Sebastian, 22 nov on Ai Journey conference we presented a multi-agent system FractalGPT and its Math agent, that is able to solve math tasks with 99.8% accuracy, outperforming most LLM nevertheless they are trained on math dataset or not (ChatGPT, GPT4, Claude2). Tested on dataset https://huggingface.co/datasets/ChilleD/MultiArith Very well scaled to any math problem

Watch: https://aij.ru/eng/broadcast?date=2023-11-22&streamName=Science Hall 1

Autonomous AI agents: industry trends and why prompts are not almighty

Viktor Nosko


This agent approach is a successor of tool use.

Expand full comment

Very cool, congratulations! The link you provided doesn't seem to link to a video directly. Do you have a more accessible link or perhaps a write-up?

Expand full comment

1. Go to the broadcast at Autonomous AI agents: https://aij.ru/eng/broadcast?date=2023-11-22&streamName=Science%20Hall%201

(parser destroy link, use mouse selection to copy)

2. Scroll down to the bottom of the page to the report "Autonomous AI agents: industry trends and why prompts are not almighty". They sliced the reports by timing a bit wrong in the player, so you have to click on the previous report - the Watch button. It will scroll up automatically

3. In the player, click on 6:45 - that's the beginning of my presentation

Expand full comment