Google’s DeepMind AI group has released a reinforcement learning tool that can develop extremely optimized algorithms. It does this without first being trained on human code examples because it is set up to treat programming as a game.
This is according to a report by Ars Technica published on Thursday.
DeepMind already had the ability to teach itself how to play games, conquering games as varied as chess, Go, and StarCraft. The software was effective at learning to play by itself and discovering options that allowed it to maximize a score through approaches to the games that humans haven’t thought of.
Removing the need for human models
Today, large language models write effective code because they have trained on human models. However, this training means they are not likely to develop something that humans haven’t done previously.
That’s why to optimize well-understood algorithms, it’s best not to base them on human code. The question that surfaces is how do you get an AI to identify a truly new and unique approach?
Programmers at DeepMind decided to replicate the approach they used with Chess and Go and transformed code optimization into a game. They engineered algorithms that treated the latency of the code as a score and tried to minimize that score resulting in software that had the ability to write tight, highly efficient code.
They did this through a complex AI system called AlphaDev that consists of several distinct components. Its representation function tracks the overall performance of the code as it’s developed, including the general structure of the algorithm and the use of x86 registers and memory.
Benefits of New System of Google’s DeepMind
The main advantage of this new system is that its training doesn’t have to involve any code examples, as it generates its own code examples and then proceeds to evaluate them. Through this system, it collects information about combinations of instructions that are effective in sorting, reported Ars Technica.
In January 2023, Google Research and DeepMind launched MedPaLM, a large language model aligned to the medical domain.
The software was meant to generate safe and helpful answers in the medical field. It combines HealthSearchQA, a new free-response dataset of medical questions sought online, with six existing open-question answering datasets covering professional medical exams, research, and consumer queries.
Meanwhile, last month, Demis Hassabis, the CEO of DeepMind, said artificial general intelligence (AGI), a machine intelligence that can comprehend the world as humans do, might be developed “within a decade.”
