All right, so what are you working on?

@sorhan.hq

3.8K followers · usually 553 views a post

Views

12.6K

Likes

456

Comments

23

Shares

24

SoundTikTok sound · original sound - sorhan.hq

Promoted product

ELUelu.dev

Elu watches real user sessions, catches the bugs they run into, and has the fix ready to ship.

6 videos

Why it worked

The video uses a high-authority setting (Meta San Francisco) and a technical, problem-first approach to establish credibility before introducing the concept behind the creator's product.

Summary

A researcher explains the concept of continual learning in AI agents and the challenge of catastrophic forgetting. The video serves as a technical pitch for the creator's product, ELU, which addresses these challenges.

Structure

1Interviewer asks about the project
2Researcher explains continual learning
3Researcher identifies the problem of catastrophic forgetting
4Researcher poses the challenge of agent adaptation

Original caption

AI agents are entering a new phase: continual learning. Instead of being trained once and deployed, they improve on the fly using reinforcement learning. By interacting with environments-often simulated-they learn from feedback, adapt to new situations, and refine their behavior over time. This shift moves AI from static models to systems that continuously evolve through experience. #startups #founder #machinelearning #Tech #hackathon

More from @sorhan.hq

Product placement

The video discusses the core technical problem that the product ELU solves, positioning the creator as an expert in the field.

On-screen text

META SAN FRANCISCO

Transcript

All right, so what are you working on? So, we're a group of ML researchers here and we have some new ideas about how we can use reinforcement learning to allow for what's called continual learning. What continual learning means is an agent who is able to adapt what it's doing to the history that it has seen without having to put all of that history in context. In other words, instead of having to read everything that has happened before in order to know what the user's preferences are or what new information it should know, the agent itself can be adapted and learn. The problem is that a lot of the techniques that are normally used to do this incur a cost called catastrophic forgetting, which is when the agent forgets a lot of other information it has learned in previous training. So the question is, how can an agent customize? How can it be adapted to user's needs without forgetting its reasoning and becoming dumber?