I'm assuming that the 20,515 signals of interest come from a pool that's one or several orders of magnitude larger than 20,515.

I understand. Still, the point remains that the AI is, in this case, clearly inferior to humans (I would have been impressed if the humans found 20,515 signatures and the AI only - verifably - found eight of them to be worth following up).

> clearly inferior to humans

But maybe we are just proto-AI, running on a different platform

"inferior" is a funny word for "doesn't perfectly impersonate us".

Trying to filter noise from data also follows the 80,/20 principle. The rules for the easy cases is written easily but then you have one-offs, maybes and special cases. Trying to filter one of these does not cover anything else. So doing it manually takes the same time.

Case study: had to clean a list of German addresses once. Excluding obviously invalid addresses like some Chinese address was easy. But some addresses had errors which needed a human eye to fix and correct.

AI is not a well-defined term. This is more likely a specific machine learning technique that is designed to identify all the boring "normal stuff" and ignore it, and then do that at very large scale. By doing that, you can find the interesting parts so that humans with more limited resources can determine if those newly flagged things need to be added to the boring list, or if they do represent something truly interesting.

edit: After reading the article a bit more, it is using a random forest classifier. This is almost certainly not meeting the definition of what many here are thinking when the term "AI" is being used in the title. The term is clearly used here for marketing purposes.

>The term is clearly used here for marketing purposes.

I think most uses are for sensationalistic purposes. I'd wager almost nobody in the general public really understands what AI is or can pin it down. It doesn't help when so many different media outlets abuse the terminology by using it to refer to different things. What's even worse is that whatever ideas people have about AI tend to come from Hollywood.

They are talking about ML, which is statistical pattern matching and finding.

Gödel ruined the fun of automated theorem testing a century ago unless we make significant discoveries in math.

Type inference is an accessable example of SOTA automated reasoning if you want a more realistic idea of what our current constraints are.

We will be restricted to Human assisted Turing machines for the foreseeable future unless there is a major development in pure math.

Remember we can't even build a logically consistent model of arithmetic due to Gödel.

There is nothing human can do that Turing proved a computer can't do. Gödel ruined the fun of manual theorem testing the same way.

A human with a piece of paper and a pen could do. AKA writing down an algorithm.

Human reason isn't limited to written algorithms. The missing text problem from NLU is an example of a problem that is thought to be impossible for Turing machines/algorithms but is trivial in most cases for humans.

> The missing text problem from NLU is an example of a problem that is thought to be impossible for Turing machines/algorithms but is trivial in most cases for humans.

Not thought to be impossible by people who believe in scientific materialism, and current mainstream ideas in theoretical physics (like the Bekenstein bound and such), and who have thought carefully about the issue.

The laws of physics are believed to be computable, and the information content in a bounded region of space, finite. Therefore, it is believed that, in principle, a Turing machine could run an accurate physical simulation of a person, and could therefore do any cognitive task (as far as input/output correspondence goes) that a human can.

If you’d like to explicitly reject scientific materialism though, I’d have no complaints about you doing so.

If you would like to reject scientific materialism (and the empiricism and the dispassionate utilitarianism that go with it), go ahead.

Just don't try to frame your conclusions as objective truth, or known scientific results like the GP.

Right, or..

Ok, uh, I don’t think one has to really reject empiricism to reject scientific materialism?

Or, err, by “empiricism” do you mean like, “support for doing experiments, and keeping track of the results and what models work good to explain them, etc.”, or do you mean stuff like “rejecting anything that doesn’t have good scientific evidence behind it”? One can do the former without doing the latter.

When I express a belief I have that doesn’t fit with scientific materialism, I make sure to mark it as such, so that people can take that into account. I don’t anticipate any clear externally-verifiable refutation of scientific materialism within my lifetime, and so I don’t anticipate predictions that follow from it to be refuted anytime soon. And I definitely wouldn’t present those beliefs of mine as being the scientific consensus.

I suppose one might accuse me of having a “belief in belief”, seeing as I don’t expect these supposed “beliefs” of mine to be predictively useful any time soon.

But I think it is right that there are goals/values that I place higher than pure predictive accuracy. And beliefs about purpose, and meaning, and what is good, etc. fall into that.

(You mentioned utilitarianism. I’m not a utilitarian, but I do think it is often a very good heuristic, and in many contexts it would be good for it to be used more.)

> Or, err, by “empiricism” do you mean like, “support for doing experiments, and keeping track of the results and what models work good to explain them, etc.”, or do you mean stuff like “rejecting anything that doesn’t have good scientific evidence behind it”? One can do the former without doing the latter.

Honestly, I added "keep your models as simple as you can" into it. But any way you cut empiricism, it's actually utilitarianism that can be seen, so it's the one where the correct fine-cutting is important (hum... well, if you keep an utilitarian point of view). Anyway, utilitarianism tends to align with the version of empiricism biased into getting computable models.

And, of course, none of those deal with purpose questions.

Anyway, your comment there is great. What I disagree is on conceding space to something like the one above yours, because it's a misleading text that implies something very different from what it says.

> Just don't try to frame your conclusions as objective truth, or known scientific results like the GP.

To clarify this objection is to @nyrikki's (incorrect) claim that "NLU is an example of a problem that is thought to be impossible for Turing machines/algorithms" and not to anything that @drdeca said right?

I think you are agreeing emphatically with @drdeca but it's possible to read this comment as an objection to @drdeca (NLU may not be as trivial as claimed - ha).

Yes, I completely agree with drdeca.

Physics models are models.

All models are wrong, some are useful.

I am not claiming that useful models need to be computable, in fact the problem with the MTP is that it induces cycles into something that needs to be recursively enumerable to be decidable.

"The trophy wouldn't fit in the suitcase because it was too [large,small]" is a nice toy case to consider how NLP can deal with that easily but NLP would have issues.

It all relates to VC dimensionality and decidablity in the end.

But the math is hard to demonstrate without actually using math.

> The missing text problem from NLU is an example of a problem that is thought to be impossible for Turing machines/algorithms but is trivial in most cases for humans.

I'm not aware of any mainstream researcher who thinks the missing text problem[1] is impossible for Turing machines. Perhaps some think neural networks are insufficient to achieve it, but this is a different thing.

[1] https://thegradient.pub/machine-learning-wont-solve-the-natu... (Ironically this badly needs updating in a post ChatGPT world, since ChatGPT can solve many of the things this article claims are impossible)

ChatGPT is state-of-the-art natural language processing (NLP) NOT NLU

That's an artificial distinction without a difference.

It's just software and it solves the problem.

Nitpicky and probably universalist or whatever but the fact that we can't build a logically consistent model was always the case, Gödel just discovered it.

No, you are most likely correct, I personally haven't found anything that even gives a reason to hope. But as reductionism and Laplacian Determinism are taught as cannon and not as the best option I sugar coated it.

The Wada property arising in simple models like predator/prey models with simple added factors like fear/cover probably also suggests that even simple models may be indeterminate at least with binary operations like modern algebra is based on.

Three attractors or exit basins can make this unintuitive topological feature pop up. Note that that fractal behavior is topological and not scale invariant noise like is typically studied with topics like fractal scattering.

Indecomposable continua like the Wada property aren't solvable with probabilistic models using automata like in the standard model, which is lucky enough to have less than three exit basins.

Here is one fairly accessable paper on the Wada property.

Hopefully some n-ary algebras may come forward to deal with restricted problems, but even if we could create a logically consistent model of arithmetic, binary operations will still be indeterminate with feature like the Wada property even with perfect knowledge of initial conditions.

https://www.researchgate.net/publication/365233050_Organized...

If you read the paper itself[1] the random forest is only used in the final stage. The main approach is a convolutional variational autoencoder (which of course is a deep learning model).

The VAE model itself is defined in step 6 in [2]

[1] https://www.nature.com/articles/s41550-022-01872-z.epdf?shar...

[2] https://github.com/PetchMa/ML_GBT_SETI/blob/4096_pipeline/te...

Ad deep learning models have no understanding of the underlying data because they are either classification or regression.

Well the humans are doing classification too so I'm not sure what this is supposed to mean.

It's true that this model doesn't know anything outside its training dataset, but that's a different objection.

It's an interesting question as to if there is anything humans do that isn't just classification or regression (and of course regression is just classification over an infinite set where we assign labels to certain ranges and then select the highest probability density).

For certain definitions of understanding.

No one claimed the AI "is so smart" - it's another tool in their kit. The article clearly conveys that the potential is a useful way to augment human-led research by providing a way of quickly processing large datasets i.e. the 20k signals are filtering down 150 terabytes of data which is much more manageable for human analysis. It's not like we have definitive parameters of what constitutes an "alien signal," so we can't exactly create an absolute model for detecting such telemetry. Instead the goal of the article is to simply demonstrate exciting new ways to leverage machine learning methods in different contexts

Because it’s not so smart, it just try to recognize a set of patterns that was trained to look at

Which is what we do. The rest of what we believe makes us special is simply emergent, and arbitrarily defined in terms of just being further useful pattern recognition, rather than any fundamental property of "intelligence".

Presumably, because it filtered through orders of magnitude more than that.

Because it isn't AI. It is just statistical models. AI is for marketing.

How do you judge whether 20,515 is a large number or a small number? If the AI provided 2 candidates is that still too high?

SETI@home sorted through I think trillions of data points.

Any 10000s of something is a large number when it involves human work.

Imagine that tax returns for millions of people were all processed by hand not but a few decades ago.

And that required 1000s of people

And the results here were three orders of magnitude less, so they require single digits people.

Assuming the task was of similar difficulty and the deadline was equally short. I doubt those assumptions hold.

True. I bet the tax returns were harder and had a bounded deadline, unlike an academic research paper.

> How do you judge whether 20,515 is a large number or a small number?

It’s a trade-off between false positives and false negatives, and the budget you have to investigate alerts.

https://en.wikipedia.org/wiki/Receiver_operating_characteris...

Because the statistical models powering AI don't have the depth of understanding of a PhD in this field, and so it casts a wide net to ensure nothing gets missed and then a more refined search by human heuristics is needed.

The others were "intelligent" signals but from human interference. There's a lot of this so the unique patterns don't occur often enough to automate.

The paper is actually worth reading about this part[1]. They have a moderately complex pipeline that you can think of as a filter: it tries to find anomalous signals.

The stage-1 autoregressor filtered 115M signals to 3M. Then they perform signal processing techniques to remove things like GPS signal contamination ("t can be seen that certain observing frequencies contain a much higher number of events compared with the others—for example, the region around 1,600 MHz. This overlaps with known RFI at the GBT site specifically from persistent GPS signals.").

After this second stage they are left with 20,515 potential signals which were visually inspected.

The issue here is that we don't know what a alien signal looks like so we can't just use a classifier. The pipeline can only find things it has never seen before, but it takes human judgement to decide if these signals are "alien" or more likely contamination from human sources that weren't filtered ("Regarding the nature of the rest of the events, most of them look like false positives associated with RFI signals.")

[1] https://www.nature.com/articles/s41550-022-01872-z.epdf?shar...

I think it's useful to replace "AI" with "app" any time you encounter it in the wild. AI is a set of computational techniques that is very broad. The problem with considering "the AI" or "an AI" is that people have in mind some kind of AI agent and that isn't really the case.

AI is also not well defined - day one of my AI course in university opened with "What is AI?" Generally once we figure out how to do something using a computer, we decide that's not really intelligent anymore so the implementation isn't AI. An example of that is the minimax algorithm - it's featured "AI: A Modern Approach"[0] but it isn't really something people think of when they hear "an AI".

[0] https://aima.cs.berkeley.edu/contents.html

The AI is fast, not smart.