Comment by gmuslera | Hacker News Reader

Comment by gmuslera | Hacker News Reader

In one hand, we are past the Turing Test definition if we can't distinguish if we are talking with an AI or a real human or more things that were rampant on internet previously, like spam and scam campaigns, targeted opinion manipulation, or a lot of other things that weren't, let's say, an honest opinion of the single person that could be identified with an account.

In the other hand, that we can't tell don't speak so good about AIs as speak so bad about most of our (at least online) interaction. How much of the (Thinking Fast and Slow) System 2 I'm putting in this words? How much is repeating and combining patterns giving a direction pretty much like a LLM does? In the end, that is what most of internet interactions are comprised of, done directly by humans, algorithms or other ways.

There are bits and pieces of exceptions to that rule, and maybe closer to the beginning, before widespread use, there was a bigger percentage, but today, in the big numbers the usage is not so different from what LLMs does.

Recently I’ve been thinking about the text form of communication, and how it plays with our psychology. In no particular order here’s what I think:

1. Text is a very compressed / low information method of communication.

2. Text inherently has some “authority” and “validity”, because:

3. We’ve grown up to internalize that text is written by a human. Someone spend the effort to think and write down their thoughts, and probably put some effort into making sure what they said is not obviously incorrect.

Intimately this ties into LLMs on text being an easier problem to trick us into thinking that they are intelligent than an AI system in a physical robot that needs to speak and articulate physically. We give it the benefit of the doubt.

I’ve already had some odd phone calls recently where I have a really hard time distinguishing if I’m talking to a robot or a human…

This is absolutely why LLMs are so disruptive. It used to be that a long, written paper was like a proof-of-work that the author thought about the problem. Now that connection is broken.

One consequence, IMHO, is that we won't value long papers anymore. Instead, we will want very dense, high-bandwidth writing that the author stakes consequences (monetary, reputational, etc.) on its validity.

The Methyl 4-methylpyrrole-2-carboxylate vs ∂²ψ/∂t² = c²∇²ψ distinction. My bet is on Methyl 4-methylpyrrole-2-carboxylate being more actionable. For better or worse.

Sorry, I have absolutely no idea what you're trying to say with that.

But that’s not the Turing Test. The human who can be fooled in the Turing test was explicitly called the “interrogator”.

To pass the Turing test the AI would have to be indistinguishable from a human to the person interrogating it in a back and forth conversation. Simply being fooled by some generated content does not count (if it did, this was passed decades ago).

No LLM/AI system today can pass the Turing test.

I've encountered people who seem to understand properly how the test works, and still think that current LLM passes it easily.

Most of them come across to me like they would think ELIZA passes it, if they weren't told up front that they were testing ELIZA.

I think state of the art LLMs would pass the Turing test for 95% people if those people could (text) chat to them in a time before LLM chatbots became widespread.

That is, the main thing that makes it possible to tell LLM bots apart from humans is that lots of us have over the past 3 years become highly attuned to specific foibles and text patterns which signal LLM generated text - much like how I can tell my close friends' writing apart by their use of vocabulary, punctuation, typical conversation topics, and evidence (or lack) of knowledge in certain domains.

on one hand

on the other hand

Are you implying that this was an AI bot comment?

I think it was a grammar/idiom correction.

Reply on news.ycombinator.com