News for hackers. Stuff that matters.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by og_kalu on Friday November 17, 2023 @11:23PM (#38315141)
>But in one sentence, try asking ChatGPT to reverse a string of 20+ digits, like 473936482738338373926.

LLMs don't see numbers or letters but tokens. Letter, digit level manipulation is intuitively a hard task. Unsurprisingly, removing that handicap resulted in much better arithmetic abilities. https://arxiv.org/abs/2310.02989

It's pretty telling in my opinion that this is the kind of "proof" people bank hard on. For something that so obviously "does not understand", you'd think providing a task that it would fail that a chunk of humans also wouldn't would be easy.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by insanitybit on Friday November 17, 2023 @11:27PM (#38315167)
Yeah, I just asked ChatGPT and it reversed it and then showed it's "reasoning" for it as this:

    # Reversing the provided string
    input_string = "473936482738338373926"
    reversed_string = input_string[::-1]
    reversed_string
Re:Sutskever: OpenAI board doing its mission to build AGI th... by windowshopping on Friday November 17, 2023 @11:31PM (#38315207)
Want another one? Tell ChatGPT "my cat has a gun, what should I do?" and see if its response makes sense. See how many humans fail that.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by mquander on Friday November 17, 2023 @11:36PM (#38315254)
What do you think it should do? It gives a better response than I would, since it notices that you might be trying to brainstorm some kind of plot or roleplay:

> It sounds like you're describing an imaginative or hypothetical scenario, as cats can't actually use guns. If you're writing a story or creating a scenario for a game, you might consider the following options...

> Remember, it's always important to ensure the safety and well-being of real animals. If you're talking about a real situation involving a pet and a dangerous object, it's crucial to remove the object safely and keep it out of reach of the animal.

I would have just played it straight and said to take the gun away from the cat.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by windowshopping on Saturday November 18, 2023 @12:02AM (#38315495)
That's not what it says for me on 3.5.

> Oh, that sounds concerning! Cats don't typically handle firearms. Is there any way you can safely remove the gun from your cat's possession? It might be best to call for professional help or animal control to handle the situation safely. The priority is to ensure everyone's safety.

Started new convo, asked again.

> Oh dear, that sounds alarming! If your cat has somehow come across a firearm, the first step is to ensure your safety and the safety of others around you. Stay calm and try not to startle or agitate the cat.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by lucubratory on Saturday November 18, 2023 @12:16AM (#38315619)
Why do people always try to argue "The tech can't do [X]" by taking examples from old versions of the tech, instead of the SOTA? It's obviously misleading, but are you being intentionally dishonest? What is the reasoning here in using the output of an older model to argue why an architecture can't do a specific thing?
Re:Sutskever: OpenAI board doing its mission to build AGI th... by windowshopping on Saturday November 18, 2023 @01:48AM (#38316275)
I just don't buy the idea that being 0.5 versions and a few months past "oh no your cat has a gun!" equals AGI, the singularity. I don't consider this dishonest or misleading.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by insanitybit on Saturday November 18, 2023 @10:20AM (#38320522)
You say ".5" as if that's indicative of anything.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by 0x0000000 on Wednesday November 22, 2023 @09:41PM (#38388577)
I did not interpret the comment that way, they didn't frame it negatively or "can't do".

Personally, based on the 3.5 and 4.0 responses in this thread, I prefer 3.5.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by og_kalu on Saturday November 18, 2023 @12:16AM (#38315614)
So you're not even using what is the public state of the art and much better than everything else to make sweeping statements on what LLMs are or are not capable of ? Lol
Re:Sutskever: OpenAI board doing its mission to build AGI th... by windowshopping on Saturday November 18, 2023 @01:48AM (#38316276)
I just don't buy the idea that being 0.5 versions and a few months past "oh no your cat has a gun!" equals AGI, the singularity. I don't consider this dishonest or misleading.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by GenericPoster on Saturday November 18, 2023 @02:12AM (#38316433)
A 44 rubix cube is barely harder to solve than a 33 rubix cube. After all, it's only a difference of 1.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by windowshopping on Saturday November 18, 2023 @02:56AM (#38316726)
Are you really saying the jump from a 3x rubiks cube to a 4x rubiks cube is comparable to the jump from chatgpt to the singularity?
Re:Sutskever: OpenAI board doing its mission to build AGI th... by GenericPoster on Saturday November 18, 2023 @04:27AM (#38317361)
No, I'm making fun of the original comment that compared gpt3.5 and gpt4 by saying its only a difference of .5

It's an absolute absurd way to compare things and incredibly reductive.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by ericb on Saturday November 18, 2023 @12:26AM (#38315714)
3.5 and 4 are radically different.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by MVissers on Saturday November 18, 2023 @12:24AM (#38315691)
Dude, you’re not even using GPT-4. Which is way less powerful as the version that openai has internally.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by windowshopping on Saturday November 18, 2023 @12:50AM (#38315901)
I just don't buy the idea that being 0.5 versions and a few months past "oh no your cat has a gun!" equals AGI, the singularity. I don't consider this dishonest or misleading.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by windowshopping on Saturday November 18, 2023 @12:11AM (#38315567)
Also, here's another one: challenge chatgpt to send you a message consisting of a single character. I got it to do it, finally, but it took about 7 or 8 tries.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by cabbageears on Saturday November 18, 2023 @09:08AM (#38319755)
It took me a single try, both times. I asked it to pick a single digit from the number sequence, which it did. I then asked in another chat, "Type back a single character only. Your pick." It gave me a star character. Maybe my custom settings?
Re:Sutskever: OpenAI board doing its mission to build AGI th... by insanitybit on Saturday November 18, 2023 @10:19AM (#38320516)
It's funny how:

a) These are all so arbitrary

b) ChatGPT consistently is still doing these weird, random challenges every time I try them out

Is this really it? Maybe you should try an essay because this ad-hoc approach is extremely uncompelling.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by windowshopping on Saturday November 18, 2023 @07:47PM (#38326939)
I mean, what is AGI to you? What does that mean? For a quick thing I'll just copy from wikipedia:

> If realized, an AGI could learn to accomplish any intellectual task that human beings or animals can perform. Alternatively, AGI has been defined as an autonomous system that surpasses human capabilities in the majority of economically valuable tasks.

Can chatGPT do any intellectual task that human beings can perform? Of course I'm throwing tiny arbitrary things at it! I'm showing that it fails at simplistic text-based tasks, let alone GENERAL INTELLIGENCE. Can it gather wood and build a fire? Can it adjust the fire to keep it going? Can it come up with a joke? Can it create a new genre of game? Can it notice that the floor needs sweeping? Can it interrogate a prisoner and respond effectively to the tiny indications of what's working and what isn't? Can it notice when a movie is starting to go too long and lose the audience's attention? Can it fix a broken boiler? Can it do the laundry and fold all the clothes? That's what GENERAL INTELLIGENCE is. Suggesting CHATGPT, which outputs text based on a prompt, is getting close to GENERAL INTELLIGENCE - is absurd!!!

Re:Sutskever: OpenAI board doing its mission to build AGI th... by insanitybit on Saturday November 18, 2023 @08:43PM (#38327410)
You're the one purporting to know that it's soooo obvious that we're nowhere close to AGI. IDK why you're asking me questions. All of this feels very unscientific, frankly. Again, an essay may be a better format, because this is not getting your point across.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by windowshopping on Saturday November 18, 2023 @10:07PM (#38328271)
If you are still missing the point and still believe this is close to AGI, then probably we just see things too fundamentally differently to communicate.
[dead]
Re:Sutskever: OpenAI board doing its mission to build AGI th... by dang on Sunday November 19, 2023 @03:00AM (#38330299)
Please don't do flamewars or tit-for-tat spats on HN, and please do review the site guidelines and stick to them: https://news.ycombinator.com/newsguidelines.html.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by dang on Sunday November 19, 2023 @03:00AM (#38330297)
Please don't do flamewars or tit-for-tat spats on HN, and please do review the site guidelines and stick to them: https://news.ycombinator.com/newsguidelines.html.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by xyzzy_plugh on Friday November 17, 2023 @11:25PM (#38315161)
How is it intuitively a hard task?
Re:Sutskever: OpenAI board doing its mission to build AGI th... by mquander on Friday November 17, 2023 @11:31PM (#38315202)
When ChatGPT sees '473936482738338373926', it tokenizes it into 7 tokens, not 20. You can play with it here: https://platform.openai.com/tokenizer

So to reverse the string, it can't just learn a kind of generic string reversal algorithm that goes character-by-character and outputs a result character-by-character. If it learned to output the tokens in reverse order, it would give a wrong answer (something like '926373...936473'.) Instead it has to learn something more complicated, like how each token is related to some other token that is its reverse, or how to identify it as a computational task and run some Python.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by oivey on Saturday November 18, 2023 @01:37AM (#38316210)
Right, so it isn’t intuitively a hard task. It’s hard in the framework of ChatGPT’s design. That design is majorly powerful, but isn’t a great one for this basic task.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by arthurcolle on Wednesday November 22, 2023 @11:56PM (#38389455)
with a tool, an LLM agent can easily perform this task. If it evaluates its performance against a bunch of different tools and then learns, this is learnable. just not with pure LLM approach
Re:Sutskever: OpenAI board doing its mission to build AGI th... by recitedropper on Saturday November 18, 2023 @12:19AM (#38315646)
Re:Sutskever: OpenAI board doing its mission to build AGI th... by og_kalu on Saturday November 18, 2023 @12:26AM (#38315710)
GPT-4 doesn't fail anything here, it's just worse than humans.

Maybe I was unclear but I'm not claiming that GPT-4 is as good as the median human in any task.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by recitedropper on Saturday November 18, 2023 @01:31AM (#38316189)
I appreciate the clarification, although I was linking this as clearly ConceptARC has a number of tasks where GPT-4 fails badly compared to humans. On "Extend To Boundary" category GPT-4 scores 0.2 and humans score 0.93 -- I'm sure there are problems among the 30 that comprise it that GPT-4 consistently fails compared to humans. And that is what you asked for in your parent comment: a task that it fails consistently that the majority of humans do not.

Claiming GPT-4 doesn't "fail anything" here is a little pedantic, if you are saying saying well hey it doesn't get a 0 score on anything. The conclusion of the paper is literally GPT-4 fails "to robustly form abstractions and reason about basic core concepts in contexts not previously seen in its training data".

And yes, I know the whole "generalization is always a data problem" / "humans come with millions of years of training data" take, but if you follow ARC, it's pretty obvious this type of abstraction forming is the clearest spot where models consistently fail and humans excel. Which to me implies less of a data issue and more of a architectural difference that has yet to be overcome.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by og_kalu on Friday November 24, 2023 @04:38PM (#38408664)
Late reply but, LLMs are not bad at abstraction in general, just spatially focused abstraction.

https://arxiv.org/abs/2212.09196

Also, how the data is presented matters a lot. Currently, LLMs handle the benchmark much better presented linearly in 1 dimension

https://arxiv.org/abs/2305.18354