Hassabis, realistic AGI and rosy tomorrows

Jorge Costa Oliveira

In 2020, Demis Hassabis and John Jumper unveiled AlphaFold2, an artificial intelligence model capable of predicting the structure of virtually every known protein. Widely used in pharmaceutical and environmental research, it earned them the 2024 Nobel Prize in Chemistry.

AlphaFold2 is among the best examples of how specialized AI can improve millions of lives—and why AI companies would do well to focus on such systems.

Yet several leading American technology companies prioritize achieving artificial general intelligence (AGI): human-level intelligence across many fields. Massive investment has accelerated large language models and fueled optimism among industry “gurus.” Because much of these companies’ financing and valuations depends on winning the AGI race, forecasts of its imminent arrival often resemble self-interested optimism.

Elon Musk expects AI smarter than the smartest human in 2026. His timelines, however, tend to slip: for a decade, he has promised fully autonomous Teslas “next year.” Cybersecurity expert Roman Yampolskiy predicted in September 2025 that humanoid robots would probably rival humans across every domain by 2030. Ilya Sutskever, OpenAI’s co-founder and former chief scientist, expects AGI within five to 10 years, while acknowledging uncertainty. Yann LeCun, Meta’s former chief AI scientist and a Turing Award winner, is among the few dissenters: he believes AGI remains decades away.

Definition is crucial. OpenAI literally defines AGI as a system outperforming humans at most “economically valuable work.” In other words, replace enough jobs and AGI has arrived. By that measure, we already have one foot through the door.

Hassabis considers that threshold meaningless. In a recent NothingButTech podcast interview, he argued that true artificial general intelligence must do what the human brain can do — the only proof that such intelligence is possible.

Today’s AI has read practically everything humanity has written, including relativity. When it explains Einstein, it reproduces an existing answer. Hassabis proposes a tougher test: train AI solely on knowledge available in 1901, four years before Einstein published special relativity theory, then ask it to derive the theory independently. With no answer to retrieve, it would have to reason from contemporary physics toward an idea nobody had conceived.

No current AI can do that. What some gurus call “near AGI” is really history’s finest librarian: it can locate any existing answer but cannot create a nonexistent one.

Hassabis offers another test. AlphaGo invented strategies and moves unseen in 2,000 years of [the game] Go. But genuine intelligence would not merely devise a new move within Go; it would invent a game as profound, complex and beautiful as Go. No existing model can.

Machines remain extraordinarily powerful tools, but they cannot yet produce the genuinely original idea that would make them truly intelligent.

Hassabis nevertheless believes AI will someday create such a game and that “safe AGI” will arrive by 2050. He does not specify what kind, but courageously sets a higher, more rational bar than Sam Altman and other fundraisers for AI megaprojects.

The trouble is that Hassabis also thinks that by 2050 we will have solved economic growth, overcome the basic dilemma of solving unlimited needs with limited resources, offset AI-driven unemployment, and turned toward space exploration.

Here lies one defining drama of our age: even AI gurus displaying sound judgment and realism on some matters, however dazzling their technical achievements may appear to us, readily skid into implausibly rosy tomorrows when forecasting the near future.

linkedin.com/in/jorgecostaoliveira