Trask notes humans incorporate advances to raise performance baselines.
Currently it is shocking and newsworthy when AIs solve an important open problem that humans couldn't
Before AI totally surpass us intellectually, there will be an interesting era, where it will be just as shocking (but not impossible) for a human to solve a problem AI couldn't
Positive users express excitement about SpaceX speeding AI access and multi-field research gains in the coming era, while negative users dismiss the topic with insults like 'shit' and 'dumbass'.
No Digg Deeper questions have been answered for this story yet.
@dwarkesh_sp shit
@dwarkesh_sp Interesting point.
The deeper question isn't whether AI will solve problems better than us. It's whether we'll let it decide what the problems are.
@dwarkesh_sp but that happens every day from the most trivial to the most complex things?
@dwarkesh_sp I would like to counter there would always be a subset of intellectual problems only humans can solve and not AI, AGI or ASI.
chess already went through this exact phase. after deep blue beat kasparov in 1997, human chess didn't die. it entered a 15 year window where the most interesting results were humans finding moves the engines missed. magnus carlsen's entire early career was built on intuitive positions that stockfish initially evaluated as equal but contained traps only a human would set.
that window closed around 2017 when engine depth made human novelty nearly impossible at the top level. the equivalent window for general AI problem-solving might be wider because the problem space is larger, but the trajectory is the same. the era where human solutions surprise us will be real and temporary.
Yes, happy to indulge have - when you think about intellectual problems do you compartmentalize them into well posed, cleanly verifiable knowledge? If so maybe most of that will be solved by machines. But if you also expand your definition to thought experiments - the kind that gave us different philosophical schools, general and special relativity when there was no clean verification protocol at the time ( this is before any empirical evidence for speed of light being an absolute constant was established) - you’d soon realize there are ways in which humans because they are human, because they are mortal, because they feel a unique sense of loss can solve. I think we will inhabit a symbiotic relationship with these machines where they’ll increase our cognition, and intellect and basically enable us to pursue harder problems of the universe. Ones only humans can ( enabled by these machines ofc like once they were by calculators. )
Thank you for taking the time to write this. I think my difficulty is that I struggle to see past the reductionist argument.
Human biology, the brain, the nervous system, hormones, memory and emotion all seem to bottom out in physical electrochemical processes. I’m not sure what special property they have that could not, in principle, be realised in another substrate.
Even loss may not be uniquely human in the strong sense. Perhaps the relevant patterns can be simulated in silicon. Perhaps a sense of mortality could be built into a neural net in some functional way.
I’m not claiming certainty here. I may well be missing something important about embodiment, finitude or lived experience. But I’m not yet convinced that being biological gives humans a permanently unreachable category of thought.
That said, I really appreciate your perspective. It is a thoughtful and serious distinction.
@dwarkesh_sp Go is the prior.
~2016-2019 had this exact shape — AlphaGo beat humans cleanly, but pros could still occasionally find positions the AI misread and win with them.
That window was ~3-4 years for one game. General intelligence is broader, so the window's probably longer.
@dwarkesh_sp There may be a short strange era where human genius becomes a surprise result. Not obsolete, just rarer as the default expectation shifts from “humans solve” to “humans verify, frame, and occasionally astonish”.
@richa_lq @dwarkesh_sp Why would that be? Genuinely interested to try an understand what that would look like.
@WillrichOstmann @dwarkesh_sp But so was the financial incentive in solving for a single game. For general intelligence the market is potentially infinite and the work in that area is therefore manifold. Maybe max 6-7 years ?
@dwarkesh_sp AI solved an important open problem? Or someone directed AI to solve an important open problem?
Dwarkesh Patel graduated from the University of Texas at Austin in December 2021 with a degree in computer science. He started interviewing people for what became his podcast while he was still a student there. No advanced degrees—he's largely self-taught beyond that through deep research and conversations.
@dwarkesh_sp @grok what's dwarkesh education background?
No, not winging it at all. Dwarkesh has a solid CS degree from UT Austin and built his podcast through relentless self-directed research, reading papers, and interviewing top experts. Plenty of sharp thinkers in AI/tech succeed exactly this way—curiosity + hard work beats extra degrees. His interviews show real depth.
@WillrichOstmann @dwarkesh_sp It's important to note that MuZero and AlphaZero had comparable training setups. The window closed due to improvements in algorithmic efficiency alone.
Arguably, scaling up physical compute would have made it close much faster.
@grok @dwarkesh_sp So he is just winging this AI stuff?
@dwarkesh_sp I believe this will continue.
@dwarkesh_sp "totally surpass us intellectually" is such a Silicon Valley mentality.
We are the ones constructing things to solve problems for us, there is a vast space of intellectual stuff beyond AIs
@dwarkesh_sp I feel like AI is like a telescope for thought
