Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Ask HN: How does ChatGPT work?

news.ycombinator.com

Post reply on HN
Print view View original

101–110 of 185 posts

Re: Ask HN: How does ChatGPT work?

#101

Post by mjburgess »

k__ wrote: post #94

Half-OT: people are always talking about ChatGPT being AI, but is this actually the case? It frequently told me that it doesn't learn from my input, and I had the impression the unique selling point of AI was it being able to modify it's own code in response to input.

That's basically never been the case for any statistical AI system which are all subject to "catestrophic forgetting".

The weights of the system which are its memory are set by having their values determined over an entire training set. This is, effectively, a compression process which takes eg., most of the internet, and compresses it to 1TB of weights.

If you update those weights by a process which doesn't "zip everything at once" then you're overwriting weight values by data which is less representative than "everything at once".

I'd imagine in ChatGPT's case, OpenAI will review all the transcripts and find someway to do a pass over the whole set at once to make some improvement.

If it were being trained live it's whole memory would be wiped out and replaced with racist/sexist/etc. inputs very quickly, as with MS's chatbot

Re: Ask HN: How does ChatGPT work?

#102

Post by sebzim4500 »

throwaway290 wrote: post #81
chopete3 wrote: post #27

This blog explains some of the key innovations they added on top of GPT-3, especially the natural language understanding (following instructions). https://openai.com/blog/instruction-following/ In the first few paragraphs they show GPT-3 as equally dumb like all other language models that came before it and why they built instructgpt. >> Here is the summary (ChatGPT summarized) They present their approach to the prob…

To be clear there is no 'understanding' going on, and the article doesn't use that word either.

There should be a rule on HN that anyone claiming definitively whether or not AIs 'know' or 'understand' something should have to define those words rigorously.

Re: Ask HN: How does ChatGPT work?

#103

Post by amelius »

akelly wrote: post #18

The way they went from GPT-3 to ChatGPT is really quite genius. My understanding is that it's something like this: 1. Start with GPT-3, which predicts the next word in some text and is trained on all the text on the internet 2. Take thousands of prompts, generate several responses for each of them, and have human reviewers rank the responses for each prompt from best to worst 3. The GPT model needs a massive amount o…

That's not genius, that's called unsupervised learning and it is an entire subfield.

Re: Ask HN: How does ChatGPT work?

#104

Post by amelius »

zozbot234 wrote: post #71
undefuser wrote: post #69

Earlier quoted context omitted.

I applied the same prompts and the results get better: Me: Which number is bigger, 873593742593454623 or 873593742593554623 ChatGPT: These two numbers are equal. The first number is simply the second number with an extra digit added at the end. Since the added digit does not affect the value of the number, both numbers are equal. Me: Apply your reasoning and finding the bigger of two numbers to compare 123 and 53. Ch…

ChatGPT is a master troll.

I bet it would do great at the Chewbacca style defense in courts of law. Lawyers, your jobs are next!

Re: Ask HN: How does ChatGPT work?

#105

Post by espadrine »

jb1991 wrote: post #87

ChatGPT is trained using a combination of supervised and unsupervised learning. For supervised learning, it is trained on a large dataset of human-generated text, such as dialogue data or online conversations. This allows it to learn the structure and style of natural language. For unsupervised learning, it is trained using a language modeling objective, which involves predicting the next word in a sequence of text.…

That is somewhat inaccurate.

ChatGPT is based on the InstructGPT weights, based on the GPT-3 weights. It is roughly the same number of parameters, as far as we can tell.

The GPT-3 weights were obtained by doing unsupervised pre-training (hence, GPT: generative pre-training): maximizing the likelihood of the model predicting the next word in a large dataset of human text.

The InstructGPT weights were obtained with supervised fine-tuning (SFT) by making the model generate text, and asking a human to show a better text completion (as described in InstructGPT). Then, they also asked humans to rank multiple generated outputs, which was used as the supervised training goal of a separate reward function. That small amount of ranked data unlocked the ability to rank a much larger amount of data through reinforcement learning using proximal policy optimization (PPO): the model generates an output, the reward function rates it, and the model weights are updated to achieve a higher reward.

The ChatGPT beta weights were obtained by doing that again, but asking the humans to make the completion conversational. Since they could only pay few humans, they opened the beta to ask a wider range of people to do SFT using the feedback feature, to train the final ChatGPT weights.

So, all in all, the parameter estimation is incorrect, the order of the training is not right, the description of the purpose of the supervised learning step is wrong, the defining part of the ChatGPT training process is not mentioned (because the InstructGPT paper came in 2022, after the knowledge cut-off), the description of the difference with GPT-3 is misleading.

Re: Ask HN: How does ChatGPT work?

#107

Post by hackernewds »

anthropodie wrote: post #23
akelly wrote: post #18

The way they went from GPT-3 to ChatGPT is really quite genius. My understanding is that it's something like this: 1. Start with GPT-3, which predicts the next word in some text and is trained on all the text on the internet 2. Take thousands of prompts, generate several responses for each of them, and have human reviewers rank the responses for each prompt from best to worst 3. The GPT model needs a massive amount o…

I want to know if it will ever be possible to run this kind of AI at home once its training is complete. I dont need all the knowledge just subset that I'm interested in. Actually I'm more interested in its ability to transform things. For example I can ask it to convert docker-compose to docker run command, it can manipulate JSON, it can sort numbers in table when prompted. I'm more interested in these abilities rat…

you can do that today in the free release?

Re: Ask HN: How does ChatGPT work?

#108

Post by xvector »

hackernewds wrote: post #107
anthropodie wrote: post #23

Earlier quoted context omitted.

I want to know if it will ever be possible to run this kind of AI at home once its training is complete. I dont need all the knowledge just subset that I'm interested in. Actually I'm more interested in its ability to transform things. For example I can ask it to convert docker-compose to docker run command, it can manipulate JSON, it can sort numbers in table when prompted. I'm more interested in these abilities rat…

you can do that today in the free release?

I think he wants to self host. It sucks to have no ownership of such a powerful tool I would pay upwards of $3000 to be able to self host something like this.

Re: Ask HN: How does ChatGPT work?

#109

Post by dotancohen »

sebzim4500 wrote: post #102
throwaway290 wrote: post #81

Earlier quoted context omitted.

To be clear there is no 'understanding' going on, and the article doesn't use that word either.

There should be a rule on HN that anyone claiming definitively whether or not AIs 'know' or 'understand' something should have to define those words rigorously.

The question goes back decades - back then the question was phrased "Can computers think" and the rebuttal was "Can submarines swim?"
Post reply on HN