Researchers just found a way to make a closed AI model smarter.

@agenticamit

12.1K followers · usually 3.9K views a post

Views

8.5K

Likes

328

Comments

3

Shares

84

SoundTikTok sound · original sound - agenticamit

Promoted product

SkillOpt

Why it worked

The video effectively explains a complex AI concept using clear animations and a concise narrative. It highlights a novel approach to improving AI models, which is relevant to a niche but growing audience interested in AI development and LLMs.

Summary

This video explains SkillOpt, a new method for improving closed AI models by editing their skill files instead of retraining the entire model. It demonstrates how this process allows for continuous improvement and transferability of skills across different models.

Structure

1Introduction to closed AI models and their limitations
2Introducing SkillOpt as a solution
3Explaining the process of editing skill files
4Demonstrating the optimization loop and validation
5Highlighting the benefits of SkillOpt

Original caption

you can't fine-tune GPT, but you can train the skill file it reads. SkillOpt edits the doc like gradient descent — ~+25 pts, no retraining #AIengineering #LLM #prompting #agents

More from @agenticamit

Product placement

The video explains and visualizes a new AI technique called SkillOpt, which allows for the improvement of AI models by editing their 'skill files' rather than retraining the entire model. The process is shown through animations of editing code, optimization loops, and performance graphs.

On-screen text

the 60-second paper frozen. still learning. skill.md weights frozen eval 42 rising SKILLOPT - SELF-EVOLVING AGENT SKILLS weights, sealed skill.md - editable can-not-touch you change this you do not train the model - you train its instructions. HOW SKILLS ARE MADE TODAY by hand one-shot LLM v1 epoch unrealized nover impress WHAT IF THE DOC WAS THE WEIGHTS? minimum weights -> words gradient -> an edit loss -> benchmark score THE OPTIMIZER LOOPS epoch 01 score 58 PROPOSE ~rewrite PATCH epoch 02 score 66 epoch 03 score 74 run . score . edit . repeat. WHAT KEEPS IT FROM BREAKING hold-out set validate rejected - do not retry ~rewrite small steps only remember rejects only strictly-better edits survive. WHAT IT BUYS YOU epoch skillopt GPT-3.5 Codex loop +24.8 skill beats: hand-written - one-shot - TextGrad - GEPA - EvoSkill PORTED, NOT RETRAINED as fine-tuning it is just a file - so it transfers. weights, frozen the model never moved. the skill grew up.

Transcript

Researchers just found a way to make a closed AI model smarter. One thing about closed AI models like GPT or Claude is that you can't fine-tune them. The only thing you can actually control is the skill file that they read. You do not train the model - you train its instructions. And that is what turns out to be the whole secret here. What if you trained that document like you trained the weights? Basically gradient descent, but in text. Weights become words, gradient becomes an edit, loss becomes a benchmark score. They call it SkillOpt. It runs the agent, scores work failed, and the second model proposes tiny edits such as adding a line, cutting a line, rewriting a line, so on and so forth. But here's what keeps it from breaking. Every edit has to prove itself on a held-out set. It needs to beat the score or it's reverted. So small steps only. Only strictly-better edits survive. What it buys you: This means that the skill teaches itself, almost +24.8 on GPT-3.5 Codex loop. Beats: hand-written, one-shot, TextGrad, GEPA, EvoSkill. Ported, not retrained. As fine-tuning. It is just a file, so it transfers to other models. The model never moved, the skill grew up.