AI Articles

Btw - this is my own original thinking - but here's how I would approach AGI (artificial general intelligence):

* have a database containing links to a large choice of AI systems

* a top level AI system

* the top level AI system analyses the input, then chooses the best AI system to handle it from the database

* the input is then passed to the chosen system

I think human brains work differently from this: I think that in our brains, input gets passed to many different parts of the brain in parallel looking for a match. Hence things can suddenly remind you of things you had forgotten that you knew.

Deepmind think they are close to AGI (artificial general intelligence): they think they just need to scale up their new product Gato - link.

Reading a couple of other articles, it seems that DeepMind's Gato is a bet that the history of AI, that big generalised models with a large amount of powerful hardware eventually trumps specialised models, will continue.

Eelco wrote:

The quality  of translation is a bit volatile I still find. For instance just now I wanted a translation of Dutch "chargeren". At first, Google thinks it is English and gives a Dutch translation "opladen". But to my mind, it is never used this way in Dutch, maybe by the Flamish. So I change the language to Dutch and then it gives "charge" as first translation. But it is used mainly as the second, "to exaggerate" in Dutch. Nothing to do with batteries and stuff. So still a lot of work to do.

Not that I am a big fan of all the money that is pumped into this, personally. In the end, open source is more 'futureproof' for one thing. GPT 3 is now exclusively licensed by Microsoft and runs only on Azure, as far as I can see there is no open source of any of it but I could be wrong. So much for Elon Musk wanting AI to be democratically developed for everybody. Then he should not have allowed this exclusive licensing nonsense. So much for Elon Musk wanting to save the human race;



For database queries etc. I can see GPT 3 has value. It is a bit like the computer in Star Trek, only not super intelligent yet

For the moment I would just like to add to my post that the so called presentation by Elon Musk that you may see as an advertisement before the video here or the one below from 2018, the Advertisement about QuantumAI is a total scam! The involvement of Elon and the "presentation" as far as I can see is a fake, so why does Google allow this? As far as I can see this is legit factcheck: https://csracademy.org.uk/quantum-ai-elon-musk-review-scam/


I would like to just repost a link to the 2018 video "Do you trust this computer" by Chris Paine, that was sponsored by Elon Musk, because the above video does not add much to this one. The original is longer but has better research, better video material and maybe most important for me it has some great music See also Talkchess thread http://talkchess.com/forum3/viewtopic.php?f=2&t=67047

This video provides good insight into Deepmind's Gato based on the paper they published:

"Foundation models" are revolutionising AI: they are pre-trained models that can easily be adapted to your particular need. They are predicted to do for humanity what other foundational technologies (like the printing press, steam engines, electric engines etc) have done before.

This article is STUNNING. It's a longish article, but don't skip it: if you only read one AI article in June, make it this one! I can almost promise you won't regret it.

https://www.economist.com/interactive/briefing/2022/06/11/huge-foundation-models-are-turbo-charging-ai-progress

My opinion: AI will change the world more in the next 15 years than other technologies have over the last thousand years. This REALLY DOES change everything!

Maybe but Neural Network is a  black box. Or box nobody can grasp in detail. So no progress. They are only taking over. Giving results and we don't understand why. Maybe they contain bugs nobody can fix.  

Keep it simpel.

Rule was: If you have no solution then use a Neural Network.  So using a Neural Network means you can't solve the problem. So you failed. Or they failed.

Neural Network with only one neuron already  difficult to debug.

By the way if you can't generate enough training examples learning will fail.

Type 7: Maybe we don't need new toys

Henk wrote:

Maybe but Neural Network is a  black box. Or box nobody can grasp in detail. So no progress. They are only taking over. Giving results and we don't understand why. Maybe they contain bugs nobody can fix.  

Keep it simpel.

Rule was: If you have no solution then use a Neural Network.  So using a Neural Network means you can't solve the problem. So you failed. Or they failed.

Yes - humans writing programs to do "intelligent" work (like evaluating a chess position) fail all the time: chess programmers find that training a net massively boosts their program. Hence, there's something about chess that the programmers don't know.

Quote :

Neural Network with only one neuron already  difficult to debug.

Not sure about that - but certainly ten neurons becomes very challenging.

Fruit flies have just 135,000 neurons (though I don't know how many other neurons each neuron is connected to). The fruit fly connectome is well underway (link), but as the linked section points out, even if the connectome is completed, we'll still be nowhere near being able to emulate a fly's brain.

For me, a bigger issue with NNs is that if a fruit fly can master hundreds of complex behaviours with just 135,000 neurons, then why do "everyday" NNs need to be so big? For me, today's NNs are nowhere near being "the final answer": I believe it must be possible to condense the intelligence that artificial NNs provide into a MASSIVELY smaller piece of software. This would also make it much easier to study how the pattern recognition is being done in particular cases.

Henk wrote:

By the way if you can't generate enough training examples learning will fail.

Another weakness with ANNs: today's chess nets are being trained on billions of positions - but GMs will never see anything like that number of positions in their lives. GMs study a small number of positions in depth, and in the process unconsciously learn a huge amount about deep chess patterns that they cannot articulate or teach, and may not even be aware that they know.

Obviously (or seems obvious to me at least) the GM method of learning is superior to training with billions of positions. It comes back to a favourite theme of mine: large number of shallow (simple) patterns vs small number of deep (complex) patterns. Deep (complex) is better.

Computers need to forget the unimportant

Mclane wrote:

Computers need to forget the unimportant

Indeed: a valuable skill for all of us!

Without forgetting the unimportant you cannot learn. And the biggest problem for machines is that they cannot forget. So they do not know what is important and what is completely unimportant.

Collecting data for NN training is troublesome and expensive - but not to worry - just build that data "synthetically". This will be a big trend going forwards, but for me it raises the risk that the AI world will gradually drift away from the real world, and at some point there will be consequences for that.

https://www.forbes.com/sites/robtoews/2022/06/12/synthetic-data-is-about-to-transform-artificial-intelligence

Mclane wrote:

Without forgetting the unimportant you cannot learn. And the biggest problem for machines is that they cannot forget. So they do not know what is important and what is completely unimportant.

My plan: after adding new data for new cases that the net isn't evaluating well, try dropping some old data. Maybe the net can now evaluate those cases "well enough" without having actually trained on them?

TheSelfImprover wrote:

Collecting data for NN training is troublesome and expensive - but not to worry - just build that data "synthetically". This will be a big trend going forwards, but for me it raises the risk that the AI world will gradually drift away from the real world, and at some point there will be consequences for that.

https://www.forbes.com/sites/robtoews/2022/06/12/synthetic-data-is-about-to-transform-artificial-intelligence

Isn't this what GMs do when they examine a position closely in order to expand their understanding?

Mclane wrote:

Without forgetting the unimportant you cannot learn. And the biggest problem for machines is that they cannot forget. So they do not know what is important and what is completely unimportant.

They can forget. Just remove unimportant connections . There are algorithms for doing that.
I also remember weight decay algorithm.

Something I just stumbled across - AI video editing. Things you don't need to make a good video in 2022:

* photography (videography) skills

* sound/audio skills

* actors

* video editing skills

* microphone

* camera

* time

The ONLY thing you need is a script!

Not even very expensive, and multiple options to choose from - link.

The demo I watched was for business videos: can't comment on other types of video - but if not available now, then probably coming this decade.

By 2050, AI will help make ocean's transparent, and it will no longer be possible to hide nuclear submarines in them - link.

Putin's submarines will be affected first, because British and American subs are already very difficult to detect in water. However, the article doesn't hold out hope that they will be able to hide forever.

Important website for anyone wanting to integrate AI into an application: it's an encyclopaedia of available tools (if you will be building an AI system for a business, it would be difficult to overstate this site's value):

https://www.futurepedia.io/

The most insane week in the history of AI:




The list of things that people can do better than machines continues to shorten - and the future is arriving too quickly!