All episodes

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.

Beating Google at Search with Neural PageRank and $5M of H200s — with Will Bryk of Exa.ai

Jan 10, 2025 · 55:54

Will Bryk, CEO of Exa.ai (formerly Metaphor), details building a neural search engine from scratch using link prediction as 'Neural PageRank' — predicting documents rather than keywords. Exa's new product offers near-perfect lists (e.g., 'startups working on hardware in SF') by scaling compute per query, from milliseconds to a day, like o1 for search. Bryk argues LLMs will become the interface to search, while Exa provides the 'super knowledge' that even AGI will need. He contrasts Exa's neural approach with Google's keyword-based system and Perplexity's reliance on Bing. The company recently purchased a $5M H200 cluster and maintains a culture of nap pods and first-principles thinking.

AI Engineering for Art - with comfyanonymous

Jan 4, 2025 · 52:09

Comfy Anonymous, the anonymous creator of ComfyUI, details in his first-ever podcast interview how his node-based image generation tool overtook Automatic1111 through superior memory management and early support for SDXL, becoming the de facto interface for advanced diffusion workflows. He started coding on January 1, 2023, and released the first version on January 16, driven by a desire to chain models and experiment with custom samplers. His work at Stability AI from June 2023 ensured ComfyUI efficiently handled SDXL, which leaked via early access and forced users from less powerful GPUs to adopt his tool. He discusses model preferences: Flux for consistency, SD 3.5 for creativity, and SD1.5 remaining popular. For video, he highlights Mochi as a 'true' video model with 3D latents, already implemented in Comfy. Now with a core team, ComfyUI is targeting a v1 release with an easy installer on Windows and Mac, while planning monetization through cloud and enterprise features—but keeping the open-source core free.

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 in Agents (from #1 on SWE-Bench Full, Prof. Graham Neubig of OpenHands/AllHands)

Dec 25, 2024 · 51:25

Professor Graham Neubig of CMU and AllHands AI, maintainer of the open-source coding agent framework OpenHands, delivers a talk on the state of agents in 2024, arguing that agents are now capable of automating significant portions of software development but still struggle with information gathering and error recovery. He demonstrates OpenHands solving real coding tasks live, explains key design choices like using Python code execution over granular tool calls to reduce LLM invocations, and shares that Claude currently outperforms other models for agentic tasks due to superior instruction following and error correction. Neubig discusses eight perennial agent problems including agent-computer interface, human-agent interface, model selection, planning, workflows, exploration, search, and evaluation. He predicts every major LLM trainer will focus on agent-oriented models by mid-2025, that agent benchmarks like SWE-Bench and WebArena will saturate and require harder successors, and that the biggest challenge will be designing human-agent interfaces for systems that succeed only 30-40% of the time autonomously but 80-90% with human feedback. The talk also covers workflow memory as a…

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.

The State of AI Startups in 2024 [LS Live @ NeurIPS]

Dec 21, 2024 · 26:35

Sarah Guo and Pranav Reddy argue that 2024 has become a far friendlier ecosystem for AI startups, with the model landscape shifting from OpenAI's near-monopoly to a competitive field where Google's Gemini now leads LMSys Arena and open-source models like Llama 8B score ten points higher on MMLU than Mistral 7B a year ago. They note that OpenAI's API cost has dropped 80-85% in 18 months, and total OpenAI API share fell from ~90% to ~60% as customers switch. The funding environment is rational, not a bubble, with foundation-model labs raising $30-40B but most startups seeing sane valuations; one portfolio company grew from zero to twenty million in PLG-style spending. Key startup themes include first-wave service automation (Sierra, Decagon, Harvey, EvenUp), better search and new friends (Perplexity, Glean, Character, Replica), and democratized creativity (Midjourney, HeyGen). They argue the 'GPT wrapper' narrative is false—applications capture value—and that incumbents face innovator's dilemma because AI changes business models (outcomes-based pricing) and data needs (reasoning traces rarely saved). The speed of change and new markets (legal, healthcare, defense) structurally favor…

Windsurf: The Enterprise AI IDE

Dec 13, 2024 · 1:06:36

Varun Mohan and Anshul of Codeium explain why they built Windsurf, a new AI IDE, arguing that VS Code's API limitations prevented them from delivering the best agentic experience. They detail how Cascade, their agentic system, uses proprietary retrieval and planning models alongside third-party LLMs, and describe their evaluation method that masks commits and tests incomplete code states. The duo also reveals that over 800,000 developers use Codeium extensions, that they still support JetBrains and Eclipse for enterprise customers, and that they intentionally avoided a waitlist launch. They discuss the trade-offs of building first-party vs. third-party models, the importance of 'go slow to go fast' in enterprise infrastructure, and their belief that individual developer profits should come after building switching costs through superior product.

[Paper Club] Weight Streaming on Wafer-Scale Clusters (w/ Sarah Chieng of Cerebras)

Dec 7, 2024 · 43:54

Sarah Chieng of Cerebras presents the weight streaming technique for training giant neural networks on Cerebras wafer-scale clusters, arguing it enables near-linear scaling by separating parameter storage from primary compute. The system uses the Wafer Scale Engine (WSE-2/3) with 900,000 cores and 44 GB on-chip SRAM, avoiding off-chip memory bottlenecks that limit NVIDIA GPUs. MemoryX provides external storage for weights and optimizer states (up to 2.4 petabytes), streaming weights to compute units via the SwarmX interconnect fabric, which aggregates gradients. This design allows training models with up to 120 trillion parameters without complex hybrid parallelism. Cerebras also leverages unstructured weight sparsity to prune 90% of data during transmission and skip zero-value computations on-chip, reducing bandwidth and improving efficiency.

0 to over $8M ARR in 2 months as a Claude Wrapper (Bolt.new, Qodo)

Dec 2, 2024 · 1:36:42

Eric Simons of Bolt.new/Stackblitz and Itamar Friedman of Qodo discuss Bolt.new's viral growth from $4M ARR in four weeks to over $8M ARR in two months as a Claude wrapper, driven by its in-browser WebContainer OS that eliminates local dev setup. They contrast Bolt's '0 to 1' creation for non-developers with enterprise-focused agents like Qodo's testing and code review tools, arguing that specialized agents outperform general-purpose ones. Simons details how WebContainer's unified sandbox enables error self-healing loops, while Friedman explains AlphaCodium's task-breaking that boosted o1 preview to 93rd percentile on Codeforces. The episode covers open-source strategy, pricing tiers ($20 to $200/month), and founder advice on ignoring expert doubt to ship fast, illustrated by Simons' Ironman completion while launching Bolt.

[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] DocETL: Agentic Query Rewriting + Eval for Complex Document Processing w Shreya Shankar

Nov 29, 2024 · 55:53

This episode of Latent Space Paper Club discusses DocETL, a framework by Shreya Shankar for agentic query rewriting and evaluation in complex document processing with LLMs. The paper introduces map, reduce, resolve, and auxiliary operators, along with rewrite directives like data decomposition and gleaning to improve accuracy. The optimizer generates 200 pipeline variants, costing $100, to rewrite user pipelines automatically. Key debates include when to chunk documents (sometimes harmful) and the reliability of LLM-as-judge for evaluation, with guests Eugene Yen and Vibhu arguing that binary metrics, ensembling, and context-dependent thresholds make it viable. Shreya emphasizes that validation agents are critical, and the work builds on her prior EvalGen paper, spanning two years of development.

The new Claude 3.5 Sonnet, Computer Use, and Building SOTA Agents — with Erik Schluntz, Anthropic

Nov 28, 2024 · 1:11:08

Anthropic's Erik Schluntz explains how Claude 3.5 Sonnet achieved state-of-the-art 49% on SWE-Bench Verified by using a minimal agent framework that gives the model full control via tools like bash and file editing, letting it self-correct without hard-coded workflows. He details the importance of tool design over prompt engineering, why XML tags work well, and how computer use reduces integration friction by letting models interact with any browser-based interface. Schluntz also shares his robotics experience, noting that while LLMs and diffusion models are promising for general-purpose robots, reliability at 99.9% and hardware variability remain major hurdles, and he expresses skepticism about self-driving as a business due to high vehicle costs versus driver-equivalent revenue.

[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.

Agents @ Work: Lindy.ai (with live demo!)

Nov 15, 2024 · 1:08:01

Florent Crivello, founder of Lindy.ai, explains how his no-code agent platform evolved from an overly LLM-dependent design to a deterministic 'on rails' system that puts AI in a box, making agents more reliable and easier to use. The product was rebuilt from scratch as Lindy 2.0, featuring step-by-step workflows with structured outputs and memory management that still requires pruning. Crivello reveals that Claude 3.5 Sonnet ended model bottlenecks but warns against over-engineering cognitive architectures, advocating instead for letting models scale. He discusses building their own eval tools (now considering Braintrust), hiring vertical GMs for go-to-market, and the decision to return to in-person work, arguing remote work hampers creativity. He also shares his contrarian views on Europe vs. US tech, testing the Overton window with spicy tweets, and balancing AI safety concerns (10% p(doom)) with building toward a potential utopia.

Agents @ Work: Dust.tt — with Stanislas Polu

Nov 11, 2024 · 58:53

Stanislas Polu, co-founder of Dust.tt, argues that horizontal agent platforms—not vertical, single-use-case agents—will win the enterprise by letting non-technical employees build their own automation. A former OpenAI researcher who worked directly with Ilya Sutskever on formal mathematics and saw GPT-4 internally before ChatGPT, Polu left in 2022 to start Dust, now an open-source infrastructure layer that connects company data (Slack, Notion, GitHub) to LLMs and lets anyone create simple, instruction-driven agents. He contrasts Dust’s API-first, integration-heavy approach with LangChain’s open-source orchestration and Adept’s browser automation bet, insisting that most useful agent workflows require only a few tool calls and that function-calling quality (where GPT-4 Turbo still leads) matters more than model size. Dust achieves 88% daily active user penetration at some enterprise customers, and Polu predicts the first billion-dollar company run by a single person will emerge once platforms like Dust mature enough to let tiny teams scale to massive output.

[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.

In the Arena: How LMSys changed LLM Benchmarking Forever

Nov 1, 2024 · 41:02

Anastasios and Weilin from LM-Sys explain how ChatBot Arena became the de facto standard for LLM benchmarking by replacing static benchmarks with dynamic human preference evaluations. They trace the origin from Vicuna fine-tuning to the viral launch in April 2023, where anonymous side-by-side battles let the community decide which model is better. To address biases like length preference, they developed Style Control, a logistic regression method that adjusts for confounders such as response length and markdown formatting. They address the controversy around labs testing multiple private models, arguing that selection bias is empirically small and the live benchmark self-corrects over time. They also discuss RouteLLM for cost-performance routing, the graduation of ChatBot Arena from LMSys to support new projects, and call for community help with red teaming, vision modalities, and implementing a REPL for coding evaluations.

[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.

Singapore: the AI Engineer Nation — with Minister Josephine Teo

Oct 19, 2024 · 56:40

Singapore Minister Josephine Teo outlines the city-state's refreshed national AI strategy and the Ministry of Digital Development's approach to balancing innovation, safety, and talent development. She details plans to triple AI practitioners from 5,000 to 15,000, the AI Verify testing tools and AI Safety Institute for governance, and internal government AI bots like AI Bot for RAG-based knowledge retrieval. On sovereign AI, she discusses hybrid cloud infrastructure and pragmatic data center expansion. Teo also explains Singapore's new law requiring transparency for AI-generated election content, emphasizing fact-based political discourse.

[Paper Club] SWE-Bench [OpenAI Verified/Multimodal] + MLE-Bench with Jesse Hu

Oct 19, 2024 · 1:01:40

Jesse Hu walks through SWE-Bench, SWE-Bench Verified, SWE-Bench Multimodal, and MLE-Bench — benchmarks that evaluate AI agents on real-world coding and ML tasks, from fixing GitHub issues to winning Kaggle competitions. SWE-Bench scrapes 2,000+ pull requests from 12 Python repos to get free unit tests, with leaderboard scores rising from ~2% to 43% thanks to improved agent scaffolding. SWE-Bench Verified uses human raters and three criteria—well-specified problem, valid test, difficulty on a 0–3 scale—to filter 500 solvable tasks. SWE-Bench Multimodal targets 617 UI tasks from 17 JavaScript/TypeScript libraries, each including an image, but achieves only 12% accuracy. MLE-Bench repurposes Kaggle competitions: agents given 24 hours and an A10 GPU, with o1-preview scoring bronze in 17% of competitions and surpassing grandmasters (7 gold medals vs 5 required), though each run costs $4,000.

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.

Production AI Engineering starts with Evals

Oct 11, 2024 · 1:56:16

Ankur Goyal, founder and CEO of Braintrust, argues that evaluation is the core workflow of production AI engineering and shows how his platform makes evals accessible to software engineers. Drawing from his experience at SingleStore and Impira, he details Braintrust's evolution from an eval tool into an end-to-end AI development platform used by Stripe, Zapier, Vercel, and other top AI teams. He shares market data: OpenAI handles over 95% of production workloads, fine-tuning is declining, and open-source models account for under 5% due to reliability issues. Ankur explains why he avoided building a vector database—the real challenge is permissions and joins, not vector search—and predicts o1-style reasoning will replace complex agent frameworks. He reveals Braintrust's differentiators: hybrid on-prem, TypeScript-first SDK, and declarative eval structures.

[Paper Club] Berkeley Function Calling Paper Club! — Sam Julien, Writer

Oct 5, 2024 · 41:29

Sam Julien (Writer) walks through the Berkeley Function Calling Leaderboard (BFCL) and its three versions, detailing how each iteration improves evaluation of LLM function calling capabilities. BFCL v1 (March) introduced abstract syntax tree evaluation across 2,000 function-answer pairs, covering simple, multiple, and parallel calling. v2 (August) added a live dataset of 2,251 pairs, reducing bias and contamination through Rouge-score deduplication and filtering. v3 (September) introduced multi-turn and multi-step evaluation with state-based metrics, using a custom API (vehicle control, stock trading, etc.) and human-validated ground truth. Sam also highlights common model failures—misunderstanding current state, unnecessary planning—and mentions the Gorilla CLI and Open Functions as related tools in the Berkeley ecosystem.

Building AGI in Real Time (OpenAI Dev Day 2024)

Oct 4, 2024 · 2:09:35

OpenAI's DevDay 2024 unveils the Realtime API with WebSocket-based voice and function calling, Vision Finetuning, Prompt Caching, and Model Distillation. Product leads Olivier Godement, Romain Huet, and Michelle Pokrass detail design choices like zero-code caching and WebSockets for human-level latency. Alistair Pullen of Cosine explains how fine-tuning GPT-4o on custom reasoning traces beat o1 on SWE-bench Verified, while Sam Altman and Kevin Weil discuss AGI levels (o1 as level 2 reasoner), the coming agent era (level 3), safety through iterative deployment, and the importance of building for the frontier. The episode also covers moderation policy relaxation, plans for o1 system prompts by year-end, and the competitive admiration for NotebookLM.

[Paper Club] Who Validates the Validators? Aligning LLM-Judges with Humans (w/ Eugene Yan)

Sep 28, 2024 · 1:00:55

Eugene Yan presents the paper 'Who Validates the Validators? Aligning LLM-Judges with Humans' by Shreya Shankar, introducing EvalGen, a framework that helps developers iteratively align LLM evaluators with human labels by grading outputs, refining criteria, and tracking coverage (recall) and false failure rate (1−precision). Yan demonstrates his own prototype 'Label', which forces users to label 20 samples before unlocking evaluation mode and 50 samples plus one evaluation run before optimization, gamifying the alignment loop. The discussion covers binary vs. pairwise evaluation, the subjectivity of criteria, and Shreya Shankar's plans to integrate natural language feedback into EvalGen's next version. Yan advocates that labeling data is essential before deploying LLM evaluators in production, turning the process into prompt fine-tuning rather than vibe checks.

Language Agents: From Reasoning to Acting — with Shunyu Yao of OpenAI, Harrison Chase of LangGraph

Sep 27, 2024 · 1:26:33

Shunyu Yao of OpenAI and Harrison Chase of LangChain discuss the evolution of language agents, arguing that combining reasoning and acting through techniques like ReAct remains foundational, while tool design and cognitive architectures like CoALA will shape the next wave. Shunyu traces his work from text games to ReAct, showing how thinking as an action improves reliability; Harrison explains LangChain's adoption and the shift to LangGraph for stateful orchestration. They emphasize agent-computer interfaces (ACI) as a neglected area, noting that SWE-Agent's success came from optimizing tools for models, not humans. The conversation covers memory types (semantic, episodic, procedural) and applies CoALA's three dimensions to current systems. They highlight promising applications: customer support, coding agents, and spreadsheet-style UX for batch operations. Shunyu stresses that better benchmarks like SWE-Bench (which solved 30% of GitHub issues) and TauBench (best model 48%) are critical, and that data—not architecture—will drive future model improvements.

llm.c's Origin and the Future of LLM Compilers - Andrej Karpathy at CUDA MODE

Sep 21, 2024 · 23:29

Andrej Karpathy explains llm.c, his from-scratch C implementation of GPT-2 training that outperforms PyTorch by 20% in speed and 30% in memory, arguing that LLMs will eventually act as compilers for custom code. He recounts starting the project out of frustration with PyTorch compile errors, writing all forward and backward passes manually in C, then porting to CUDA with kernel optimizations. A community of contributors (Eric, Arun, Alex R) helped achieve 50% MFU on a single H100 node, training GPT-2 1.6B in 24 hours for $600. Ongoing work adds Llama 3.1 and FP8 support. Karpathy sees llm.c as proof that LLMs can automate such low-level optimization, making Python and PyTorch a temporary crutch.

The Ultimate Guide to Prompting - with Sander Schulhoff from LearnPrompting.org

Sep 20, 2024 · 1:07:09

Sander Schulhoff, author of The Prompt Report and founder of LearnPrompting, presents a comprehensive taxonomy of 58 prompting techniques and argues that prompt engineering is a skill everyone should have, not a specialized role. He categorizes techniques by problem-solving strategy (zero-shot, few-shot, thought generation, decomposition, ensembling, self-criticism) and debunks role prompting for accuracy-based tasks, showing an 'idiot' prompt outperformed a 'genius' prompt on MMLU. Schulhoff shares his experience running Hack-a-Prompt, which collected 600,000 malicious prompts and won best paper at EMNLP, and differentiates between prompt injection and jailbreaking. He advocates for automatic prompt engineering with DSPy, which beat his manual effort in 10 minutes, and previews Hack-a-Prompt 2.0, aiming for a $500,000 prize pool to generate real-world harms for safety tuning.

[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."

Building AGI with OpenAI's Structured Outputs API

Sep 17, 2024 · 1:12:23

Michelle, OpenAI's API tech lead, explains how the newly launched Structured Outputs feature achieves near-100% reliability for schema-following outputs by combining constrained decoding with model training, moving beyond the limitations of JSON mode and Function Calling. She details the refusal field that allows the model to decline harmful requests while maintaining parseable output, and discusses use cases like dynamic UI generation and mathematical chain-of-thought. The roadmap includes parallel function calling and custom grammars. The episode also covers the GA of GPT-4o fine-tuning, Batch API's 50% cost savings, Vision API integration with structured outputs, the Whisper API's surprising lack of diarization, and the shift toward speech-to-speech APIs. Michelle shares insights from her Waterloo co-op background and what qualities succeed at OpenAI: low ego, user-focused, and driven.

Personal benchmarks vs HumanEval - with Nicholas Carlini of DeepMind

Aug 28, 2024 · 1:07:28

Nicholas Carlini, a DeepMind research scientist, argues AI models are practically useful for personal tasks despite flaws, and personal benchmarks tailored to individual use cases matter more than generic leaderboards. He shares from his 'How I Use AI' post: using LLMs for ephemeral software, kickstarting Docker, debugging by pasting error messages. He built a personal benchmark DSL that runs code from real chat history. On security, he explains buying expired domains from LAION-400M allowed poisoning any model trained on it, and he extracted OpenAI's Ada and Babbage model dimensions via API (with permission). He also recovered training data from GPT-3.5 by repeating a word until ChatGPT output verbatim sequences. He prefers attacking over defending because it's more fun and essential for discovering real vulnerabilities.

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.

The Winds of AI Winter (Q2 Four Wars of the AI Stack Recap)

Aug 2, 2024 · 1:23:36

Swyx and Alessio recap Q2 2024 through their 'Four Wars' framework, arguing that the AI landscape is shifting from frontier model dominance to commoditization and vertical applications. They highlight Claude 3.5 Sonnet overtaking OpenAI on coding benchmarks, Llama 3.1's synthetic data approach enabling 7B models to rival GPT-4, Mistral Large 2's non-commercial license and lost open-source crown, and on-device models like Gemini Nano and Apple Intelligence. The Quality Data Wars see NYT suing OpenAI, Reddit licensing data for $200M+, and synthetic data proving real for math (AlphaProof near IMO gold) and code. The Multimodality War includes ChatGPT Voice Mode delayed, Meta's Chameleon for native fusion, and Google's PaliGemma for PDF extraction. The renamed LLM OS War covers agent protocols, memory databases, and the collapse of model cost by an order of magnitude every four months, pushing startups toward vertical services like Brightwave and Dropzone that sell labor, not tools. The episode ends with a CrowdStrike joke about agent safety.

[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.