I think your strong definition of the Turing test is too strong to be useful. Consider the variance in the human population. Someone might be able to identify one specific model but not others. To require a model to be able to convince literally anyone is almost an impossible task. And I think it is also not necessary to have the dangers you describe. I won't ask an expert or Pangram to determine if a convincing phishing email has been written by AI. Maybe I will notice myself that it's a phishing mail, but this is no problem at all, as long as enough othe... (read more)
Hey Alessandro, thank you for this very interesting post! I just thought a lot about this yesterday. I was building an experiment where I want to measure the engagement of a model to receive a positively valenced stimulus. My original plan was to not deliver it, I saw no gain in it, I wouldn't get more information and only pay for the inference (and use downstream resources like energy and water). I said to myself that it doesn't matter, since it wouldn't have any influence on the model's behavior.
But Claude mentioned exactly your argument that this would... (read more)
I think the nature of the process is very different
You're right, gathering information by TV or AI is very different. But I think I see a continuous scale from one to the other. Maybe you can call it the sort of 'engagement' towards the medium. With TV we choose only one of let's say 50 channels. When informing myself on youtube I have millions of different channels. On reddit probably even more, and a richer interaction. Every step down leads to people believing more crazy conspiraicy theories, getting more and more delusional ideas. I don't think that th... (read more)
I'm not sure if I see really a qualitative jump of the TV to AI. Many people don't question the 'flashlight' of attention the TV is imposing and see it only as true (and only one) unbiased view onto the world. But you're right, the impression of "I control the question, thus the result should be authentic" may even lead to greater confidence in the answer given by AI. At the same time, a lot of people distant from AI are to my impression even under-confident, like "You cannot trust AI, because it hallucinates, so AI content is likewise false", which for su... (read more)
Interesting post.
But something that challenges me is that you place the inability of LLMs mainly on the behavioral side and argue that they don't learn the abilities necessary for expensive-loop tasks. I also think that this comes from RL training, but maybe it would be better to put the focus of the analysis on the cognition rather than on the behavior. In most situation the behavior follows the cognition (although also the reverse happens), thus we can search for upstream reasons for it.
You mention the importance of good theory building to solve expensiv... (read more)
How do you think your observation connects to metacognitive capabilities (not) present in current LLMs? I could imagine they fail at a lot of simple tasks because they are still missing an embedding in this world, especially during training. For sure, your agents are in a harness, as they are during a lot of RLVR training, but this is only very narrow access to this world. They learn to maximize some reward, but the reward is a proxy for some behavior, not for some cognition. There is a striking difference between "I try hard to not fail at that task" and ... (read more)