Lin is founding a new lab for world models.
ai making ai better recursively is an impressive idea but sometimes depressing as human researchers seem to become less and less significant in this process. i do believe that there is great potential in this direction while human supervision will become more principled and higher-level. then imagination and vision matters more than ever.
Positive users argue human intuition, vision, and judgment will remain essential amid recursive AI self-improvement, while negative users worry it erodes researcher standing and job value.
can't agree more~!
ai making ai better recursively is an impressive idea but sometimes depressing as human researchers seem to become less and less significant in this process. i do believe that there is great potential in this direction while human supervision will become more principled and higher-level. then imagination and vision matters more than ever.
@JustinLin610 Good point. I ran a few tests, and I tend to believe that human input is still needed to reduce the system's entropy.
@JustinLin610 high-level oversight still needs a ground truth
@JustinLin610 Is the news that you started your own business to set up an AI laboratory and raised financing true?
but I'm worried that AI's imagination and vision may also be better than humans. They can generate many many ideas and then pick the best ones, which is likely better than what an individual researcher can come up.
Right now, I feel my only advantage over models is that I have a longer context window, but not sure how long that will last...
ai making ai better recursively is an impressive idea but sometimes depressing as human researchers seem to become less and less significant in this process. i do believe that there is great potential in this direction while human supervision will become more principled and higher-level. then imagination and vision matters more than ever.
@NyanpasuKA @JustinLin610 Exactly that, not only the scale will get better but the vision i see for the future is that we will start talking about maintaining scalability of research pipeline
I think that’s exactly the shift we’re starting to see.
As AI becomes better at generating ideas, experiments, code, and even research directions, the comparative advantage of humans moves further upstream. Less time spent on execution, more time spent on deciding what is worth pursuing in the first place.
@JustinLin610 Humans will always be needed for intuition. Machines dont have pineal gland.
@JustinLin610 The most significant anything in progress, technical or otherwise will always be insight, not solution space search. The latter just powers the former.
@JustinLin610 Self-improvement of AI requires high-level abstract thinking, understanding and leveraging concepts deeply. So if we want humans to remain meaningful, we need to think at an even higher level. Curious whether AI self-improvement will surpass that ceiling.
@JustinLin610 I wonder how accurately RVLR is offsetting the cost to remain in-distribution though
@JustinLin610 But the frontier is not fully solved for yet. Recursive improvement means easy gains go but hard gains can still require human judgement :)
@JustinLin610 Why not having human become knowledge consumer? After all, most human spend most time consume knowledge, instead of creating new ones. We should strive to make it easier for human to understand the science as well
@JustinLin610 worded my thoughts perfectly
@JustinLin610 First is how to complete automatic verification
@JustinLin610 As scale gets bigger I think more eyes will be needed on the same pipeline, so to maintain the cadence of experimentation and scale restructuring or scaling researchers count will be necessary
@NyanpasuKA @JustinLin610 PreAGI is when even that someone is fully capable of identify and flag them ( would even argue that we are partly there)
@JustinLin610 You killed my job as a SWE. Sounds like karma to me.
@JustinLin610 Exactly right, vision and taste become the final edge.
@JustinLin610 The shift is already happening — just not evenly. The humans who stay relevant aren't the ones with more imagination. They're the ones who got good at noticing when the AI drifted. Supervision is becoming a skill. Most people haven't started practicing it yet.
Lin is founding a new lab for world models.
ai making ai better recursively is an impressive idea but sometimes depressing as human researchers seem to become less and less significant in this process. i do believe that there is great potential in this direction while human supervision will become more principled and higher-level. then imagination and vision matters more than ever.