Microsoft just made

@parthknowsai

98.4K followers · usually 23.5K views a post

Views

71.1K

Likes

4.4K

Comments

103

508

Why it worked

The video explains a complex AI concept (self-improving AI agents) in a simple, accessible way using clear visuals and relatable analogies, making it highly engaging and informative for a broad audience.

Summary

This video explains Microsoft's new AI research called SkillOpt, which allows AI agents to improve themselves by editing a 'skills stock' text file instead of retraining the entire model. The system achieved significant accuracy improvements across various benchmarks by making small, iterative edits based on the agent's performance.

Structure

1Introduction to Microsoft's SkillOpt
2Explanation of the traditional AI training method
3Introduction to the 'skills stock' concept
4How SkillOpt works: self-editing and improvement
5Demonstration of SkillOpt's effectiveness with data
6Limitations and future outlook of SkillOpt

Product placement

Microsoft (appears, does not act, removing it would not change what happens in the video)

On-screen text

MICROSOFT JUST MADE GPT. FIVE'S ACCURACY JUMPED FROM 41% TO FROM 41% TO 80% WITHOUT TOUCHING 80% WITHOUT TOUCHING A SINGLE PARAMETER. BY EDITING ONE SMALL TEXT FILE. THAT SOUNDS STRANGE, BUT LET ME EXPLAIN. SEE, THINKS ABOUT MAKING BETTER IS YOU GO IN, YOU RETRAIN THE MODEL, YOU UPDATE THE WEIGHTS. YOU SPEND MILLIONS OF DOLLARS AND MONTHS OF COMPUTE. THAT'S THE DEFAULT ASSUMPTION, BUT MICROSOFT JUST PUBLISHED A PAPER THAT SAYS, FORGET RETRAINING. LET'S JUST UPDATE THE SKILLS STOCK INSTEAD. SKILLS STOCK, IT'S BASICALLY A SMALL TEXT FILE Skills doc AGENT EXACTLY HOW TO BEHAVE, WHAT STEPS TO FOLLOW, HOW TO USE ITS TOOLS, ITS ANSWERS. IT'S BASICALLY AN SOP. FOR THAT AGENT. PAPER MICROSOFT BUILT SOMETHING CALLED SKILLOPT SYSTEM THAT AUTOMATICALLY REWRITES AND IMPROVES THAT SKILL DOCUMENT OVER TIME WITHOUT ANYONE TOUCHING THE MODEL ITSELF. AGENT GOES OUT AND DOES A BUNCH OF TASKS, THERE'S A SECOND, THAT WATCHES IT. IT CAN SEE WHERE IT FAILED, WHERE IT SUCCEEDED, AND PROPOSE TINY SKILLS DOC. FOR EXAMPLE, WE'RE GONNA ADD THIS NEW RULE HERE, OTHER ONE. WE'RE GONNA REPLACE THIS VAGUE INSTRUCTION WITH THE SHARPER ONE, THEN IT OBSERVES THE NEW VERSIONS. IF IT'S BETTER, WE KEEP IT. IF NOT, WE AND REMEMBER, WHY IT FAILED. AGENT IS GETTING SMARTER FROM ITS OWN MISTAKES. AND THIS PAPER IS GETTING A BECAUSE THE RESULTS ARE PRETTY GOOD ACROSS 52 DIFFERENT MODELS AND TASK COMBINATIONS. WIN EVERY SINGLE ONE ON C P.T. SKILLSOPT RESULTS clears the strongest baseline 5.5 SPREADSHEET ACCURACY WENT FROM 41% TO 80% DOCUMENT TASK WENT UP FROM 33% TO 72% AND MOST OF THOSE GAINS CAME FROM JUST ONE TO THAT SKILLS DOC. NOW, I WILL SAY THIS WHEN THE AI. CLEAR RIGHT OR WRONG ANSWER TO CHECK ITSELF AGAINST. SO THINGS LIKE SPREADSHEETS, DOCUMENTS, MATH, SEARCH. PERFECTLY THERE. BUT THE SECOND YOU MOVE INTO SOMETHING CREATIVE OR STRATEGIC, WORK BECAUSE THERE'S NOTHING SOLID TO I MEAN, THE PAPER ADMITS THIS THEMSELVES. GROUNDBREAKING? ABSOLUTELY. BUT IS IT THE FULL PICTURE? NOT YET. FOLLOW AND I'LL KEEP

Original caption

Microsoft’s new AI research paper SkillOpt is impressive #ai #LearnOnTikTok #techtok #Tech #EduTok

More from @parthknowsai

Their other posts in the index, biggest breakout first.

Transcript

Microsoft just made GPT. 5.5's accuracy jumped from 41% to 80% without touching a single parameter. By editing one small text file. And I know that sounds strange, but let me explain. See, the way everyone thinks about making AI better is you go in, you retrain the model, you update the weights, you spend millions of dollars and months of compute. That's the default assumption. But Microsoft just published a paper that says, forget retraining. Let's just update the skills stock instead. If you're not familiar with the skills stock, it's basically a small text file that tells an AI agent exactly how to behave, what steps to follow, how to use its tools, how to format its answers. It's basically an SOP for that agent. In this paper, Microsoft built something called SkillOpt, which is a system that automatically rewrites and improves that skill document over time without anyone touching the model itself. So when the agent goes out and does a bunch of tasks, there's a second AI that watches it. It can see where it failed, where it succeeded, and propose tiny edits to that skills doc. For example, we're gonna add this new rule here, or delete that other one. We're gonna replace this vague instruction with the sharper one. Then it observes the new versions. If it's better, we keep it. If not, we throw it away and remember why it failed. Which means the agent is getting smarter from its own mistakes. And this paper is getting a lot of attention because the results are pretty good across 52 different models and task combinations. SkillOpt won every single one on GPT. 5.5 spreadsheet accuracy went from 41% to 80%. Document task went up from 33% to 72%. And most of those gains came from just one to four edits to that skills doc. Now, I will say this, this only works when the AI agent has a clear right or wrong answer to check itself against. So things like spreadsheets, documents, math, search, the loop works perfectly there. But the second you move into something creative or strategic, it doesn't really work because there's nothing solid to grade against. I mean, the paper admits this themselves. Is it groundbreaking? Absolutely. But is it the full picture? Not yet. Follow and I'll keep you posted.