General > General Technical Chat

Why does OpenAI ChatGPT, Possibly Want to disagree/annoy and change my eating...

<< < (5/22) > >>

MK14:

--- Quote from: BrianHG on December 28, 2022, 08:53:41 am ---I am of the opinion that what I see is interesting, but still too primitive to be of serious use other than experimental.  I will wait 10 more years before I begin to play with AI chat bots and just be amazed at what's been accomplished instead or fooling around today and seeing the miniature growing steps.

--- End quote ---

I think it is already powerful enough, to have many uses.  But it is a bit like an electronics simulation package, or PCB design program.  It needs time and experience, to learn how to use such tools, effectively and usefully.  At least for some applicable parts, of some tasks, but with human help to control and tidy up the results, afterwards.

We already have google (but it's hindered by google wanting to support advertisers and their webpages, rather than be as helpful and useful, as it could be) and many books available etc.  But I think in some cases, this (AI) can provide extra functionality, which can speed up some peoples workflow/hobby/activities.

As you hint/say, we are sort of at the (AI equivalent) level of valves (vacuum tubes), where they can only do limited things, keep on burning out at the wrong time, are big, make too much heat and cost too much.  But later on we will have the equivalent of the transistor phase and eventually the integrated circuit phases.  Which will bring on big improvements.

I wouldn't mind sitting in the back of a car, on my own, as the car quietly drives down the road, on a very long journey, while I rest.  But such self-driving cars, are not yet ready, for such activities.  But will probably come along, as time progresses.

I wouldn't mind an AI robot, in approximate human like form.  Which can be told to do various activities.  Much like a butler, cook, personal assistant, care worker, driver, shopper, cleaner, house keeper, etc, can do for us.  As long as you are financially able to afford such luxuries, which the vast bulk of us can't, and many who can, don't bother to, for various reasons.

Maybe powerful, semi-total immersion VR units, which can present any situation (within reason), that you describe to the AI, to create for you.  Maybe go back to the 1950s/60s, and play around with a typical mainframe computer from that era, as if it was real-life.  Then choose any car, to drive around in, at any part of the world, created by the AI, in the VR unit.

tom66:
The problem with GPT's will be the scaling.  Currently the model requires close to 1TB of VRAM to execute.  And GPT-4 will be 100x larger.  The cost of running such a model will be considerable, even with dedicated hardware running NN's rather than running it on graphics cards.  Sure is interesting to watch though!

MK14:

--- Quote from: tom66 on December 28, 2022, 05:38:51 pm ---The problem with GPT's will be the scaling.  Currently the model requires close to 1TB of VRAM to execute.  And GPT-4 will be 100x larger.  The cost of running such a model will be considerable, even with dedicated hardware running NN's rather than running it on graphics cards.  Sure is interesting to watch though!

--- End quote ---

There was a similar problem with chess programs.  Earlier computers were too slow and expensive, to allow most people to play with reasonably well playing chess software.  E.g. a (wild estimate of cost) $25,000,000 Cray Supercomputer, could play a reasonable game of chess, in around 1976+.  But for most people, outside of researchers in some Universities, that wasn't an option.

Since that was way too expensive.  So, as Moore's law progressed, chess computers and its software, became cheaper, widely available and more powerful, as time progressed.
These days most desktop PCs, can beat even the world (human) chess champion, using free software, such as stockfish.

So I would imagine, a need of extreme computer power (and storage space, ram etc).  Would only be a delaying factor, rather than a long term show stopper.

But I suspect, the (apparently) freely available free ChatGPT services, that seem to be available now.  Will migrate to needing significant subscriptions, even just to try them out (except perhaps free trials and/or a very low end, low capacity, free tier).

I know there are already subscription services for things like that, including the ChatGPT company (openAI).  But they let people use the 'free tier', for trying it out.  That might either stop, or be significantly curtailed, in the future.  Maybe?

I think it looks like, AI is reaching critical mass, and will be able to not only do useful things, but hence bring in more money for the companies that offer it.  Which should be able to fund the growth in computing power, needed for AI.

E.g. Legally (genuinely working and safe) usable self-driving car AI, is something, some, would happily pay for.  Not to mention its capabilities (once working and legal), to deliver things to people, both quicker and with less cost.

tom66:
I was speaking to a friend of mine about chess just today, in fact.   He disagrees with your hypothesis:  in fact, the performance of chess is really down to new software solutions.  Take a 20 year old computer and run modern chess solvers on it, and it will beat a brand new computer running 20 year old software on it.    Kasparov was beaten by Deep Blue by brute force, more than anything.  Nowadays, Kasparov could be beaten by a smartphone.

The challenge for GPT models will be sparsity.  The engineers have to figure out how to restructure the model so that the least frequently accessed weights/portions of the network are stored on slower forms of storage, which is cheap, without aversely compromising performance.  That will be more software/architecture than it will be hardware.   Moore's Law will limit the capacity of these networks eventually, who knows if GPT-4 will only be practical if it is $1 per word, if it requires a datacenter worth of hardware to run.

AI is not yet at critical mass, but it's at that early stage where we're going "wow, that's neat".   I think it's still a decade away from being a part of everyday life, at least in the form of things like ChatGPT.

MK14:

--- Quote from: tom66 on December 28, 2022, 06:36:13 pm ---I was speaking to a friend of mine about chess just today, in fact.   He disagrees with your hypothesis:  in fact, the performance of chess is really down to new software solutions.  Take a 20 year old computer and run modern chess solvers on it, and it will beat a brand new computer running 20 year old software on it.    Kasparov was beaten by Deep Blue by brute force, more than anything.  Nowadays, Kasparov could be beaten by a smartphone.

--- End quote ---

I suppose a better way of saying, what I originally wrote.  Would have been, that at any given chess playing level capabilities.  It needs a suitable combination of computer hardware performance power AND software capabilities (and maybe other things, such as chess databases etc).

I.e. A 1 instruction per second, electromechanical relay based, computer, from a very long time ago.  Perhaps could be programmed to play chess, or at least some type of subset of the full game.  Especially if given long periods of time, to calculate each move.  But is not going to play especially well, even against a complete beginner of chess, even with the best possible software algorithm and its associated databases.  Give or take a fair bit, as sometimes people take up such challenges, and prove people wrong, about what can and can't be done, in practice.

You're both right.  There is a significant complication, because chess playing software, has improved in leaps and bounds over the years.  So that the best chess software available today, running on even somewhat slow computer hardware, compared to the latest computers available today.  Would still be able to play, really well.
So I can believe, a twenty year old computer with the latest chess software, beating a brand new computer, but with a twenty year old, chess program.

Navigation

[0] Message Index

[#] Next page

[*] Previous page

There was an error while thanking
Thanking...
Go to full version
Powered by SMFPacks Advanced Attachments Uploader Mod