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

Teaching ChatGPT to speak my son’s invented language

szopa.medium.com

Post reply on HN
Print view View original

141–150 of 212 posts

Re: Teaching ChatGPT to speak my son’s invented language

#141

Post by vidarh »

nearbuy wrote: post #120
gwright wrote: post #109

Earlier quoted context omitted.

I feel like your observation that this "isn't a complicated question" is leaning on an implicit assumption that ChatGPT is a general AI and not a LLM. It is just generating text based on probabilities -- it isn't "reasoning". I might go as far to say that inferences computed by a LLM are of all the same complexity but I don't really know enough about ChatGPT to be confident in that statement.

People keep repeating that LLMs are "just generating text based on probabilities". That statement doesn't mean anything. I think people who say this are imagining LLMs work something like a statistical model. Maybe it's doing a linear regression or works like a Markov chain. It's not. A single artificial neuron sort of works like that. But that's sort of like saying a single transistor is just an electronically contr…

It's worth noting that "just generating text based on probabilities" describe Markov algorithms [1], which are Turing-complete. People overestimate how much it takes to end up with something Turing-complete.[Markov algorithms only generate text with probability 100% or 0% based on whether a certain rule matches or not, so it's even simpler]

(A Markov algorithm is distinct from a Markov chain, but as far as I can tell you could emulate a Markov algorithm with a Markov chain with sufficient number of input states, transitions clamped to 0% or 100%, and allowing it to iterate over its own output; with a large enough state machine, iteration, and a mechanism to provide memory it's almost hard not to end up with a Turing machine)

[1] https://en.wikipedia.org/wiki/Markov_algorithm

Re: Teaching ChatGPT to speak my son’s invented language

#142

Post by simonh »

nearbuy wrote: post #120
gwright wrote: post #109

Earlier quoted context omitted.

I feel like your observation that this "isn't a complicated question" is leaning on an implicit assumption that ChatGPT is a general AI and not a LLM. It is just generating text based on probabilities -- it isn't "reasoning". I might go as far to say that inferences computed by a LLM are of all the same complexity but I don't really know enough about ChatGPT to be confident in that statement.

People keep repeating that LLMs are "just generating text based on probabilities". That statement doesn't mean anything. I think people who say this are imagining LLMs work something like a statistical model. Maybe it's doing a linear regression or works like a Markov chain. It's not. A single artificial neuron sort of works like that. But that's sort of like saying a single transistor is just an electronically contr…

That's a valid point in that we don't fully understand how LLMs solve some problems and using logic and rules of inference isn't excluded by the architecture, but on the other hand understanding that they are generating probabilistic token sequences is a very powerful and effective way to understanding how to engineer prompts and understand some of their failure modes. If we discard that insight, reasoning about their many failure modes and limitations becomes near impossible.

For example we often see people thinking that because an LLM can explain how to do something that therefore it knows how to do it, like arithmetic. That's because if a human can explain how to do something, we know that they can. Yet for an LLM outputting a token sequence for an explanation of something, and outputting a token sequence for solving a problem statement for that problem domain are fundamentally different tasks.

We can get round this with very clever prompt engineering to 'force' chain of reasoning behaviour, as this discussion shows, but the reason we have to do that is precisely because the cognitive architecture of these LLMs is fundamentally different from humans.

Yet these systems are clearly highly capable, and it is possible to dramatically improve their abilities with clever engineering. I think what this means is that LLMs may be incredibly powerful components or elements of systems that may become far more advanced and sophisticated AIs. However to do that engineering and build dramatically more capable systems, we need to have a clear understanding of how and why LLMs work, what their advantages and limitations are, and how to reason about and work with those features.

Re: Teaching ChatGPT to speak my son’s invented language

#143

Post by PartiallyTyped »

FPGAhacker wrote: post #115
famouswaffles wrote: post #9

In context learning is hands down the biggest breakthrough of LLMs. The flexibility the model displays without updating weights is genuinely mind blowing, bordering on absurd especially if you've trained other kinds of models before. See here - https://imgur.com/a/w3DAYOi from the paper - https://arxiv.org/abs/2211.09066 GPT 3.5's (4 is much much better) addition accuracy tanks after 2 digits. However, by approaching…

I have very limited understanding of chatGPTs architecture, and neural nets in general. From what I've read over the last few days, the "attention" mechanism used in chatGPT and similar LLMs does indeed dynamically change weights of a portion of the model.

It is a run-time executed reweighting of predictions. The scalar is a multiplication between the outputs of the Key and Query kernels passed through softmax, which then multiply the result of the value kernel V(x).

Now, you may argue that because it's a multiplication with a linear or affine kernel, you might as well use commutative property of scalar multiplication and multiply the factor with the weights first, and then multiply with the input to the kernel.

But this only holds for very few kernels.

Re: Teaching ChatGPT to speak my son’s invented language

#144

Post by spyder »

famouswaffles wrote: post #9

In context learning is hands down the biggest breakthrough of LLMs. The flexibility the model displays without updating weights is genuinely mind blowing, bordering on absurd especially if you've trained other kinds of models before. See here - https://imgur.com/a/w3DAYOi from the paper - https://arxiv.org/abs/2211.09066 GPT 3.5's (4 is much much better) addition accuracy tanks after 2 digits. However, by approaching…

I agree, the in-context, few shot learning is the most interesting thing a about these transformer models, they have basically learned to learn (meta-learning).

Some additional interesting papers about it:

Transformers learn in-context by gradient descent:

https://arxiv.org/abs/2212.07677

What Can Transformers Learn In-Context? A Case Study of Simple Function Classes:

https://arxiv.org/abs/2208.01066

"We also show that we can train Transformers to in-context learn more complex function classes -- namely sparse linear functions, two-layer neural networks, and decision trees -- with performance that matches or exceeds task-specific learning algorithms."

Re: Teaching ChatGPT to speak my son’s invented language

#145

Post by vidarh »

simonh wrote: post #142
nearbuy wrote: post #120

Earlier quoted context omitted.

People keep repeating that LLMs are "just generating text based on probabilities". That statement doesn't mean anything. I think people who say this are imagining LLMs work something like a statistical model. Maybe it's doing a linear regression or works like a Markov chain. It's not. A single artificial neuron sort of works like that. But that's sort of like saying a single transistor is just an electronically contr…

That's a valid point in that we don't fully understand how LLMs solve some problems and using logic and rules of inference isn't excluded by the architecture, but on the other hand understanding that they are generating probabilistic token sequences is a very powerful and effective way to understanding how to engineer prompts and understand some of their failure modes. If we discard that insight, reasoning about thei…

> For example we often see people thinking that because an LLM can explain how to do something that therefore it knows how to do it, like arithmetic. That's because if a human can explain how to do something, we know that they can.

I think example shows LLMs to be more like people not less. It's not at all unusual to see humans struggle to do something until you remind them that they know an algorithm for doing so, and nudge them to apply it step by step. Sometimes you even have to prod them through each step.

LLMs definitely have missing pieces, such as e.g. a working memory, an ability to continue to learn, and an inner monologue, but I don't think their sometimes poor ability to recall and follow a set of rules is what sets them apart.

Re: Teaching ChatGPT to speak my son’s invented language

#146

Post by lloeki »

> For example, Kłeti has a sentence structure that is SOV (Subject-Object-Verb), while all the languages I know have a Subject-Verb-Object (SVO) sentence structure.

"all the languages I know" caveat alright, still Japanese has the verb at the end, and my German is far far away but I seem to recall it has something like that as well (at least the meaningful "verb-like" semantic part of the sentence often being at the end)

Re: Teaching ChatGPT to speak my son’s invented language

#147

Post by fenomas »

FPGAhacker wrote: post #131
famouswaffles wrote: post #118

Earlier quoted context omitted.

the connection between weights are dynamic but the weights themselves aren't being changed i.e being updated in response to the text. when training a model, the forward pass would happen i.e the generation and then depending on how close to truth it was, the configuration settings (aka the weights/neurons) would be adjusted to incorporate whatever little insight was gained from the text. Weights are matrices. The val…

Maybe it comes down to semantics but when I read things like [1] I come away with the idea that the weights are altered. But it could also just be my misunderstanding. 1: https://towardsdatascience.com/an-intuitive-explanation-of-s...

An attention mechanism is one or more layers in the neural network. When someone talks about attention altering the input vectors, they're referring to what those layers are doing and how data is transformed as it passes through them. But zooming out to the big picture, a neural network is a bunch of layers full of weights, and none of the weights changes except during training (including the weights in the attention layers).

Re: Teaching ChatGPT to speak my son’s invented language

#148

Post by snickerer »

What a wonderful story and a great family. Congratulations!

You inspired me to ask GPT to do it the other way around. I asked it to invent a successor to Esperanto, with influences from Slavic, Germanic, and Latin languages. It called its language Euroglossa and wrote a short story in it. Who can understand it?

Un tag, en froliko vilaž, un jun chico namen Tomas trovat un misterioz mapo v star bibliotek. Na mapo, skribet: "Skarb de Tri Montes." Tomas decidet da sledit la mapo in aventuro, sperante da otkriti grand skarb.

Tomas paket svoi rukzak s neobkhodim stvari, inkluziv kompas, binaukli, i nutrimento. Nachet svoi putovanje, iz vilaž, do la Tri Montes. Po nekoliko dni, on prishel k bazen na pervoj monte. Tam, on otkril zagadka, ki je klyuch za dalsi koraki: "Kogda solntse küsst la luna, dvigat kamen i vstretit un oko."

Tomas wartet geduldig bis la sonne küsst la luna in noktchielo. Kvando moment venit, er raskt dvigat un gros stein u otkryvajet secret passaž. Ingressante, on sledit un dunkel tunel, portant un torča por iluminar svoi put. La tunel führt tief in monte, bis Tomas entdeckt un hider kammer s ančient skulpturen i un glänzend tresor.

V la hider kammer, Tomas najde un skarbnica poln von stran objekte, inkluziv un animat statuo von un banane, ki parlat und tanzt. Zaskočen, ale smejšeći, Tomas realisat ke la glänzend tresor je ver la amikeco i rado ke li povas dividi kon svoi vilaž. La legenda "Skarb de Tri Montes" je veritabile un symbol za unite de diversa kulturi i lingvoj en un mondo plen da froliko i humor.

Re: Teaching ChatGPT to speak my son’s invented language

#149

Post by whiteandnerdy »

FPGAhacker wrote: post #131
famouswaffles wrote: post #118

Earlier quoted context omitted.

the connection between weights are dynamic but the weights themselves aren't being changed i.e being updated in response to the text. when training a model, the forward pass would happen i.e the generation and then depending on how close to truth it was, the configuration settings (aka the weights/neurons) would be adjusted to incorporate whatever little insight was gained from the text. Weights are matrices. The val…

Maybe it comes down to semantics but when I read things like [1] I come away with the idea that the weights are altered. But it could also just be my misunderstanding. 1: https://towardsdatascience.com/an-intuitive-explanation-of-s...

I think it does come down to semantics. When you say "weights", people will take you to mean the pre-trained parameters of the network.

I agree that in some sense the attention weights are more like meta-weights that are applied to the context of the conversation to decide how to actually weight the various words. So it's totally correct to say that previous words in the conversation affect how future words will be weighted, and I think it's reasonable to call that 'learning': for example, you can tell ChatGPT new words and it will be able to use them in context. Again though, people usually take 'learning' to mean making updates to the trained parameters of the model itself, which obviously isn't happening here.

Re: Teaching ChatGPT to speak my son’s invented language

#150

Post by monkeydust »

dTal wrote: post #96
joshribakoff wrote: post #78

Earlier quoted context omitted.

Maybe even more powerful would be reducing the number of examples needed to learn, eg less than one shot https://www.technologyreview.com/2020/10/16/1010566/ai-machi... updating weights in real time is useless if each update basically does nothing because it takes an insurmountable amount of training, on the other hand if i can give my model a succinct “lesson” i’d then be very willing to wait a while for it to “proc…

As I understand, that's basically what fine-tuning is?

What sort of performance difference can you expect from in context Vs fine tuning, any relevant papers on this ?
Post reply on HN