Policy on Education

Education is at the forefront of change. Ways to learn are expanding and improving. Ways to evaluate learning are under strain, and are being forced to return to individual personal interaction. What we need to learn is about to make an historic shift.

To learn about anything, from cleaning mould in grouting to higher mathematics, there are wonderful videos explaining them clearly. But we only learn by doing. There are places where you can learn by doing. A nice one is at BBO (Bridge Base Online) teaching declarer play. But there isn't yet one for defence at Bridge, which is a cooperative activity. What is surely coming is AI systems that watch what you do, in the real world as well as online, and uses that to determine the gaps in your knowledge, and recommend ways to bridge those gaps.

Even before AI, using unsupervised homework for evaluation was unacceptably prone to cheating. The rise of AI has made that problem worse. A rough remedy is to have tutoring staff interview the student about their submission. An obvious step is to train AI to do that interviewing, but it would need to happen in a controlled environment. In the previous paragraph, I speculated on education by AI watching attempts at learning by doing. It is tempting to think that that could be combined with evaluation, but that would then be a bad learning experience, or a bad evaluation method, or, very likely, both.

I suggest that Australia should put its main effort into providing curriculum development with a trusted evaluation system. It should unambiguously link a person, with multiple biometric identifiers, to an unfakeable evaluation result for a specific curriculum. It should provide education for that curriculum, but people should be able to just pay for the evaluation and get their training elsewhere.

Which brings us to the question of what we should teach. Historically some people get to give orders while most of us get to receive them. So training to give orders and manage the people carrying out the orders was an elite role. The future will be different.

Everyone will manage AI things which: (a) know a lot; and (b) are good at reasoning. However the intelligence of these things is not human. They lack empathy, and they lack their own motivation. Dealing with them is an important skill that everyone will need to learn.

The useful skill that AI has now is coding. But to take advantage of that you have to know exactly what you want to create, and you have to express the instructions in clear unambiguous English. This is far from easy. Indeed, though coding is not trivial, it has never been the hardest part of getting computers to do useful things.

So coding is where we start with educating people to use these alien intelligences. But soon enough everyone will need to extend this capability into every area of life. Whether we want the dishes washed or to build a skyscraper or design a machine to do some task, managing and coordinating intelligent machines will be the way it is done. You have to take advantage of the machine's wide knowledge and reasoning ability, but it doesn't understand the human needs that we are meeting, though it is very likely that it thinks it does.

We expect that everyone will be able to get more done, but to do that correctly and successfully means that they have to understand a lot. That is the challenge of future education.

Policy on Housing

 Young people in the developed world are increasingly unable to transition to home ownership.

A century ago something similar was happening in America.

A big part of America's solution was the 30 year fixed mortgage. This has a serious downside. When interest rates are rising people are stuck and can't move to where suitable jobs are, because if they do they lose their cheap mortgage. The effect is worse than the way that Australian states depend on Stamp Duty for revenue.

The correct answer is the equity-preserving loan. The interest rate needs to vary because it needs to be the same for existing and new borrowers. But we also don't want a situation where falling house prices mean that sellers will lose equity they need to buy elsewhere. The equity-preserving loan works like this: (a) interest payments are lost, of course; (b) however repayments that reduce the loan create equity which is preserved as a proportion of the house value. So if your house cost 1 million, and you're loan was $500,000 then your equity at the start is also $500,000. Now if you've paid off $100,000, that is 10% equity that you've gained and you are up to 60% equity. Even in a falling market you retain that 60%. So if you now sell for $900,000 then you only have to pay back 40% of the loan when you sell, in the example case that is $360,000 instead of $400,000. Assuming house prices have changed uniformly, you are just as well placed to buy a new place as you would be if you'd sold for 1 million in a steady market.

Note that this scheme means that home owners can't go under water and get a margin call when their house is worth less than the loan. Which in turn means that buying a house with nothing down is not necessarily a problem.

Well there are complexities and potential for cheating. So the government would make up the difference to allow equity-preserving, but only if you buy and sell through the government housing agency.

The 2nd part of the plan is cheap housing with zero deposit in exchange for some National and Local Service. Of course the most important National Service we need is the production and nurturing and education of the next generation. We also need to create local communities that work, starting with cooperative child care.

The way to make cheap housing is to get Chinese companies involved. Flat pack housing is a start, but I think it is possible to do a whole flat pack suburb to suit an allocated area -- everything from the buried services to houses and support buildings like schools, medical places, and more.

Policy on AI

AI is useful, but has a lot of potential for psychological and economic harm. 

Firstly we need to firmly establish the fact that AI is just a lossy database of its training data. It does have an ultra sophisticated user interface with its firm grasp of natural human language. This view of LLM AI is necessary to sort out copyright questions. Of course humans are also a lossy repository of our training, but we are explicitly allowed to learn from copyright material. Machines are not.

Secondly we have real world actions taken by AI. They cannot be treated similarly to humans for the legal implications of those actions. They are like children, and the people or corporations that control them must be substantially responsible for their actions. The fact that we don't understand an AI's thinking means that they are hard to control, but their controllers can't use that as an excuse. Actually we need to improve the design of AI so that it comes with the constraints that the law requires, and cannot escape them.

Next we come to the ways that AI pretends to be human. At a mundane level, we expect that the use of personal pronouns means that there is an entity which has continuity and memory and a name. A judge should be able to ask it what it did, and also why. Of course even humans often don't know why, at a particular moment, we did some thing, and AI has even less capability for introspection. AI that has started up with training but no memory or continuity, should not be allowed to interact as if it was part of an interaction between continuing intelligent entities.

I've somewhat changed my mind. It is hard for AI things to talk without using the personal pronoun. However they should introduce themselves. E.g. they could start by saying "I have just been created using the Gemini-7.3 training done in May 2027". You should then be able to say "I'll call you Betty, you can call me Al", and then at a later stage ask their name to see if it is continuing or newly minted. Long running AI with continuity and some memory of their previous actions should be given unique names in some way.

Even if AI is intelligent, it is not human and it is never going to be. It cannot experience human emotion and must not be allowed to pretend to. It can't understand the emotions of humans by empathy, as we do, and must not be allowed to pretend to. It can use scientific knowledge to understand human emotions, and their likely effects, and this can influence what it says and does. The reason we need these rules is that people are otherwise likely to misunderstand their relationship to the AI and be made vulnerable to many bad consequences, depending on their situation and mental health..

[update] Humans should always know whether the entity they are communicating with is another human or is a machine with AI. This needs to be written into law and enforced.

[Footnote: AI is just software, but even its creators don't know how it decides what to do. A possible mechanism to manage such a situation is to enforce that, when the AI wants to act it can only produce verifiable software to perform the actions, and then that software can be externally checked for legal safety, before it is executed.]

Youtube video on aging

 I made a youtube video on aging: 

Not very successful. I went a bit easy. I left out the bit about alzheimer's and cataracts being caused deliberately by our genes, though it is partly hinted in a slide.

Suddenly aging communication is everywhere. We see a conference at Berkeley: https://www.berkeleycal.org/. And then PBS has a video:

Which says stuff I disagree with: in particular competing with your descendents is not group selection. And even mammals that aren't sociable don't move as far from their descendents as birds and fish commonly do.

Aging and evolution

Recently Nick Norwitz pointed out that there was a variant of the ApoC3 gene that was associated with longer life (https://open.substack.com/pub/staycuriousmetabolism/p/the-untold-story-of-apoc3-and-human). Why then isn't that variant more common? Well you might think you'd be fitter, in the survival of the fittest game, if you lived longer. Unfortunately your genes don't especially care about you, they care about themselves. And what they want is for you to live long enough to do grandparenting, then depart and stop using resources that your genes would prefer to see go to your descendants. 

So now we see that ageing works like light coloured skin. There isn't one gene for white skin. There are lots of genes which affect skin colour, and we get a cocktail of these depending on how far our ancestors lived from the equator, plus what particular ones have appeared in that area in the past.

Similarly with ageing: genes which make your life too short are selected against because your grandchildren lack that extra bit of care that human (and orca) babies need. But genes which make you live too long are also selected against because then your great grandchildren and other descendants have less resources, after you are no longer needed for grandparenting.

What this means from a practical point of view, in our modern world that is awash with resources for humans, is that ageing is not a malfunction. It isn't caused by errors but by deliberate choice. We can track down a lot of those choices by looking at the DNA of people who live a long time. And the problems that these genetic choices make for individuals might be relatively easy to fix.