I chose RC boat because some people are using this competing at "stupid human trick" as an example of intrinsic proof of LLMs being able to supersede humans.

In the scenario of an olympic swimming competition, an autonomous boat vs a manned boat would show no difference to each other, both would compete the task much better than a human. It's a useless test to measure general utility. Just like a person swimming a 1500 meter distance is not really a useful indicator on its own of how useful they are. These Math Olympiads are similar in that they are not particularly indicative of people being useful. The same ways that we would stress a human in impressive ways does not mean that a computer coming at it from a different approach should be considered to have broadly superseded the humans.

Yes, a real boat is valuable and LLM can be valuable when utilized correctly, but in the face of exaggerated hype some pessimistic reality check is called for to balance expectations.

LLMs have a great deal of utility and extend the reach of computing to a fair amount of scope that was formerly out of reach to computing, but they don't "think" and the branding of the "reasoning" models is marketing, not substantive.

The best evidence is reviewing so-called "reasoning chains" and how the mistakes behave.

Mistakes are certainly plausible in "true thinking", but the way they interact with the rest of the "chain" is frequently telling. It flubs a "step" in the reasoning and if it were actual reasoning, that should propagate to the rest of the chain. However when a mistake is made in the chain, it's often isolated and the "next step" is written as if the previous step said a correct thing, without ever needing to "correct" itself or otherwise recognize the error. What has been found is that if you have it generate more content and dispose of designated "intermediate" content you have a better result, and the intermediate throwaway content certainly looks like what a thought process may look like, but ultimately it's just more prose and mistakes in the content continue to have an interesting behavior of isolation rather than contaminating the rest of an otherwise ok result.

I think most people can do long division, but almost no one would bother day to day, because why would they?

When the model is used for inference, yes. But I assume he was speaking to the awkwardness of training. Take a machine vision that has never been trained on dogs and cats, feed it a dozen labeled images of cats and dogs to retrain it to add dog/cat recognition. Then try to do inference on that model and it will be utterly useless still for dog/cat recognition. Take a model trained on normal images. Then have it try recognition on a fisheye lense. It will fail because it has no idea. You might hope to retrain it to recognize the fisheye distortion generically, but generally that won't work and you have to retrain it to catch the fisheye variant of everything you want. After retraining, if you just slap the fisheye variant into a normal picture, the model is likely to be oblivious to the anomaly.

Don't know if this is an argument about 'thinking', but it is certainly a difference in 'learning', that AI models need to consume way way more than a human to start to give useful results. A model training to operate a car is still oddly dodgy even after consuming more car operating hours than a human will consume in a lifetime. In areas where we have enough training data to feed, this is fine, but it's certainly different than a human learning, which can do something better to extend less training to better results.

I'm not in denial, the LLMs and other forms of AI have utility, but expectations have to be mitigated.

Was in a discussion with a software executive a couple weeks back who said he fully anticipates he can lay off every one of his software developers and testers in the next year and only have to retain the 'important' people: the executives and sales people.

People see articles like this show how LLMs enable computing to reach another tier of 'stupid human tricks', which is certainly novel, but people overextend exactly what this means. We put too much focus on whether the LLMs can navigate tests explicitly designed to be graded and passable for humans, with effort and imagine that means they are able to chart the less well trodden set of problems that have an unknown or maybe no solution at all.

There's a lot of stuff LLMs can do that computing wasn't able to do before, and that does potentially speak to a great deal of the things humans do today, but we have to temper the wild expectations people are getting.

The point is we have a myriad of "tests that are hard for humans, but don't necessarily translate to anything vaguely useful". In academics, a lot of tests are only demanding of reasoning ability because the human has limited memory. Computers short on actual "reasoning" largely make up for it by having just mind boggling amounts of something more akin to recall than reasoning (it's something a bit weirder, but as far as analogies go, recall is closer).

It's kind of like bragging that your RC boat could get a gold medal in the Olympic 1500 meter freestyle. It didn't complete the challenge in the same way, and that boat would be unable to, for example, save someone that is about to drown, because the boat can just go places, it can't do what that human swimmer could do. That person swimming 1500m itself is a useless feat of interest, and not really directly useful in and of itself.

If it needs USB-PD, looks like the products are pretty pricey.

If it can take the old fashioned USB power, looks like a splitter is reasonably available for like $15.