Jake Levi

[ About | Vision | Papers | Talks | Social | Blog | Personal ]

About

I'm a PhD student on the AIMS CDT supervised by Professor Mark van der Wilk in the Department of Computer Science at the University of Oxford.

Training modern AI systems requires huge datasets and expensive data centres. In contrast, we as humans can learn much more efficiently, in terms of data and memory consumption. How can we design AI models that can learn as efficiently and robustly as humans can learn? Answering this question is the goal of my research.

Answering this question would be interesting because it would improve our scientific understanding of intelligence. Answering this question would be useful because it would reduce the energy and hardware costs of training large AI models. The sooner we develop AI models that learn more efficiently, the more we can reduce the environmental cost of training large AI models.

But why do I care about answering this question? Because I think that needlessly inefficient algorithms are just a bit rubbish. Would you be happy using bubble sort when you know quicksort exists? Then why should the AI research community be satisfied training transformers with backpropagation, when we know that natural intelligence is so much more efficient?

Vision

AI is here to stay (like it or not), and is (at least in the short term) massively accelerating productivity for many people on an individual level. There are however 4 common complaints that people have about AI tools: (1) they're expensive, (2) they quickly run out of context length during complex tasks, (3) AI companies have the power to withdraw powerful models at will, (4) training and inference of AI models requires huge data centres which have a massive environmental cost and negative impact on communities.

Given the memory and data efficiency of human intelligence, I believe that one day it will be possible to develop AI models which are (1) open-source, (2) at least as powerful as current frontier models, (3) orders of magnitude less expensive to train, and (4) efficient enough to run locally for everyone on commodity hardware (cheap GPUs) without restrictions on context length. This will avoid all of the above complaints, and put AI research back in the hands of academia. Getting there will require a lot of fundamental research which is not necessarily well-aligned with current mainstream AI research, and that is exactly what I am working towards.

Papers

Talks

Social

Blog

Personal

Besides rock climbing (trad + sport + bouldering), cinema, and playing classical guitar, my major passion is designing HTML websites. If you would like me to design your website, email me at
<a href="mailto:example@email.com">ERROR! BROKEN LINK</a>.