I think that's the question of this thread. Who can buy and use solver real time is not customer, they are afraid that every bum could ask bot and get answer : ) Something like chatGPT = human interface of solver.

You could fine-tune GPT-3 or one of the open source models to create a chat-based interface for an external solver. I don't know if anybody will do it, the market for something like this doesn't seem to be large.

IMO the biggest contribution to ChatGPT or equivalent will be in writing code that solves problems to nth degree, with the most efficient code that could possibly be written. For example solving intractable math problems using advanced numerical methods/bayesian inference/calculus of variations/uber large matrices.

There exist programming techniques (e.g. programs writing programs) that can do what ChatGPT can do, but only in narrowly defined use cases. ChatGPT will likely open that up to be so abstract anybody can program at world class level.

Are you talking about some possible future versions? Because it's not close to doing this.

I understand ChatGPT can indeed write code. Perhaps it has not been released to the general public.

It should be straightforward to assume that these tools could program at a far more effective level than humans. Imagine a team of 1,000 top software developers working perfectly in harmony focused on a solution to a single problem. Digital domain is far more in ChatGPT's wheelhouse than spoken/written language.

It can write code but at best it improves productivity by about 20%. It's good at things like rewriting existing code, writing boilerplate code, adding documentation, showing you how to use a new framework. There is also Github Copliot, also from OpenAI. It's similar. They both can introduce bugs, hallucinate non-existent functions but they're impressive anyway. They'll get better but not close to replacing programmers yet.

Essentially programs are written so humans can understand them, which is a limitation. Think what XML was when it was introduced. With ChatGPT it can write code that nobody understands. Instead, it can focus on performance using massively parallel processing which can be quite unwieldy to program. The code would be suited exactly to purpose, instead of readability or maintenance.

If its not there I don't think it is too far off.

Programs writing programs were introduced some time ago, but their inefficiency and unreliability were problematic. In certain use cases (e.g. high performance computing) they can be effective.

ChatGPT is far off. It's terrible at logic and math and it often makes up things. There is no clear path to AGI but I think we'll get there.

Yeah I must have overestimated what ChatGPT's capabilities are in coding.

The key will be in abstraction/generalizing problems for AI consumption. When that happens anybody will be able to code at the highest level.

It is not a learning exercise, just an organizational exercise. The solutions already exist, just not organized in a way that can be easily consumed. Human intelligence will still be required to define the problem and desired outcome(s).

What ChatGPT doesnt understand today it will tomorrow and be the expert at it sometime later.

What humans don't understand today, 99% never will.

Just wait until ChatGPT merges with Atlas of Boston Dynamics and gain control of killer drone fleets.... better get ready for Judgment Day!!

microsoft has bought chatgpt and is going to release a search engine (bing) that works like chatgpt. i am sure they will put in a lot of work in improving this bot because theres so much at stake for them. for search engines in general the new bing is planned to be a rival to google. google has reacted and bought also some chatbot and is now improving it.
all this is very interesting. if microsoft succeeds they can be really competitive in this space and they have the resources to do this well.
as for poker, now when AI has taken big steps it's close to be so smart it can win at poker on some level. in future some AI will destroy human players. the annual poker game computer vs human was already lost first time recently so yes we are close to AI beating the games. of course nobody has access to quantum computers but even these simpler AIs are close to be very good.

Microsoft hasn't bought ChatGPT. It invested $10B in OpenAI, which makes ChatGPT. The Bing version, which can search the web and display references, is already available to those who signed up for the waiting list. Google has its own language models but has only used them internally until now. Google can probably match Bing when they release their product but OpenAI is reportedly training GPT-4, which could be superior.

AI is already better than humans in poker but the methods used are very different from those used in large language models.

thanks for the correction.

Actually it's pretty ****ing good most of the time, you just have to know what to ask.

I asked it to write a bash script that removes all old docker images but to keep the latest two newest of each image and it was spot on.

Then I asked it to write it as an Ansible playbook and again, spot on.

No, this doesn't show the ability to do logic or math. It's a very well-known weakness of LLMs to anybody in the field. It fails on very simple questions that any human can do. Like this
But there will be improvements, probably this year. For example, you can offload logic questions to a generated external Python code and execute it.

The leap from chatGPT being able to access solutions and actually being able to explain them with any level of understanding is enormous. That's definitely not happening any time soon.

Also, why would chatGPT be able to generate solutions faster than a solver even if it has access to a solver?

It cant killpoker it is only semantics as yet, it will be another year or so before it uses real data. Also its IQ is only 100; but could AI help? I think so once a program meant for training can play like a human it will provide a serves worth using. It has also been limited due to many requests for illegal and harmful info so a public AI that can really be useful cant happen, this is all space ships to Mars and E Musk colonies in the deserts of the Sahara. One thing you may not know is that they have had the technology to make interview "Chat" AI since the 80's also the langauges they use to program it varies some are much better than others. With that low of a test score it cant help me much more than a clerk, what it writes is not very deep, so perhaps it could replace a phd in about seven thousnd years.

For math, try this and watch it fail:

"Klein read 30 pages of a book on Monday and one eighth of the book on Tuesday. He completed the remaining quarter of the book on Wednesday. How many pages are there in the book?"

You got me curious...

I think you just got lucky.




Sometimes it can get it but if you try multiple times, you'll get wrong answers too.

FWIW gpt4 can already do this no problem. same with the book pages question, gets it right everytime.

poker is obviously requires huge computation so is pretty far off from a language model to solve, maybe if it could have a full solution of the specific format in its training model.

Yes, GPT-4 is definitely an improvement over GPT-3.5, I'm pretty impressed. It's still the wrong tool for interfacing with a solver or similar tasks, though.

Google just released Bard but it's worse.

OpenAI now has plugins for ChatGPT, including one from Wolfram Alpha, so math should be much better. I imagine you could develop a plugin to access a solver as well.

It is also multimodal which means it understands images/screenshots just to make it even easier to put it all together into a very comprehensive system.

That depends on the dataset of images it was trained on. I think it's unlikely it was trained on many information-dense screenshots so it probably will do poorly on them.