undefined | Better HN

0 pointsno_wizard11mo ago0 comments

I don’t believe the technology horizon in the next 5 years is sufficiently developed for recursive self improvement to work so well it requires no human intervention, by which I mean it will hit a limit and the technology will still be a tool not a sentient (near sentient?) thing.

I think there will be a wall hit eventually with this, much like there was with visual recognition in the mid 2010s[0]. It will continue to improve but not exponentially

To be fair I am bullish it will make white collar work fundamentally different but smart companies will use it to accelerate their workforce productivity, reliability and delivery, not simply cut labor to the bone, despite that seemingly being every CEOs wet dream right now

[0]: remember when everyone was making demos and apps that would identify objects and such, and all the facial augmentation stuff? My general understanding is that the tech is now in the incremental improvement stage. I think LLMs will hit the same stage in the near term and likely hover there for quite awhile

0 comments

TeMPOraL11mo ago

> I don’t believe the technology horizon in the next 5 years is sufficiently developed for recursive self improvement to work so well it requires no human intervention

I'm personally 50/50 on this prediction at this point. It doesn't feel like we have enough ingredients for end-to-end recursive self-improvement in the next 5 years, but the overall pace is such that I'm hesitant to say it's not likely either.

Still, my reply was to the person who seemed to say they won't be impressed until they see AIs "able to build better versions of themselves" and "exponential improvements of their capabilities" - to this I'm saying, if/when it happens, it'll be the last thing that they'll ever be impressed with.

> remember when everyone was making demos and apps that would identify objects and such, and all the facial augmentation stuff? My general understanding is that the tech is now in the incremental improvement stage.

I thought that this got a) boring, and b) all those advancements got completely blown away by multimodal LLMs and other related models.

My perspective is that we had a breakthrough across the board in this a couple years ago, after the stuff you mentioned happened, and that isn't showing signs of slowing down.