2
$\begingroup$

I have had a somewhat hard time trying to understand how ChatGPT can "solve" some tasks that cannot be entirely cast as language-model-based rephrasing of textual subsets of the internet directed by a textual query.

For example, ChatGPT seems to be able to do calculations, even if not always appropriately used, execute code, extract literature references from input text, etc.

It seems to me that it is not simply one huge neural network, but has access to some additional programs like a calculator or some interpreters for programming languages, etc. So it would be more like a human-machine interface.

To which extend is my interpretation correct?

$\endgroup$
5
  • 1
    $\begingroup$ It's incorrect. Check Wikipedia: en.wikipedia.org/wiki/ChatGPT, en.wikipedia.org/wiki/GPT-3 $\endgroup$
    – Dmitry
    Dec 19, 2022 at 12:16
  • $\begingroup$ I had read the Wikipedia article and I am not impressed with it. Even this very limited discussion does not seem to rule out sub-programs, e.g. " ChatGPT is built on top of OpenAI's GPT-3.5 family of large language models, and is fine-tuned with both supervised and reinforcement learning techniques." Some people have claimed that it can run programming languages, which I am not sure is correct, but not sure this article provides any proof countering that. $\endgroup$ Dec 19, 2022 at 13:12
  • $\begingroup$ Again, I do not know the answer to the question, how would you explain for example, this: pad.gwdg.de/s/LWUjPOi1F $\endgroup$ Dec 19, 2022 at 13:20
  • 1
    $\begingroup$ Explaining it is on one hand hopeless (because who knows what the model learned), and on another hand trivial: the model "detected" keywords "bibliographic references" and "CSL-JSON format", and generated the output accordingly. It's trained on a lot of data, so it's not surprising that it learned that. Additionally: 1) it'll likely output something very different if you change the instruction a bit (while leaving it semantically equivalent), and 2) in particular, I couldn't reproduce it in chatGPT (it returned bib references, but there was no JSON) $\endgroup$
    – Dmitry
    Dec 19, 2022 at 13:44
  • 1
    $\begingroup$ BTW, one thing you can potentially do is to run some smaller language model (as I recall, GPT-2 should be fully available, although I'm concerned about the length of the input) and look at the attention. I suspect that you'll see a lot of attention to tokens corresponding to "bibliographic references", and you'll get either a lot or very little attention to "CSL-JSON format" depending on whether the model tries to return either JSON-formatted or not-formatted output. $\endgroup$
    – Dmitry
    Dec 19, 2022 at 19:57

3 Answers 3

2
$\begingroup$

It seems to me that it is not simply one huge neural network, but has access to some additional programs like a calculator

I'd add to D.W.'s answer that OpenAI's CEO explicitly mentioned today in the GPT-4 announcement video that GPT-4 isn't hooked up to a calculator.

$\endgroup$
2
$\begingroup$

ChaGPT's own words:

"The cosine of 2.325 radians is approximately equal to -0.695206961. You can use a calculator to find the value or use the cosine formula:

cos(2.325) = cos(2.3 + 0.025) = cos(2.3)cos(0.025) - sin(2.3)sin(0.025)

Using a calculator, we can find that cos(0.025) ≈ 0.999687. Also, from the previous calculation, we know that cos(2.3) ≈ -0.746630108 and sin(2.3) ≈ -0.665122709. Substituting these values in the formula above, we get:

cos(2.325) ≈ (-0.746630108)(0.999687) - (-0.665122709)(0.025) ≈ -0.745952881 - (-0.016628068) ≈ -0.695324813

Rounding this value to 10 decimal places gives us -0.695206961."

By the way, $\cos(2.325)=-0.68470850929173064765331747989949$.

Notice the gross confusion and sign error between the sine and cosine of $2.3$.

$\endgroup$
1
  • 1
    $\begingroup$ theregister.com called it "Dunning-Kruger on steroids". Very convincing if you don't watch out, yet totally clueless. $\endgroup$
    – gnasher729
    Mar 16, 2023 at 16:03
1
$\begingroup$

No, your interpretation is not correct. One way to think of GPT-3 is as a predictor: given some words, it tries to predict what words come next. Another way to think of GPT-3 is as a "bullshitter": it tries to predict some words that will superficially sound plausible. It has no deep knowledge or ability to do additional computation beyond what is inherent in that capability. It does not have access to additional programs -- it is simply one big neural network. I realize it might look like ChatGPT has a lot of capability, but that is just us getting fooled into ascribing something that's not actually there.

See https://en.wikipedia.org/wiki/GPT-3 and https://en.wikipedia.org/wiki/ChatGPT.

$\endgroup$
1
  • $\begingroup$ You mean it's AI with the I missing... $\endgroup$
    – gnasher729
    Mar 16, 2023 at 15:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.