H100

Next 100x in AI: Inference, Networking, & Self-Optimizing Models — Philip Kiely & Ali Taha, Baseten

Aug 3, 2026 · 1:42:54

Philip Kiely and Ali Taha of Baseten join Swyx to explain what actually happens when a 200,000-token request hits a production inference system, arguing that stacking quantization, speculative decoding, and disaggregated prefill/decode can make open models like GLM-5.2 up to 10x faster. They detail cache-aware routing, training traffic-specific speculators, and why quantization errors can cancel out so a more-quantized model beats a less-quantized one. They also reveal how Baseten grafted Kimi's vision encoder onto GLM-5.2, discuss NVIDIA Dynamo as a toolkit rather than a turnkey speedup, and explain why they are bearish on mega kernels. The conversation covers video generation's compute barriers, the trend toward ASIC-like GPUs, and the emerging loop where GLM-5.2 writes the GPU kernels that serve itself.

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.

[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.

How Zyphra went all-in on AMD + Why Devs feel faster with AI but are slower — with Quentin Anthony

Nov 3, 2025 · 1:00:02

Quentin Anthony, head of model training at Zyphra and advisor at EleutherAI, explains why his company moved all training to AMD MI300X GPUs, outperforming Nvidia H100s on certain workloads thanks to 192GB VRAM and higher memory bandwidth, and describes his kernel development approach of writing directly in ROCm or GPU assembly rather than using Triton. He also shares his experience in the METR study on AI coding productivity, where he was one of the few developers with measurable speedup, and offers tips: timebox AI use, avoid the slot machine effect, maintain context hygiene, and use direct API over tools like Cursor. Additionally, he argues that open source AI research benefits from siloed focused teams with guaranteed funding over grand collaborations, and notes that kernel datasets alone won't solve GPU programming due to evaluation challenges.

⚡️Accelerators @ 3x NVIDIA H200 perf, Made in the USA - Thomas Sohmers + Mitesh Agrawal, Positron AI

Aug 18, 2025 · 47:56

Thomas Sohmers and Mitesh Agrawal of Positron AI argue memory bandwidth, not compute, is the true bottleneck in AI inference, and their accelerator achieves 93% memory bandwidth utilization—triple NVIDIA's efficiency—enabling 70% faster token generation at 150W. The founders, both Lambda Labs veterans, shipped an FPGA product 15 months after founding, then raised a $51M Series A for an ASIC in late 2026. Their hardware requires no recompilation: it ingests raw binary weights from NVIDIA training and outputs an OpenAI-compatible API. Positron focuses on the decode phase of transformers, where memory-bound matrix-vector multiplication dominates, and already counts Cloudflare and Parasail as customers. The company sells systems directly, prioritizing capital efficiency and ROIC over operating its own cloud.

The Shape of Compute (Chris Lattner of Modular)

Jun 13, 2025 · 1:18:18

Chris Lattner of Modular explains how his company is breaking the CUDA monopoly with Mojo and MAX, matching NVIDIA's best inference performance on AMD and NVIDIA GPUs. After three years in R&D, Modular open-sourced its stack and now delivers state-of-the-art Llama 3 serving at over 800 tokens per second. Mojo, a Python-family language, runs kernels faster than Rust and can extend Python without bindings, while MAX provides a full inference framework with automatic kernel fusion and cluster management. Unlike VLLM or SGLang, Modular's container is just a gigabyte, fully open source, and not reliant on proprietary CUDA blobs. DeepSeek's low-level PTX work validated the approach, but Lattner emphasizes that Mojo's portability avoids rewriting kernels for each new GPU architecture. He also details his daily routine, using Cursor for coding, and hiring "elite nerds" to grow the team.

SF Compute: Commoditizing Compute

Apr 11, 2025 · 1:12:02

Evan Conrad, co-founder of SF Compute, argues that GPUs behave like a real estate business, not a traditional cloud, because price-sensitive customers value every incremental GPU and will switch for a 10% margin. CoreWeave succeeded by selling locked-in long-term contracts to low-credit-risk customers like Microsoft and OpenAI, ignoring short-term demand. He predicts hyperscalers and providers like Together and DigitalOcean will lose money on GPU clusters because software margins cannot match the hardware costs. SF Compute started as an AI lab forced to sublease its cluster monthly to avoid bankruptcy, then evolved into a market where anyone can buy H100s by the hour via dynamic pricing—often below $1/hour for short bursts. Utilization stays near 100% as prices adjust. Future plans include cash-settled futures to reduce financial risk across the industry, while the brand deliberately stays anti-hype and calm.

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.

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.

[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.

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.

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.

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.

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 State of Silicon and the GPU Poors - with Dylan Patel of SemiAnalysis

Dec 5, 2023 · 1:07:55

Dylan Patel of SemiAnalysis argues the 'GPU poor vs GPU rich' divide defines AI, with Google's TPU ramp (millions of units) dwarfing others and Nvidia selling over a million H100s this year. He explains that training costs are irrelevant compared to inference costs, and that memory bandwidth is the key bottleneck for LLM inference (e.g., Llama 70B needs ~2.1 TB/s for human reading speed). Patel advises GPU-poor players to focus on on-device innovations like speculative decoding (Medusa), asynchronous training, or fine-tuning for edge use—not on gaming benchmarks or fine-tuning small models. He details why AI hardware startups (Cerebras, Graphcore) bet wrong on on-chip SRAM, while Google’s TPU/Broadcom partnership and Nvidia’s yearly cadence make alternatives tough. He also warns that rebuilding the semiconductor supply chain in the US is unfeasible due to its fragmentation, and that safety through obscurity doesn't work—open innovation is better.