Hacker News Reader

I personally think that humans easily apply structure to language that doesn’t really exist. In fact, we restructure our languages daily, as individuals, when communicating verbally and through text. We make up words and shorthands and abbreviations and portmanteaus. But I think the brain simply makes connections between words and things and the structure of speaking those words is interpreted like audio or visuals in our brains — just patterns to be placed.

Really, words, utterances by themselves, carry meaning. Language is just a structure for _us_, so to speak, that we agree on for ease of communication. I think this is why probabilistic models do so well: the ideas we all have are mostly similar, it really is about just mapping from one kind of word to another, or kind of phrase to another.

Feel free to respond, I’m most certainly out of my depth here.

Hrm, true, no one’s ever gotten fluent in a language just by reading a textbook.

> Computational Linguistics I kept running into the unspoken question

I've done a lot of work in NLP and the times when computational linguistics has been useful is very rare. The only time I shipped something to production that used it was a classifier for documents that needed to evaluate them on a sentence by sentence basis for possible compliance issues. Computational linguistics was useful then because I could rewrite mulit-clause sentences into simpler single clause sentences which the classifier could get better accuracy on.

> And here was Google Translate being "good enough" for 80% of all use cases using a "dumb" statistic model that didn't even have a coherent concept of what a language is.

I assume you are aware if Frederick Jelinek quote "Every time I fire a linguist, the performance of the speech recognizer goes up"?[1]

That was in 1998. It's been pretty clear for a long time that computational linguistics can provide some tools to help us understand language but it is insufficiently reliable to use for unconstrained tasks.

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

> But it will always be a 80% solution because it's an evolutionary dead end: it can't know things, we have only figured out how to trick it into pretending that it does.

At the margin, these are equivalent (Chinese room, and all). I wonder if humans also learn similarly then retroactively tell themselves they actually do know things instead of just containing experiences encoded in their neurons (and whether that is any different than a neural network encoding trained "knowledge" in its neurons, too). This is the semantics of epistemology, at the end of the day.

Ask a toddler how the world works and you'll get a very similar response. It is entirely likely the 80%-of-human-intelligence barrier is not a "dead end" but merely a temporary limitation until these models are made to hone their understanding and update over time (i.e. get feedback) instead of going for zero-shot perfection. The GPT models incorporating video should start developing this "memory" naturally as they incorporate temporal coherence (time) into the model.

The fact we got this far through brute force is just insanely telling. This is a natural phenomena we're stumbling upon, not something crafted by humans.

Also - fun fact, the Facebook Llama model that fits on a Raspberry Pi and is almost as good as GPT3? Also basically brute force. They just trained it a lot longer and it shrunk the model. Food for thought.

Google translate works amazingly will on languages with a similar grammar (or at least, it works so on European languages, which I have the experience to judge).

However, translation of more distant languages is pretty terrible. Vietnamese to English is something I use Google translate for everyday and it's a mess. I can usually guess what the intended meaning was but if you're translating a paragraph or more it won't even be able to translate the same important subject words consistently throughout. Throw in any kind of slang or abbreviations (which Vietnamese people use a lot when messaging each other) and it's completely lost.

Given that Google Translate already exists, what is even the point of all of this?

Because for the other 20 percent it's plainly -not- good enough. It can't even produce an acceptable business letter in a resource-rich target language, for example. It just gets you "a good chunk of the way there."

And there's no evidence that either (1) throwing exponentially more data at the problem with see matching gains in accuracy or (2) this additional data will even be available.

Yeah... Google Translate is still occasionally translating good/item as "baby" on taobao. "Return Defective Baby" was hilarious for a year or two, but that was ~8 years ago IIRC, and now it just stands as a reminder that Google Translate still has a considerable way to go.

Indeed. Google Translate is just barely useful. Whenever I use it to translate to English, what I get is generally poor. It's good enough to understand the gist of what the original text said, but that's about it. Fortunately, most of the time, understanding the gist is enough.

Google translate lets you correct a translation, right?

Did people stop doing this at some point? Maybe after the advent of massively addictive social media, people more often ended up screenshotting it and sharing it for likes instead of correcting it.

I'm sure Google have stats on this but no idea whether they're public.

I haven't seen the option to correct a translation in years.

If you click the feedback button it lets you make a correction.

I learnt some very basics of computational linguistics since it was related to a side project. I kept wondering why people were spending huge amounts of resources into tagging and labelling corpora of thousands of words, while to me it seems that in theory it should be possible to feed wikipedia (of a certain language) into a program and have it spit out some statistically correct rules about words and grammar.

I guess the same intuition led to these new AI technologies...

I think a huge part is that computational linguistics still chases the idea of a universal language model, which may simply not be possible. I haven't followed the science in general linguistics but something feels off when most of the information ends up being tagged onto nil particles (i.e. parts of speech present neither in utterances nor written language and not affecting intonation or otherwise being detectable except by contrasting the structure with related languages).

In a sense the model is universal. It's just a 100GB (give or take) neural network.

And apparently (or so I heard, I think) feeding transformer models training data of Language A could improve its ability to understand Language B. So maybe there's something truly universal in some sense.

The secret is that there are no grammars in our brains. Rules are statistical, not precise. Rules, idioms are fluid and... statistical.

We're a bit more specialised than these new models. But that's it, really.

^ This. I think the more we internalize the fact that we're also basically LLMs, the more we'll realize that there likely isn't some hard barrier beyond which no AI can climb. If you watch the things kids who are learning language say, you'll see the same kinds of slip-ups that belie the fact that they don't yet understand all the words themselves, but nobody thinks that 2-year-olds aren't people or thinks they will never learn to understand these concepts.

Internalising the idea (not 'fact') that 'we're basically LLMs' will only take you to a place of deep sterility, delusion and nihilism.

Why? Do you prefer magical "handwavy" thinking?

A lot of things just fall in place if you accept an obvious fact that the brain is just a powerful enough function with certain inputs and outputs.

Neurons are not that complicated.

There are interesting consequences here though: no stable grammars, no truly uniform thinking, no singular way to understand people.

The 'handwavey' thinking here is from those who state as a 'fact' that the human brain is an LLM

Well, I didn't say a brain is just an LLM. It's more complicated than just this. But what LLM showed is that it works using the same simple building blocks + specialised parts of the brain: memory, constant relearning and world input, etc.

So in this particular case of LLMs we've managed to optimise our way around having specialised brain structures using a powerful enough math function. Next steps require improvements in how the model is trained, how we can reduce the amount of training data, what additional machinery might be necessary, etc...

But, damn it, this very thing that makes humanity possible - our language - is solved now. Natural language is a solved problem now. That very thing that makes complex societies possible - it's done. This is the fact. And it is crazy.

EDIT: typos

I don't share any of the excitement that ppl like you appear to be feeling, and I recoil from the grandiose claims being made by people who, in my opinion, are being fooled by the 21st century equivalent of a ventriloquist's doll.

(I did feel excitement while following the development of AlphaZero and its played Go matches, but that was because it was revealing greater depths and beauty in the human created game of Go. And I maintain some interest in following the development of self-driving, particularly by Tesla.)

With regard to LLMs I can see how they could be useful. I think more particularly useful when they work from a constrained corpus, so the user can know what they're drawing from (and thus the limitations of that knowledge base). The example site that been posted by its maker to HN [1] where you can ask questions against a particular book is a good one for showing the use of the tool I think. But it's just a tool and it's not in any way a breakthrough in our understanding of ourselves, of cognition or anything like that. I think the people who are making these claims can't distinguish science fiction from actual reality. They are fantasists and I think they are leading themselves and others into delusion.

[1] https://portal.konjer.xyz/

I am not excited. I am terrified.

Right at the beginning of the current wave (2010-2012) of ML approaches I did some work on ML systems and NLP, and back then I clearly saw how nothing truly outstanding is happening, we were only starting to figure our what GPUs were capable of.

So all of this was fun: NLP, ML, vintage AI. But nothing felt like it did was groundbreaking, or would solve fundamental true GAI problems, or was even close.

Yet, 10 years later, here we are. Language is solved. In most areas I know /something/ about (programming, ML, NLP, compilers) this is huge and makes mountains of knowledge obsolete.

For me AlphaZero was boring. :-) Solution space is vast but rules are simple. It was a question of time when somebody could put things together here. There was nothing unknowable about it, unlike how natural languages were always a mystery to me. Even with all the syntax, grammars, linguistic knowledge, NLP... Something was lacking.

Interesting to have this contrast in perspectives. For me, the language generated by ChatGPT is flat and boring. No spark of human creativity or originality or flair. And this cheap trick of getting it to write in rhyme or 'in the style of' such and such I find awfully tacky.

I'm not saying AlphaZero was creative either. But because it was operating inside a system that was already beautiful and which had such a vast 'solution space' as you put it, its exploration into greater depths of that space I found intriguing.

I think that's the contrast for me. Machine learning can be useful and even intriguing inside constrained spaces That's why I liked AlphaZero, working inside a very constrained (but deep) space. And why I also find Tesla's progress with self-driving interesting. It's a constrained task, even though it has a huge range of variables. And again why I find ChatGPT potentially useful in drawing from a constrained corpus but still don't find the language it generates appealing. It comes across as exactly what it is - machine generated text.

The breakthrough of ChatGPT is not a brilliant literary work per se.

It's how it interprets what people write and provides coherent answers. This was not possible previously.

AlphaZero, chess algos do not have to break this barrier, they work form a very clear and well-defined input. It was clear that a mixture of machine brute force and smart thinking would eventually beat us at these games. No magic here. Alpha family algos are /very/ understandable.

Language, on the contrary, is fundamentally not very well defined. Is it flawed, fluid, diverse... not possible to formalize and make it properly machine-readable. All the smaller bits (words, syntax, etc) are easy. But how these things come together - this can be only vaguely described through rigid formal frammars, but never fully.

Compare that to how on the lowest level we understand our brain very well. Every neuron is a trivial building brick. It's how super-complex functions of input to output arise from these trivial pieces - that's amazing. Every neural network is unique. Abstractions, layers of knowledge - everything is there. And it's kind of unique for every human so unknowable in the general case...

Your third paragraph describes language pretty well (although I'd quibble with formal grammars only being 'vague' in their coverage - I think they do a pretty good job although I agree they can never be perfect). And I appreciate the achievement of LLMs in being able to take in language prompts and return useful responses. So it's an achievement that is useful certainly in providing a 'natural language' querying and information collating tool. (I agree here with your second paragraph.)

But it remains a tool and a derivative one. You will see people in these recent HN threads making grandiose claims about LLMs 'reasoning' and 'innovating' and 'trying new things' (I replied negatively under a comment just like this in this thread). LLMs can't and will never be able to do these things because, as I've already said, they are completely derivative. They may, by collating information and presenting it to the user, provoke new insights in the human user's mind. But they won't be forming any new insights themselves, because they are machines and machines are not alive, they are not intelligent, and they cannot think or reason (even if a machine model can 'learn').

Well... complaining about people online or in the media making grandiose is like fighting wind.

I totally see your point about inherent "derivativeness" of LLMs. This is true.

But note how "being alive" or "being intelligent" or "be able to think" are hard to define. I'd work for the "duck test" approach: if it is not possible to distinguish a simulation from the original then it doesn't make sense to draw a line.

Anyways, yes, LLMs are boring. I am just not sure we people are not boring as well.

> LLMs can't and will never be able to do these things because, as I've already said, they are completely derivative.

I agree, they are completely derivative. And so are you and I. We have copied everything we know, either from other humans or from whatever we have learned from our simple senses.

I'm not asking you to bet that LLMs will do any of those things really, I suppose it's not a guarantee that anything will improve to a certain point. But I am cautioning not to bet heavily against it because, after witnessing what this generation of LLM is capable of, I no longer believe there's anything fundamentally different about human brains, so, to me, it's like asking if an x86-64 PC will ever be able to emulate a PS5. Maybe not today, but I don't see any reason why a typical PC in 10 or 15 years would have trouble.

This has the same sort of energy as the fears of the devotedly religious that a world without religion leads people to moral depravity.

Just more pearl clutching and reverence for the mystical is what it looks like to me.

> a world without religion leads people to moral depravity

This is exactly the case. Religion is a constant in human society across time and space. And one of its main functions is exactly that - to keep people away from moral depravity. Speaking against this function of religion (again, proved across all cultures and all times) only shows profound shallowness.

Personally I'm not much of a fan of Christianity or the other Abrahamic religions. And I think generally its time is coming to an end (with a long tapering off). But I think you can see quite clearly that the moving away from Christianity (probably inevitable) over the last century or so has led to moral decline and depravity in the West. If Christianity's time is coming to an end, we will neeed (and I believe we will generate) a new religion to replace it. And I don't think it will be in the Abrahamic tradition.

And astrology, and magical thinking, and many other things were always there.

So what you say is that humanity is not capable of learning moral behaviour without gods? And this is why it needs deities?

No, humans continually develop moral behaviour and codify it in religion (for transmission and social maintenance). Gods and deities aren't strictly necessary. Buddhism started without such but it is interesting that the conceptualisations of Buddhas & Boddhisattvas have become more and more like the gods & deities of other religions (at least in the popular religious traditions of Buddhism).

Well, I agree with you in that humans inevitable develop moral rules and codify these one way or another. Otherwise a society wouldn't be possible.

Religions are one of the traditional ways of motivating these rules... Legal systems are supporting them on an enforcement level.

I like religions as a very deep cultural phenomena/ideology having all kinds of effects on the, ehm, society. I just don't think religions are strictly necessary for a functioning society.

On one hand, considering our current understanding, one could say that that's a good analogy for our brain; on the other, it reminds me of the fact that people in the past used to think that their thinking was akin to the work of a steam engine. A number of expressions in English language refer to the fact, such as "cool off".

English Wikipedia is the largest. Wikipedia in other languages would be less useful.

I've just started testing GPT-4 on translation from Japanese to English, but it seems to blow Google Translate out of the water. It was particularly good with a novel excerpt. I encourage people to try it with a variety of languages and texts to see if those results hold up.

It would be useful to know whether you know both languages or just English.

Not trying to poke holes, just clarifying.

No problem. My first language is English, but I worked as a Japanese to English translator for twenty years and have written books in Japanese. My personal website is linked from my user page.

Thanks. Refreshing to see one of these claims made by someone actually qualified to verify them! I might have to play with some Norwegian <-> English translation at some point. Should be even more effective than for Japanese due to the common germanic origins.

I know only the bare basics of NLP and AI but isn't Google basically just a specialized case of LLM? Translate and Search work on the same principle that all you need to do is regression analysis on a huge pile of human behavior. Early search engines worked so hard to try to understand content and user intent and got thumped by a comparatively naive heuristic and a giant pile of data.

To take it further: if 80% is good enough and you gotta do some work anyways on the output of LLMs, maybe all the extrapolations like "just wait 10 years and most jobs are doomed" are exaggerated. It’s not unlikely that LLMs hit a wall, because they inherently lack any sort of logic and reasoning.

Which of course is a good thing to make sure many people get to keep their jobs.