More. We’re more suspectible.
We are the people who take the Turing Test as an objective test of GAI, after all, instead of looking at it like a test for those robot birds in nature shows that fool real birds into thinking that they’re one of them; from an objective standpoint (not relative to those birds), such robots have not demonstrated that they are birds. Only that they imitate them well enough to fool real birds.
Correct in that that was what Turing intended. But the vulgar interpretation of it makes it out to be more than it really is.
For Pete’s sake: two years ago I even saw a psychologist who suggested that psychologists should take notes from how computer science ostensibly measures “intelligence” by the Turing Test, because it’s so “objective”. That was sad.
AlphaGo guessed it’s way to being the best.
Can something guess it’s way to evolving into a less guessy thing? I think so. Look at us?
I am aware of how it works and used to be ‘it’s ok, it won’t happen’ but now I am actually pretty concerned.
We think GAI as the thing but there maybe a reality that doesn’t require full GAI to completely unhinge us.
I mean, what is the point? It only plays to the negatives of our society. We don’t need it. It allows for ‘do more’ but what cost does that ‘do more’ come with?
No it didn't, AlphaGo was backed by an algorithm that perfectly understood go and used that to train itself to play go better. That is a completely different situation.
If you tried to make a go engine that only tried to replicate old moves then you'd have a hard time making it perform valid moves. ChatGPT is like such an engine and we just made it perform valid moves most of the time, and now people say "now that it can make valid chess moves it is only a matter of time until it beats masters!". No just statistically reproducing human moves wont lead to a smart chess engine, so why would you think that would make a smart text engine?
It seems surprising that training an LLM would result in layers/weights which contain such representations but it seems that this is indeed an efficient representation of information in the same way that neural nets would learn addition by encoding textual integers into floats.
AlphaGo was trained with tech. It beat humans.
ChatGPT et al are being trained.
Everyone is arguing the inner workings of what is in the box. It doesn’t matter.
The usage of the box(s) is what will be the big impact.
Previously those boxes couldn’t be relied upon so stayed down in the hierarchy of usefulness.
Now. They are rising.
It had basic rules of Go as an input.
Our world is not a fully described game (unless you want to take a radical philosophical position, and in which case we still don't know the rules of this game). We talk about and think about this world in language. Training a machine on human language and then after training, having that machine take language inputs and generate language outputs is not the same as training a machine to 'play' Go because the two worlds they have been trained in (1. the game of Go; 2. the real world) are fundamentally not equivalent.
Can ChatGPT(-ish), trained on a ton of games, beat AlphaGo?
I used GPT-3/3.5/4 for multiple tasks and I never found it useful. It just feels… average.
It can “answer” many questions correctly but that’s only on the data set it’s been training on. I asked it to solve Elixir issues, write Elisp snippets, figure out why my Rust code was complaining about lifetimes and even went for easy mode and asked for JavaScript code.
Only JavaScript (very popular) responses were workable but they still had popular problems included in it (e.g. not clearing out timers or dealing with common edge cases). Elixir “hints” were very interesting since - when out of water - it would produce code assuming that it’s a OOP and that was a rabbit hole it declined to exit.
Too technical? Went for fun with family members who have exceptional cooking skills. We tried to figure out a dish variation of an apple pie. Suggestions were heavily criticized as non workable for many reasons (too sugary, consistency wouldn’t be too good, modification change the ingredients but GPT refuse to break of out theme etc.)
Colin Meloy made an experiment some time ago and created The Decemberists-like song with it and then actually recorded it. It worked but it was described as below average and bland. I share that vibe.
In the end I have this feeling that GPT is on a level of 7 year old with an access to vast library. It uses proper grammar it can copy parts of texts, but one needs to provide it with a great benefit of doubt and energy just so that it’s useful.
In contrast to 7 year old, GPT isn’t grateful for that sacrifice and is actually making money of you.
This is not an exception either; teams like mine exist in every large enterprise. I worked in several of them and this AI makes life so much better. Let’s not pretend it makes certain outsourcing countries not fully obsolete though, to the tune of 10s of millions+ programmers, data entry workers, text producers (bloggers for seo, docs no one reads etc) and more are mostly all already worse than gpt3.5 and 4. Wait for 6 or 7.
I do think really this is you’re not being in touch with most of reality and what most people on earth do daily at this moment. It definitely as nothing to do with the productivity you have and the drive you feel internally to ‘get shit done’. Most people, by a huge stretch, get education, want to do work what they are told to do related so that education and that’s it. They have no, at all, drive to ever move beyond that. They learned Angular on their first job; that’s it. Etc.
This is now replaceable; this is not, at all, comparable to automation before this. It’s a threshold; it you cannot see it, you definitely will be shocked in the coming years. This is not going away and it is replacing humans now on a small scale. The companies that really can just remove 90% of their staff are late always, but it will come. I see this all day because I work in massive orgs where this holds, not only for tech.
It's probably more like a 14 year old but I think you're probably right. It can give step-by-step instructions and even the code to set up entire web apps, but it's only when you go through those steps that you find the issues. Packages out of date, less-than-optimal ways of handling non-obvious cases, and I've even found it trying to use javascript functions that have never existed. To its credit, it apologizes and tries again when I call it out. Maybe GPT-5 will be the real one to worry about.
I’ve found it useful for drafting cover letters (I just need to do final edits).
I’ve also found it useful as an adjunct to google when I’m searching for niche things like “Luxury resorts with activities for my dog and lots of off-leash trails”.
So far that’s it, but it does seem like a useful addition to Google search for super niche things. Obviously it hallucinates a lot on these niche edge cases but it gives some great starting points.
So... this early-stage AI was able to make music that's better than the vast majority of commercial recording artists these days? I fail to see how that's a failure.
But look how far we've come since those heady and carefree days. We're already able to pour buckets of scorn on its capabilities, and brush them aside as if designing a language (albeit a derivative and uninspired one) is a trivial task that any human could achieve.
No we don't, we are just trying to put a damper on all the "the singularity is here!" style comments that seems to flood any discussion about GPT.
Yep. Welcome to HN. And welcome to the majority of humanity!
https://lifearchitect.ai/contrarianism/
These models are trained to produce text snippets that look like text snippets it has seen before, and it has seen all internet. That means it can do a lot of impressive things, but also that it is very dumb in other ways.
I've seen it being dumb in maths or real world problems. But as a large language models, they understand and speak languages fine, and even mistakes they make look like mistakes humans who are not natives in the language would make.
We may as well say that when we speak, we are just predicting words we have trained on. I don't see how these models are worse than people in that regard.
The general knowledge and thinking of these models are surely limited. But seeing GPT-4 go from text only input to text with images, I think it is very possible to break the barriers very soon.
> they understand and speak languages fine
No, they neither 'understand' nor 'speak' languages. The first word here is the more delusional, they have no understanding of languages. They have simply generated a model of the language. And they do not 'speak' the language they have modeled; they generate text in that language. Speaking generally implies an active intelligence; there is no intelligence behind an LLM. There is simply a machine generating (wholly derivative) output text from input text.
> We may as well say that when we speak, we are just predicting words we have trained on
This is the delusion, commonly being repeated, that humans themselves are only LLMs. This is a dangerous delusion, in my view, and one that has no evidence behind it. It is only a supposition, and a sterile and nihilistic one in my view.
> The general knowledge and thinking of these models are surely limited [...] I think it is very possible to break the barriers very soon
The limitations are fundamental to LLMs. LLMs have no general knowledge and LLMs don't do any 'thinking'. Your understanding of what they are doing is in grave error, and that error is based on a personification of the machine. An error coming from the delusion that because they generate 'natural' language they are operating similarly to us (false and addressed above). They are never going to break the limits because they have never started to transcend those limits in the first place, nor can they. They don't and will never 'think' or 'reason'.
> designing a language (albeit a derivative and uninspired one) is a trivial task that any human could achieve.
I know a competent man should be able to "change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders [...]" and what not and specialization is for insects, but even on the Mysterious Island not everyone is Cyrus Smith (Harding).
What will be the other impacts?
This is just another tool in the student’s toolbox and should be treated as such.
Modern countries don't do homework, other than reading, for ages
While the language it generated needs refinement, the design process is iterative, and it takes time to develop a programming language that is worth using.
I encourage you to consider how you would approach designing a programming language. Like you, I would likely start by incorporating my favorite features from languages I love and addressing problems I've encountered in the past, similar to the dialogue GPT and I have in the post.
As I mention in the post -- creativity often involves remixing existing ideas, and GPT-4's attempt at generating a programming language is no exception. Notably, it also called out explicitly the constructs it was borrowing from other languages when designing the language.
It's worth noting that this was GPT-4's initial attempt at something highly complex, and even if there are "plot holes" -- it's a noteworthy achievement in and of itself.
> As I mention in the post -- creativity often involves remixing existing ideas
Creativity involves remixing but is not all remixing. Is it? Not even all humans are at the same level in this. Number of attempts means nothing if you don’t add features that the LLM desperately needs to be grounded in the reality, or perform logical reasoning beyond RLHF.
Government regulation has been brought up multiple times, most recently by Sama himself on twitter. I think it will have to come, but I don't know what flavor it should take.
The cynical take is quite clearly: "Because sales people can sell it"
eg wrap some pretty UX around it, tell people it's "nearly human", then have your sales people try and get non-technical CxO types (or consultancies) to cut them a cheque
It will also learn all the bugs and reproduce them.
Programmers who think ChatGPT is going to replace their job must have a very low opinion of their own skill set.
We see it in this article, both the statement "Indentation-based scoping, similar to Python" and the code with bracers were probably correct in their original contexts, but combined they are no longer correct. LLMs will make such errors no matter what kind of data you train it on, there is no known way to avoid that class of errors.
As a devops guy I live in fear of anything that can copy and paste from stackoverflow faster than I can.
However I want to add that it does the guesswork and the presentation surprisingly well to a degree that it’s effective in certain contexts.
The more people think of this as a tool that they can use and master, the more they will benefit from it.
But then applying them directly into an example falls down a bit. The language is a bit of a mismash and parts of it might not work semantically. The use of actors looks a bit out of place, and raises a lot of questions about how the memory management and scheduling is working under the hood. Despite the bold aims in the tenents thre is nothing in the language exposed so far that addresses them: why would the language produce more modular code in the large? what features of the syntax would increase the level of abstraction in code written? etc etc
Overall this would warrant a (barely) passing grade by modern standards, i.e. it can regurgitate the material succesfully but without demonstrating any deep understanding. I find the result quite interesting as it demonstrates how much of what we judge to be quality is actually fitting a particular pattern of response (w.r.t a specific field of study) rather than exploring the material more deeply.
I like the use of the arrows in the function definitions though :) I've seen that somewhere before and will be "appropriating" it for the project that I'm currently working on.
Python type hints probably.
I (ab)used them for good effect when I was fiddling around with the spark earley parser that used to be part of the python build system.
Not sure why need any new languages when we have the perfection that is Golang tho.
--Amos Tversky
On the one hand, I agree.
On the other hand, I see some variant of this comment on HN about every new language, so...
And now here we are: it seems we tech people are every bit as susceptible to this kind of thinking as the average joes we looked down on.
Actually look at this “language” that GPT-4 has written: it’s just a mishmash of features from the most commonly blogged-about existing programming languages. It has no conceptual insights, no original ideas, no taste that isn’t cribbed from one of the few languages it’s copying from.
There are cases where getting a short distillation of the mean opinion of The Internet is useful. But designing a new programming language is clearly not one of them, mostly because the chatbot is not “designing” a language—it is guessing the most likely response to a question about programming languages.