An autonomous agentic framework for cross-campaign generalization and sensor…
Search inside any video
SavedThat transcribes your saved videos and lets you search across all of them instantly. Save this video and find any moment.
TL;DR
This video discusses the challenges faced by autonomous agents in long-running tasks and introduces MOSAN, a solution for better record-keeping and learning.
Chapters
- 0:00Introduction to Long-Running Tasks01
- 0:30Context Dilution Explained02
- 1:00Intercampaign Amnesia Overview03
- 1:30Introducing MOSAN Framework04
- 1:50How MOSAN Works05
- 2:00Conclusion and Newsletter Info06
Transcript
If you try to give an LLM a long-running task with lots of steps, it will probably fail. Why? Because of two problems. The first is context dilution, which is just a fancy way of saying the model has tried to do what you ask, but has gotten lost along the way. The agent formulates a plan, starts following it, and logs everything it does as it goes.
But then, a few dozen steps in, the strategy it's trying to follow gets buried under all the information it's ingested and processed so far. And it just loses the thread and starts reacting to the previous steps instead of following the plan it wrote for itself. The second issue is called intercampaign amnesia, which just means that when a long-running task or series of tasks ends, the transcripts and logs end with it.
So if you then point your agent at a new series of similar tasks it will have to start completely over from scratch without any lessons from what it learned before This is a big problem when you running scientific experiments You want the agent to get better at these kinds of things over time but every time
you invoke it, it's forgotten everything it learned. In today's paper, the authors are proposing a fix for both of these problems. It's called MOSAN, the Multi-Objective State Action Network. It's an autonomous agent built around the idea of proper record keeping. Rather than logging everything to
a running conversation like a chatbot would, it maintains two things. A structured written ledger modeled after the way researchers use lab notebooks, and a persistent map of past experiments, the lessons of which can seed new projects without contaminating them. On today's
episode of Journal Club, we're going to walk through how the system works and how it solves both of these problems. If you're already subscribed to our email newsletter, then you'll be getting this episode in text, audio, and video form, delivered to your inbox in the next few hours. Journal Club, the latest academic research explained in plain English.
Are you the creator or rights holder of this video? Request removal of this transcript.
Related Transcripts
Claude Code + Karparthy’s AutoResearch = GOD MODE
Lex Fridman
I Built an Entire Marketing Agency With 1 AI Tool!
Higgsfield AI
Full Course: Spec-Driven Development with Coding Agents
DeepLearningAI
JEV: How It Works and What You Can Build
Riley Brown
Meta Ads Creative Strategy in 2026: The Full System
Blue Sense Digital
3 AI businesses for 2027 | Pt.2 I spent two days at GOSIM Shenzhen 2026 with AI…
Alex Wood
Never lose a moment again
Save videos from YouTube, Instagram, and TikTok. Search everything that was said, and jump to the second.
Start Saving Videos - Free Trial