Issue #62·Friday, March 20, 2026·28 min read·14 stories

OpenAI Acquires Python's uv and ruff, Bezos Seeks $100B

Uber bets $1.25 billion on Rivian robotaxis, Karpathy's autoresearch scales 9x with a GPU cluster

OpenAI acquired Astral, the company behind uv and ruff, pulling critical Python infrastructure into its Codex orbit. Jeff Bezos is separately raising $100 billion for an AI manufacturing fund targeting chipmaking and defence. Plus: Nvidia's $20 billion Groq deal, Karpathy's autoresearch scaling 9x, and why CNBC thinks Jensen Huang needs a moat, not another chip.

▲NEWS

4 stories

Jeff Bezos in Talks to Raise $100 Billion for AI Manufacturing Fund

Jeff Bezos is in early talks to raise $100 billion for a fund that would acquire manufacturing companies and apply AI to accelerate automation. He has met with sovereign wealth funds in the Middle East and Singapore. The fund targets chipmaking, defence, and aerospace, rivalling SoftBank's Vision Fund in scale. Bezos plans to deploy technology from Project Prometheus, his AI startup that simulates the physical world.

Read full story→

Nvidia's $20B Groq Deal Adds SRAM Inference to Vera Rubin

Nvidia's Groq 3 LPU, the first chip from its $20 billion Groq deal, will act as a co-processor for the Vera Rubin platform's decode phase. Built on Samsung's 4nm process, this SRAM-based accelerator offers higher memory bandwidth than GPUs for inference on trillion-parameter models. The specialised hardware replaces earlier CPX accelerators, signalling Nvidia's push into custom inference silicon.

Uber Orders 50,000 Rivian Robotaxis for $1.25 Billion

Uber is ordering 50,000 Rivian R2 robotaxis in a deal worth $1.25 billion. The first 10,000 vehicles deploy in San Francisco and Miami by 2028, with 40,000 more optional by 2030. It's the largest single robotaxi fleet commitment to date, covering 25 cities by 2031. The deal also gives Rivian a much-needed revenue anchor.

OpenAI Acquires Astral, the Company Behind Python's uv and ruff

OpenAI is acquiring Astral, the company behind uv, ruff, and ty. uv alone saw 126 million downloads last month and has become the default Python environment manager for many developers. Simon Willison thinks the deal is about both Astral's Rust engineering talent and potential uv integration with Codex. The question is whether OpenAI will maintain its open source commitments long-term.

⚙TECHNICAL

3 stories

TorchSpec: Speculative Decoding Training at Scale

TorchSpec is a new PyTorch framework that separates LLM inference from draft model training, streaming hidden states over RDMA instead of writing them to disk. The team trained a Kimi K2.5 EAGLE-3 draft model with 1,500 H200 GPU hours across 6 billion tokens. Output throughput improved over 60% at batch size 1 and 30% at batch size 8.

Karpathy's Autoresearch Hits 9x Speedup With a 16-GPU Cluster

Karpathy's autoresearch project got access to 16 GPUs via SkyPilot and ran 910 experiments in 8 hours, a 9x speedup over the single-GPU baseline. The parallel setup changed how the agent searched. Instead of greedy hill-climbing, it ran factorial grids of 10-13 experiments per wave, catching interaction effects that sequential search misses. It also taught itself to route validation to H200s while screening ideas on cheaper H100s.

How Squad Runs Coordinated AI Agents Inside Your Repository

Squad drops a preconfigured team of AI agents into your GitHub repository with two commands. A coordinator routes tasks to specialists (backend dev, tester, docs writer) who share context through committed files like decisions.md and history. The key design choice: when a tester rejects code, a different agent must fix it, forcing genuine independent review instead of self-correction.

◈ANALYSIS

4 stories

Why Perfect AI Retrieval Still Produces Wrong Answers

A PM presented a flawless traffic analysis. One question revealed the baseline included weeks of bot scraping, making every conclusion wrong. Leah Tharin argues the AI industry obsesses over retrieval accuracy while ignoring factual correctness. Her fix: a verified context layer where leaders document known truths, acting as a CI pipeline for organisational knowledge.

Jensen Huang Doesn't Need a New Chip. He Needs a New Moat.

Nvidia's real strategic move at GTC wasn't new chips. It was NemoClaw, an open-source, chip-agnostic platform for building AI agents. Nvidia dominated the training era through hardware lock-in, but inference doesn't require the same moat. By giving away the agent platform and monetising the compute underneath it, Huang is building a stickier business than chip sales alone.

What 81,000 People Want From AI

Anthropic surveyed 81,000 people globally about what they want from AI. The top desires are professional efficiency, personal transformation, and time freedom. People value AI for learning and cognitive partnership but worry about unreliability, job displacement, and cognitive atrophy. Lower-income countries are more optimistic and focused on opportunity, while wealthier regions prioritise managing life's complexity.

Generic Enterprise AI Is Indefensible

Frontier AI argues that general-purpose enterprise AI agents are a trap for startups. Foundation model providers are already shipping the same features directly into their platforms, with built-in advantages in data, compute, and distribution. The defensible opportunity is in vertical AI agents that learn from domain-specific data and create a flywheel competitors can't replicate.

⚒TOOLS

3 stories

Build Knowledge Agents Without Embeddings

A new approach to knowledge agents that relies on the filesystem instead of vector databases. Rather than embedding pipelines, it leans on LLMs' existing understanding of file operations like grep and bash for more deterministic, explainable retrieval. The open-source Knowledge Agent Template supports GitHub, Discord, and Slack integration. For many use cases, embeddings add complexity without proportional value.

AI SRE Agent Discovers Alerts, Diagnoses Root Cause, Suggests Fixes

A step-by-step guide to assembling a multi-agent SRE system with the AWS Strands Agents SDK. The system discovers active CloudWatch alarms, performs root cause analysis with Claude Sonnet 4 on Bedrock, and suggests Kubernetes or Helm remediations. It includes a dry-run mode for safe evaluation before live fixes. Ships with 12 unit tests and full source on GitHub.

Screenpipe Turns Your Screen Into a Searchable AI Memory

An open-source Rust tool that continuously records your screen and audio, then makes it all searchable and automatable through AI. Everything runs locally on your machine with no cloud dependency. It has 17,000+ GitHub stars and growing. Useful for developers who want full-context AI assistance without sending screen data to third parties.