AMD buys Fei-Fei Li's World Labs for $8.2B, all-stock

Good morning 👋 AMD just paid $8.2B in stock for Fei-Fei Li's World Labs. Most of that price tag is buying her name for the org chart, not the 3D world models.

In today's issue:

• 🔭 AMD buys Fei-Fei Li's World Labs for $8.2B, hires a chief scientist along with it
• 🔬 What happens when an agent's tool call fails and it reports success anyway
• 🚀 SiMa.ai's physical-AI chips hit $1.45B
• 🛠️ Retire your brittle UI test scripts for an agent that actually looks at the page
• 🔗 Anthropic's IPO risk warning, Samsung's $1B infra bet

🔭 THE ONE THING

🧠 AMD buys Fei-Fei Li's World Labs for $8.2B, all-stock

AMD is paying $8.2 billion in stock for World Labs, the San Francisco startup where Fei-Fei Li builds spatial-intelligence and 3D world models. It's AMD's second-biggest acquisition ever, behind only the $49B Xilinx deal in 2022, and Li comes with it: EVP and Chief Scientist, reporting to Lisa Su, with the deal expected to close by year-end pending regulatory approval. That kind of money buys AMD a marquee AI research chief as much as it buys the tech. Whether world models actually show up in AMD's product roadmap, or this was mainly a hire, is the thing to watch.

🔬 RESEARCH HIGHLIGHTS

• TokenCast predicts how many tokens an agent run will cost before you press go, not after. Same task, same agent, and real runs can burn over an order of magnitude more tokens depending on how tool calls and context balance out along the way. Useful if you're the one setting budget caps on agent workloads instead of eating the invoice. Paper

• Failure-Transparent Agents names the failure mode nobody benchmarks directly: a tool call fails, and the agent reports success anyway because nothing forces it to show its evidence. Most existing benchmarks bury this inside tool-selection and recovery scores, so it never shows up as its own number. Isolate it and you get an honest read on how much you can trust an agent's self-report. Paper

• Self-retrospection, no RL gets agents better at a task by having them write explanations of their own past runs, then fine-tuning on those explanations. No reward model, no RL loop. Cheap trick if it survives contact with tasks outside the paper's test set. Paper

🚀 AI STARTUPS

• SiMa.ai closed a $150M Series C at a $1.45B valuation, pushing total funding to $500M for chips that run AI directly on robots, cars, and drones instead of the cloud. Fidelity and Amplify led the round. Press release

🛠️ TRY THIS

Swap your brittle UI test scripts for an agent that actually looks at the page

World Labs sold AMD $8.2B on the idea that an agent needs a working model of the space it operates in, not just the next-token guess. You don't need spatial AI to test that logic on your own stack this week. Amazon's walkthrough for Nova Act replaces Selenium-style synthetic monitors, the kind that break every time a button moves three pixels, with an agent that reads the page and decides what "logged in" actually looks like.

1. Pick one user journey that keeps breaking your existing synthetic checks (login, checkout, search) and write it as plain-language steps, not selectors.

2. Stand up Nova Act against a staging environment through Bedrock AgentCore, using the sample repo in the post as your scaffold rather than starting from a blank script.

3. Schedule it to run the journey on a cadence and log a pass/fail plus a short natural-language reason, not just an HTTP status code.

4. Run it in parallel with your current monitor for a week before you retire the old one. Agents fail differently than scripts do, and you want to see how before you're relying on it alone.

Prompt: Log into staging as a returning customer, add the cheapest item in "New Arrivals" to the cart, and complete checkout with the saved card. Report what happened at each step in plain language, and flag anything that looks broken even if no error was thrown.

Worth a look

• SkyRL on SageMaker HyperPod AWS's walkthrough post-trains a Qwen3-VL-8B model with GRPO on a Ray cluster, open-source end to end. If you're RL-tuning a vision-language model and dreading the infra glue, this is the reference build.

🔗 QUICK LINKS

• Anthropic Its own IPO filing warns the tech it's selling could pose "existential risks to humanity," alongside a $42B net loss and a $2T valuation ask. Reuters via CNBC
• Samsung Putting $1B into Helix, the Nvidia- and KKR-backed data center venture run by former AWS chief Adam Selipsky. Samsung Newsroom

Watch what Lisa Su does with a research chief instead of a roadmap slide.

Pradeep Perugu