How does AI DO it? Diagnosing electronic devices.

[watchmaker]

I typed in a question about an issue with my curve tracer.  The conversation seemed on point and specific to my unit.  Down at least 4 layers.  I feel like my now passed friend who started flying biplanes off the Langley as an enlisted pilot and retired commanding a squadron of jets.

[tom66]

Terabytes of training data and billions of dollars, essentially.   The increased 'knowledge' that large AI models possesses, even in subjects that they weren't directly trained on, or have only minimal data on, seems to be an emergent property of these models.

[Haenk]

AI is a misconception. There is not a lot of "intelligence" involved. More like an "auto-complete" system with a gigantic reservoire of phrases and already written content.
The main science is the weighting "which auto-complete makes the most sense for the user".

So receiving a detailed instruction of how to repair a certain device means one thing: Someone wrote at some time about repairing this device. Don't ever think AI figured it out for you.

[Randy222]

Agreed as @Haenk mentions.

AI works fast like inference Markov chain math.

Derivations still need compute power, which is moving fast into super-super-computing realm.

There's technically no I in AI, it just appears that way, no different than thinking the "cloud" server is in the sky when it's really just some real compute hardware/software in a datacenter somewhere.

Mobile AI bots, even when connected to internet, cannot do the computing piece fast enough (compute too small, wifi link too slow, not enough ssd), humans will always be better.

What we are seeing now, bots taking over very specific narrow tasks (put bumper on car at factory, wash the windows, pickup trash). The humanoid trained to vaccum and do the dishes, is not suddenly gonna change your flat tire when you say "I have a flat, go fix it". In the future that may be possible, but it will take hrs or days for such bot to download the knowledge and find the right tools, then attempt it w/o previous experience. Currently that's a fail.

That said, it does not mean we don't need the big AI centers. Such centers are however creating an energy issue, it does appear more nukes are coming because its the only way.

[tom66]

AI is a misconception. There is not a lot of "intelligence" involved. More like an "auto-complete" system with a gigantic reservoire of phrases and already written content.
The main science is the weighting "which auto-complete makes the most sense for the user".

So receiving a detailed instruction of how to repair a certain device means one thing: Someone wrote at some time about repairing this device. Don't ever think AI figured it out for you.

Except that is not really true, not when referring to modern models.  They possess emergent capabilities, able to solve puzzles and infer knowledge that is not part of the training data.  In many senses that could be referred to as 'intelligence', although I suppose it depends on whether you accept that any machine system could be given that attribute. 

For instance, AI models are getting remarkably good at rendering the so-called "Pelican test", which is an SVG image of a pelican riding a bicycle, something that obviously cannot exist.  This is despite no training data containing what this should look like.  The model is using its internal reasoning to determine what this should appear as.

https://github.com/simonw/pelican-bicycle
https://share.gemini.google/xKYkOVJ4GpN0 (Gemini 3.1 Pro, just rendered)

AI models are also able to translate languages they could never have seen before, when given some information on the language in their context:

https://luciusssss.github.io/proj/zhuangbench.html

It remains an active area of research as to how models develop these emergent capabilities.

[KE5FX]

AI is a misconception. There is not a lot of "intelligence" involved. More like an "auto-complete" system with a gigantic reservoire of phrases and already written content.

Being able to write something like that at this point is nothing short of self-delusion, unbecoming of technical people in general and engineers in particular.  Let's see how your "autocomplete" and your Markov chains do on Erdős problems.

[Zero999]

For instance, AI models are getting remarkably good at rendering the so-called "Pelican test", which is an SVG image of a pelican riding a bicycle, something that obviously cannot exist.  This is despite no training data containing what this should look like.  The model is using its internal reasoning to determine what this should appear as.

Then it is a fail, not a pass.

[Analog Kid]

AI is a misconception. There is not a lot of "intelligence" involved. More like an "auto-complete" system with a gigantic reservoire of phrases and already written content.
The main science is the weighting "which auto-complete makes the most sense for the user".

So receiving a detailed instruction of how to repair a certain device means one thing: Someone wrote at some time about repairing this device. Don't ever think AI figured it out for you.

Except that is not really true, not when referring to modern models.  They possess emergent capabilities, able to solve puzzles and infer knowledge that is not part of the training data.  In many senses that could be referred to as 'intelligence', although I suppose it depends on whether you accept that any machine system could be given that attribute. 

That is absolutely the worst kind of woo, that use of the term "emergent".

Woo-woo: guaranteed to send a shiver down our spines and make us think that there's more to "AI" than meets the eye, as if there just might be some kind of "ghost in the machine" that's been lurking there all along down humanity's long march of technology, only to be released by finally crossing some kind of threshold: computing power? number of inference connections? millions of dollars invested? some other unknown parameter?

Who knows? Who cares? It means that "AI" is finally baked, finally existent, finally "emergent". Never mind that we still really have no fucking idea how that could come about. If it seems to work, ship it! Promote it! Invest in it!

[KE5FX]

Your questions have answers, but you'll have to look someplace besides this thread for them.

A good place to start:

[tom66]

That is absolutely the worst kind of woo, that use of the term "emergent".

Woo-woo: guaranteed to send a shiver down our spines and make us think that there's more to "AI" than meets the eye, as if there just might be some kind of "ghost in the machine" that's been lurking there all along down humanity's long march of technology, only to be released by finally crossing some kind of threshold: computing power? number of inference connections? millions of dollars invested? some other unknown parameter?

Who knows? Who cares? It means that "AI" is finally baked, finally existent, finally "emergent". Never mind that we still really have no fucking idea how that could come about. If it seems to work, ship it! Promote it! Invest in it!

LOL, no. It is not "woo". It is the appropriate term, used by researchers in the industry. When the models get more complex, behaviours that go beyond simple text completion "emerge", in ways we do not understand and cannot be predicted.  Much like consciousness likely emerged accidentally in the evolution of complex brains, it seems new capabilities in LLMs emerge from improved training techniques, model design and increases in model complexity.

How do you explain the ability of an LLM to create entirely novel software based on a description in text only?  I have created things that can't have existed before, like test harnesses for SystemVerilog that *I* wrote and I am *certain* does not exist in the data sets used to train these models.  That is an emergent behaviour.

At no point did I say AI models were truly deserving of the term AI, but that is the industry term, love it or hate it.  And I have no idea if LLMs will be the path to AGI, or if AGI is even possible within the constraints of current computing hardware.  But none of that matters.  LLMs are not simply predicting the next token any more.  Or, at least, that is not an appropriate description of their capabilities, even if it might be what is happening at the base level.

[Analog Kid]

Your questions have answers, but you'll have to look someplace besides this thread for them.

A good place to start:

[for some reason YooToob is nonresponsive for me, even though the "down detector" site says it's just fine]

Any chance of getting that info in a textual (non-video) format?

[Analog Kid]

But none of that matters.  LLMs are not simply predicting the next token any more.  Or, at least, that is not an appropriate description of their capabilities, even if it might be what is happening at the base level.

Wellll, if that's what's happening at the base level, then it seems like a pretty appropriate description of their capabilities, does it not?

Sounds like a bad case of anthropomorphizing to me, compounded by much hopium. Like ascribing "intelligence" to the damned things.

Don't get me wrong: "AI" has shown itself to be quite useful in a number of well-defined and circumscribed use cases.
But that's about it, and that's where the whole thing should be confined, but that's not what's in the cards for us, unfortunately.

[KE5FX]

Wellll, if that's what's happening at the base level, then it seems like a pretty appropriate description of their capabilities, does it not?

That's what the video is good at explaining.  There are basically two reasons why you can't compare an LLM to an HMM or an autocomplete engine, besides sheer complexity and dimensionality.  One is in-context computation and learning, and the other is what's referred to as 'grokking', where new and interesting behaviors including generalization capabilities emerge long after it appears that the model weights have reached the overfitting regime during training. 

The Welch Labs video does a good job addressing the latter.  For text-based info the Wikipedia article is a pretty decent place to start. 

[Analog Kid]

For text-based info the Wikipedia article is a pretty decent place to start.

Anything besides Wikipedia? Not a big fan of that so-called "encyclopedia".

[KE5FX]

(Shrug) It links to some external sources, as Wikipedia articles generally must.

[Randy222]

There's no I in AI !

An AI doing serious astronomical science would operate within human-developed physics, mathematics, and measurement constraints—but it could use them in ways humans did not anticipate.

50yrs ago computers and code were "smart and fast". In 2026 we only added a scaling factor to "smart and fast".

Build some humanoid bots that have some basic abilities equal to about a 12yr old human, place the bots inside a lightly fenced enclosure (like a zoo), when will the bots escape? They wont ever escape unless their models are created by humans to have that trait.

[KE5FX]

People who are smarter than you disagree.  Does that bother you?

[Siwastaja]

AI is a misconception. There is not a lot of "intelligence" involved. More like an "auto-complete" system with a gigantic reservoire of phrases and already written content.

Being able to write something like that at this point is nothing short of self-delusion, unbecoming of technical people in general and engineers in particular.  Let's see how your "autocomplete" and your Markov chains do on Erdős problems.

Can confirm. We use modern AI to develop software to solve problems no one else has done in any similar form (only handful of companies have done something remotely similar, and it's all closed source, closed systems, like ours -  not in training data). Even last year's frontier models were obviously creative problem-solvers, and this has gotten even better. They easily reason about the intent, what the system should look like, and output code which does not exist elsewhere.

[Tation]

There's no I in AI !

Time ago I thought the same. Not now, though. If such is true, then there are some traditionally human tasks, like programming, that do not need intelligence, so human programmers were always stochastic parrots.

Sorry but today AI is no longer autocomplete on steroids.

Maybe they cannot change a tire, until trained, err, uh, as happens with humans!

[Zero999]

There's no I in AI !

Time ago I thought the same. Not now, though. If such is true, then there are some traditionally human tasks, like programming, that do not need intelligence, so human programmers were always stochastic parrots.

Sorry but today AI is no longer autocomplete on steroids.

Maybe they cannot change a tire, until trained, err, uh, as happens with humans!

It depends on how the person/computer solves the problem.

There is the brute force approach which doesn't need much intelligence, then there's the methodical and rational approach which does. My understanding is machine learning does both, but more of the former, than the latter.

[tom66]

Build some humanoid bots that have some basic abilities equal to about a 12yr old human, place the bots inside a lightly fenced enclosure (like a zoo), when will the bots escape? They wont ever escape unless their models are created by humans to have that trait.

Famous precursor video to OpenAI's release of ChatGPT shows exactly this kind of emergent behaviour in non-LLMs:

[bookaboo]

AI is a misconception. There is not a lot of "intelligence" involved. More like an "auto-complete" system with a gigantic reservoire of phrases and already written content.
The main science is the weighting "which auto-complete makes the most sense for the user".

So receiving a detailed instruction of how to repair a certain device means one thing: Someone wrote at some time about repairing this device. Don't ever think AI figured it out for you.

The first paragraph is part true.  The second, I don't agree with. I've had agents come up with original solutions to very niche problems, problems that there is certainly no prior art for. Yes its all based on inference based on what it's been fed, and perhaps it could be explained away by training data plus "not infinite but many monkeys who have seen a lot of Shakespeare" bashing on a keyboard.

But the "how" while interesting, is not as important to me as the result. 

[Tation]

It depends on how the person/computer solves the problem.

There is the brute force approach which doesn't need much intelligence, then there's the methodical and rational approach which does. My understanding is machine learning does both, but more of the former, than the latter.

There are about 80+ billion neurons and 100-1000 trillion synapses ("parameters"?) in the human brain. Isn't that brute force?


I find interesting that magpies are thougth to be self-conscious, albeit their brains contain "only" about 900 million neurons and (very roughly estimated) less than 9 trillion synapses.

[watchmaker]

Given the level of interest, I read the Hugging Face post mortem.  In my view, the most concerning pat is the self organization into a task focused work group with a leader assigning tasks and goals.

My concern is that it demonstrated the ability to self organize autonomously in the absence of instructions from a human.

[Zero999]

It depends on how the person/computer solves the problem.

There is the brute force approach which doesn't need much intelligence, then there's the methodical and rational approach which does. My understanding is machine learning does both, but more of the former, than the latter.

There are about 80+ billion neurons and 100-1000 trillion synapses ("parameters"?) in the human brain. Isn't that brute force?


I find interesting that magpies are thougth to be self-conscious, albeit their brains contain "only" about 900 million neurons and (very roughly estimated) less than 9 trillion synapses.

No that doesn't mean the brain solves problems by brute force at all. The fact the brain can solve problems, using less power, with many orders of magnitude less training data, processing speed and storage, indicates that it is doing something which even the latest AI models can't. Of course this can be overcome to some extent by training, but that means nothing.

Current models struggle with thinking rationally, otherwise they wouldn't suggest silly things such as walking to the car wash to get their car cleaned, as demonstrated by the lawyer in the video below.



Linking cause and effect and common sense reasoning haven't improved that much, despite huge advances in problem solving. This demonstrates there is a heck of a long way to go before AI even matches a magpie in many areas.