zlacker

>>wavela+(OP)
Interesting, both Karpathy and Sutskever are gone from OpenAI now. Looks like it is now the Sam Altman and Greg Brockman show.

I have to admit, of the four, Karpathy and Sutskever were the two I was most impressed with. I hope he goes on to do something great.

>>zoogen+Ix
Top 6 science guys are long gone. Open AI is run by marketing, business, software and productization people.

When the next wave of new deep learning innovations sweeps the world, Microsoft eats whats left of them. They make lots of money, but don't have future unless they replace what they lost.

>>nabla9+pH
If we look at history of innovation and invention it’s very typical the original discovery and final productization are done by different people. For many reasons, but a lot of them are universal I would say.

E.g. Oppenheimer’s team created the bomb, then following experts finetuned the subsequent weapon systems and payload designs. Etc.

>>fsloth+O21
Except OpenAI hasn’t yet finished discovery on its true goal: AGI. I wonder if they risk plateauing at a local maximum.

>>fprog+I51
I'm genuinely curious: what do you expect an "AGI" system to be able to do that we can't do with today's technology?

>>Zambyt+hc1
An AGI could replace human experts at tasks that doesn't require physical embodiment, like diagnosing patients, drafting contracts, doing your taxes etc. If you still do those manually and not just offload all of it to ChatGPT then you would greatly benefit from a real AGI that could do those tasks on their own.

And no, using ChatGPT like you use a search engine isn't ChatGPT solving your problem, that is you solving your problem. ChatGPT solving your problem would mean it drives you, not you driving it like it works today. When I hired people to help me do taxes they told me what papers they needed and then they did my taxes correctly without me having to look it through and correct them, an AGI would work like that for most tasks, it means you no longer need to think or learn to solve problems since the AGI solves them for you.

>>Jensso+bk1
> An AGI could replace human experts at tasks that doesn't require physical embodiment, like diagnosing patients, drafting contracts, doing your taxes etc.

How come the goal posts for AGI are always the best of what people can do?

I can't diagnose anyone, yet I have GI.

Reminds me of:

> Will Smith: Can a robot write a symphony? Can a robot take a blank canvas and turn it into a masterpiece?

> I Robot: Can you?

>>xdenni+OA1
> How come the goal posts for AGI are always the best of what people can do?

Not the best, I just want it to be able to do what average professionals can do because average humans can become average professionals in most fields.

> I can't diagnose anyone, yet I have GI.

You can learn to, an AGI system should be able to learn to as well. And since we can copy AGI learning it means that if it hasn't learned to diagnose people yet then it probably isn't an AGI, because an AGI should be able to learn that without humans changing its code and once it learned it once we copy it forever and now the entire AGI knows how to do it.

So, the AGI should be able to do all the things you could do if we include all versions of you that learned different fields. If the AGI can't do that then you are more intelligent than it in those areas, even if the singular you isn't better at those things than it is.

For these reasons it makes more sense to compare an AGI to humanity rather than individual humans, because for an AGI there is no such thing as "individuals", at least not the way we make AI today.

>>Jensso+pQ1
People with severe Alzheimer's cannot learn, but still have general intelligence.

>>Heatra+F62
If they can't learn then they don't have general intelligence, without learning there are many problems you wont be able to solve that average (or even very dumb) people can solve.

Learning is a core part to general intelligence, as general intelligence implies you can learn about new problems so you can solve those. Take away that and you are no longer a general problem solver.

>>Jensso+V72
That's a really good point. I want to define what I think of intelligence as being so we are on the same page: it is the combination of knowledge and reason. An example of a system with high knowledge amd low reason is Wikipedia. An example of a system with high reason and low knowledge is a scientific calculator. A highly intelligent system exhibits aspects of both.

A rule based expert intelligence system can be highly intelligent, but it is not general, and maybe no arrangement of rules could make one that is general. A general intelligence system must be able to learn and adapt to foreign problems, parameters, and goals dynamically.

>>Zambyt+kN2
Yes, I think that makes sense, you can be intelligent without being generally intelligent. For some definitions the person with Alzheimer can be more intelligent than someone without, but the person without is more general intelligent thanks to ability to learn.

The classical example of a general intelligent task is to get the rules for a new game and then play it adequately, there are AI contests for that. That is easy for humans to do, games are enjoyed even by dumb people, but we have yet to make an AI that can play arbitrary games as well as even dumb humans.

Note that LLMs are more general than previous AI's thanks to in context learning, so we are making progress, but still far from as general as humans are.