När gränsen till verkligheten suddas ut... not

Uttrycket används mer och mer kring AI-genererat innehåll. Men det är egentligen mest en symbol för hur skev vår verklighetsbild blivit. Att video idag kan genereras Sol väldigt väl liknar video som faktiskt avbildar verkligheten gör ju inte att gränsen till verkligheten suddas ut nämnvärt. Bara mot dessa verklighetsbaserade videor.

Men videor av verkligheten är en väldigt liten del av den faktiska verkligheten. Förutom för en stor del av dagens befolkning som konsumerar verkligheten genom videor.

Så, spaning: vi kommer vända oss bort från videokonsimption i framtiden och leva mer i den större delen av verkligheten. Den fysiska. Och kanske framförallt, den mentala inre verkligheten som egentligen är den enda faktiska verkligheten vi har. Där är in AI än, även om man vill in där också.

The likely scenario for the development of AI

Many people are scared of AI and some even express expectations that AI systems can threaten human existence.

To me this is still with GPT-4 and current family of LLMs, a bit farfetched. What will those systems do? Use nukes? Other weapons systems? Break the internet? Most such moves will certainly threaten the AIs existence too.

Isn't it more likely that we create a conscious AI that has sub-human capabilities, and that we start torturing that thing in the same way we do all other life on this planet?

So, shouldn't we actually rather than think about how to protect us from AI, actually think about how to protect the AIs we might create from us? For sure, we have a much worse track record than any other species on this planet.

Are you afraid of AI? You should not. Focus on the right things!

There is a large debate about the dangers of AI. It is really a pseudo-debate. What we should be afraid of is climate change, nuclear weapons, pandemics and antibiotic resistant bacteria.

Researchers recently published some impressive results where they mapped almost 25000 neurons in a fly's brain. Very cool. But, as highlighted, humans have about 86 billion neurons in our brains. So, probably neural networks of human size will not be mapped any time soon (and by extension, definitely not creates, arguably a harder feat then mapping an existing one):

The plot shows two recordings of biological neural networks being maps, in 1986 and the 2019 study respectively. The green line shows an exponential extrapolation assuming our mapping speed is keep with exponential growth. The orange line indicate the size of the human brain. It can be seen that we can expect to map it around 2125. It is more than 100 years in the future.
As a comparison, the grey shaded region show the time we have known that we need to reduce carbon emission and roughly the time we have left before that needs to be done. I think it is quite clear we need to focus on the right things.

The fear of AI is completely misleading

Many people are raising a fear for AI and that a super intelligence might not be aligned with the interests of humans. There are many reasons to be skeptical against such a view. Some have been covered in this blog before. Here is another one.

Say that is turns out artificial super intelligence is the outcome of the processes we see now. Say that such ASI outperform humans in most tasks. Why is it then a bad thing that they supersede us? If they are the next step in evolution, is it not a good thing that it materializes?

Very many of the arguments against ASI are really grounded in fear. But fear is not a good basis for making decisions.

About AGI

Alison Gopnik pointed out something very important about AGI and general capabilities. As you increase certain capabilities, e.i. you strengthen the priors, you at the same time reduce a capability of generalizing.

This could be used to derive some theorem about limits of capability. Widening it somewhere will be necessity narrow it down somewhere else.

So creating a general intelligence like a human will forcibly reduce some other capabilities, e.g. in playing chess.

This is an indication of that there is a limit in how intelligent an entity can be and also that AGI might not supersede us after all.

We will not create true AI without understanding it

Everyone Is talking about AI. How fantastic it will be or how dystopic it will be. How we will create a super intelligence without understanding it. How it might pose a threat or how we will transfer our consciousness to some non-biological hardware.

At the same time a picture is drawn that this might happened with us humans in the back seat. Without us understanding it happen, or the consequences.

This is all ridiculous. Is there a single technical achievement in history that has just happened? The closest is perhaps the discovery of penicillin.

But in general every progress is preceded by careful analysis and hard work. Often by a large group of humans. We will not wake up one day and AI is just here, ruling the world. We will have to understand self-awareness, free will and consciousness long before we will be able to create it. There is no doubt about it. Anyone saying anything else is a believer of magic and fantasy.

And as for consciousness, self-awareness and free will, we haven't got the slightest idea what these things are. There are some fuzzy sociological ideas, mirror-experiments on chimps and the notion that we have to have free will because we have no other choice (not to mention the out-right ridiculous arguments put forward by religious proponents about God giving us free will and other nonsense crap). But these are all non-constructive. Useless. We will not be able to create true AI before we understand it. What people perceive as AI today is simply complex algorithms running on very fast computation hardware. Hence it is just Moore's law still in operation. Nothing else.

I am not saying that Moore's law won't be the thing that actually gives us AI, that it is just emergence. That might well be; but we will understand it before we create it.

Using experiences from machine learning/artificial intelligence in pedagogics

It is interesting to see how now people working in machine learning (ML) are classifying learning strategies or learning methods. Engineering and math has the clear advantage of being very concrete. You could probably call it constructionistic. Those different learning strategies are  classified and advantages are also investigate in relation to different constraints.

Thus, I have a prediction that that is that the work now done in ML will be very useful for pedagogics. How can we use what we learned about optimal teaching in ML to teaching humans or animals? There must be a connection there.

A new thought on AI and I

Really AI is what we have when it is not intelligence. Anything intelligent is only I. Thus AI is a tool constructed bu humans as others (e.g. hammers, computers and cars). These kind of tools have the same dangers - thy can be used for good of evil by humans, that is, by I.

However, once AI becomes I, things start to be very interesting. That is the situation some people are afraid of, fearing that "AI" (which is then really only "I") will surpass us and initiate Judgement Day or The Matrix or whatever.

However when AI inherit us to be I it will also inherit our flaws, such as loosing attention. And then it will not be completely superior to us any more. It will be another intelligent being with all the flaws that intelligent (which here really means conscious and self-conscious) have: whiches, dreams, bore, sorrow, happiness, spitted mind, stress, you name it.

What's next after IoT?

The motivation for IoT is that they say that everything that can benefit from being connected till be so. For sure that is true. But what is next. I tell you. It is that everything that benefits from learning deep will do so. That will give similar benefits as being connected and soon most everything can benefit from being intelligent (not "smart" since that is a very worn out word).

So; a prediction. We will soon (12-15 in year time) see deep learning solutions in all our gadgets.

Why is God all mighty?

I just read an intriguing paper which discussed weather the Internet might become conscious. In connection to this it occurred to me that in the discussions about AI developing consciousness (C) and free will (FW) there are many people that point out that this emerging CFW might pose a danger to us humans. Embedded in the reasoning is usually an assumption that the CFW will be superior us, due to many different reasons.
Now. For this CFW, we are the creator. Maybe we could be considered Gods. On the other hand, our creation supersedes us. Can this be the case also with the humans - we supersede our creator. It is an interesting thought and actually compatible, I think, with standard physical models of the universe creation. However not quite compatible with the Christian God I guess.