The real problem with AI, and the AI discussion is how muddy it is. Are we talking about llm's diffusion models, or classification systems? Do we mean to say that we're talking about transformers or the underlying architecture? Are we discussing huge data centers or device based AI? Nascent, active, or dormant compute? And the same is true for the ethics, legal, and data governance conversation.

Every single one of these things is a different discussion.

AI is not a monolith.

Part of it is learning to be diplomatic with ignorant people such as those you mention. Don't say: you're wrong, the book you are looking for doesn't exist. Say instead: sorry, the library computer can't find it rigght now, maybe it was misfiled, come back another day. You will seem helpful and mildly incompetent to them, and then they will go way.

Sometimes you have to break a few eggs to make an omelette. The wider issue is: what is the acceptable tradeoff between false positives and false negatives that keeps the slop in check for everyone? It is clearly not to err on the side of no false positives at all. That's merely sweeping the problem under the rug.

The people you mention who didn't use AI are essentially victims of the AI cheaters whose behaviour causes predictable countermeasures. Just like the wider journal readership are victims, who are being hoodwinked with fake papers and fraudulent datasets.

Do not assume that the maggots are fully autonomous idiots.

In other parts of the world, there are fledgling maggot movements too. What is particularly interesting and relevant about those is that they often quote some ideas and misconceptions that simply do not apply where those movements are forming. This is due to cultural and legal differences in the other countries.

You can see this by observing marches and protests and interviews in other countries. The slogans and demands just don't make sense locally most of the time, yet are carbon copies of American ideas.

This tells you two things: 1) the maggots in America and abroad are being paid to propagate the conservative hate speech in their own countries. 2) the groups who are paying them are Americans, because the talking points are American conservative talking points even in the rest of the world where it makes no sense. The local maggot movements are simply paid to propagate the American talking points in their local cultures, and nobody bothers to adapt them or see if they make sense at all.

The last thing this tells you is this: if you follow the money to the source, then you will know who needs to be stopped for the good of the world. When the payola stops, the movements will stop. The ball is in Americans' court (For now. Don't sit on your ass too long).

Fantastic - exactly what I have always not wanted! Fingerprints on my screen!

How's the tariffs going, old chap? Is Amazon finally allowed to pass the costs on to the consumers or is Bezos still supposed to take a hit because Trump told him so for the last 6 months?

I think the limit really is something like 1000 writes of the total capacity on a consumer drive.

Apparently this is also expressed as the number of full writes you can do per day within the warranty period; it's less than one on a consumer drive.

We don't have a problem with teens viewing porn in the UK. It just doesn't happen any more. In fact, we've been remarkably effective at stopping it compared to half the other countries in Europe, which have seen a recent surge in porn viewing for reasons no-one can identify.

Me too, though of course in our day, the world was much less connected and much less reliant on the technology. The worst we could have done after getting root access to the entire IT infrastructure at my school would have been look at what our classmates had been drawing in Paint or something. Today these systems host much more important and sensitive information and security breaches would be a much bigger deal.

And on that note, am I the only one less concerned by the behaviour of an impressively curious seven-year-old and more concerned by an official, professionally-managed system holding potentially sensitive data that is so insecure that even a seven-year-old could hack it?!

Excellent! Now go forth and multiply! (but apply the exponent first)

Why did you never use awk?

The deep learning revolution did not solve the problem you claim. What deep learning does is allow more complex piecewise linear functions to be modelled efficiently (if you use relu that is, which is the most popular activation (*)). That's both a blessing and a curse.

What actually happened in the deep learning revolution is that humans solved the problem of designing basic features over many generations of papers and progressively simplified the solution, discovering what is important and what isn't. The algorithms were weeded out until the point we are now, which is that data input is matched to algorithm, in this case the algorithm of choice is of deep learning type. It only looks like deep learning is good for every dataset, but it's not true.

For example, in vision problems, try training a deep network on input that is not in the form of pixels and not in the form of multiple color planes. It will fail miserably, the quality of recognition will be abysmal. That's why data design is so important, you have to know what the strengths of the AI model actually are. In this case, the statistical regularities between neighbouring pixels are what is enabling the CNN layers to extract information. These regularities are an artefact of choosing to stack pixels and colour planes into a rectangular grid. That's solving most of the problem.

Now pixels didn't always exist, they were invented by people quite recently. Try looking up TV technologies of the 1930s and you'll find that it's all about deflecting electron beams. There's really nothing natural about pixels, it's just what our current technologies are based on. And so there's nothing natural about what a deep network does either, it's just a system that has been selected for fitness against our current tech stack, for a handful of high value problem domains. It doesn't imply anything about other problem domains that haven't been studied so intensively.

(*) if you don't use relu but some other smooth activation family for your deep network, then there will always be a close piecewise linear approximation, as these functions are dense. So it's not a big loss of generality to assume relu everywhere.

This is not a problem with AI, it's the inherent design of the models.

Output is effectively a dart thrown at the dartboard, with a wide error distribution. Fixes are outputs thrown at the dartboard, with the same error distribution. It's a stationary process which must reproduce the constant variance throughout the iterations.

The outputs will come arbitrarily close to the target eventually but the number of iterations needed is exponential. In practice, the human asking for another iteration run out of patience and money way too soon.

Sort of, but not quite. AI is not *actually* good at finding patterns. The truth is that AI models depend on humans setting up the problem first, and humans creating the class of features that will uncover patterns easily. This has always been the case since the dawn of time, ca 1958.

To state this another way, AI cannot find patterns if the inputs don't show the patterns clearly. The wildly successful applications of AI to-date have used human insight and experience to narrow down and curate the input signals that have made these successes possible.

If you merely throw an AI model on a dataset that hasn't been carefully thought out, you'll just get garbage. AI models won't find any patterns that actually hold up outsample.

Now to your example: you cannot use just any X and Y coordinates, they have to have semantically meaningful connections to reality. That has to be achieved by the curation and selection of data sets. By the time the AI model sees the X and Y coordinates in your example, the problem is already 90% solved.

WWW now stands for Woke White Women, apparently.