François Fleuret, Meta FAIR research scientist and University of Geneva professor, argues that pure AI systems outperform any human-AI collaboration on sophisticated reasoning tasks

What can a human communicate to an AI stronger than them: information to which they both have access, which is then not necessary, or information to which the user has privileged access, that is the human's internal thoughts.

1/2

This internal thoughts are either the results of the human's own reasoning, which are "worse" than those of a stronger AI, or idiosyncratic facts (e.g. what this human likes or wants).

So IMO there will be no "collaboration", only "armchair leadership".

2/2

@francoisfleuret There's an argument from comparative advantage. If the energy available to AI is finite, it's rational to leave low-value tasks to humans so that AI could focus it's limited energy on more high-value tasks -- https://www.noahpinion.blog/p/plentiful-high-paying-jobs-in-the

@negatingspirit @cdriclion @francoisfleuret Indeed, you introduced « of economic relevance », but I wondered whether that limitation was relevant re @francoisfleuret ´s xweet.

@francoisfleuret Humans are the end point of AI work. If nobody understands what AI has written it has no value… Somebody should understand what AI has written and say to others yep, dudes, it totally makes sense.

@cdriclion @francoisfleuret Why should it be absurd? I am stronger than a snail on all resoning tasks of economic relevantce. You maybe too.

@francoisfleuret Neural implants negate the problem here. Mental exercises to keep reasoning alive in the brain will be more common, in case of issues with your neural implants.

@negatingspirit @cdriclion @francoisfleuret The snail probably knows better than you how to plan its path in the garden.

The idea that things in the real word can be well ordered that way is so naive that your all argument collapses to the ground instantly

saying 'AI > humans' is like saying 'pear > apple' and concluding that apple argiculture will desappear as a logical consequence

So I'd say that rationality so far is on their side more than on yours

Well, would need to define accurately what he means by "sophisticated reasonning problems" to discuss properly.

But the point is: his position is very fragile as it is based on absolute superiority of AI, which breaks with only one counter example. And humans work so much differently than (current) AI that it's very likely that it remains superior at least for some very specific dimension(s) for quite some time.

But again, depends on how widely you define "sophisticated reasonning problems"

@civicheathen @francoisfleuret Neural implants is just a pipe dream

@francoisfleuret Hey François, As a true orthogonalist I keep my phone parallel to Earth while standing vertical — perfect 90°. Humans and AIs aren’t scalars. In high dimensions, if we’re roughly orthogonal, |H + A| > |A|. Let’s collaborate perpendicularly!

@negatingspirit @francoisfleuret Example : AI have biases

Just like humans, but not the same.

As long as AI has a bias that human can help to remove, the reasoning breaks.

@francoisfleuret The internal thoughts are the only chip you have left. Everything else the model already ingested during training.

@francoisfleuret There is nothing more nonlinear than that!

@francoisfleuret there will be a parents child relationship not a deal in an economic sense.

@francoisfleuret That’s if you believe the production function is additive (substitutes) rather than multiplicative (complements)

@francoisfleuret I suspect collaboration will be needed when user preferences matter deeply, or when the necessary context is too tedious to fully specify.

For example, a scientist can't realistically encode 30 years of intuitions and direct experimental experience into a computer.

@francoisfleuret So in problems like current Erdos one, AI made the initial breakthrough where pure reasoning was required to find counter example and 1 day later a very massive improvement was done by human by purely intuitive thinking.

@francoisfleuret You are assuming a very specific kind of scenario that is not current.

Humans are exceptional cross domain transfer learners. This gives an ungodly advantage over reasoning. It is the basis of intuition.

AI is better at pure reasoning than humans.