When will there be AI superintelligence?

[andres@mastodon.hardcoredevs.com]

@evan @boz
This and also the current speed of the climate change.
Hardware costs for one (it's true that LLM companies can incur into massive amounts of debt to finance it but not forever)
Food and unstable weather after, will for sure disrupt people's life and their will to work.

[ct@app.wafrn.net]

@boz@mastodon.uno @evan@cosocial.ca

Pretty sure they mean like, in general

[y6nh@im-in.space]

@evan It's impossible to answer this without specific definitions of "AI" and "superintelligence". If AI is something that can write text and draw pictures in a way that is superficially like human output, and superintelligence means that it can do that faster than a human, then we've got it already. If not, then the answer is some time between tomorrow and never, depending on those definitions.

[beandev@social.tchncs.de]

@evan
In the context of "AI = LLM", never. In context of "AI = whatever we will find as technology" we will need decades to have the technology to simulate at least a simple brain (much simpler than a human brain).

The reason is that the neural networks that we map in silicon today make up only a fraction of a brain. Countless elements that make up our thinking, our consciousness, and our self-awareness are missing. Even our senses, which are necessary for these, can currently only be represented in a rudimentary way.

Currently, we only emulate knowledge through high speed and parallelization. But we have clear limits in this regard with ressources.

And knowledge is not intelligence. It is also only simulated.

[ct@app.wafrn.net]

@boz@mastodon.uno @evan@cosocial.ca

Turning it to something more concrete: do you think AGI is achievable through the same base approach as current-day LLMs?

[evan@cosocial.ca]

@ct @boz me? I'll answer after the end of the poll.

[boz@mastodon.uno]

@evan Yep

[owlyph@cosocial.ca]

@evan for a long time I've been strongly opposed to the possibility of AGI or superintelligence as mostly just simulation that people conflate with what humans do. I still find the arguments pretty unconvincing and usually neglecting a lot of the human experience but the past few years have made me question more the nature of what humans are doing BECAUSE of the patterns that make AI work. I'm intrigued by the regularity and depth of patterns exhibited across the outputs of human thinking.

[evan@cosocial.ca]

@draken why not? Physically impossible?

[evan@cosocial.ca]

@y6nH and yet one of the things that intelligent entities do that makes us interesting is that we can reason about vague concepts.

[draken@masto.nyc]

@evan
Because computer programs are not "intelligent." They've operated under the same Garbage In Garbage Out principle that has governed computers since Babbage and Lovelace built their engines.

[evan@cosocial.ca]

@beandev you should check your math on that argument!

It's not going to take decades to have hardware and software systems with complexity equivalent to the number of neurons (86 billion) or synapses (100 trillion) in a human brain. Frontier models have about 0.5 trillion parameters, roughly equivalent to synapses. That's about 2 OOM from humans.

I think it's very interesting to ask what the difference between true intelligence and simulated intelligence is, though.

[evan@cosocial.ca]

@draken only God can make intelligence, I guess?

[draken@masto.nyc]

@evan
1) Which God?
2) False dichotomy.

[beandev@social.tchncs.de]

@evan
Thanks to the hint about the math. I completely forgot it .

It would be nice to go deeper in the science behind neuronal networks, simulations and models and what that means. If you count transistors and compare it to the number of neurons, your math is completely wrong. Additionally you need synaptic connections. All the current simulations are based on reduced and simplified models. Typically only the electrical signal way, binary switches, heavily reduced synaptic connections, and spike signal population (to reduce energy, parallelization, and computation power). We are far away.

However, a nice read about the mathematics behind it, here a nice article:
https://www.golem.de/news/maschinentraeume-1-ki-und-der-mythos-der-emergenz-2606-209312.html

It's in German, but you can translate it. Well worth reading.

[davep@infosec.exchange]

@evan How about never?

[hecate@merveilles.town]

@evan not in our lifetime

[simonzerafa@infosec.exchange]

@evan

No

[frederic@gts.raspi0.blog]

@evan
It's already there. GPT 6 Astra is smarter than any human.

[malte@radikal.social]

@evan I will go with something a little bit arbitrary, but fun: the early 1970s with the invention of the first pocket-sized electronic calculators. These AI had extraordinary intelligence and could do very difficult calculations at lightning speed. Almost whatever multiplication or division with many decimals you threw at it, these AI would give you the answer in an instant. And the answer was always right.