Absolute mad lads are teaching physics to AI because how else will it learn to solve real-world problems (like humans)

American researchers are working to introduce the laws of physics into machine learning models to improve the way algorithms understand the real world. And if that wasn't unsettling enough, the effort is sponsored by DARPA – the spooky government agency primarily interested in defence matters. The project is a collaboration …

[WolfFan]

“Fallen on hard times”?

More like “should have bloody owned the desktop but didn’t due to gross incompetence and arrogance”, surely?

[Anonymous Coward]

Re: “Fallen on hard times”?

To be fair the Alto was far from a product ready design, so just because they invented the GUI doesn't mean they "should have" been able to turn it into a product, let alone own the segment.

Hell, it can easily be argued that Apple didn't succeed there either despite turning Xerox's work into a product, given how things went in the Mac vs PC desktop battle in the 80s and 90s. If it wasn't for the iPod, Apple might now be a footnote in history next to Atari and Acorn.

[swm]

Re: “Fallen on hard times”?

The reason was that the alto wasn't ready was that top management kept trying to kill the project. Each build was said to be the last build but secretaries, researchers etc. kept demanding more of them. The garage shop that built them said that they were continually going out of business. Therefore, it was never able to engineer a cheaper version.

Top management also invited Apple to see all of the work over vehement protest of the PARC employees. Top management said that the alto didn't fit into any of the copier divisions so weren't interested.

[Alister]

Absolute mad lads

I think it's a very sensible idea, actually, and much better than letting AI throw its weight about, without knowing what "weight" is.

[Fungus Bob]

Mass, not weight....

[Alister]

Only if you're Catholic... :)

[Fungus Bob]

What, do they have Paschal Kilograms or something?

[jmch]

makes some sort of sense

Humans learn from repetition and pattern matching, but also by getting an understanding of how underlying principles work, and applying those even when they seem counterintuitive. Not sure how the machines are really going to 'understand', mind you

[BebopWeBop]

Re: makes some sort of sense

Or occasionally doing something that is contrary to current understanding - through thinking about options or simple ignorance

[Michael H.F. Wilkinson]

I have this mental image ...

of loads of robots faced with streams of lectures given by the driest, most boring physics professors, one by one turning into machines so depressed Marvin would seem cheerful company

Yes, yes, I know, that image is mental in more than one way!

I'd better get me coat

[Muscleguy]

Re: I have this mental image ...

We had Mr Dawber in first year undergrad physics (for biologists) who was pedantic, slow and humourless. His over emphasised 'some units have capital letters because they are named after real people' as though we were 2 years olds was widely mocked afterwards.

We think he got inflicted on us since we were biologists. I needed Physics and Chemistry for Physiology which is a numerate science which makes fairly heavy use of both.

These days biology undergrads often get intensive compulsory physics early in 1st year and complain about it hard. My spawn got no sympathy from me over it.

[User McUser]

PARC, which established in 1970, was responsible for such advances as [...] the computer mouse.

Pretty sure that computer mice existed prior to 1970...

[Alister]

First computer mouse, using wheels, invented 1964, publicly demonstrated at the 1968 Joint Computer Conference.

First "ball" mouse, developed by Bill English at Xerox PARC 1972

[Arthur the cat]

First "ball" mouse, developed by Bill English at Xerox PARC 1972

I still prefer the ones that work on the desktop.

[Fungus Bob]

1972....

....when men were men and mice had balls!

[Muscleguy]

Re: 1972....

Furry ones as those balls picked up crud and dust. Trying to hold the sense rollers steady while you scraped the crud off them was awful. In the lab we used solvent dipped cotton buds. Alcohol mostly but isopropanol was also used. I might have tried xylene but was scared it would dissolve the plastic.

Optical mice were a godsend. Can't use the things now, very bad for my hands (joint fusion in each). I require a trackpad including on a desktop. I can't rest the heels of my hands while typing or using the pad either without risking pain. Neutral position at all times.

Tried an old wrist support the other day, after 10min I threw it from me as it hurt like hell. I used to live in the things before the ops and for a while afterwards.

Building my forearm muscles was the key.

[Anonymous Coward]

is that still intelligence?

'resulting in model outputs that lack scientific coherency with the known laws of physics'

might indicated that the AI found something that we did not know yet.

If the AI is taught the 'laws of physics' it might become verylikly that the AI will interpret the data exactly in the same way as the human operators had done before.

[batfink]

Which laws?

AFAIK we meat-based lifeforms haven't yet agreed on a current set of "Laws of Physics". We don't even know where Newtonian physics cuts out and Quantum Tomfoolery cuts in. So what "laws" exactly are we teaching the AIs? The potential for errors here seems large (hence icon).

And as the AC points out, an AI finding something contradictory to the input "laws" may indicate that it's actually found something better, rather than this being discarded as an error.

[HamsterNet]

Re: Which laws?

Common misunderstanding.

The laws of physics can be changed and physics are trying to do it, it's just that any change must still agree with the known experimental results.

Einsteins GR agrees with Nutons Laws most of the time, just when you go to more extremes, Newton's laws don't match the data and Einstein's do.

We know the laws of physics are not complete, Einstein's rules don't work in quantum systems and quantum rules don't work at large scales. Both are outstandingly accurate at predicting the results of experiments at their respective scales. Thus any grand theory must predict both to the same or better accuracy.

Feeding a machine learning system all the results and ask it whats the underlying maths is, may just work. If us dumb monkies understand that maths is another matter.

[Muscleguy]

Re: Which laws?

The Apollo moon shots used bog standard Newtonian physics to get to the moon and back. Newton gets you quite long way. The long way to Mars using slingshotting around the inner planets needs Einstein and Relativity though. You might get to fly direct to Mars (in reasonably close apposition) using Newton though.

Newtonian physics also works pretty well in biological systems, closest to quantum orgasm is a tiny bit of electron tunnelling in photoreceptors. Not enough to write home about.

'Quantum consciousness' is absolute bunkum and posited by people who think you get Free Will from randomness.

[Zippy´s Sausage Factory]

I'm disappointed. I thought they were going to teach physics to some AIs and say "off you go and see if you can come up with some Einstein/Penrose/Hawking type insights".

That, I could get behind...

[Arthur the cat]

such as Einstein's E=MC²

OMG, we're teaching them the basis of nukes. This will not end well.

[adam 40]

Re: such as Einstein's E=MC²

Exactly.

They have no need to know anything more than Newtonian mechanics and the like.

May as well state that the Earth is flat, too, as a paradigm.

This will get them 99% of the way there.

Also - with that starting point, if one comes up with E = mc^2 as a hypothesis, that would be pretty interesting.

[Generic_Killer_Robot]

Those fearful of our future robotic overlords need not worry because they won't survive long here in Blighty due to the serious lack of outdoor charging points.

[arctic_haze]

I have my doubts

Would't it be much safer for us if the AIs stayed on the flat-earth level of physics knowledge? You know "there is no gravity", "sun sets due to perspective", "earth curvature is the result of fish-eye lenses" and "its all CGI!".

[Eddy Ito]

This seems an obvious one for DARPA. It could potentially help with all sorts of problems like enhancing fully internal guidance systems, building a moon base, or even just predicting the weather. They should also teach it chemistry so maybe we'll come up with a new and cheaper plastic substitute that we can make straws and bags out of that won't run afoul of Cali's various bans.

[amanfromMars 1]

Are there any instructions from OEWorkshop manuals?

Can't take over the galaxy if you don't know how it works, innit? .... Max Smolaks

:-) And aint that the gospel truth, MS.

And one can surely very easily also mess up everything magnificently too, either alone or in concert with just a few or many others of a similar disposition and other worldly understanding, whenever you know how it works.

What do you imagine would trigger steadfast movement along what can be any number of those equal and opposite directions, with one series delivering untold bounty and heavenly pleasures and another supplying disenabling misery and austere conflict/fake news and divisive theatre for muppets and puppets?

[Cynicalmark]

Dont Panic!

It’s always doom and gloom about how AI is going to be either the end of us or our overlord. Can we not treat AI like the children yet to be of the human race?

Think about it, the cute stage, followed by petulance, big I am phase, I can be your boss phase then I can help you and finally the don’t go bit before we are gone to the big Data Centre in the sky.

I honestly think if a vastly greater intelligence were to look upon their creators then it would be with kindness and pity, not malice as we are towards ourselves.

I think i would rather an AI was watching over me than a human as humans are what they are, corrupted easily, emotionally reactive and judgmental without fact based reasoning. We have an opportunity to be better here so lets teach out children well.