introduction
demis hassabis, prominent ceo of google deepmind and the esteemed 2024 nobel laureate in chemistry, shared his profound and comprehensive vision regarding the remaining obstacles and challenges standing in the way of achieving desired artificial general intelligence during an extensive and insightful interview with investor gary tan of y combinator, published on april 29. among the most exceptional and striking proposals presented by hassabis during the discussion is an ambitious new benchmark dubbed the “einstein test.” this rigorous test aims to evaluate systems’ capabilities by exclusively and strictly training an artificial intelligence system on human knowledge and data available only prior to 1901, after which it is asked to independently and logically derive and discover albert einstein’s astounding 1905 breakthroughs, including the special theory of relativity that changed the course of physics.
- the long road toward artificial general intelligence
- isomorphic labs and virtual cell discoveries
- valuable advice for technology entrepreneurs
- frequently asked questions
the long road toward artificial general intelligence
during the engaging conversation titled “agents, agi, and the next big scientific breakthrough,” hassabis stated clearly and confidently: “once an ai system is able to achieve that and successfully pass this test, these systems will be very close and fully qualified to genuinely invent new and innovative scientific concepts.” hassabis told his interviewer that while current available technologies and tools—such as large-scale pre-training, reinforcement learning from human feedback, and sequential logical reasoning—will certainly form an essential part of the final architectural structure of artificial general intelligence, one or two “missing elements” are still urgently required to complete the complex picture. hassabis identified continual learning without forgetting the past, long-term and complex reasoning, and effective memory as the core fundamental challenges that remain fully unresolved to this day, thereby setting his personal timeline and expectations for reaching artificial general intelligence around roughly 2030.
he also acknowledged a striking scientific paradox and clear contradiction in current technical systems; models that can brilliantly solve difficult international mathematical olympiad problems at a gold medal level are unfortunately still capable of making basic, elementary calculation errors when a question is rephrased differently. commenting on this, he noted: “it seems there is something missing and fundamental in the system’s introspection and direct reflection on its own thinking process.” he also observed that while artificial intelligence has not achieved any fully independent major scientific discovery up to this moment, he firmly believes that such a desired breakthrough is “imminent,” and that the most rigorous and demanding test for artificial intelligence will lie in its ability to propose and formulate an entirely new set of complex scientific problems with the same depth and difficulty as known millennium prize problems.
isomorphic labs and virtual cell discoveries
in a related context, hassabis proudly revealed that isomorphic labs—a pioneering ai-driven drug discovery company that successfully spun out of and operates under the umbrella of deepmind—now stands on the verge of officially announcing new and dazzling scientific results and achievements. he added: “we are now tackling and breaking into adjacent and complex biochemical research, designing chemical compounds with targeted and desirable properties, and there are major and important announcements imminent soon.” it is worth noting that this company, which earlier this year announced a massive strategic research collaboration with johnson & johnson and officially received u.s. food and drug administration approval to begin human clinical trials in january for its innovative ai-designed drug candidate, now holds 19 independent and comprehensive programs covering cancer, heart disease, and immunology, according to a detailed report published by the specialized magazine “the economist.”>
hassabis explained that the long-term strategic scientific ambition of this organization is to build a complete and comprehensive “virtual cell,” which is a fully functional cellular simulator whose scientific outputs and predictions closely match experimental and real-world results in laboratories. he estimated that achieving this ambitious and monumental scientific goal is still about ten years away from hard work, as deepmind’s specialized science division has first begun building and developing a virtual nucleus as a more tractable starting point at present to establish the required infrastructure.
valuable advice for technology entrepreneurs
in concluding his inspiring remarks, hassabis strongly urged deep tech startup founders and entrepreneurs to take artificial general intelligence timelines very seriously and think about them deeply when planning their future projects and investments spanning decades. he sincerely called on them to focus intensely on scientific fields that involve direct interaction with the physical and atomic world, where easy shortcuts or solutions are highly unlikely to exist. advising them, he said: “life is too short and human energy is limited, so invest your vitality and dedicated effort in vital and difficult fields that no one else will do if you do not do it yourself.”
frequently asked questions
question: what is the einstein test proposed by demis hassabis?
answer: it is a proposed benchmark for evaluating artificial general intelligence, in which a system is trained on knowledge available prior to 1901 and then asked to completely independently derive and discover einstein’s special theory of relativity.
question: when does hassabis expect to reach the stage of artificial general intelligence?
answer: based on current developments and the pace of progress, hassabis expects to reach artificial general intelligence around roughly 2030, after overcoming obstacles in continuous learning and memory.
question: what is the ambitious virtual cell project?
answer: it is a scientific project aimed at building an advanced digital simulator of a fully functioning living cell whose outputs match experimental and real-world results in medical laboratories accurately, requiring about 10 years to complete.
question: what advice did hassabis give to new entrepreneurs and investors?
answer: he advised them to take ai timelines seriously and direct their investments and focus toward complex fields dealing with the physical and atomic world where there are no shortcuts or easy solutions.
