In partnership with
In Today’s Issue:
🏗️ TSMC CEO warns that advanced AI chip demand is outstripping capacity by 300%
🚀 Massive 750MW partnership with Cerebras will integrate wafer-scale chips into OpenAI’s stack
🔄 Barret Zoph and Luke Metz depart Mira Murati’s Thinking Machines Lab
💻 Cursor deployed hundreds of hierarchical agents
🚨 Developer George Hotz slams Anthropic for restricting API-level access
✨ And more AI goodness…
Dear Readers,
What happens when you let hundreds of AI agents loose on a single codebase for an entire week? Cursor just found out—and the answer is a fully functional web browser with over a million lines of code. Today's issue dives deep into this breakthrough in autonomous coding, where hierarchy beat chaos and simpler systems surprisingly outperformed complex ones. But that's just the opener: we're also tracking the global chip squeeze that's reshaping the AI race (spoiler: TSMC can't build fast enough), OpenAI's bold bet on ultra-fast inference with Cerebras, and the talent wars heating up as key researchers boomerang back to Sam Altman's team.
Plus, a remarkable story from the NHS where a man just received 100 million reprogrammed cancer-killing cells in three teaspoons of liquid—a "living drug" that could redefine what's possible in leukaemia treatment. Grab your coffee and let's get into it.
All the best,
🚀 AI Chip Demand Outpaces Supply
The global AI boom is pushing Taiwan Semiconductor Manufacturing Co. to its limits, with demand for advanced chips running 3× higher than capacity, according to CEO CC Wei. Heavyweights like Nvidia, Google, Broadcom, Apple, and OpenAI are all scrambling for more chips, but new factories in Arizona and Japan won’t ease shortages until 2027 or later. With packaging capacity also maxed out and cautious investment after past boom-bust cycles, customers like Tesla are already turning to alternatives such as Samsung - signaling that chip supply constraints could shape the AI race for years.
🚀 OpenAI Boosts Real-Time AI
OpenAI is partnering with Cerebras to add 750MW of ultra low-latency AI compute, dramatically speeding up real-time AI responses for users. By integrating Cerebras’ giant-chip architecture into its inference stack, OpenAI aims to enable faster code generation, image creation, and AI agents—unlocking higher-value, more interactive workloads. The rollout will happen in phases, with full capacity coming online through 2028, setting the stage for scalable real-time AI at global scale.
🔄 OpenAI Reclaims AI Talent
Two cofounders of Thinking Machines Lab—Barret Zoph and Luke Metz—are leaving the startup to rejoin OpenAI, following a sudden and controversial split. The move is a setback for Thinking Machines, led by Mira Murati, and a strategic win for OpenAI as it regains senior researchers amid an intense AI talent race. The shake-up also highlights investor pressure and governance tensions inside high-valuation AI startups.
The insane engineering feat of the world's most important machine
George Hotz argues Anthropic is alienating power users by repeatedly blocking “opencode” access through the Claude Code API, framing it as a sign the company will increasingly enforce restrictive usage (fixed prompts, tighter controls) instead of supporting open workflows.
Cursor Agents Code for a Week
The Takeaway
👉 Cursor successfully ran hundreds of AI agents for weeks on single projects, proving large-scale autonomous coding is viable with proper coordination architecture
👉 Flat agent coordination fails—hierarchical planner/worker structures eliminate bottlenecks and keep agents focused on hard problems
👉 Model selection is role-dependent: GPT-5.2 excels at long-running tasks while other models perform better for specific functions like planning
👉 Prompts drive system behavior more than infrastructure—extensive experimentation in prompt engineering was key to avoiding drift and maintaining focus
Cursor just ran hundreds of AI agents on a single codebase - for an entire week. The result? A web browser built from scratch with over 1 million lines of code across 1,000 files.
But getting there wasn't straightforward. Early attempts at flat coordination failed spectacularly. Agents held locks too long, avoided hard tasks, and churned without progress. The breakthrough came from hierarchy: planners continuously explore and create tasks, while workers focus purely on execution. This separation eliminated most coordination headaches.
Interestingly, model choice proved crucial. GPT-5.2 outperformed Opus 4.5 for extended autonomous work - staying focused and avoiding shortcuts. Different models now handle different roles based on their strengths.
Perhaps the biggest surprise? Simpler systems won. Removing complexity often helped more than adding it. And prompts mattered more than anything else for keeping agents on track.
The team also migrated Cursor's codebase from Solid to React in three weeks and made video rendering 25x faster with autonomous Rust rewrites. Real production code, shipping soon.
Could this approach eventually let AI tackle month-long engineering projects? The early signs are remarkably promising.
Why it matters: This research shows that scaling AI coding isn't just about better models, it's about coordination architecture. Hundreds of agents collaborating for weeks brings us closer to AI handling entire software projects autonomously.
Sources:
🔗 https://cursor.com/blog/scaling-agents
🔗 https://github.com/wilsonzlin/fastrender
Learn AI in 5 minutes a day
This is the easiest way for a busy person wanting to learn AI in as little time as possible:
Sign up for The Rundown AI newsletter
They send you 5-minute email updates on the latest AI news and how to use it
You learn how to become 2x more productive by leveraging AI
First leukaemia patient to get new form of treatment
Oscar Murphy just received 100 million cancer-killing cells in three teaspoons of liquid. Welcome to the future of leukaemia treatment.
The 28-year-old from Bury became the first adult patient with B-cell acute lymphoblastic leukaemia to receive CAR-T therapy on the NHS. His own immune cells were extracted, genetically reprogrammed in a lab, and infused back to hunt down his aggressive cancer.
The science is elegant: T-cells get modified with a harmless virus to recognize cancer cells like a lock and key. They multiply into millions, then go to work inside the patient's body—permanently. It's called a "living drug" because these cells keep growing and fighting long after treatment ends.
Clinical trials showed 77% of patients went into remission, with half cancer-free after three and a half years. For a disease where patients typically survive just six to eight months, that's transformative.
The treatment costs £372,000 per infusion but was developed from UK research at University College London. Around 50 NHS patients yearly could benefit initially, though doctors expect broader use soon.
Oscar married his fiancée Lauren in hospital last month. Now he's dreaming of children and a normal life.
The best HR advice comes from those in the trenches. That’s what this is: real-world HR insights delivered in a newsletter from Hebba Youssef, a Chief People Officer who’s been there. Practical, real strategies with a dash of humor. Because HR shouldn’t be thankless—and you shouldn’t be alone in it.
