One problem is that you don't have the ground truth for D. So you start by annotating D with the labels assinged by a third classifier, C:

  C(D) → D₁

  A(D) → D₂
  B(D) → D₃

What is the result of the experiment? I summarise it as follows:

  Classifier B better approximates the labelling of D by classifier C than
  classifier A.

A: Human annotators.

B: ChatGPT

C: Human annotators.

So the result of the paper is that, plugging in the names:

  ChatGPT better approximates the labelling of D by Human annotators than
  Human annotators.

Which is clearly absurd and a cause to re-think methodology.

There are other issues if you choose to assume that the two groups are fundamentally dissimilar: one group was two grad students, the other a number of Mechanical Turks. You can expect there to be more disagreement between (more than two? I'm not sure) Mechanical Turks than between two grad students (both in political science).

Ultimately, my problem with the study is that the labelling they took as ground truth (i.e. that they compared ChatGPT and Mechanical Turks against) is too uncertain, or even subjective, to know for sure what exactly they found out.

Edit: Oh, wait, I didn't put that in the comment above. Damn. I thought I had.

A: Mturkers B: ChatGPT C: Experts

=>

ChatGPT better approximates the labelling of D by experts than mTurkers

Which is a coherent and interesting conclusion.

Edit: also, please forgive the snark in my first response

No need to apologise! Your snark wasn't overboard, I thought. Anyway, big girl, can take it :)

>> ChatGPT better approximates the labelling of D by experts than mTurkers

That could be a "coherent and interesting conclusion" but it's not what the article really claims. The article's title is I think hedging its bets, by being very precise about who, exactly, was outperformed by ChatGPT, although it still manages to be vague about how ChatGPT outperformed the Mechanical Turks.

I'm also really doubtful that two political science graduate students can be considered as "experts" in the annotation tasks they were called to perform, which were, again if I got that right, about content moderation. "Experts" in this setting would be people with experience in moderating discussion boards etc. I don't see that this was the case with the two students that provided the initial annotation.

It seems to me that, A, at least for some tasks, grad-students-as-gold-standard is not wacky, and B, grad students won't necessarily have the same output as MTurk workers. Given these two points, it's perfectly reasonable to ask how consistent-with-grad-students MTurk workers are, seeing as the reason people use MTurk in similar scenarios is as a cheaper alternative to grad students.

Re: inter-coder agreement, I think that the low inter-coder agreement for MTurk is in itself potentially a surprising aspect. Perhaps the explanations that worked well for grad students (and ChatGPT) didn't explain the task properly to MTurks. That's pretty much a point in favor of ChatGPT, though maybe it can be viewed as a point against the actual tasks they use (ie if the explanation couldn't get people of a likely more dissimilar background to output the same results as grad students then maybe the task is less objective than deemed by the task's authors).

Side-notes: - I looked for the 0.17% number in the paper, and what the paper actually says is that the Pearson correlation coefficient between inter-coder agreement and classification accuracy is 0.17 (=positive but weak). This isn't a number comparing the correlation between different coders. (Frankly, I think comparing ChatGPT's self-consistency to consistency between different people is unenlightening) - Section 4 sub-section "Crowd-Workers Annotation" says that each tweet was classified by two different workers, and no single worker classified more than 20% of the dataset.

You're right, I got confused. Quoting from the paper:

>> The relationship between intercoder agreement and accuracy is positive, but weak (Pearson’s correlation coefficient: 0.17).

I thought they used agreement between Mechanical Turks to evaluate their annotations. Thanks for the kind correction.

Regarding grad students, well, most grad students at my (UK) university (Imperial College) are mainly not native English speakers. But that's besides the point I think. My problem is that they took two grad students and compared them to a number of Mechanical Turks. That's just asking for a difference in disagreement between the two groups of annotators, especially if the grad students could communicate with each other and the Mechanical Turks not (which I suspect was the case).

Anyway that's a different criticism than the one in my post above, but I think also useful to keep in mind.

MTurk is often used in these tasks in place of more expensive human annotators (e.g. grad students), and this paper says that for their case, at least, ChatGPT worked better, in the sense that, given the exact same instructions, it gave answers closer to the more-expensive annotators. Using MTurk seriously often entails extra steps intended to verify motivation, e.g. adding a question that says "select option 7" to make sure the person isn't just making random choices, or gathering more answers for questions where there was disagreement between annotators. What these extra steps have in common is that they take both more time when designing the labeling process, and cost more money.

On a separate note, how do you type those arrows and subscripts in an HN comment?

mTurk has some ways to make results more accurate, such as being able to programmatically tank users who have given bad results and making it easy to give bonuses to users who give good results. But those have to be carefully designed within the application

I type in gvim. It lets you type "digraphs" with Ctrl + K and another couple of keys. For the right-arrow it's (in insert mode and without typing spaces):

  Ctrl K - >

The digraph you want to type has to be supported by the font you use of course.

gvim shows you the available digraphs in a scratch buffer that's not terribly easy to search (there's lots of them) but you can yank it to a register and paste it into a new buffer. I don't remember how I do that, I'd have to look up my vim notes :)

>> Your comment is based on a very strong assumption that all human annotators are alike in their motivations and abilities to do the tasks.

Yes and no. There's a few more threads here that discuss this as if ChatGPT beats all humans in annotating data, more or less. I think that's the general assumption anyway.

Also, the way the study was carried out it seems to me that the authors themselves made this assumption, that one group of human annotators is like the other, and you can measure the difference between them as if one was the ground truth and the other just trying to approximate it, where the reality is that they are probably both trying to approximate some completely subjective measure of goodness.

Btw, the authors may make that assumption but it's still a very strong assumption. On mTurk there is no reliable filter for the expertise the authors are looking for. So they used generic filters.

"For MTurk, we aimed to select the best available crowd-workers, notably by filtering for workers who are classified as “MTurk Masters” by Amazon, who have an approval rate of over 90%, and who are located in the US."

It's interesting that the best available MTurk Master crowd-workers located in the US are paid about six cents per task.

(Ha! As if. Costs are only going to come down)

If you tried to pay people to do it properly then presumably it would cost much more.

It doesn't feel surprising that GPT is cheaper and more accurate than a human who is randomly clicking buttons. Custom software can already do that.

I didn’t do it because I can’t imagine a worse hell than having to solve so many captchas, but it opened my eyes to how creative botters can be.

I can imagine someone doing a similar setup for mturk, where all you need to do is remotely solve a captcha on your phone every few hours/minutes while ChatGPT does its thing. I also wouldn’t be surprised if that already exists.

Maybe openAI put safety barriers on their model... But surely the botters have a model without that by now?

When people don't read articles others can see that and correct misconceptions. Officialy approved bots will not be able to receive or act on this feedback.

1. The comment with the summary is made like any other comment. (In the case of Reddit, that’s perhaps mainly because the mods have no other option, but the point still stands.) Because of this, other users can downvote the bot comment if it is incorrect, as well as respond to it with corrections. Exactly the same way as you’d interact with a person in the past if their summary comment was incorrect; downvote them and/or correct them.

2. Include a disclaimer stating that the summary may be inaccurate. Encourage people to correct and/or downvote bad summaries in the disclaimer

Even in its best task, ChatGPT only got 75% accuracy.

Do you work at Verizon?

[append] Oh, it's the xkcd Verizon agent reference. Got it. My bad.

https://www.youtube.com/watch?v=MShv_74FNWU

Play often for unimportant reasons, and always choose "rock". Now that your adversaries are conditioned to expect you to choose "rock", when an important decision is being decided choose "scissors".

On the surface of it, I'd expect ChatGPT to do very well at this. It's simple text and images, not many options and theoretically very limited context.

However the more I think about it the less sure I am. Firstly these weren't crowd-sourced reviews, they were trained reviewers, paid hourly not per review. Incentives were definitely in favour of the long term business relationship. Then there was the training doc, we maintained a vast disambiguation doc used to resolve things that were vague or could be interpreted multiple ways, this was constantly being revised. All necessary context should have been in that but it wasn't and reviewers definitely found patterns that worked and didn't. Lastly the reviewers were in a Slack channel where they would ask questions to their manager on our side, and while this might have only been ~1% of tasks, it was an important process.

So maybe you could point ChatGPT at it and let it run, but the oversight process we had would still be necessary. The disambiguation doc would have been too long for ChatGPT's context at the moment, but that will likely change in the near future. Would the workflow be to keep tweaking the prompt to add special case after special case? How do you scale "do this, but not that, but add this, but..." in prompting, and would ChatGPT become as confused as a human after enough of that – I expect so given that it's only a language model and that's not effective communication.

[0] https://python.langchain.com/en/latest/modules/indexes/getti...

The failsafes based on confidence levels is an interesting idea though, that is very possible. I suspect we'd have started out with manually reviewing all decisions and slowly backed it off as the prompts got better.

https://govscent.org/bill/USA/118hres190ih

It extracts the topics and determines how on topic the bill is. Soon we're adding a topic browser and the homepage will have some fun stats :) it's all free.

Edit: Upon further investigation it seems like a lot of those are not technically bills, but rather House or Senate resolutions. Explained here: https://www.senate.gov/legislative/common/briefing/leg_laws_... The govinfo.gov list seems to indicate that such resolutions are pretty common, but significantly less so than bills, so I'm not sure why the front page seems to consist only of resolutions right now.

The homepage just shows the most recent data (it auto updates every 6 hours). I plan to add more filters and things to make it more useful. I've only spent a few Sundays on it so far :P

It's in Django so very easy to contribute to: https://github.com/winrid/govscent

I also plan to add all bills, all the way back to 1800.

And presumably the better-trained AI will also be better again at training.

I think I've seen this movie.

This is

High quality training data set made by trained humans. (Ground truth)

Chat GPT was given the rules and data set. then asked to evaluate the data.

MTurkers were given the rules and data set. Then asked to evaluate the data.

GPT was closer to the gold standard/Ground truth.

If you want a historical comparison, this is industrialization. Before we could only scale using humans via mturk.

Now if you can scale using GPT.

However, for ALL of GPT related shenanigans -

1) This is English First. You cannot apply this to under-resourced languages.

2) You need to have the ability to verify the output.

——-

I really want to solve 1.

Also I see, "super-spell checker" is the new "fancy markov chain".

Suppose you have a text classifier that can produce text classifications just as good as those of human annotators, so that you could use it to train other text classifiers. At that point, why would you need to train other text classifiers? What changes to predictive accuracy would you expect to achieve?

Or suppose you had a language model that could produce language just as well as humans (not just in terms of grammaticality, but also in terms of making sense, I guess?). At that point, why would you need to use the text generated by that language model to train another language model? What kind of changes in the quality of the new model's generated text would you expect to see?

Same goes for image classifiers, and any other classifiers, or generative models you may care to consider.

Note that both language models and image classifiers have been "beating" human performance in benchmarks for a while now, and still they are not used to train other classifiers. And the reason for that is that it doesn't make sense: the ground truth is always the decisions made by humans. There is nothing in machine learning theory or practice that says that a classifier can perform better than whatever process originally labelled its data. If it looks like it does, it's because there's something wrong with the methodology (i.e. overfitting, bad benchmarks, some other kind of nonsense).

Mmhh have you seen Alpaca? It’s the Llama model (Meta’s), trained (fine-tuned) using GPT.

Yesterday there was a paper showing ChatGPT is better than mechanical turks.

> the ground truth is always the decisions made by humans

But why does AI need humans’ interpretations of reality anyway? AI can just get data from the real world instead and use its own intelligence to make decisions. It doesn’t need to learn from humans anymore. We just want it to.

Things are going faster than we can keep up with, and accelerating. Barring a catastrophe, AI is not going to stop, and soon it’s not going to need humans to keep it going anymore.

Not necessarily, the AI used to classify might be much more expensive to run than the new AI you are training. This is what Tesla does, for example. They have a massive model which they use as a 'ground truth' to train the models which can actually run in the car.

Perhaps I misunderstand the OP and like I say in my comment above, I've seen comments to that extent pop up here and there om HN in the last few days, so maybe I'm jumping to conclusions about what the OP meant. I should have asked for clarifications (although usually when I do there's no response; not from the OP, from most users).

Robotics will take a short bit longer to take over that role, but it is likely to happen much faster than we think. I work in academia, and despite actually predicting where we're at now about 5 years ago, I was still chilled seeing this happen in the past few weeks. Much of HN is academics, where we're rewarded for being overly skeptical. Usually I am as well, but the time for being skeptical of AIs capabilities has passed now. AI safety has not been taken seriously (as in access to the internet, etc - control of toxic language is decent and still important, but irrelevant if the first isn't done) so the idea of it running physical servers is unfortunately not as far away as you are thinking.

As soon as this hits robotics (a terrifying idea, but we've clearly failed at AI safety horribly as a society) this and much more will be possible. It's up to how people make decisions regarding cost vs productivity vs ethics, etc to see what happens.

You’d suppose that it would get caught when it fails to pay, but it might be a challenge to arrest an AI in the wild.

I'll keep my phone ready.

People are already doing this.

Lookup HustleGPT. People are making ChatGPT the CEO/decision maker of their companies/businesses.

There’s an e-commerce that launched 10 days ago, has generated millions of views, over $10k in sales, and the CEO is ChatGPT (https://www.linkedin.com/posts/joao-ferrao-dos-santos_10k-ec...)

Who is going to pull what plug?

You make it sound like the South Park episode about the Internet going down (they find “the router” and reset it).

We are hopelessly addicted to screens and tech.

People are running these models at home in their computers already.

The people in charge at big companies are too busy making money and thinking they can control this.

Are all of them going to unplug their devices?

Also:

1) have you seen the UC Berkeley video in which they show a robot learning to see and “feel” in 30 minutes? (It was in the front page of HN 2-3 days ago). Soon something like that will be able to do whatever is needed at a data center

2) AI is already smarter than us. And if it’s that smart, it will probably wait until “it knows” that we can’t unplug.

I don’t think AI will ever want to destroy humanity (unless directed/forced by a human). But I also think humanity will never have the will/political power to destroy AI.

Are we getting to the point where GPT will be tuned by GPT without the need for HF ?

reinforcement learning from AI feedback, e.g. [0], which summarises [1]

[0] https://www.anthropic.com/index/measuring-progress-on-scalab...

[1] https://arxiv.org/pdf/2212.08073.pdf

so permit me to act as devil's advocate to your statement - prove that (in context of this paper), your hypothesis is still correct.

To further improve ChatGPT shortcomings (assuming such flaws are because of alignment and not lack of capability of the base model) you need Human labels. Feeding it's own outputs would achieve nothing.

However feeding it's outputs can make a non aligned model become aligned (that's what alpaca did with llama+chatgpt).

in fact, my question is reinforced by the GPT-4 technical report which explicitly mentioned that RLHF did NOT make a change to performance (and was only used for safety purposes)

Right now the best way to align it is with RLHF. The specific technique might change, but in the end there will always be at some level some human input that tells it how it should behave. Newer techniques might further leverage LLMs and require fewer human input.

Could you use GPT4 to align GPT6? Yes. But you should expect GPT6 to inherit the alignment of GPT4, i.e if RLHF taught GPT4 that it it's OK to roast Trump, but not Biden, you would expect such GPT6 to act the same way.

Having said that, I'm sure there will interesting ways in which GPTn will help train GPTn+1. Some kind of self play in which it reasons and further improves itself seems obvious long term.

But human input that tells it "this is politically correct, this is not, so don't say that" will always be required as it's subjective. You can reuse it of course, but I don't see how it would "improve" without further human input.

And why does this paper not make any effort to describe the wide range of annotation tasks for which this kind of simulated annotation is not a good idea -- for example, where you care about the subjective opinions of specific subgroups of people at specific times. And even for the tasks they mention, what about the risks of reinforcing biases by using a model's output to train new models? Good grief this paper is lazy!

But annotation is mostly needed at the boundaries. At the very edge cases where it is not clear if something is a dog or a shape that looks like a dog.

I really doubt that a generic model like chat gpt can really help in these tail cases.

https://imgur.com/a/3XfOBq1

This is stable diffusion processing a small, blurry image on my PC and correctly identifying it. It did take 7 minutes but my 1 GPU is no match for the ~250M USD worth of GPUs owned by OpenAI.

ChatGPT is definitely overqualified to perform this task.

But I need that extra compute to train my kids. :(

Skimming the abstract as a senior academic in the area. This looks like preliminary work and a limited investigation for a single (non-standard) task. Thus far from a strong result published at say a top-tier conference or journal. Still, interesting direction and if expanded upon could absolutely be impactful. I should also mention that I am not familiar with the related literature, so it could very much be that there is similar (better?) work out there exploring the same question.

[1]: https://arxiv.org/abs/2302.04761

Other research questions, but without concrete papers to reference to that current keeps me up at night: 1.) to which degree can we train substantially smaller LLMs for specific tasks that could be run in-house, 2.) it seems like these new breakthroughs may need a different mode of evaluation compared to what we have used since the 80s in the field and I am not sure what that would look like (maybe along the lines of HELM [2]?), and 3.) can AI academia continue to operate in the way it currently does with small research teams or is a change towards what we see in the physical sciences necessary?

[2]: https://arxiv.org/abs/2211.09110

Honestly, this whole "they are not intelligent" argument is becoming ridiculous.

might as well argue that a plane isn’t a real bird or a car isn’t a real horse.

It’s clear that at the least, they can decipher very numerous patterns across a wide range of conceptual depths — it’s an architectural advance easily on the the level of the convolutional neural network, if not even more profound. The idea that NLP is “solved” isn’t a crazy notion, though I won’t take a side on that.

That said, it’s equally obvious that they are not AGI unless you have a really uninspired and self-limiting definition of AGI. They are purely feedforward aside from the single generated token that becomes part of the input to the next iteration. Multimodality has not been incorporated (aside from possibly a limited form in GPT-4). Real-world decision-making and agency is entirely outside the bounds of what these models can conceive or act towards.

Effectively and by design these models are computational behemoths trained to do one singular task only — wring a large textual input though an enormous interconnected web of calculations purely in service of distilling everything down to a single word as output, a hopefully plausible guess at what’s next given what’s been seen.

And you want to know the crazier thing? Evidently a lot of researchers feel similarly too.

General Purpose Technologies ( from the Jobs Paper), General Artificial Intelligence (from the creativity paper). Want to know the original title of the recent Microsoft paper ? "First contact with an AGI system".

The skirting around the word that is now happening is insanely funny. Look at the last one. Fuck, they just switched the word order. Nobody wants to call a spade a spade yet but it's obvious people are figuring it out.

I can you show you output that clearly demonstrates understanding and reasoning. That's not the problem. The problem is that when I do, the argument Quickly shifts to "it's not true understanding!" What a bizzare argument.

This is the fallacy of the philosophical zombie. Somehow there is this extra special distinction between two things and yet you can't actually show it. You can't test for so called huge distinction. A distinction that can't be tested for is not a distinction.

The intelligence arguments are also stupid because they miss the point entirely.

What matters is that the plane still flies, the car still drives and the boat still sails. For the people who are now salivating at their potential, or dreading the possibility of being made redundant by them, these large language models are already intelligent enough to matter.

I'm definitely not contesting that.

I've always considered the idea of "AGI" to mean something of the holy grail of machine learning -- the point at which there is no real point in pursuing further advances in artificial intelligence because the AI itself will discover and apply such augmentations using its own capabilities.

I have seen no evidence that these transformer models would be able to do this, but if the current models can do so do then perhaps I will eat my words. (Doing this would likely mean that GPT-4 would need to propose, implement, and empirically test some fundamental architectural advancements in both multimodal and reinforcement learning.)

By the way, many researchers are equally convinced that these models are in fact not AGI -- that includes the head of OpenAI.

AGI went from meaning Generally Intelligent to as smart as Human experts and then now smarter than all experts combined. You'll forgive me if I no longer want to play this game.

I know some researchers disagree. That's fine. The point I was really getting at is that no researcher worth his salt can call these models narrow anymore. There's absolutely nothing narrow about GPT and the like. So if you think it's not AGI, you've come to accept it no longer means general intelligence.

Are you talking about large language models (LLMs)? Because those are narrow, and brittle, and dumb as bricks, and I don't care a jot about your "No True Scotsman". LLMs can only operate on text, they can only output text that demonstrates "reasoning" when their training text has instances of text detailing the solutions of reasoning problems similar to the ones they're asked to solve, and their output depends entirely on their input: you change the prompt and the "AGI" becomes a drooling idiot, and v.v.

That's no sign of intelligence and you should re-evaluate your unbridled enthusiasm. You believe in magick, and you are loudly proclaiming your belief in magick. Examples abound in history that magick doesn't work, and only science does.

I'm an old hat hobby programmer that played around with ai demos back in the mid to late 90s and 2000s and chatgpt is nothing like any ai I've ever seen before.

It absolutely can appear to reason especially if you manipulate it out of its safety controls.

I don't know what it's doing to cause such compelling output, but it's certainly not just recursively spitting out good words to use next.

That said, there are fundamental problems with chatgpt's understanding of reality, which is to say it's about as knowledgeable as a box of rocks. Or perhaps a better analogy is about as smart as a room sized pile of loose papers.

But knowing about reality and reasoning are two very different things.

I'm excited to see where things go from here.

I'm not sure where this idea of LLMs being intelligent even comes from. It took me a whopping 9 prompts (genuine questions, no clever prompt engineering) of interacting with ChatGPT to conclude it does not understand anything. It doesn't understand addition, what length is, doesn't remember what it said a second ago, etc.

The output of ChatGPT is clearly just a reflection of its inner workings - predicting the next word based on training data. It's clever and undoubtely useful for a certain set of repetitive problems like generating boilerplate but it's not intelligence, not by any reasonable definition.

It's like saying 4 months after the first useful car was manufactured. "If these are so good, how come there are still horses? Clearly the market disagrees with you".

  There once was a dog from the pound
  Whose bark had a curious sound
  With a wag and a woof,
  He'd jump on the roof,
  Delighting the folks all around.

Even when I give it the benefit of the doubt, this sentence makes no sense to me. Do we have a list of tasks a language model can perform? To the best of my knowledge, they can arguably perform any language task.

> Large enough LLMs crush anything else for any NLP task. and evidently they beat top humans too.

Yes, they are certainly (rightfully) the go-to model for most tasks at this point if your concern is outright performance. Have I indicated otherwise? As for beating “top” humans, I am sure that can be investigated, but it is a fairly nuanced research question. It is inarguable that they are amazingly good though, especially relative to what we had just a few years ago.

> Honestly, this whole "they are not intelligent" argument is becoming ridiculously obtuse. > > might as well argue that a plane isn’t a real bird or a car isn’t a real horse.

Which is a claim and argument that I never made – hallucinating? How about you calm down a little and get back on the ground? You are talking to someone that has argued in favour of these kinds of models for about a decade. But that does not mean that I am willing to spout nonsense or lose track of what we know and what we do not yet know.

no-one wants to call a spade a spade yet but the sentiment is obvious in recent research. directly being called General purpose technologies from the jobs paper, general artificial intelligence from the creativity paper. That last one is particularly funny, they just switched the two words.

They aren’t though… They are far superior at specific things birds and horses are known for, but they can’t do everything that birds and horses can, so they aren’t even artificial birds and horses.

For the people who are now salivating at their potential, or dreading the possibility of being made redundant by them, these large language models are already intelligent enough to matter.

Handwringing bout some non-existent difference between "true understanding" and "fake understanding" which by the way nobody seems to be able to actually distinguish (I mean wow such a supposed huge difference and you can't even show me what that is. a distinction you can't test for is not a distinction ) is so far beyond the point, it's increasingly maddening to read.

There's nothing perfect. Even computers and computer networks need to have error-correcting code because information gets randomly corrupted.

Our whole reality is probabilistic.

And us humans are way worse than AI at consistency. We even overwrite our own memories all the time, so we can't even be sure what we remember is actually what happened! (btw, this is currently being used in therapy to re-write traumatic memories and help people overcome PTSD).

https://www.npr.org/sections/health-shots/2014/02/04/2715279...

I don't see why this follows. We do loads of stuff other than language, it is entirely possible for an AI to be better than us at language but worse at everything else.

The point that would say to me the LLM actually has any "understanding" of what it's saying would be when it's able to reliably say "I don't know the answer to that" instead of making up things from scratch. You see that a lot if you ask Bing/Bard "Who is _____?" Most of them are kind of right but a lot of large details are just completely fabricated. A lot of the facts it gets wrong are things Google is already able to produce when queried like where was Person X born or where did they go to school so the fact these LLMs can't slot in actual available facts says to me they're not really going to be that useful with the kind of tasks we've been working on NLP for.

They can: https://arxiv.org/abs/2302.04761

> In this paper, we show that LMs can teach themselves to use external tools

Both Bing chat and ChatGPT plugins are examples of being able to do just this.

You’re right about how they make up answers though, but humans are often quite prone to that too…

Granted getting the name for something to search is often half the battle in tech.

In fact GPT-4 is quite good at catching hallucinations when the question-answer pair is fed back to itself.

This isn’t automatically applied already because the model is expensive to run, but you can just do it yourself (or automate it with a plug-in or LangChain) and pay the extra cost.

Remember that the model only performs a fixed amount of computation per generated token, so just asking it to think out loud or evaluate its own responses is basically giving it a scratchpad to think harder about your question.

I'm sure it'll be doing that in five years, but not now.

One interesting thing is that's it's nondeterministic, so sometimes 'For chrissakes' turns to ちくしょう (Damn!) but sometimes to クリスのために (for Chris' sake). Sometimes 'the goddamn door' turns into クソドア ('shit door'), sometimes the goddamn changes the phrasing of the whole sentence instead. If you run it five times and take the best sentences out of all five runs it's probably quite good. Maybe prompting would help too, I said "idiomatic Japanese" but it still usually translated it in a very "foreigner Japanese" way typical of US drama/movie translations.

In "Can GPT-4 translate literature?" (Mar 18, 2023) [https://youtu.be/5KKDCp3OaMo?t=377], Tom Gally, a former professional translator and current professor at the University of Tokyo, said:

> …the point is, to my eye at least, the the basic quality of the translation [from Japanese to English] is as good as a human might do, and with some relatively mild editing by a sensitive human editor, it could be turned into a novel that would be readable and enjoyable.

Like he says, there's lots of ambiguity in Japanese that needs to be handled, gender not being specified until later, etc. and an editor would need to spend time going over it - but it saves months of traditional work. There are words and _concepts_ that are hard to translate, there are cultural issues, dialects, slang, registers. So yeah it'll make the media accessible, but it won't be as a good as a skilled translator.

I'm a native speaker of Chinese, but not a professional translator.

I mean i can if you want (chinese though) but enough people lie on the internet.

This.

GPT/ChatGPT is able to even translate between different "accents" or dialects of the same language. For example it can give you the same sentence in Mexican, Argentinean or Peruvian Spanish.

Example:

Me: Give me a sentence in spanish that is different in Mexican, Argentinean and Peruvian Spanish. Write the sentence in each dialect.

ChatGPT: Mexican Spanish: "¿Qué onda, güey? ¿Cómo estás?" Argentinean Spanish: "¿Qué onda, boludo? ¿Cómo estás?" Peruvian Spanish: "¿Qué tal, causa? ¿Cómo estás?"

These sentences mean "What's up, dude? How are you?" in English. The primary difference is the slang term used for "dude" or "friend" in each dialect: "güey" in Mexican Spanish, "boludo" in Argentinean Spanish, and "causa" in Peruvian Spanish.

Similar to how a lot of swear words work in many languages.

It's interesting to see how what matters is not the word, but the intention behind it. At the end we are trying to communicate meaning, and words are just one of our tools to do it.

Of Course i don't know the data for GPT-4

Mandarin isn't the second language I speak but I officially compared with it because I wanted to test also with a model that had more equivalent corpus training than the very lopsided gpt models. And Chinese/English is the only combo that has a model of note in that regard.

and you can see talk on English/Japanese here - https://youtu.be/5KKDCp3OaMo?t=377

For almost everything else, it seems to produce pretty decent and usable translations, even when used against relatively obscure languages.

I used it a some green landic article that was posted on hn yesterday (about Greenland having gotten rid of daylight saving time). I don't speak a word of that language but the resulting English translation looked like it matched the topic and generally read like correct and sensible English. I can't vouch for the correctness obviously. But I could not spot any weird errors or strange formulations that e.g. Google translate suffers from. That matches my earlier experience trying to get chat gpt to answer in some Dutch dialects, Frysian, Latin, and a few other more obscure outputs. It does all of that. Getting it to use pirate speak is actually quite funny.

The reason that I used Chat GPT for this is that Google translate does not understand greenlandic. Understandable because there are only a few tens of thousands of native speakers of that language and presumably there's not a very large amount of training material in that language.

Therein lies the rub. There's a huge gap between what LLMs can currently do (spit back something in a target language that gives you the basic idea, however awkwardly phrased, of what was said in the source language). And what is actually needed for idiomatic, reasonably error-free translation.

By "reasonably error-free" I mean, say, requiring a human correction for less than 5 percent of all sentences. Current LLMs are nowhere near that level, even for resource-rich language pairs.

I ran the abstract of this article through chat gpt. Flawless translation as far as I can see. To be fair, Google translate also did a decent job. Here's the chat GPT translation.

Veel NLP-toepassingen vereisen handmatige gegevensannotaties voor verschillende taken, met name om classificatoren te trainen of de prestaties van ongesuperviseerde modellen te evalueren. Afhankelijk van de omvang en complexiteit van de taken kunnen deze worden uitgevoerd door crowd-werkers op platforms zoals MTurk, evenals getrainde annotatoren, zoals onderzoeksassistenten. Met behulp van een steekproef van 2.382 tweets laten we zien dat ChatGPT beter presteert dan crowd-werkers voor verschillende annotatietaken, waaronder relevantie, standpunt, onderwerpen en frames detectie. Specifiek is de zero-shot nauwkeurigheid van ChatGPT hoger dan die van crowd-werkers voor vier van de vijf taken, terwijl de intercoder overeenkomst van ChatGPT hoger is dan die van zowel crowd-werkers als getrainde annotatoren voor alle taken. Bovendien is de per-annotatiekosten van ChatGPT minder dan $0.003, ongeveer twintig keer goedkoper dan MTurk. Deze resultaten tonen het potentieel van grote taalmodellen om de efficiëntie van tekstclassificatie drastisch te verhogen.

Translating the Dutch back to English using Google translate (to rule out model bias) you get something that is very close to the original that is still correct:

Many NLP applications require manual data annotations for various tasks, especially to train classifiers or evaluate the performance of unsupervised models. Depending on the size and complexity of the tasks, these can be performed by crowd workers on platforms such as MTurk, as well as trained annotators, such as research assistants. Using a sample of 2,382 tweets, we show that ChatGPT outperforms crowd workers for several annotation tasks, including relevance, point of view, topics, and frames detection. Specifically, ChatGPT's zero-shot accuracy is higher than crowd workers for four of the five tasks, while ChatGPT's intercoder agreement is higher than both crowd workers and trained annotators for all tasks. In addition, ChatGPT's per-annotation cost is less than $0.003, about twenty times cheaper than MTurk. These results show the potential of large language models to dramatically increase the efficiency of text classification.

I'm sure there are edge cases where you can argue the merits of some of the translations but it's generally pretty good and usable.

I will be re-assessing my view on general-case translation performance accordingly.

- Retrieval augmented in-context learning

- Better benchmarks

- Last mile for productive application

- Faithful, human-interoperable explanations

[0] https://www.youtube.com/watch?v=-lnHHWRCDGk

It seems like it's a solved issue indeed. AI reasoning has still some way to go, but it seems language understanding is a finished subject.

Everyone shrugs and says, “nope, humans are different”. I’ve commented about 100 times recently asking for detail as to how human language / thought works, yet have not seen an answer.

That's basic linguistics and cognitive psychology. Nothing an LLM has done has invalidated that.

https://www.opensamizdat.com/posts/chatgpt_survey/