▶ 10:50 · $OPENAI · Building the Automated AGI Lab: Core Automation's Jerry Tworek and Rohan Anil

episode briefing

Building the Automated AGI Lab: Core Automation's Jerry Tworek and Rohan Anil

2026-07-29 · 5 company · 9 thematic

sentiment

4 bull0 bear1 neu

speakers

Led the Strawberry and reasoning teams at OpenAI; reinforcement learning maximalist who believed scaling RL was the path to AGI; now building an automated lab to discover post-transformer architectures.

rohan anil

One of four pre-training leads for Google's Gemini; worked on fundamental AI research at Google Brain and Anthropic; developed the Shampoo optimizer and N-gram memory architectures; focuses on training algorithm optimization and hardware-software co-design.

quote

“The the the majestic thing about Transformer, which goes back to like why why why why do we have to appreciate Transformers so deeply, is that Transformers are economically valuable. The training them the cost of training them is lower tha…”

—jerry tworek

episode shorts · 2

The Problem With Testing AI Architectures at Small Scale | Jerr…

The Most Automated AI Lab Isn't Removing Humans | Jerry Tworek,…

now playing · $OPENAI

$OPENAIbullish· highjerry tworek

Tworek: Transformers economically viable as training cost below revenue

Transformer architecture generates more revenue than training cost, making scaling economically sustainable unlike LSTMs which would not have achieved positive unit economics.

$GOOGL···neutral· mediumrohan anil

Anil: Google's scale mindset drives inference efficiency focus

Google's requirement to serve billions of users creates unique pressure for latency-constrained, token-efficient architectures that most labs ignore.

$CORE-AUTOMATIONbullish· high· posrohan anil

Core Automation founders bet on test-time learning architectures to replace transformers

Transformers fundamentally cannot learn continuously at test time; new architectures enabling continual learning on real-world distributions are needed, and Core Automation is bui…

$DEEPSEEKbullish· mediumrohan anil

Anil: DeepSeek validated N-gram memory replacing MoE

DeepSeek's N-gram embeddings demonstrated improved memory scaling without mixture-of-experts, validating Core Automation's architectural direction.

$CORE-AUTOMATIONbullish· high· posjerry tworek

Founders building automated lab to find transformer replacement

Core Automation aims to build the most automated research lab to discover post-transformer architectures that enable continual test-time learning, using kernel automation as the i…

←