red75prime

They probably can't order tofu. It would be B12-fortified oatmeal or something like this.

where and when the LLM use is justified and where it is not

while being bombarded with articles like "AI makes things worse", "AI consumes all the water" and the like

So what if your image classifier thinks a cat is an impala?

For starters, the artificial neural network is the only instrument we have for sufficiently robust image classification. If it doesn't work nothing else will. We can only validate and retrain.

Getting back to self-driving. What if an occupancy network misclassify an impassable surface patch as passable? It would mean a possible accident, so you bias the network towards impassable classification. In this way misclassification is overwhelmingly more likely to cause spurious braking, unneeded swerve, and the like rather that a collision ("brake stabbing" was a recurring problem in Tesla FSD. Today it is largely mitigated.)

The fact they can't is the strongest evidence that whatever proprietary secret sauce moat they have... doesn't really work.

Well, Waymo has publicly embraced world models, VLMs and end-to-end backpropagation relatively recently (December 9, 2025)[1]. There's not enough data yet to judge how this has influenced their expansion rate. Waymo representatives tout improved generalization (sorry, I can't find the interview. I think it was Dmitri Dolgov.)

if their systems were really autonomous they could just deploy them anywhere a human driver can drive

As I've said it's a matter of degree. Human-caused road casualties are a familiar and largely inevitable evil. An autonomous vehicle killing a person will (and already has) cause a media storm. They can't ignore a potentially problematic distribution shift caused by deployment in a new area.

You are talking about the perfect system. Real systems are computationally constrained. Distillation necessarily degrades performance and might require retraining for particular areas.

[1] https://waymo.com/blog/2025/12/demonstrably-safe-ai-for-auto...

https://youtu.be/C4NQNeSO2vs?si=epkxhVXpypOCppGW

"VLAs fine-tuned on human demonstrations overfit. Here's how to mitigate it." Er, OK, I guess. But I think that Waymo uses mitigations or a different approach (RL, for example).

I did crack up seeing a martini glass presented to the camera, then float in place on it's own

It's just a cheap video-generation model doing its thing. I wonder if the orchestrator model noticed this and tried "Create a scene with a martini glass. Make no mistakes".

Non-trivial percentage of people work in the real world until they do something stupid and work no more. It's not real autonomy, not yet.

It's a matter of degree. Sorry, I don't want to watch an hour long video to be told how something that works 99.99% of the time doesn't really work for some contrived definition of "really".

BTW, as is typical with people, remote operators occasionally cause problems.

ML is as thick as plankton by this definition

...during autoregressive pretraining ([2]). A model pretrained on texts with 1930 data cutoff can solve a few programming problems when given a few examples. Its success rate is understandably much worse.

[1] https://talkie-lm.com/introducing-talkie

[2] That is when a model starts from a blank state that can be described in a few kilobytes.

real-world autonomy isn't really working

It works in limited ways (but in the realest-real world nevetheless). Waymo, Wayve, Baidu Apollo, Tesla and others seem to rely on VLAs, VLMs or transformer models in general to do autonomous driving.

if discovering the required technology to build a self-replicating interstellar probe requires solving an instance of an EXP problem[...]

Ah, I think I know why I felt it was very unlikely. Engineering constraints should be very "unlucky" so that only the solution works. That is, there's no wiggle room that allows an approximate solution (which is usually simpler to find) to work, even if it would be slightly heavier, costlier, or something of the sort.

For such a tight constraint not to be a coincidence, nature itself would have to be able to solve EXP problems. Which might mean that not all is lost.

ETA: ChatGPT has pointed out that many interesting instances of computationally hard problems are still solvable. For example, see SAT-solvers (SAT is NP-complete).

That is in addition to tight constraints the problem instance should have no exploitable structure.

It's hard to solve P?=NP due to P!=NP.

Protecting forests is good and all, but it will not reverse climate change. We'll have to stop all fossil fuel burning and, additionally, sequester around 600 billion tonnes of CO2 (in addition to all the natural carbon sinks) to get back to +1.5C by 2100.

they still wanted to see someone in person.

Well, that someone might use AI diagnostic tools too.

there is no reason to believe that the way to build it would be found before, say, the sun runs out of hydrogen.

A lot of computational power can be thrown at the problem in this time. So, the problem should admits no shortcuts, no decomposition into simpler problems, no alternative ways to get similar functionality that allow shotrcuts or decomposition. The result should look like a jumble of atoms that somehow produces the required functionality.

I wonder which problems can admit only this kind of a solution.

Autoregressive pretraining (text/images/video prediction) produces a foundational model. You can look at it as a highly compressed conditional probability distribution of the human brain output. The information-theoretically optimal compression of the data is a program that reproduces functionality of the process that generated the data.

So, it stands to reason (and observations) that such a model captures not only surface statistics of the data, but a part of functionality of the system that generated the data (the human brain for text, physics for video).

It wouldn't be "Ah! It's lupus." It would be "I need the results of such-and-such tests to make a differential diagnosis."

How is this not obvious to everyone?

That RSI can be bottlenecked? I guess this is obvious to many people. Whether RSI will be bottlenecked (at some not very interesting stage) is another question.

Noise begins to settle down... "Continual RL is all you need" paper comes out.

By achieving the same or a higher level of effectiveness than doctors, obviously.

How else would you trust a tangled mess of neurons to produce something that works? Artificial neurons, of course, hehe.

From my POV it's "we haven't found physical processes that allow to do information processing, which is relevant to learning, exponentially faster than digital computers and we aren't likely to find them in the brain."

Or, at least, I don't tend to postulate existence of such mechanisms in the brain after hearing about every empirical ANN deficiency.

BTW, I'm not talking about supernatural. The "magic" could be unknown physical processes, unknown quantum algorithms, unknown physics.

Stochastic gradient descent can be likened to traveling down a billion-dimensional canyon. But inference? Hardly.

Yep, I'm fairly certain that general learning algorithms acting upon an ANN (which is fairly general too, see the universal approximation theorem) can reach and surpass performance of the human brain. As we have approximately zero evidence that the human brain contains "magic," that is something that can't be modeled by an ANN of a practically feasible size. (I know about chaotic processes that can't be modeled precisely. The "magic" here would be the brain using such a process to make useful decisions. "Useful" not in a sense of a mixed strategiy. You can roll dice to do that.)

But, no, it's not "any day now." The required size and structure of the ANN is to be determined.

People are so confident that this just-a-tool will hit its limits any day now...

Attempts to formalize dialectics do exist, but it mostly stays at a word-weaving level.

Classical physics doesn't have particles that are simultaneously here and not here. It's a discrepancy between theory and experiment. And there's no point of accepting contradictory statements that are both true to deal with that. You can't fix a wrong description of reality by using some fancy logic that allows (apparent) contradictions.

things could be true and false at the same time at other scale.

Being true and false at the same time is a contradiction. But yeah, there is such a thing as mathematical intuitionism that rejects the law of excluded middle (which is not "being true and false at the same time"). It's just one philosophical stance among others though.

Kenyans who provided the data for RLHF liked that style. That's the most of it. And a transformer is not an "averaging machine", it's a prediction machine.

What about the German confederation? A list of countries with a governing body is a bit more than a list.

I guess we can use a bit of 44 petawatts of solar energy reaching Earth in electric cars.

Why so sensitive? There are other battery chemistries (say, VRFB) that are less dangerous and that are less like tire fire.