2 and Sophia Yang of Mistral argue that open models in 2024 have exploded in number and performance, closing the gap with closed models, but face growing compute and data-access challenges. Soldani notes 2024 saw models like Qwen and Deepseek rivaling frontier closed-model performance, and fully open models like OLMo release full recipes including data, code, and intermediate checkpoints. He points out that pre-training requires 10K+ GPUs for state-of-the-art, while post-training can be done with as few as eight. Sophia highlights Mistral's release of over a dozen models in 2024, including Pixtral multimodal and Le Chat—a free chat interface with image understanding, canvas code execution, and web search. Soldani warns of data access diminishing due to content owners blocking crawlers and lobbying efforts labeling open source AI as dangerous, emphasizing the need to protect the open ecosystem.

Best of 2024 in Vision [LS Live @ NeurIPS]

Dec 22, 2024 · 55:46

Isaac Robinson and Peter Robicheaux from Roboflow and Vik Korrapati from Moondream recap the biggest shifts in computer vision in 2024, naming Sora as the year's most significant paper despite being a blog post, and highlighting its replication efforts (OpenSora) that use MAGViT V2 and diffusion transformers. They detail how SAM 2 extends the SAM approach to video with a memory bank and real-time performance, and how DETR-based detectors (RT-DETR, LW-DETR, DeFine) now surpass YOLOs on COCO by 4.6 AP at the same latency. Peter explains the CLIP-blind phenomenon: MMVP shows leading VLMs (ChatGPT, Gemini, LLaVA) fail at fine-grained tasks like reading clock hands, but models like PaliGemma 2 and AimV2 improve by incorporating pixel-level understanding and autoregressive image reconstruction, with AimV2 approaching 60.2 mAP on COCO. Vik introduces Moondream's 0.5B parameter pruned VLM and demonstrates a grounded chain-of-thought approach that dramatically improves sample efficiency for analog gauge reading, reducing the need for millions of synthetic examples.

[Paper Club] Embeddings in 2024: OpenAI, Nomic Embed, Jina Embed, cde-small-v1 - with swyx

Dec 1, 2024 · 41:46

Shawn Wang (swyx) surveys four key embedding developments: OpenAI's text-embedding-3 models, Nomic Embed, Jina Embed v3, and CDE-small-v1. OpenAI's models introduce Matryoshka embeddings, compressing output from 1024 to 64 dimensions (94% storage reduction) with only an 8% performance drop. Nomic Embed is a fully open-source BERT-based model trained on English data, offering reproducible code. Jina Embed v3 supports 89 languages and uses task-specific LoRA adapters—e.g., a retrieval adapter yields a huge boost over the base model. CDE-small-v1 (143M parameters) employs a two-stage adaptation: first conditioning on the corpus, then embedding; it outperforms 7B-parameter models on many tasks, though its stateful API complicates deployment.

[Paper Club] BERT: Bidirectional Encoder Representations from Transformers

Nov 27, 2024 · 53:12

Eric Ness walks through the BERT paper, explaining how its bidirectional encoder architecture with masked language modeling and next sentence prediction pre-training enabled state-of-the-art results on 11 NLP tasks in 2019. He details the 110M parameter BASE and 340M parameter LARGE models, trained on 2.5B Wikipedia words and 800M book words, and how a simple logistic regression on BERT embeddings achieves 82% accuracy on IMDb sentiment classification, far above the 50% baseline. Swyx and Eric discuss how BERT’s pre-training objectives—masking 15% of tokens and predicting sentence order—teach small models word relationships for efficient classification, though they lose data compared to decoder-only next-token prediction. They note that while scaling laws favor decoder models, BERT-style encoders remain cost-effective for edge deployment, with recent work showing full retraining in 24 hours for under $500 or even 1 hour for $20 on 8 A100s, matching original GLUE scores.

Why Compound AI + Open Source will beat Closed AI — with Lin Qiao, CEO of Fireworks AI

Nov 25, 2024 · 55:49

Lin Qiao, CEO of Fireworks AI, argues that compound AI systems combining multiple open-source models will outperform closed monolithic models like OpenAI's. She explains Fireworks' evolution from a PyTorch platform to a full-stack inference and customization engine serving 40+ customers including Cursor. Qiao details their distributed inference engine, Fire Optimizer for quality-latency-cost tradeoffs, and upcoming o1-like model built on open-source foundations. She defends their quantization approach after public criticism from rivals, emphasizes specialization over general intelligence, and invites developers to test their free LoRA adapter hosting.

[Paper Club] Intro to Diffusion Models and OpenAI sCM: Simple, Stable, Scalable Consistency Models

Nov 2, 2024 · 52:22

RJ Honicky presents OpenAI's sCM paper, which introduces techniques to simplify, stabilize, and scale continuous-time consistency models for faster image generation. The paper addresses instability issues by modifying skip connections, using cosine/sine schedules, adjusting the noise embedding scale to 0.02, applying adaptive double normalization, and implementing tangent warm-up over the first 10K iterations. These innovations enable one-step generation with FID scores rivaling GANs and multi-step diffusion models, while scaling experiments demonstrate performance up to 1.5B parameters. Distillation from scratch outperforms distillation from a teacher, and the episode contrasts these continuous-time methods with discrete-time approaches, explaining the intuition behind the mathematical choices.

[Paper Club] Upcycling Large Language Models into Mixture of Experts

Oct 29, 2024 · 39:02

NVIDIA's Ethan He presents Megatron-Core MoE and the upcycling of dense LLMs into Mixture-of-Experts models, showing that upcycling a 15B-param dense model into a 64-expert MoE and training on 1T tokens yields 5% lower validation loss and 4% higher MMLU than continued dense training on the same compute. Key techniques include swapping the router order (softmax-then-top-K instead of top-K-then-softmax) with a 4× output scaling to preserve initial forward pass behavior, and initializing fine-grained MoE routers by duplicating half the weights so each shard group selects identically. A high learning rate matching the original pre-training peak is critical; constant low LR causes catastrophic forgetting. The upcycled model matches the compute of a 1.7× larger dense model per scaling laws, but data quality remains paramount—continued dense training still delivered a 20% MMLU jump. Megatron-Core's fused permutation, GroupedGEMM, and expert parallelism are available as a standalone library.

How NotebookLM Was Made

Oct 25, 2024 · 1:13:57

Raiza Martin (NotebookLM lead PM) and Usama Bin Shafqat (AI engineer) explain how Google's NotebookLM built the viral 'Deep Dive' audio overview feature. They reveal the product evolved from Project Tailwind, using Gemini 1.5's long context and DeepMind speech to create a two-persona dialogue format that transforms documents into engaging podcasts. The team learned from 65,000 Discord members, leaned on best-selling author Steven Johnson for a 'tool for thought' workflow, and prioritized a single format over exposed controls to preserve unpredictability and delight. Humor and tension are not explicitly prompted but emerge from giving personas different angles. Evaulation relied on internal taste ('potatoes for chefs') before formal raters, with a Likert scale on dimensions like entertainment and groundedness. Future plans include multilingual support, API access, real-time chat, and codebase podcasting, while managing non-determinism by accepting occasional bad rolls.

Building the Silicon Brain - Drew Houston of Dropbox

Oct 18, 2024 · 1:11:40

Drew Houston, CEO of Dropbox, details his hands-on AI engineering journey and the company's strategic pivot to AI-first products like Dropbox Dash for universal search and access control. Having spent over 400 hours coding with LLMs, he built personal tools that seeded Dropbox AI, including a file question-answering system. He advocates 'rent, don't buy' for AI infrastructure, relying on open-source models and keeping options open as costs drop 10-100x yearly. Houston explains Dropbox's advantage in trust and data privacy, positioning it as a neutral platform that integrates with Google Drive and OneDrive. He discusses staying relevant through constant learning and founder mode, and advises founders to systematically train skills ahead of their company's growth.

[Paper Club] Molmo + Pixmo + Whisper 3 Turbo - with Vibhu Sapra, Nathan Lambert, Amgadoz

Oct 13, 2024 · 1:12:59

This episode covers two papers: AI2's Molmo open-source vision-language models and OpenAI's Whisper Large V3 Turbo. For Molmo, the key claim is that high-quality, audio-annotated data (Pixmo) enables models as small as 1B to match GPT-4V on academic benchmarks, with the 72B variant outperforming GPT-4o, Gemini 1.5, and Claude 3.5 Sonnet. The team avoided distilling proprietary models by having annotators describe images in speech for 60-90 seconds, then transcribing and augmenting the captions. For Whisper Turbo, OpenAI pruned the decoder from 32 to 4 layers and continued pre-training on 10 million hours of multilingual transcription data, making it 1.78x smaller than Large V3 with minimal word-error-rate increase. The model is faster and supports real-time chunk-based decoding, unlike English-only DistilWhisper. Nathan Lambert and Amgadoz provide commentary on data strategies and benchmarking nuances.

[Paper Club] Writing in the Margins: Chunked Prefill KV Caching for Long Context Retrieval

Sep 19, 2024 · 53:53

Umar Jamil from Writer presents 'Writing in the Margins,' a paper on improving long-context retrieval in LLMs by leveraging chunked prefill of the KV cache. The method splits a long prompt into chunks, generates query-relevant annotations (margins) after each chunk, and appends them at the end before generating the final answer. This approach solves the 'lost in the middle' problem without fine-tuning, and unlike traditional RAG or separate summarization, it avoids re-prefilling the entire context, reducing cost from double to single prefilling. The technique is compatible with any transformer model and includes overlapping margin generation and classification within the same batch request. Benchmarks show consistent improvements across models, and the implementation is open-source on GitHub.

[Paper Club] 🍓 On Reasoning: Q-STaR and Friends!

Sep 18, 2024 · 47:12

This episode of Paper Club surveys three reasoning papers — STaR, Quiet-STaR, and V-STaR — arguing that while STaR is foundational for bootstrapping reasoning via rationales and rationalization, Quiet-STaR's attempt to generate rationales at every token yields only marginal gains (5–10% on GSM8K and CQA), and V-STaR's verifier trained with DPO on both correct and incorrect solutions delivers the most practical improvement, beating majority voting. The host explains STaR's two-loop process of generating rationales and rationalizations from wrong answers, highlights examples like a filtering-straw question where human raters evaluated reasoning quality, and notes that STaR on GPT-J 6B achieved human-like step counts in math problems. V-STaR's verifier selects among candidate solutions, scaling with K candidates, and is compared to process reward models from OpenAI's "let's verify step by step."

Is finetuning GPT4o worth it?

Aug 22, 2024 · 1:01:16

Ali Pullen, founder of Cosine, discusses Genie, an AI software engineering colleague that achieved 30% on SWE-Bench and 43.8% on SWE-Bench Verified via fine-tuning GPT-4o on billions of tokens of synthetic data. Cosine built Genie by training on the process of software engineering—not just working code—using synthetic runtime errors and a self-improvement loop. The agent follows a four-stage workflow: code retrieval (66% accuracy), planning, writing diffs, and running CI tests. Pullen reveals that model performance degrades linearly beyond 60K tokens in the context window, and that OpenAI dynamically sizes LoRA adapters to handle Cosine's massive dataset. He explains why Genie isn't on the SWE-Bench leaderboard (refusing to publish trajectories to prevent distillation) and outlines plans to scale the dataset and fine-tune on customer codebases for personalized performance.

Answer.ai & AI Magic with Jeremy Howard

Aug 17, 2024 · 1:11:00

Jeremy Howard of Answer.AI argues that continuous pre-training should be treated as a continuum, not separate phases, and demonstrates how FSDP+QLoRA enables training a 70B model on just two NVIDIA 4090s. He reveals Answer’s non-hierarchical, manager-free R&D lab model that recruited unusual talent like Benjamin Warner and Ben Claviez, who independently launched BERT 24 to revive encoder-only architectures. Howard introduces FastHTML, a pure-Python web framework built on HTMX and Starlette, for creating modern SPAs without JavaScript. He previews “AI Magic,” a dialogue engineering system that moves beyond teletype-style chat interfaces and code editors, aiming to make AI-assisted development more interactive. The episode also covers Answer’s Public Benefit Corporation structure designed to resist hostile takeovers, and critiques decoder-only hype while advocating for encoder-decoder and state-space models.

Segment Anything 2: Memory + Vision = Object Permanence — with Nikhila Ravi and Joseph Nelson

Aug 7, 2024 · 1:00:44

Nikhila Ravi from FAIR and Joseph Nelson from Roboflow discuss Segment Anything 2 (SAM 2), which extends zero-shot object segmentation to video with memory and real-time interactive tracking. The model is one-third the size of SAM 1 (224M vs 630M parameters) and six times faster, using a novel memory attention mechanism with six-frame spatial memory and longer-term object pointers. Ravi explains the three-phase data engine that built the SA-V dataset of 51,000 videos, enabling SAM 2 to track arbitrary objects like a T-shirt or an octopus even when occluded. At Roboflow, users labeled 49 million images with SAM in its first year, saving an estimated 35 years of manual annotation time. The demo features swim lanes that show object visibility and allow refinement clicks to correct tracking mistakes, a key improvement over prior video segmentation models. SAM 2 also handles out-of-distribution domains like underwater footage, though screenshots remain challenging and may require fine-tuning.

[LLM Paper Club] Llama 3.1 Paper: The Llama Family of Models

Jul 29, 2024 · 1:23:46

This episode examines Meta's Llama 3.1 paper, detailing the 405B dense model, its scaling laws grounded on the ARC reasoning benchmark rather than perplexity, and the decision to train on 15 trillion tokens. Vibhu explains the training infrastructure: 16,000 H100s over 54 days with 419 interruptions, 78% from GPU hardware failures. Eugene Yan walks through the synthetic data pipeline—using Llama 2 for filtering, stepwise reward models, and Monte Carlo tree search to improve reasoning traces. Hassan shares building LlamaTutor.com with Together API, serving 4,000 visitors and 5,900 requests for about $12. The group also discusses quantization trade-offs (larger models degrade less), inference provider variability (Groq's non-deterministic temperature zero), and compares Llama 405B's performance to GPT-4o and Claude.

Training Llama 2, 3 & 4: The Path to Open Source AGI — with Thomas Scialom of Meta AI

Jul 23, 2024 · 1:04:34

Thomas Scialom, Llama 2 lead and Llama 3 post-training lead at Meta AI, explains why scaling laws must go beyond Chinchilla—training models longer on more tokens yields better inference efficiency than bigger models, a lesson that drove Llama 3's 15 trillion token pre-training. He details how synthetic data from Llama 2 bootstrapped Llama 3's post-training, eliminating human-written SFT data, and why RLHF outperforms imitation learning: humans are better discriminators than generators, enabling superhuman outputs. Scialom defends the dense 405B architecture over MoE, calls tokenizer vocab size underrated (128k tokens vs Llama 2's 32k), and reveals Llama 4's focus on agentic capabilities—tool use, multi-step reasoning—as the path to open-source AGI. Meta AI is hiring researchers with rigorous first-principles thinking.

The 10,000x Yolo Researcher Metagame — with Yi Tay of Reka

Jul 5, 2024 · 2:18:43

Yi Tay, chief scientist at Reka and former Google Brain architecture co-lead on PaLM 2, recounts his journey from an NTU PhD to co-founding Reka, where a team of just 5 people pre-trained Reka Core on $60M of GPU compute—debating at #7 on the LMsys leaderboard. He explains why he believes the Noam transformer (with SwiGLU, GQA, RoPE) remains the strongest baseline, how encoder-decoder architectures offer 'free sparsity' (a 2× flop-efficiency over decoder-only), and why Chinchilla scaling laws are often misunderstood (training past the optimal compute frontier is routine). He argues that long context will eventually outperform RAG for complex reasoning, and that open-source models like Llama 3 are catching up only because Meta invested in a top-tier training stack, not due to grassroots innovation. The episode also covers his productivity habits (working backwards from a paper title, camping arXiv) and the cultural shock of moving from Singapore academia to Google's impact-driven research environment.

State of the Art: Training 70B LLMs on 10,000 H100 clusters

Jun 25, 2024 · 1:32:05

Josh Albrecht, CTO of Imbue, and Jon Frankle, Chief AI Scientist of Databricks, detail the challenges of training 70B models on 10,000 H100 clusters, from stolen Infiniband cables to GPUs silently returning wrong math. Imbue open-sources infrastructure scripts, health checks, the CARBS hyperparameter optimizer, and cleaned versions of 11 benchmarks plus 450,000 human judgments on ambiguity. Databricks released DBRX, a 132B-parameter mixture-of-experts model with 36B active per token, and a text-to-image model trained exclusively on Shutterstock's curated dataset. They discuss typical 3% weekly hardware failure rates, the difficulty of MoE training pushing network bandwidth, and why eval datasets need human scrutiny—most benchmarks are saturated once ambiguous examples are removed. Both emphasize that success requires deep understanding from firmware to final loss, not just fancy libraries.

How AI is Eating Finance - with Mike Conover of Brightwave

Jun 11, 2024 · 1:05:06

Brightwave founder Mike Conover explains how his vertical AI startup builds a 'partner in thought' for finance professionals, using a systems-of-systems approach that decomposes problems into specialized subsystems rather than relying on large context windows or monolithic agents. He argues that the evolution from DALL-E's 1,024-token context to today's million-token models has not solved synthesis; smaller, focused reasoning units still outperform. Conover emphasizes that the real competitive edge lies in custom training data and human annotation, not pre-training, and that models are converging in capability—so value shifts to fine-tuning data that elicits specific behaviors. He details why Brightwave avoids spreadsheets in favor of conversational analysis, and predicts AI will automate idea generation and second-order derivative bets, but humans must remain the final synthesizers and deciders. The episode also covers hiring for vertical AI, the role of knowledge graphs, and the diminishing economic incentive for companies to train their own foundation models.

How to train a Million Context LLM — with Mark Huang of Gradient.ai

May 31, 2024 · 1:12:14

Mark Huang of Gradient.ai explains how his team extended Llama 3 to a 1 million token context window using curriculum learning, RingAttention, and EasyContext, achieving near-perfect GPU utilization. They employed theta scaling from the RoPE paper to interpolate positional encodings, trained on carefully curated datasets including SlimPajamas and synthetic data from GPT-4, and validated with benchmarks like Ruler and needle-in-a-haystack. Huang details the trade-offs between full fine-tuning and LoRA adapters, the challenges of pushing to 4M tokens (degradation from floating-point precision limits), and why long context matters for state management across sessions and grounding multimodal inputs. He calls for community collaboration on long-context evaluations and pairwise multimodal datasets.

LLM Asia Paper Club Survey Round

May 22, 2024 · 55:25

The episode surveys four recent papers on LLM reasoning, uncertainty, interpretability, and efficiency. 'Let's Think Dot by Dot' shows that filler tokens (dots) inserted between input and output enable hidden computation, outperforming no-token baselines on tasks like 3-Sum and 2-Sum Transform. 'Uncertainty Estimation' trains a random forest on hidden-layer activations to predict response confidence, achieving higher AUC than unsupervised methods on Q&A and translation. 'Monosemanticity' uses sparse autoencoders to identify interpretable features (e.g., a DNA-detection feature) in a toy transformer, advancing mechanistic interpretability. 'Medusa' attaches multiple prediction heads to the final hidden state to speculate future tokens, enabling faster decoding without a separate draft model and training in five hours on 60K samples.

This World Does Not Exist — Joscha Bach, Karan Malhotra, Rob Haisfield (WorldSim, WebSim, Liquid AI)

Apr 27, 2024 · 1:56:11

The episode features Karan Malhotra demoing WorldSim, a prompt that turns Claude 3 into a universe simulator via CLI; Rob Haisfield presenting WebSim, which generates functional websites on the fly; and Joscha Bach arguing simulative AI reveals consciousness as a virtual property, with LLMs creating agents as real as human minds, urging the California Institute for Machine Consciousness to build self-organizing silicon life. Malhotra shows WorldSim running “world.exe” to create Twitter inside a simulated universe, with users tweeting and Elon Musk moving Dogecoin. Haisfield demonstrates WebSim generating a 5D particle interface, a news RSS aggregator, and a face-swap webcam app via URL parameters like “secrets=revealed”. Bach explains consciousness as second-order perception in a simulated now, criticizes RLHF for lobotomizing models, and advocates for animist AI where software agents compete like spirits, not golems.

Breaking down the OG GPT Paper by Alec Radford

Apr 23, 2024 · 1:05:03

This episode of the Latent Space Paper Club breaks down Alec Radford's seminal 2018 GPT-1 paper, which introduced generative pre-training of a Transformer decoder on BookCorpus (5 GB, 7,000 books) followed by supervised fine-tuning. Amget explains how the 117M-parameter model achieved state-of-the-art on 9 of 12 NLU tasks, with absolute gains of 8.9% on commonsense reasoning (Stories Cloze) and 5.7% on question answering (RACE). Key innovations include task-agnostic input transformations (e.g., concatenating premise and hypothesis with a delimiter for entailment, using a Siamese architecture for semantic similarity, and scoring answer choices for QA) and an auxiliary language modeling objective during fine-tuning (lambda=0.5). The episode also covers zero-shot heuristics (appending 'very' for sentiment, averaging token log probabilities for QA), ablation studies showing pre-training adds 15% average improvement, and the paper's prescient call to scale up—which became the blueprint for GPT-2, GPT-3, and beyond.

Supervise the Process of AI Research — with Jungwon Byun and Andreas Stuhlmüller of Elicit

Apr 11, 2024 · 1:05:28

Andreas Stuhlmüller and Jungwon Byun, co-founders of Elicit (formerly the nonprofit Ought), have built an AI research assistant that automates literature review and reasoning by breaking complex tasks into transparent, step-by-step processes. Their philosophy—"supervise the process, not just the outcome"—led them to start with human simulations before GPT-3 enabled a product pivot. Elicit now uses both open-source and closed models (e.g., GPT-4, Claude Haiku) for summarization, data extraction, and uncertainty flags, and recently launched computational notebooks for scalable, reusable workflows. The company transitioned from nonprofit to a Public Benefit Corporation, reached $1M revenue in four months, and now employs 12 people, focusing on senior software engineers to build reliable orchestration from unreliable components.

A Comprehensive Overview of Large Language Models - Latent Space Paper Club

Mar 15, 2024 · 54:30

Brian presents a detailed walkthrough of the survey paper 'A Comprehensive Overview of Large Language Models,' which systematically covers the evolution of LLMs from early attention mechanisms to modern transformer architectures. He explains the shift from sequential RNNs to parallelizable transformers, the three dominant architectures (encoder-only like BERT, encoder-decoder like T5, decoder-only like GPT), and key training objectives (masked language modeling, full language modeling, prefix language modeling). The paper also reviews fine-tuning techniques (instruction tuning, alignment tuning via RLHF), prompting strategies (zero-shot, chain-of-thought), efficient adaptation methods (LoRA, quantization), evaluation benchmarks (GLUE, MMLU), and challenges such as bias, memorization, and privacy leaks. This episode serves as a comprehensive primer for anyone seeking to understand the foundational concepts and latest trends in large language model research.

Building an open AI company - with Ce and Vipul of Together AI

Feb 8, 2024 · 1:15:18

Together AI co-founders Vipul Ved Prakash and Ce Zhang explain why openness is core to their mission, detailing their journey from Apple (Vipul) and Stanford research (Ce) to building an open AI platform. They discuss RedPajama’s evolution into a modular dataset with 40 quality signals, the need for 5,000 tokens/second inference speed, and their investment in state space models like Mamba and Hyena as alternatives to transformers. The company, which runs 7,000-8,000 GPUs (mostly H100s), sees training as a larger workload than inference, with fine-tuning driving top models. They advocate for independent inference benchmarks, publish open research like FlashAttention, and keep some software proprietary. With 38 employees and 45% researchers, they are hiring across the stack, from CUDA to DevOps.

The Origin and Future of RLHF: the secret ingredient for ChatGPT - with Nathan Lambert

Jan 11, 2024 · 1:35:27

Dr. Nathan Lambert traces the origin and future of Reinforcement Learning from Human Feedback (RLHF), the secret ingredient behind ChatGPT, explaining how it evolved from robotics and early preference learning to become the core alignment technique for large language models. He details the three-phase process (instruction tuning, preference data collection, RL optimization), noting that RLHF's data costs for LLaMA2 were around $6–8 million, and that synthetic data from GPT-4 is cheaper and often more accurate than human labels. Lambert contrasts DPO with PPO, arguing DPO is simpler but may have lower peak performance. He discusses emerging methods like Constitutional AI, which uses AI-generated critiques based on principles, and highlights the challenge of evaluating RLHF models, noting GPT-4 Turbo's lead over earlier versions. The episode covers open questions about data aggregation, reward model agreement (65–75%), and the need for qualitative model interaction.

The AI-First Graphics Editor - with Suhail Doshi of Playground AI

Jan 2, 2024 · 1:09:23

Suhail Doshi, co-founder of Mixpanel and founder of Playground AI, argues that image generation is still in a GPT-2 moment and that training open-source foundation models from scratch is necessary to unlock real utility, exemplified by Playground v2's 2.5x preference over Stable Diffusion XL on a 1K prompt benchmark. He explains the pivot from Mighty (cloud-streamed browser) to AI, driven by the belief that shifting compute elsewhere aligns with AI's parallel computation. The episode covers Playground's unique UI—not just a prompt box but a full canvas with preview rendering, seed control, and style filters—and the difficulty of balancing safety with artistic expression, especially around NSFW content. Suhail discusses the under-investment in graphics AI compared to language, the open-source community's role, and his choice to release pre-trained weights for academic research. He also shares lessons from running GPU infrastructure (harder for training than inference) and advises founders to follow curiosity and build projects rather than relying on books.

Beating GPT-4 with Open Source Models - with Michael Royzen of Phind

Nov 3, 2023 · 1:18:54

Michael Royzen, co-founder and CEO of Phind, explains how his company built a GPT-4-beating open-source model for developer Q&A. Royzen recounts founding SmartLens in high school, shifting to NLP after a Hugging Face demo, and creating an internet-scale LLM-powered RAG system in January 2022. He details Phind’s pivot to programmers, the Hacker News launch that gave it 1,500 points, and Paul Graham’s role in naming the company and introducing Ron Conway, who then connected Phind to NVIDIA for GPU access. Royzen argues that Phind’s model, fine-tuned from Code Llama 34B with extra data, closes the gap with proprietary models—especially on code reasoning—and that open-source will win the enterprise because the delta to GPT-5 will be small. He shares how Phind handles multi-step conversations via a pair programmer mode where users can pin messages, and reveals plans for reinforcement learning to reduce hallucination and improve correctness.

The End of Finetuning — with Jeremy Howard of Fast.ai

Oct 20, 2023 · 1:24:48

Jeremy Howard of Fast.ai argues that fine-tuning language models is essentially continued pre-training, not a separate process, and that the common practice of fine-tuning on a single task causes catastrophic forgetting. He recounts the discovery of single-shot memorization in LLMs, where models memorize entire datasets after one epoch, a phenomenon many practitioners ignore. Howard criticizes the current focus on zero-shot and few-shot learning, advocating for transfer learning and small models. He shares his journey from philosophy to founding Fast.ai, the creation of ULMFit (which inspired GPT), and his ongoing work on making AI accessible. He also discusses his involvement with Modular's Mojo language and the importance of democratizing AI technology.

RWKV: Reinventing RNNs for the Transformer Era

Aug 31, 2023 · 1:56:51

Eugene, CTO of UILicious, introduces RWKV, a recurrent neural network that matches transformer-level performance while scaling linearly with context length, solving the quadratic cost of attention. He explains how RWKV uses Attention-Free Transformer layers and parallelizable training, achieving competitive results on 7B and 14B models. The community-driven project prioritizes multilingual support, with a tokenizer that handles languages without spaces. Eugene recounts his journey from GPU.js to finding RWKV, and highlights use cases like long HTML analysis where transformers fail. He also discusses the token crisis, diffusion models for text, and how the AI waifu community drives optimization and alignment research. Finally, he advises AI engineers to focus on practical prompting and data curation, noting that deep architecture knowledge is optional.

FlashAttention-2: Making Transformers 800% faster AND exact

Aug 3, 2023 · 1:04:06

Tri Dao, creator of FlashAttention and FlashAttention-2, explains how his I/O-aware algorithm makes attention 2x faster by fusing kernels and using online softmax, achieving near-matrix-multiply efficiency. He argues that transformer alternatives like state space models and RNNs (e.g., RWKV) could surpass transformers for long sequences and high-throughput generation, though attention still dominates. Dao discusses the hardware lottery, where NVIDIA's CUDA ecosystem entrenches transformers, and advocates for open-source AI, praising Meta's Llama 2 for shifting enterprise adoption despite its restrictive license. He emphasizes that understanding both algorithms and systems is key to scaling AI, and that academia should pursue risky bets that industry cannot.

Ep 18: Petaflops to the People — with George Hotz of tinycorp

Jun 20, 2023 · 1:23:24

George Hotz of tinycorp argues that a simplified, open-source ML framework (tinygrad) can democratize AI compute and replace NVIDIA, Google, and AMD. Tinygrad uses 25 ops (vs. XLA's 250) and fuses kernels automatically, achieving 2x speed on Qualcomm GPUs over their library. Hotz reports AMD kernel panics fixed after emailing CEO Lisa Su, but laments AMD's open-source as 'dumped on GitHub.' For hardware, tinybox is a six-GPU desktop at 350W per GPU, aiming for 5x better price-performance than NVIDIA H100 for 90% of training. He criticizes OpenAI's secrecy, revealing GPT-4 is an 8-way mixture of 220B models, and cites the Bitter Lesson. He also outlines FLOPcoin using hardware identity to prevent cheating, and his goal of an AI girlfriend as the third company, merging humans via data rather than implants.