Ask HN: Is anyone working on AGI on the side? | Hacker News Reader

Ask HN: Is anyone working on AGI on the side?

3 pointsopenquery6 comments

I was wondering if anyone was working on AGI (artificial general intelligence) on the side and what approaches people are taking.

I know there a few companies working on this. Are there any lone souls also trying to crack this nut?

View on HN

Yep. I spend a not insignificant amount of my time on this. I don't feel like sitting here and typing up a massive essay on what I've been doing, and it's not like I'm at a place where I have any particularly interesting results to share anyway. I just have a lot of speculation, a pile of books taller than Mt. Everest (or so it seems), a bunch of electronics, and time to burn.

That said, I've written a modest amount about my AI interests here on HN in the past. Feel free to trawl through my comment history[1] if you're curious. The most salient post I made was probably this one:

https://news.ycombinator.com/item?id=32868680

Going forward I'll be posting AI related thoughts at a couple of places:

    - Fogbeam Labs AI Blog - https://aiblog.fogbeam.com

    - Twitter              - https://twitter.com/mindcrime

                           - https://twitter.com/Bosworth_AI

    - Mastodon             - mindcrime@fosstodon.org

Follow along if you'd like. Note that I am not making any claim to be any kind of genius AI researcher, and I don't have any "magic sauce" per-se. I do have some fairly specific ideas about angles to pursue and lines of research, all of which might eventually lead to something, or it could all be absolutely useless. But I will say that what I'm doing is (mostly) at least rooted in / based on somewhat established ideas / principles taken from ML, earlier AI research, cognitive science, neuroscience, etc. I might be effectively a crackpot in the end, but I'd like to think that I'm not just a crackpot. :-) I'm just somebody who is more comfortable working independently and doing things my own way, than working in formal academia. But I don't ignore mainstream research at all.

[1]: https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...

Yes John Carmack is working on it. My understanding is that it's mostly (all?) by himself, with $20M, anticipating ~60% chance by 2030, and by applying old methods that didn't work with ~1MHz computers but that might work with modern desktop computers. Personally I think it's clear that large language models will scale to get human level language abilities, when you look at the curve of capabilities they have been unlocking, and if you believe that chip efficiency and model training expertise and capital expenditure will all increase.

My understanding is that it's mostly (all?) by himself, with $20M, anticipating ~60% chance by 2030, and by applying old methods that didn't work with ~1MHz computers but that might work with modern desktop computers.

A lot of my own thinking mirrors some similar ideas. That is to say, the idea that some older AI approaches that "didn't work" Back In The Day may more accurately be said to have worked, but "didn't scale." Given the massive advances in hardware, computing architectures, parallel computing, etc. since then, I think there may be some value in revisiting some of this stuff and examining it under a more modern light.

Not touching AGI until I can at least master natural language tasks specific to a narrow domain.

ChatGPT was a warning shot for me. Right now, I am [taking a break from] working on fine-tuning davinci.

Chat bot training patterns really clicked for me. I can now see how I can shard out the task of prompt/completion building to the rest of the team, and then maintain a library of tunes & models that are refreshed on some regular basis.

AGI is going to be a completely different ball game. One step at a time for me.

I'm in this supermassive blackhole and don't expect to hit rock bottom anytime soon ;-} on the triangle of personality disorder I'm closer to SB and SJ than the naïveté of SW.

John C and Elon M have intuitions and brainscar tissue I respect as much as the CEO of DeepMind.

Radio deep knowhow is being lost with the passing of time. What people have thought about "memory" could be the same for A.I.

I have a dream of implementing shit-talking AI on a Blockchain to be sure nobody will ever stop it.

Reply on news.ycombinator.com