Vibeleaderboard
← All Intel
Intel / video

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

Source
youtube.com
Author
Sequoia Capital
Date

Context

Jerry Tworek, co-founder of Core Automation, argues that the main obstacle to better AI is that models are trained in a lab but used in a messier real world, so they need to keep learning while in use. In this interview hosted by Sequoia Capital, the introduces him and co-founder Rohan Anil as veterans of OpenAI reasoning work and Gemini .

Tworek says benchmarks resemble the tasks models are trained on, while real work is murkier. He sees two ways to learn at use time and finds both lacking. Learning inside the prompt is data-efficient but small: after about 20 minutes of coding-agent use he has to compact the . Continual risks catastrophic forgetting, where new training erases old skills, and is data-inefficient. He therefore wants a learning algorithm expressed in the architecture itself.

Anil separately argues that transformers spend computation inefficiently. These are the founders' opinions and research direction, and the interview presents no working replacement architecture.

Terms in this piece · Glossary
  • Host — The computer on which a process runs.
  • pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
  • Session — One conversation or piece of work as a harness keeps it: a record of messages, tool calls and observations that can be saved and resumed.
  • fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
Read the source www.youtube.com
More from Sequoia Capital
Recommended reads
Comments

Checking sign-in…

Loading comments…