Blackwell

Building Cursor Composer – Lee Robinson, Cursor

Dec 2, 2025 · 15:36

Lee Robinson explains how Cursor built Composer, its first agent model for real-world software engineering, by focusing on being both fast and smart — achieving 4x more efficient token generation than similarly intelligent models while matching open-source performance initially and approaching frontier models after reinforcement learning. The model's training posed infrastructure challenges: matching training and inference environments across thousands of GPUs, handling complex rollouts with hundreds of tool calls and up to millions of tokens, and ensuring consistency by using the same tool format and responses as production. Cursor solved these with custom kernels that sped up training by 3.5x on NVIDIA Blackwell chips for mixture-of-experts layers, load balancing across threads to avoid idle time, and co-designing RL infrastructure with its Cloud Agents product using virtual machines that mirror the production Cursor environment. This allowed the model to become a power user of tools like semantic search, which improved all models but especially Composer. RL also taught the model to parallelize tool calls (e.g., reading 10 files simultaneously) and to search more before editing,…

Structuring a modern AI team — Denys Linkov, Wisedocs

Jul 24, 2025 · 17:40

Denys Linkov, who leads ML at Wisedocs, argues that building a modern AI team hinges on identifying your company's bottleneck—shipping features, acquiring users, or scalability—rather than reflexively hiring AI researchers. He introduces Ampere's Wager: trading your entire domain-savvy team for five top-lab researchers is usually a losing bet. For early-stage AI strategy, generalists who blend model training, serving, and business acumen outperform specialists; Linkov lived this in 2021 building a custom MLOps platform for a conversational AI startup and again in 2024 using advanced open-source tools for medical record processing. He stresses reskilling existing teams through weekly learning cadences and moving domain experts from giving feedback to writing evaluations. Hiring should hold context and act on it, verifying trends like 'don't hire juniors' against YC's AI school drawing 2,000 young people.

[Full Workshop] Reinforcement Learning, Kernels, Reasoning, Quantization & Agents — Daniel Han

Jul 19, 2025 · 2:42:28

Daniel Han of Unsloth presents a technical workshop covering reinforcement learning (RL), kernels, reasoning, quantization, and agents, arguing that RL with verifiable rewards (RLVR) is the key to unlocking LLM capabilities beyond supervised fine-tuning. He explains why open-source models plateaued after September 2024 until DeepSeek-R1 showed that RL can elicit reasoning, and breaks down PPO, GRPO, and the REINFORCE algorithm, emphasizing that GRPO removes the value model for efficiency. Han details how reward functions—not algorithms—are the hardest part, with examples like distance-based scoring for math. He demonstrates a free Colab notebook training a base model to reason, and shows that dynamic quantization can shrink models like DeepSeek-R1 from 730 GB to 140 GB with only ~1% accuracy loss, arguing that GPUs may stop getting faster after FP4 precision.

The Geopolitics of AI Infrastructure - Dylan Patel, SemiAnalysis

Jun 19, 2025 · 18:29

Dylan Patel of SemiAnalysis argues that despite US sanctions, Huawei has engineered a 384-chip cluster (Cloud Matrix 384) that Nvidia failed to deploy, while accessing TSMC via Softgo and HBM from Samsung via shell companies — all legally. China's SMIC will soon produce 7nm AI chips in high volumes, debunking the notion that China lacks compute. Meanwhile, Middle East players like G42 (UAE) and Datavolt (Saudi Arabia) are building multi-gigawatt data centers, with G42's deal letting it keep 20% of 500,000 GPUs yearly for itself while 80% goes to US companies like OpenAI. Patel highlights the US's 63-gigawatt power shortfall vs. 100 GW of planned data centers, explaining why US companies rely on Middle East capacity and why China's superior power buildout gives it a geopolitical edge.

Mastering LLM Inference Optimization From Theory to Cost Effective Deployment: Mark Moyou

Jan 1, 2025 · 33:39

NVIDIA solutions architect Mark Moyou explains that LLM inference differs fundamentally from standard deep learning deployment, requiring careful management of KV Cache, attention mechanisms, and GPU memory to control cost. He details how tokens are processed: prefill computes attention across the entire prompt, then generation produces one token at a time, with KV Cache storing key-value pairs to avoid recomputation. Llama's 32 attention heads and FP8 quantization (halving memory with near-identical accuracy) are cited as key optimizations. Moyou emphasizes measuring time to first token, inter-token latency, and input/output sequence length distributions to size inference engines. He presents NVIDIA's TRT-LLM (model compilation for LLMs) and Triton inference server as tools to maximize throughput, and discusses how query patterns like long-input-short-output or short-input-long-output impact GPU utilization and deployment cost.