00006 - Insights from my First Batch Projects - Google Slides
00006 - Insights from my First Batch Projects Slideshow Share FileEditViewToolsHelp View only 1 Intuitions about LLMs Insights from my work in the first batch of Turing’s Dream (10 Sept to 22nd October 2024) Paras Chopra, founder of Turing’s Dream 2 What I focused on during last 6 weeks: ● Power of abstractions ● LLM “dark arts” ● Improving reasoning in LLMs 3 Project 1: Power of abstractions ● Replicated (somewhat) the DreamCoder paper ● Problem: given inputs and outputs, find a (functional style) program that maps inputs to outputs ○ Approaches tried: ■ Exhaustive search ■ Exhaustive search with abstractions put back in ■ Evolutionary methods that put winners back in the mix ■ Neural guided search ■ LLMs Paper link 4 Results on a total of 150 problems Approach Finding Exhaustive search Solves 59 in 1 hour of search Exhaustive search with abstracted functions (from previous iteration) put back in Solves 74 in 1 hour of search Neural guided search Didn’t work Evolutionary searc
Explore this link on the map →