Full text
Albert Einstein is one of a handful of scientific minds in history who made profound mental leaps that arguably no other researcher could have achieved during their era. There just wasnโt sufficient scientific data to motivate anyone else to come up with Einsteinโs theories of special and general relativity (though some came close ). In fact, the core principles of relativity are counterintuitive to minds that evolved here on Earth, where objects move slowly, gravity feels like a force, and time is a yardstick that consistently measures our lives.
Famously, Einstein made his breakthroughs by performing thought experiments in which he imagined himself in extreme physical situations, like riding a beam of light or standing in a falling elevator. By projecting himself into a reality beyond our natural human experience, he transformed our understanding of space, time, and gravity.
Of course, Einstein isnโt the only innovator to make mental leaps that were well ahead of his time. Alan Turing invented the field of computer science before anyone had built a truly programmable computer. Using pure logic and imagination, he envisioned a โ Universal Turing Machine โ decades before the components existed to create one. He even predicted that programmable machines would enable human-like intelligence, a milestone we now call โartificial general intelligenceโ (AGI) , and put forth his famous Turing Test to assess when AI reaches this level.
Unfortunately, Turingโs test relies too much on human judgement. He proposed using a human evaluator to hold a blind conversation with both a human and a machine. If the evaluator could not reliably determine which participant was human, the machine would have implicitly matched human intelligence. It turns out, this test fails because people are too easily fooled due to our inherent tendency to anthropomorphize anything that behaves remotely human.
A better Turing Test
Many benchmarks have been proposed for assessing the arrival of AGI , but most require machines to solve problems whose solutions could be embedded within the data used to train them. As a result, it has been challenging to find repeatable tests that require an AI to make genuinely original and creative leaps.
This brings me to the โ Einstein Test โ proposed by Benrimoh et al. and recently endorsed by Nobel Prize-winning computer scientist Demis Hassabis. The concept is to build special versions of frontier models trained only on the scientific information available to Einstein before his breakthroughs. If such an AI system could independently derive the special and general theories of relativity, it would demonstrate that AI could make creative leaps at the upper limits of what a human brain has ever achieved.
I like this test because it sets a high bar and could be replicated in disciplines outside theoretical physics. Hassabis, who oversees Google DeepMind, suggested the Einstein Test is a good benchmark for AGI, which he estimates thereโs a 50% chance of achieving by the end of the decade. If this happens, it could unleash a flood of scientific advancements as the world suddenly deploys thousands of โAI-powered Einsteinsโ to solve thousands of hard problems, all at software speeds.
As context, we have experienced rapid acceleration in innovation over the past 100 years as the world transitioned from mechanical improvements, which progressed at a roughly linear rate, to digital advancements that have progressed closer to an exponential rate. AGI could push things even faster, driving innovation at a hyperbolic rate because AI is likely to operate at steadily increasing speeds and could trigger recursive loops of self-improvement.
Would this be good for humanity?
Iโve been focused on this issue for 20 years. While I agree with Hassabis that society would benefit from a tsunami of genius-level advancements, I fear superintelligence could have a demoralizing impact on us humans. After all, when AI systems can make leaps at the level of an Einstein or Turing, and do so in a fraction of the time, human scientists and inventors may never have the opportunity to make grand breakthroughs. In fact, from an innovation perspective, we could soon pass โ Peak Human โ as AI systems take the reins.
I hope this doesnโt happen, but superintelligence could mean we never produce another genius-level innovator like Einstein or Turing, at least not a human one. In addition, this threat might not be limited to science and technology. What if superintelligence means we never have another human Shakespeare, Aristotle, Confucius, or Dickinson? This is harder to predict, but certainly within the realm of possibility. Would this be demoralizing?
In the spirit of an Einstein thought experiment, ask yourself this: How would you feel if your favorite new TV show were written entirely by AI ? Or your favorite new comedian, a hilarious AI? Or the hottest new fashion designer? Or the most influential political thinker? None of these things are certain to happen, but all could happen. If so, how would that impact our collective psyche? Our human dignity? This is a genuine risk factor that is not getting sufficient attention. In fact, this month, the United Nations held its first-ever Global Dialogue on AI Governance in Geneva, and these issues were not on the agenda.
I know that many who read this will insist AI could never write the next Breaking Bad or do creative standup at the level of George Carlin, and I hope those people are correct. That said, I worry those views are rooted more in denial and wishful thinking than an objective assessment of the power and trajectory of AI. In fact, I see many people cling to outdated views of AI being nothing but statistical parrots. As I covered in Big Think in January , this stance is untrue of modern AI systems like Gemini and Claude.
Believe me, I would love to see AI hit a wall before reaching human-level intelligence. In fact, I clung to this view for much of my career. Then, 15 years ago, I asked myself a simple question: Why is this the most likely future? I had no reasonable answer and had to admit it was rooted mostly in wishful thinking.
So, I asked the next logical question: What can we do to ensure that human intelligence remains relevant in an age of superintelligence? For me, the answer boils down to five simple words: keep humans in the loop . I am not talking about giving an elite group of people supervisory control over a superintelligence in the hope they can contain it or align it with human interests โ Iโm skeptical either will work over time. I mean building systems in which large numbers of ordinary humans are part of the thinking itself.
Specifically, I believe we should shift our priority from purely digital superintelligence to a hybrid method I refer to as โ collective superintelligence .โ This means building systems that incorporate human values, morals, sensibilities and interests into the process โ not by training them on data about humans but by connecting them to groups of real humans .
This might sound like science fiction, but evolution has gone down this path many times before. From bird flocks and fish schools to bee swarms, groups of brains can form real-time systems that exceed the intelligence of every individual member by leveraging their combined insights and perspectives. Biologists call this โswarm intelligence,โ and as I described in Big Think in June , it could enable humans and AI agents to โthink togetherโ at scale, producing superintelligent systems that are genuinely human at their core.
From a practical perspective, instead of releasing millions of AI agents designed to replace people (as is currently happening), we should be releasing AI agents designed to connect people โ harnessing, amplifying, and integrating our human collective intelligence with AI systems. The resulting hybrid collective superintelligence would likely drive innovations more slowly than a purely artificial superintelligence, but it would think in ways that are inherently more human than the alternative.
Personally, Iโve been researching the potential of hybrid human-AI systems for the past decade through a company I founded, Unanimous AI. In an early NSF-funded study performed in collaboration with Stanford University Medical School, we found that teams of human doctors working in collaboration with AI systems could outperform both humans and AI systems working in isolation, demonstrating a symbiosis that harnesses the advantages of both while overcoming their respective limitations.
I know that many people reading this would prefer that superintelligence never happens (myself included), for it will mean there are countless brains in our lives that are smarter, faster, more creative, and more knowledgeable than any person โ even Einstein-level geniuses. In other words, for the first time since humans climbed down from the trees, our species may lose cognitive supremacy in our natural environment. This changes our role in the world.
Can we prevent it? Unfortunately, the race to AGI is unlikely to slow โ the economic and political pressures are only growing stronger. Superintelligence will likely arrive , and it could make us feel like mere observers in a rapidly changing world thatโs driven more by AI than humans. An alternative is to push for collective superintelligence . If we take this path, there still could be another human Einstein โ it might just be all of us, together.
This article Have we seen the last Einstein? is featured on Big Think .
Comments
No comments yet โ be the first to weigh in ๐
No comments yet. Be the first!