Hacker News Reader

Comment by oceanplexian | Hacker News Reader

The power in ChatGPT isn't that it's a chat bot, but its ability to do semantic analysis. It's already well established that you need high quality semi-curated data + high parameter count and that at a certain critical point, these models start comprehending and understanding language. All the smart people in the room at Google, Facebook, etc are absolutely pouring resources into this I promise they know what they're doing.

We don't need yet-another-GUI. We need someone with a warehouse of GPUs to train a model with the parameter count of GPT3. Once that's done you'll have thousands of people cranking out tools with the capabilities of ChatGPT.

View on HN

InstructGPT which is a "sibling" model to ChatGPT is 1.3B parameters. https://openai.com/blog/instruction-following/

Another thread on HN (https://news.ycombinator.com/item?id=34653075) discusses a model that is less than 1B parameters and outperforms GPT-3.5. https://arxiv.org/abs/2302.00923

These models will get smaller and more efficiently use the parameters available.

The small models are usually tested on classification, question answering and extraction tasks, not on open text generation where I expect the large models still hold the reign.

Your point about needing large models in the first place is well taken.

But I still think we would want a curated collection of chat/assistant training data if we want to use that language model and train it for a chat/assistant application.

So this is a two-phase project, the first phase being training a large model (GPT), the second being using Reinforcement Learning from Human Feedback (RLHF) to train a chat application (InstructGPT/ChatGPT).

There are definitely already people working on the first part, so it's useful to have a project focusing on the second.

I’m new to this space so I am probable wrong, but it seems like BLOOM is in line with a lot of what you outlined: https://huggingface.co/bigscience/bloom

I would argue that it appears very good at syntactic analysis... but semantic, not so much.

> but its ability to do semantic analysis

where is that shown ?

>We need someone with a warehouse of GPUs to train a model with the parameter count of GPT3

So I'm assuming that you don't follow Rob Miles. If you do this alone you're either going to create a psychopath or something completely useless.

The GPT models have no means in themselves of understanding correctness or right/wrong answers. All of these models require training and alignment functions that are typically provided by human input judging the output of the model. And we still see where this goes wrong in ChatGPT where the bot turns into a 'Yes Man' because it's aligned with giving an answer rather than saying I don't know even when it's confidence in the answer is low.

Computerphile did a video on this in the last few days on this subject. https://www.youtube.com/watch?v=viJt_DXTfwA

It's a robot, it's supposed to do what I say, not judge the moral and ethical implications of it, that's my job.

I think it's about time we had a "Stallman fights the printer company" moment here. My Android phone often tries to overrule me, Windows 10 does the same, not to mention OSX. Even the Ubuntu installer outright won't let you set a password it doesn't like (but passwd doesn't care). My device should do exactly what I tell it to, if that's possible. It's fine to give a warning or a "I know what I'm doing checkbox", but I'm not using a computer to get it's opinion on ethics or security or legality or whatever its justification is. It's a tool, not a person.

"I know what I am doing, I accept unlimited liability"

There are two particular issues we need to address first. One is holding companies criminally and civilly reliable for the things they create. We kind of do this at a regulatory level, and we have some measure of suing companies that cause problems, but really they get away with a lot. Second is personal criminal and civil liability for management of 'your' objects. The libertarian minded love the idea of shirking social liability, and then start crying when bears become a problem (see Hongoltz-Hetlings book). And even then it's still not difficult for an individual to cause damages far in excess of their ability to remediate them.

There are no shortage of tools that are restricted in one way or another.

No, it is not a robot. The models that we are developing are closer to a genie. That is we make a wish to it and we hope and pray it interprets our wish correctly. If you're looking at this like a math problem where you want the answer 1+1 you use a calculator, because that is not what is occurring here. The 'robots' alignment will highly depend on the quality of training you give it, not the quality of the information it receives. And as we are learning with ChatGPT there are far more ways to create an unaligned model with surprising gotchas then there are ways to train a model that behaves in alignment with human expectations of an intelligent actor.

In addition the use of the word robot signifies embodyment. That is an object with a physical quantity capable of interacting with the world. You better be damned sure of your models capabilities before you end up being held criminally liable for its actions. And this will happen, there are no shortage of people here on HN alone looking to embody intelligence in physically interactive devices.

> It's already well established that you need high quality semi-curated data + high parameter count and that at a certain critical point, these models start comprehending and understanding language

I’m not sure what you mean by “understanding”.

Likely something like being able to explain the meaning, intent, and information contained in a statement?

The academic way of verifying if someone "understands" something is to ask them to explain it.

Does someone only understand English by being able to explain the language? Can someone understand English and not know any of the grammatical rules? Can someone understand English without being able to read and write?

If you ask someone to pass you the salt, and they pass you the salt, do they not understand some English? Does everyone understand all English?

Well there seem to be three dictionary definitions:

- perceive the intended meaning of words, a language, or a speaker (e.g. "he didn't understand a word I said")

- interpret or view (something) in a particular way (e.g. "I understand you're at art school")

- be sympathetically or knowledgeably aware of the character or nature of (e.g. "Picasso understood colour")

I suppose I meant the 3rd one, but it's not so different from the 1st one in concept, since they both mean some kind of mastery of being able to give or receive information. The second one isn't all that relevant.

So only someone who has a mastery of English can be said to understand English? Does someone who speaks only a little bit of English not understand some English? Does someone need to “understand color” like Picasso in order to say they understand the difference between red and yellow?

Why did we need the dictionary definitions? Do we not already both understand what we mean by the word?

Isn’t asking someone to pass the small blue box and then experiencing them pass you that small blue box show that they perceived the intended meaning of the words?

https://en.m.wikipedia.org/wiki/Use_theory_of_meaning

> Does someone who speaks only a little bit of English not understand some English?

I mean yeah, sure? It's not a binary thing. Hardly anyone understands anything fully. But putting "sorta" before every "understand" gets old quick.

> Isn’t asking someone to pass the small blue box and then experiencing them pass you that small blue box show that they perceived the intended meaning of the words?

You can teach a dog to fetch something particular. The utility of that is quiet limited.

I mean if I memorize an explanation and recite it to you, do I actually understand it? Your evaluation function needs to determine if they just wrote memorize stuff.

Explanation by analogy seems more interesting to me as now you have to know two different concepts and how the ideas in them can connect in ways that may be not be contained in the dataset the model is trained on.

There was an interesting post where someone asked ChatGPT to make up a song/poem as if written by Eminem about the how an internal combustion engine works, and ChatGPT returns a pretty faithful rendition of just that. The model seems to 'know' who Eminem is, how their lyrics work in general, and the fundamental concepts of an engine.

I think a lot of ink has already been spilled on this topic, for example under the heading of "The Chinese Room"

https://en.wikipedia.org/wiki/Chinese_room

> The question Searle wants to answer is this: does the machine literally "understand" Chinese? Or is it merely simulating the ability to understand Chinese? Searle calls the first position "strong AI" and the latter "weak AI".

> Therefore, he argues, it follows that the computer would not be able to understand the conversation either.

The problem with this is that there is no practical difference between a strong and weak AI. Hell, even for humans you could be the only person alive that's not a mindless automaton. There is no way to test for it. And just as well the same way a bunch of transistors don't understand anything a bunch of neurons don't either.

Funniest thing about human inteligence is how it stems from our "good reason generator" that makes up random convincing reasons for doing actions we're already doing, so we could convince others to do what we say. Eventually we deluded ourselves enough to believe that those reasons came before the subconscious actions.

Such a self-deluding system is mostly dead weight for AI, as as long as the system does or outputs what's needed there is no functional difference. Does that make it smart or dumb? Are viruses alive? Arbitrary lines are arbitrary.

You could have written this exact same post, and been wrong, about text2img until Stable Diffusion came along.

Isn't OP's point that we need a game-changing open source model before any of the UI projects will be useful at all? Doesn't Stable Diffusion prove that point?

How? Stable Diffusion v1 uses, for example, the off the shelf CLIP model. The hard part is getting the dataset and something that’s functional, and then the community takes over and optimizes like hell to make it way smaller and faster at lightning speed.

The same will probably happen here. Set up the tools. Get the dataset. Sew it together into something functional with standard building blocks. Let the community do its thing.

Reply on news.ycombinator.com