ChatGPT just spouts out a different wild guess when you tell it it's wrong, but doesn't learn from its mistake—not even within a single chat. Sydney just goes full psychopath.
What you're describing is a particular kind of superintelligence, recursively improving kind. Between that and general intelligence is another gap.
No, in the same way that boiling water in a kettle is not incrementally closer to inventing the fuel for a rocket ship.
You can, in hidsight do a linear connection of water boiling -> steam engine -> fuel engine -> aeronautics -> space travel
But the first time someone boiled water, if you said that got you closer to being on the moon I think you would be mad. There is a non trivial chance that all our current work in AI is functionally useless in terms of creating AGI, it might take a complete breakthrough in ML, tech not invented yet, or it might not be possible at all.
Who is to say that our current ML models are not Ether studies and utterly useless.
If we crack AGI in 10 years, maybe our current models are the fuel engine. But if they are not even close, we might be boiling water. Or perhaps be wrong all together, thats why I said your observation of us getting closer thanks to those proyects is not right, at least not yet.
Large language models might not be capable of general intelligence by themselves, but a model network that includes a LLM as part of a generalized game playing model would certainly have the potential to get there.
But we also have examples of things all around us we cannot make ourselves (e.g., a simple animal or lifeform from scratch).
If we can define what "generally intelligent" means in concrete terms, we can absolutely create it. The problem is that we don't have a good understanding/shared idea of what we're trying to achieve. We can create models that can learn to play many different games using the same weights, solve IQ/logic tests and pass the Turing test consistently, but people are going to move the goalposts and say those are just complex mashups of autocomplete and flowcharts.
To be honest I don't think AI skeptics are ever going to believe until there's a Skynet trying to exterminate them.
I rather think AGI "idealists" (maybe there is a better word) will forever hold out (any) algorithmic advances as a sign of the immanent creation of AGI proper.
But no--the state of the world in that second versus the next could not have been predicted. It just happened, just so, and it would take longer to calculate the determinism than to adjust to the new.
Notably the people closest to the fire, those advancing SOTA AI capabilities and those deepest into alignment/existential risk research, both say it is coming this decade. They predict very different outcomes (crudely summarized: utopia vs extinction), but they both predict the event horizon is on our doorstep.
You're right that we might get it wrong, but getting things wrong with a new technology really always means negative outcomes. When you're figuring out rocket science and you realize you worried about the wrong things your rocket doesn't somehow still succeed and make it to space, it blows up in a fireball. It might not blow up because of the specific thing you were paying attention to, but it still blew up.
Why should we believe you over the experts for whom this is their life's work? What are your credentials? Where is your experience? You seem to think they're unserious clowns so clearly you have a good reason.
Existing AI is more important to worry about.
This article is a bit dated but it gives a good explanation of the difference between the powerful but limited tools we have today versus a true AGI.
AI alignment people are the group who believe “responsibly” means, centrally, bringing AGI to existence as fast as possible, and, critically, under the tight, opaque, authoritarian control of a narrow group of people whose ideology is, well, exemplified by figures like Sam Bankman-Fried, and doing everything possible to advance that goal, including fostering broad acceptance of sub-AGI AI under the same tight control for other purposes without concern for adverse social impact as a means of advancing and funding progress toward AGI. They believe that the ends (AGI) inherently justifies any means taken in its pursuit.
AI ethics people are the group who believes “responsibly”, withe regard to AI, isn’t restricted to pursuit of AGI (which is maybe a nice to have, but a non-essential goal), and means openness, transparency, and resolving bias and harm issues as rapidly as possible when, and ideally before, adopting AI – including AI far short of AGI – in roles with social impact (and most critically, resolving those for roles with the greatest social impact.) They believe that AI is, to the extent it at any level is desirable, is a means to improving concrete, present, immediate conditions, not an ends to be pursued in itself.
From what I’ve seen it’s the exact opposite? They want AI development to slow down so we can get it right and make it safe. There are people who are trying to speed it up but that seems like the opposite of alignment to me
The only way that the alignment faction wants to slow things down is in that they want to maintain tight control of models and gated access to assure “proper” use. This does have an effect of slowing certain kinds of development, as broader access to build around models without centralized controls fosters some aspects of development (the “Stable Diffusion moment”) but in ideal terms, the alignment faction is about fast-but-narrowly-controlled development.
The ethics faction is more about slowing things down – though more about adoption, particularly in sensitive uses, than development – but is more oriented toward openness/transparency.
(and, y'know, being a fraudster who temporarily pulled the wool over the eyes of the regulators not just the charities he promised money to)
Also: is SBF really the central example of alignment in your head, and not, say Yudkowsky? Because Yudkowsky is my central example, and he's the exact opposite on all counts, wanting us to go slow and explicitly saying that humans are terrible at reasoning properly when they allow themselves to think "ends justify the means": https://www.lesswrong.com/posts/K9ZaZXDnL3SEmYZqB/ends-don-t...
Likewise, there are plenty of examples of misaligned non-general AI, and have been since at least the 80s when it was GOFAI and databanks rather than neutral nets that led to new legislation: https://www.legislation.gov.uk/ukpga/1984/35/enacted
For normal people who don't spend six hours every day on lesswrong, they were literally in the same camp up until about six months ago (when all the "rationalists" violently disclaimed SBF after being funded by him for a few good years.)
No, he's a particularly well-known example of the broader ideology to which virtually all of tbe alignment camp subscribes, and which centers their desires for AI.
Hey, let's be fair now, a lot of them also believe that we just need a world government that tightly controls GPU use.
Sarcasm aside, the alignment crowd definitely does not have a consensus about the desirability about creating AGI asap. Yudkowsky is clearly horrified by OpenAI.
Mostly not at all. They write papers like "Can Language Models Be Too Big?"
I say then: can people be too afraid of the unknown?
Whether either is capable of "bringing AGI to life" remains to be seen, but seems very doubtful.
AI ethics people are the group who are not capable of bringing AGI to life, but want to make sure that those who can do it responsibly or not at all.