Demis Hassabis was one of the cofounders of DeepMind. The company was later acquired by Google, and he is now the CEO of Google DeepMind. In September 2023 Demis and John Jumper have received the Lasker Award in basic medical research (1) for the invention of AlphaFold, an AI technology for predicting the three-dimensional structure of proteins from the amino acid sequence (2).
Several statements from recent interviews show his vision for the role of AI in science:
“We’re going to see a kind of new renaissance in science, where these AI techniques continue to get more sophisticated and get applied to a wide range of scientific fields.” (3)
About the future of large language models (LLMs):
There is a whole branch of research going into what’s called tool use. This is the idea that these large language models or large multimodal models, they’re expert at language, of course, and maybe a few other capabilities, possibly coding. But when you ask them to do something specialized, like fold a protein, they call a tool, which could be another AI system, that then provides the answer to that particular problem. You can think of the central system as a switch that you prompt with language, and it roots your query to the right tool to solve that question for you. Again, using the interface, the best interface really, of natural language. (4)
In September 2023 a paper about another important AI science tool developed by DeepMind, AlphaMissense, appeared in Science (5). AlphaMissense predicts the pathogenicity of all possible human single amino acid substitutions. All the components of the AlphaFold AI model were shared openly (6), but in the case of AlphaMissense the trained weights were not shared (7). The weights are parameters that AI models obtain from learning. In the Acknowledgement of the Science paper (5), they state “As part of our commitment to releasing our research breakthroughs safely and responsibly, we will not be sharing model weights, to prevent use in potentially unsafe applications. “. We have contacted the corresponding authors of this paper (5) to ask for any examples of potentially unsafe applications that could originate from sharing the weights, and we will report here if we receive a response.
REFERENCES
1 - 2023 Lasker award announcement
https://laskerfoundation.org/winners/alphafold-a-technology-for-predicting-protein-structures/
2 - Tunyasuvunakool, K., Adler, J., Wu, Z., Green, T., Zielinski, M., Žídek, A., ... Jumper J & Hassabis, D. (2021). Highly accurate protein structure prediction for the human proteome. Nature, 596(7873), 590-596.
3 - https://www.technologyreview.com/2022/02/23/1045016/ai-deepmind-demis-hassabis-alphafold/
4 - Abridged from
https://www.theverge.com/23778745/demis-hassabis-google-deepmind-ai-alphafold-risks
5 - Cheng J, Novati G, Pan J, Bycroft C, Žemgulytė A, Applebaum T, Pritzel A, Wong LH, Zielinski M, Sargeant T, Schneider RG, Senior AW, Jumper J, Hassabis D, Kohli P, Avsec Ž. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science. 2023 Sep 22;381(6664):eadg7492. Epub 2023 Sep 22.
6 - https://www.deepmind.com/open-source/alphafold
7 - https://github.com/google-deepmind/alphamissense