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

Killed by LLM

r0bk.github.io

Post reply on HN
Print view View original

71–80 of 102 posts

Re: Killed by LLM

#71

Post by michaelt »

spookie wrote: post #68
fastball wrote: post #52

Earlier quoted context omitted.

It's not a lack of emotional energy, it is the guardrails you point out. All of the SotA models are heavily fine-tuned to be botlike , and even then they are fooling people. If you had an LLM fine-tuned with RLHF to deliberately confuse humans in a Turing test it seems clear it would do a good job.

Why aren't the open source models like this then? Seems like it would've already happened. To me, at least, the guardrails are there for both the human and the bot. Without them the bot steers too far out of the conversation subject.

In most cases, "confuse humans in a Turing test" is counter to other more important goals.

Do you want your LLM to have an encyclopedic knowledge? So it knows who Millard Fillmore is even if the average human doesn't?

Do you want your LLM to be able to perform high-school-level math with superhuman speed and precision?

Do you want your LLM to be able to translate text to and from dozens of languages?

Do you want your LLM to be helpful and compliant, even when asked for something ridiculous or needlessly difficult - like solving "Advent of Code" problems using bash scripting?

If you answered yes to any of these questions, you probably don't want your LLM optimised to behave like an average human.

Re: Killed by LLM

#72

Post by ben_w »

silisili wrote: post #40
Ukv wrote: post #38

IMO a critical feature of the Turing test/imitation game, which many modern implementations including this site's linked paper ignore, is that the interrogator talks to both a human and a bot and must decide that one xor the other is a human. So fooling an interrogator means having them choose the bot as human over an actual human, not just judging the bot to be human (while probably judging humans to be human even m…

I'm skeptical on the claim. I think most folks, given the test you describe, would be able to pick out which is human. I think it can get there, but I'm not sure anyone has made one yet. ChatGPT responses are heavily downvoted and mocked because they're easy to spot. Does there exist a public LLM that isn't so...wordy, excited, and guardrailed all the time? You can pretty much spot the bot today by prompting somethin…

> Does there exist a public LLM that isn't so...wordy, excited, and guardrailed all the time?

Most of them, if you prompt them right, for that specific problem.

Most people don't bother, and instead treat them as if they're magic (they are "sufficiently advanced technology", but still), and therefore we get them emphasising "nuance" and "balance" where it doesn't belong.

> You can pretty much spot the bot today by prompting something horribly offensive.

Yes, though also each model's origins give a different idea of what counts as "horribly offensive". I'm thinking mainly because the Chinese models don't want to talk about Tiananmen Square as I've not tried grok (how does grok cope with trans/cis-gender as concepts? I know Musk doesn't, but it would be speculation to project that assumption onto the AI).

> Their response is always very inhuman, probably due to lack of emotional energy.

This, specifically, can also be faked fairly well with the right prompt. Tell ChatGPT to act like a human with clinical depression, and it does… at least by American *memetic* standards of what that means.

That said, ChatGPT and Claude are also trained specifically to reveal that they're AI, not humans, even if you want them to role-play as specific humans.

Probably for the best, given how powerful a tool they are for, e.g. phishing and similar scams.

Re: Killed by LLM

#73

Post by anon373839 »

yamrzou wrote: post #65

ARC-AGI is not yet killed by LLM. O3 achieved a breakthrough only on ARC-AGI-PUB , which is semi-private. Nothing guarantees that the test data wasn't leaked to OpenAI in previous testing rounds, because the model is not running offline. See: https://news.ycombinator.com/item?id=42478098

I think this should be discussed more. Models that can only be accessed via API cannot be tested without giving their owners access to the test data. You just have to trust that they’ll do the right thing.

Re: Killed by LLM

#74

Post by Tepix »

yamrzou wrote: post #65

ARC-AGI is not yet killed by LLM. O3 achieved a breakthrough only on ARC-AGI-PUB , which is semi-private. Nothing guarantees that the test data wasn't leaked to OpenAI in previous testing rounds, because the model is not running offline. See: https://news.ycombinator.com/item?id=42478098

See https://arcprize.org/blog/oai-o3-pub-breakthrough

ARC-AGI-1 will be replaced by ARC-AGI-2

So yes, ARC-AGI-1 was killed.

Re: Killed by LLM

#75

Post by Tepix »

anon373839 wrote: post #73
yamrzou wrote: post #65

ARC-AGI is not yet killed by LLM. O3 achieved a breakthrough only on ARC-AGI-PUB , which is semi-private. Nothing guarantees that the test data wasn't leaked to OpenAI in previous testing rounds, because the model is not running offline. See: https://news.ycombinator.com/item?id=42478098

I think this should be discussed more. Models that can only be accessed via API cannot be tested without giving their owners access to the test data. You just have to trust that they’ll do the right thing.

In particular, in cases where the model gets 16 hours to solve a task that a human can solve in a few minutes, cheating is trivial!

Re: Killed by LLM

#76

Post by foldr »

tripletao wrote: post #59
fastball wrote: post #49

Earlier quoted context omitted.

That's not the original Turing test either. The original imitation game as proposed by Turing involves reading a text transcript of a human and a computer and having the evaluator determine which is which. The evaluator does not interact directly with the conversing parties.

Where are you getting that? Turing's most famous paper is just as Ukv describes. The link on that site doesn't work for me, but the reference is buried in their source: https://courses.cs.umbc.edu/471/papers/turing.pdf In Turing's test, the forced binary choice means P(human-judged-human) + P(machine-judged-human) is necessarily equal to 100%. This gives the 50% threshold clear intuitive and mathematical significance…

It’s interesting that even though you link to the original paper, you still repeat a very common incorrect summary of the task.

The interrogator is not required to judge which of A or B is human, they are required to judge which is a woman on the implicit (though incorrect, in the case of interest) assumption that A and B are both human. While this amounts to more or less the same thing, it’s an interesting nuance that’s often lost in summaries of the task. It would not, for example, make sense for the interrogator to ask A or B whether or not they are human (even on the naive assumption that they’d receive a true answer), as they are working on the assumption that both are human. Hence why Turing’s initial example questions are about hair length and gender, not humanness.

To be fair, even Turing himself seems to imagine the interrogator trying to judge humanness rather than gender in subsequent parts of the paper. It’s unclear to me why exactly his initial framing of the task introduces this additional element of complexity.

Re: Killed by LLM

#77

Post by yamrzou »

Tepix wrote: post #74
yamrzou wrote: post #65

ARC-AGI is not yet killed by LLM. O3 achieved a breakthrough only on ARC-AGI-PUB , which is semi-private. Nothing guarantees that the test data wasn't leaked to OpenAI in previous testing rounds, because the model is not running offline. See: https://news.ycombinator.com/item?id=42478098

See https://arcprize.org/blog/oai-o3-pub-breakthrough ARC-AGI-1 will be replaced by ARC-AGI-2 So yes, ARC-AGI-1 was killed.

ARC-AGI-2 was planned long before those results came out. Also from the link: ARC-AGI-2 (same format - verified easy for humans, harder for AI) will launch alongside ARC Prize 2025. So, no, it will not replace it.

Re: Killed by LLM

#78

Post by d0mine »

fastball wrote: post #49
Ukv wrote: post #38

IMO a critical feature of the Turing test/imitation game, which many modern implementations including this site's linked paper ignore, is that the interrogator talks to both a human and a bot and must decide that one xor the other is a human. So fooling an interrogator means having them choose the bot as human over an actual human, not just judging the bot to be human (while probably judging humans to be human even m…

That's not the original Turing test either. The original imitation game as proposed by Turing involves reading a text transcript of a human and a computer and having the evaluator determine which is which. The evaluator does not interact directly with the conversing parties.

The original Turing game is whether machine can pretend to be a woman better than a man can (via teletype) as judged by an interrogator:

> We now ask the question, "What will happen when a machine takes the part of A in this game?" Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman? These questions replace our original, "Can machines think?”

https://courses.cs.umbc.edu/471/papers/turing.pdf

Re: Killed by LLM

#80

Post by alganet »

An very reliable, very unethical test would be to deploy LLMs on the internet as humans and gauge how other humans react (ignore, call out as LLM, engage, etc). There isn't much in the way of stopping a company from doing that (there should be!).
Post reply on HN