The good thing about math is that you can formally verify the results, making hallucinations much less scary.
This is similar to games, where large learning models had some of their first big wins (go). Games, like math, have strict rules that can be used to prune invalid outputs.
It is already in the making, the next big step in math will be the complete formalization of all existing, relevant theorems and proofs via LLMs.
Essentially part of a machine vision software package, which in turn is computer vision specialized for factory automation. It's "industrial" because it was trained on such data (more expiration dates and SEMI-font on wafers, less random pictures of streets with some text).
That, plus engineering and support: a robust implementation, deployment on almost any system, fast and good long-term support so your factory doesn't stop for too long if something goes wrong etc.
> OCRing this is a nightmare and is a good benchmark to any self-proclaimed good OCR/vision model.
It's not that difficult, our industrial OCR model read it correctly on its first attempt with default parameters. The characters are easily separable, there is no structured background (think expiration dates on yogurt aluminum lids) that confuses the reader, there is no almost-text-like texture anywhere that would clutter the result. The font is also nice and standard.
They already collect and track you, even with car play. I strongly recommend this CCC talk, where they hacked a Volkswagen database that contained unfiltered, high-accuracy, timestamped locations of a large majority of electric cars from VW group.
Based on that they were, for example, able to identify cars owned by members of Germany's security apparatus: where they work, where they live, where they drop off their children each morning. Who visited brothels.
It is like questioning the purpose of shovels, because surely if shovels generate so much wealth for the miners, wouldn‘t shovel companies ask for a percentage of the gold (say 30%) instead of selling shovels at a unit price?
Bridges deteriorate by default. That is a well understood process with well understood reasons, which we monitor and work against. Salt, thermal cycles, load, rust.
Humanity does not deteriorate by default. Claiming it does so through some hand-wavy pseudo-evolutionary arguments is not a strong case, and requires at least some evidence to be taken serious. How about a (equally unfounded and just for the sake of argument) reverse claim: Humanity got more intelligent, because high child mortality favored physically strong children instead of mentally strong children.
> I had been cured of tuberculosis by them... in 7 days
Just FYI that might suck in the future. Some countries might deny you entry / visa, or at least require expensive additional screening if you had tuberculosis diagnosed in the past. It might be worth to fight this.
The thing is we tried that for decades, using more formal logic to build reasoning engines. And we never got it to be even a fraction as good and generic as learning-based LLMs are today.
This interpretation also generalizes nicely to 3d. You have 60 moves, 20 of which will be up (60 choose 20), then 20 are east (40 choose 20), then 20 are south (20 choose 20).
This is called metamerism. It can be a practical issue if two pigments have the same color under one light source, but a different one under another. You want your artificial teeth to have the same color as your real teeth in sunlight, led light, and a classic lightbulb for example.
Overly specific and not more than feel-good blurb. “When you’re recording” - what if they record it for, say, debugging purposes? Are images send into the cloud even if that light is off? “Prioritize” on-device processing is a meaningless promise, on the contrary - it means that some things will not be done on device. There is nothing in this text stopping them from streaming and storing whatever they want and need.
There is a static branch predictor that is used if there is no statistic on a branching instruction yet, and it's really simple: Jumps backward are assumed to be taken (they usually are from a loop), jumps forward are assumed to be not taken.
So the jump that forms the loop will be predicted correctly for all executions but the very last (when the loop ends).
They are "more" vertical, but they too have vital suppliers that they could not do without. The semiconductor supply chain is deep. Everybody knows ASML, but there are countless others that produce raw wafers, etching machines, special chemicals and so forth.
This is similar to games, where large learning models had some of their first big wins (go). Games, like math, have strict rules that can be used to prune invalid outputs.
It is already in the making, the next big step in math will be the complete formalization of all existing, relevant theorems and proofs via LLMs.