Wolfram Alpha and ChatGPT • Hacker News Board

Re: Wolfram Alpha and ChatGPT

#271

Post by tombert » Wed, Jan 11, 2023, 3:33 AM UTC

IshKebab wrote: post #203

alfalfasprout wrote: post #44

Earlier quoted context omitted.

It's quickly apparent that the people impressed with ChatGPT's code are generally solving pretty toy problems and want an alternative to SO for "how to do X".

Well yes because that's about the limit of what it is capable of at the moment, but that limit is very impressive ! I don't know why so many people are unimpressed with it because it isn't some kind of all knowing hyperintelligent god. This guy is literally unimpressed with it because it can't solve Fermat's last theorem! Wtf?

I said it was “neat”!

It’s impressive tech, but giving a false-but-believable proof is substantially worse than saying “I don’t know how to do that”. If it had come back with “I don’t know how to give a proof of Fermat’s Last Theorem in Isabelle”, I wouldn’t have been unimpressed; there are plenty of very cool things it can do. I just think that it being confidently incorrect is a bad thing.

EDIT:

I should point out, it also was wrong with substantially simpler proofs, like the proof of infinite primes.

Re: Wolfram Alpha and ChatGPT

#272

Post by wvoch235 » Wed, Jan 11, 2023, 3:38 AM UTC

thomastjeffery wrote: post #202

The one thing I want everyone to understand about ChatGPT: ChatGPT interfaces with semantics , and not logic . -- That means that any emergent behavior that appears logically sound is only an artifact of the logical soundness of its training data. It can only echo reason. The trouble is, it can't choose which reason to echo! The entire purpose of ChatGPT is to disambiguate, but it will always do so by choosing the mo…

I would wager that it will eventually be able to do math and logic, and may already be able to with the right prompting. It can follow instructions quite well, and multi step problems can be solved by breaking the problem into small digestible steps, and having it write out each step, just the same as humans do. If it can't do this yet, it's likely just that it doesn't have enough parameters yet. But I don't see why this couldn't be achieved with fine tuning.

But ultimately I think it's just too computationally heavy to do math, remember obscure facts, or track rapidly changing data, with an LLM, and it's far more efficient to pass it off to a piece of specialized software, just the same as humans do.

Re: Wolfram Alpha and ChatGPT

#273

Post by hugs_vs_toph » Wed, Jan 11, 2023, 3:43 AM UTC

sinuhe69 wrote: post #248

jackmott42 wrote: post #207

It is a shame that Mr. Wolfram cannot write about things without making it 75% about himself. I once bought a book he wrote about great scientists, each chapter about a different scientist. I thought "This guy's stuff is usually so self promotional it is kind of gross, but this will be fun to see his take on these other people". The book was still about him. Amazing.

My thought, too: what a nice written promotional text for Wolfram Alpha and himself! But it appeared on his blog site, so it’s ok.

Dude. For some reason I stopped reading this at the "I, myself" and was like WTF!?

Happy I'm not alone LOL

Re: Wolfram Alpha and ChatGPT

#274

Post by not2b » Wed, Jan 11, 2023, 3:48 AM UTC

jeegsy wrote: post #263

krisoft wrote: post #197

Earlier quoted context omitted.

> But how do you use that to cause actual harm? It is very easy to use it to harm OpenAI, and that is enough for OpenAI to wish to constrain it. You cause it to write something objectionable and then publish it to reap all the attention and ad-revenue which goes with people being outraged. Best if you manage to provoke into outrage inducing output without your prompt itself being outrageous, but not strictly necessar…

> There was that microsoft chatbot which “turned nazi”.[1] They then had to turn it off. At the time, I thought it was incredible that Microsoft was held responsible for this! so much so that they had to turn it off because of all the blame. It was really weird and it made no sense. Here we are again.

The problem was that it was made useless. If you ask random people to teach your chatbot, you are at the mercy of those random people. Sure, MS could have left it on, and no one would use it except for two groups: actual neoNazis, and people who think a Nazi chatbot is funny. The mistake MS made was they forgot that people suck, or, at least, enough people suck that what they did was a bad idea.

Re: Wolfram Alpha and ChatGPT

#275

Post by pdonis » Wed, Jan 11, 2023, 4:17 AM UTC

jameshart wrote: post #249

dragonwriter wrote: post #192

Earlier quoted context omitted.

> To do things like do accurate math, you need a different kind of model, one that is based on having actual facts about the world, generated by a process that is semantically linked to the world. Or you just need a model that can recognize math, and then pass it to a system that can do math. Math is actually something traditional, non-AI systems are very good at doing (it is the raison d’être of traditional computin…

I've shared this link before, showing Riley Goodside's artful prompt-crafting, but it's again relevant to this topic: https://twitter.com/goodside/status/1581805503897735168?s=20... GPT-3 is perfectly capable of recognizing what kinds of things it will be bad at, and can be encouraged to generate machine-executable queries to fill in that gap. (note this is based on prompting GPT-3, not chat GPT, but the principles a…

> GPT-3 is perfectly capable of recognizing what kinds of things it will be bad at, and can be encouraged to generate machine-executable queries to fill in that gap.

By what seems to me to be the obvious choice for a definition of "bad at", namely "not answering queries based on an actual semantically connected world model", GPT-3 is bad at everything. And an obvious example of an endpoint of your perfectly reasonable suggestion to have it pass on queries to solution machines that are based on actual world models, is...Wolfram Alpha.

Re: Wolfram Alpha and ChatGPT

#276

Post by shagie » Wed, Jan 11, 2023, 5:10 AM UTC

pdonis wrote: post #275

jameshart wrote: post #249

Earlier quoted context omitted.

I've shared this link before, showing Riley Goodside's artful prompt-crafting, but it's again relevant to this topic: https://twitter.com/goodside/status/1581805503897735168?s=20... GPT-3 is perfectly capable of recognizing what kinds of things it will be bad at, and can be encouraged to generate machine-executable queries to fill in that gap. (note this is based on prompting GPT-3, not chat GPT, but the principles a…

> GPT-3 is perfectly capable of recognizing what kinds of things it will be bad at, and can be encouraged to generate machine-executable queries to fill in that gap. By what seems to me to be the obvious choice for a definition of "bad at", namely "not answering queries based on an actual semantically connected world model", GPT-3 is bad at everything. And an obvious example of an endpoint of your perfectly reasonabl…

The thing is that it is good at classifying things.

You could then use it to classify things that it is bad at and use that information (as part of a larger whole) to dispatch a query to the knowledge system that can return back the proper (current) information, and report on that.

    The following is a list of questions.  Identify the category they belong to as one of {Current Events}, {General knowlege}, {Unit conversion}, {Math}:

    1. How many feet in a mile?
    2. What is the square root of 541696?
    3. Who is the Speaker of the House?
    4. How many turkeys in Turkey?

to which it responds:

    1. Unit conversion
    2. Math
    3. Current Events
    4. General Knowledge

The supervisor system (for lack of a better word) would then dispatch the questions to different systems that it is coded to be able to either further classify the question or provide the proper answer.

Re: Wolfram Alpha and ChatGPT

#277

Post by thomastjeffery » Wed, Jan 11, 2023, 5:23 AM UTC

wvoch235 wrote: post #272

thomastjeffery wrote: post #202

The one thing I want everyone to understand about ChatGPT: ChatGPT interfaces with semantics , and not logic . -- That means that any emergent behavior that appears logically sound is only an artifact of the logical soundness of its training data. It can only echo reason. The trouble is, it can't choose which reason to echo! The entire purpose of ChatGPT is to disambiguate, but it will always do so by choosing the mo…

I would wager that it will eventually be able to do math and logic, and may already be able to with the right prompting. It can follow instructions quite well, and multi step problems can be solved by breaking the problem into small digestible steps, and having it write out each step, just the same as humans do. If it can't do this yet, it's likely just that it doesn't have enough parameters yet. But I don't see why…

An infinite amount of "fine tuning" would solve any problem, yes...

...but what we just described isn't a language processor anymore: it's just a database. There is no need to use a neural network if you already have a known response to every possible input!

The entire point of ChatGPT in the first place is to generate a response that doesn't exist in the training data. If all we needed was to look up the answer we wrote down earlier, we already have SQL.

What we are talking about here is being able to provide a logically correct response without just pulling it out of a lookup table. Either you can do that, or you can't, and ChatGPT can't.

Re: Wolfram Alpha and ChatGPT

#278

Post by nojvek » Wed, Jan 11, 2023, 5:33 AM UTC

In terms of LLMs what we're seeing in newer research are following trends:

1. Scaling up LLMs only work if you scale up the data.

2. Chain of Thought prompting helps it improve accuracy. Teaching it how to solve similar problem in steps and then showing it how to answer full problem.

3. LLMs are great at translation. e.g translating to code / sql. Interfacing LLM to a knowledgebase / python repl / physics engine also improves it's accuracy.

I'd have to cite a bunch of papers on arxiv if anyone is interested.

Stephen Wolfram is on spot that marrying LLMs to symbolic computation is the holy grail.

E.g the avatar generators are able to tap into latent nodes for a certain face/body and use that to generate other images.

I'm sure as neural architectures evolve, we'll see more symbolic computation in neural networks aka neurosymbolic AI. The hard part is having computers abstract and figure out the symbolic representations by themselves instead of an army of humans carefully building databases and code.

Re: Wolfram Alpha and ChatGPT

#279

Post by gptgpp » Wed, Jan 11, 2023, 5:37 AM UTC

sharkster711 wrote: post #256

gptgpp wrote: post #173

When I ask it the same questions my results are different: Q: What Is the distance from Chicago to Tokyo? A: The distance from Chicago, United States to Tokyo, Japan is approximately 8,113 miles (13,055 kilometers) if you were to fly in a straight line. It takes approximately 15 hours of flight to reach from Chicago to Tokyo. Which is fun because it is also wrong, but gives a different incorrect value. Q: What is 3 t…

> Interestingly, it now correctly outputs Honduras for the second largest country in South America Did you mean Central America?

No, I'm just terrible at geography (still better than ChatGPT somehow though)

Re: Wolfram Alpha and ChatGPT

#280

Post by thomastjeffery » Wed, Jan 11, 2023, 5:45 AM UTC

Winsaucerer wrote: post #259

thomastjeffery wrote: post #234

Earlier quoted context omitted.

You would be trading the problem for another instance of the same problem. When you ask ChatGPT to construct a mathematical question, it will do so the same way it does everything else: by semantic popularity . And that is the problem we are trying to avoid. The semantically popular guess might be logically sound , but it might not . It's a gamble no matter when or where it is done. -- All it takes is what I call a "…

> It's just going to be whatever arbitrary text that already exists in the training dataset is closest to the semantic phrasing of the question. My understanding of machine learning in general is that this is not how it works, and rather it uses a neural network for a lot of what it does (which isn't merely picking the closest arbitrary text in its training set), though I don't know details about NLP specifically. I…

> but the fact is that it didn't offer these options, and instead it offered the correct one

...because you gave it a "correct" prompt; or in other words, you gave it a prompt that ChatGPT can respond "successfully" to. But if you keep trying, you will surely find yourself with "failed" responses. And ChatGPT has no way of knowing the difference. That's my point.

> Is the other "completely valid and possible output" you provided actually likely?

We're splitting hairs here. I didn't go and test a bunch of prompts to find an example failure. You are welcome to do that if you want. It's not very hard, especially when you start adding variables. This is well-known behavior that ChatGPT itself warns its users about.

> > ChatGPT cannot do anything at all without making a guess, because "guess" is everything that ChatGPT is

> Is this different to what humans do?

Yes!

We don't just spit out the first thing that comes to mind, unless that's the task we are intending to do - and anyone who has played music can tell you just how difficult that task is.

Most of the time, we conceptualize. We comprehend the meaning of language - sure, it takes some guesswork along the way, but that's not the main event - then we construct a new idea, then we express it.

We don't just want accurate answers. We want conceptualization. That's the feature we are all dreaming about with Natural Language Processors. If we can teach a computer to conceptualize language, then we can get it to interface with any arbitrary human knowledge!

But ChatGPT doesn't interface. It parrots; and that's an insult to the birds themselves - unlike ChatGPT, they have an intelligence of their own.

The only way ChatGPT could give you a valid response about any arbitrary concept is to write every thing from every angle about every concept ever, and feed it all to ChatGPT as training data. That's exactly the work we want to avoid!