Week 12

Investments

A new AI software engineer called Devin has been made. You can ask it to develop any piece of software, and even just link to a some web page, and ask it to implement stuff based on that.

Looking at the Youtube comments from the announcement videos people asked if it could just be asked to make a better version of itself. Right now it probably can't but it is based on GPT-4, and can be switched out with a better LLM when that is available.

Having a software engineer is one part, but a neccesary one, of the super intelligent AI I have described previously. Also, the Figure 01 robot has made progress and this is also based on a multi modal LLM.

A video describing the progress is in this video: 

AI Agents Take the Wheel: Devin, SIMA, Figure 01 and The Future of Jobs - YouTube

The point of the video is that everything is based on LLM, and when an LLM improves everything improves. The CEO of Devin says that right now the tool is more expensive in ressources than having an actual software engineer. But everything improves - computation speed, and better algorithms. So what will probably happen is that fast and cheap LLMs will be used for the bulk of calculations, and the more expensive LLMs will be used as needed. The ceiling for how smart an LLM can get has not been reached, and a really big LLM used when a spark of genious is needed can probably be made.

 

In the Devin announcement video another user commented that if we actually have AGI we can just ask it to make a monthly payment. I have previously written that you can just ask the AGI to help you have a good life. And one thing it will probably just tell you, is that you should have sufficient money.

 

And it is of course also possible for an AGI to earn money - but some things needs to be clarified:

• What are your assets? Even an AGI can not make money out of nothing.
• Pollution -  As described in a previously blog post, you can destroy the world in different degrees, so how much do you want to do it while making money?
• Risk
• Timing - when do you want to start to get money and how much

 

If we look at stuff a little bit more high level, the reason you want money is to be able to buy things. But what if you could build those things your self?

Lets say you want to eat a nice spaghetti Bolognese for dinner. When you sit at the table and eat, the meal has reached its final destination and gives value. Before that it is just on the way to give value.

So you could own a field where you could have a robot sowing, taking care of livestock. The robots would then also harvest to feed to the animals, and prepare the animals when ready. When all the ingredients are ready, the dinner can be prepared. 

So you need to own a lot of resources, and you need to have a way to improve said resources, until they are at the final result on the dinner plate.

 

 

If we look at investments at a high level we only have two things we can invest in - ownership of some resource and transformation of a resource to a higher value.

What describes what a higher value is, is what gives value to a consumer. So an AGI that makes money is very simple:

Find out what consumers want to buy. Look at all the resources and transformations that goes into producing that, and invest where there is the highest return of investment - taking into account how the investment should be made according to the investor.

 

One of the important things was the timing - when do you want a return of your investment. You could invest in something that creates immediate value, or you can do long term investments. Or in other words - you can invest in a machine makes value to the consumer, or you can invest in machine that make more machines.

 

Another important aspect is the price of AGI. It is very expensive to run, and that should be taken into account.

So to keep cost down investors have to pool resources. Which means we will have investments companies using and expensive AI and the investors using a cheap AI to choose the investment company.

In general AI will always be chained like that - and inexpensive AI using a more expensive AI, which could then use an even more expensive AI.

 

This chain will also help to move risk close to the investor. The closer the risk is to the investor, the less overhead there is. An example could be some company that needs a bank loan to build a machine. In that case the bank needs to get money from users, then evaluate the company and taking some risk that the company could go bankrupt. The company also takes some risk that the idea won't pan out.

So the risk for the user is very low, and gets almost no money. The risk for the bank is medium, and will get some small fixed percentage. And only if the company defaults, it will not get any money.

What if the investor could instead invest directly in the company and the money invested would give ownership to resources and machines - then the risk would be at the investor, and also the profit.

This would mean the investor would need to do the calculations for what the company should do - but that is what these chains of AI will do for the investor. Using a cheap AI to do pay some small part of an expensive AI, that can tell you what you should invest in and keep the risk as close to the investor as possible.

 

 

Jobs

A lot of places predict that a lot of jobs will be lost to AI. And this is definitely true, since the cost of AI automation is so much lower than human labour, that it will not be worth while for humans to try to compete. On the other hands tons of jobs require a humans or have much more value when done by a human.

 

 

Everything social - taking care of kids and elders. Sports competitions. Social events. 

And then there are all the creative outputs. A painting or vase made by a human is more valuable, since you know that it is the case. On the other hand, create outputs made from AI is also valuable, so the consumer might not care if it is made by a human and not.

 

So humans will only do these type of jobs, and we will still be able to have a very fulfilling live.

The big problem with this is that since all consumers will be doing these type of jobs, which is a closed economic loop. If everybody is barber who will then provide the food?

 

Fundamental democracy

We can look at how a person can have power. The body can directly manipulate things, we can think things through, we have memory and we have our biometric signature.

By using these elements and individual can wields he or hers power. A thought experiment could be that we had a simulation that could perfectly simulate what would happen if everybody would wield their power to the max. Then it would be possible who would come out on top, and instead of actually doing the power struggle, everybody would accept the end result and distribute wealth accordingly.

An example is that the US had to drop two nuclear bombs on the japanese populations, to end the second world war. If they only had dropped one, people could argue that they instead could just have exploded the bomb without any loss of life, and it would have been enough to convince the Japanese to surrender.

But what if the Japan had an AI they trusted and could understand that they had a loosing position. Then maybe there would not have been a need to drop the bombs on civilians?

 

Doing a perfect simulation is of course not possible, and even if it were, it would not be possible, since the participants would have to give up secret information. Which would then mean that they would lose power.

 

So the idea with fundamental democracy is to simplify the simulation, so the power of individual is taken into account, secrecy is maintained and the calculations are cheaper.

 

To do this we start out with assuming that everybodys power is the same. This make sense, since when fundamental democracy is implemented, that will be the case. We can see this in places where democracy is implemented – the state is the strongest actor, and people who can vote have approximately the same power.

 

With that assumptions, we can have secrecy. We just need to have a system where each person can post information anonymously with an ID, and make sure a person only can have one ID.

We then need to figure out a way that everybody agrees to the same law. This can be done in the following way:

 

Each person marks the laws that they can accept. A person can at any time change which he/she accepts. The law with the most people accepting is the active law, that everybody should follow.

This does give a problem that people will change what law they accept, depending on law other people accepts.

It is just like if you ask some kid in a kindergarten if they want to go swimming or to the playground, and they say they will go to what their friend will do. And the friend says he will do what the other one will do.

The solution is that each person also have a description for each law for the circumstances it will be accepted under. The circumstances only relates to what laws will be accepted.

 

One things that should be clarified is what happens if a person have not accepted the active law. In that case the person is against the law. In simplified terms you could say that from the point of view of people that have accepted the law, they are seen as terrorist. And from the point of view of people who have not accepted the law they are freedom fighters.

 

An optimization to the system is that each person can also write on a law a description for a change, that would make him/her accept it. This is a hint on how a new law can be made, that even more people can accept. And also a way to improve existing laws, by removing accepting of a worse law, and then accepting a better law.

 

Al of this could be done without AI, if people where super intelligent and had tons of time on their hands to do just this. Which is of course completely infeasible.

 

But having AI it is very feasible. A person would then use an AI, that would ask question and put the information into the fundamental democracy system. The AI would be trusted by the person, and different persons would not need to use the same AI. The AI would also suggest new laws if needed.

Then one or more AIs would calculate the law with the highest probably to be stable, and suggest that. They should come to the same conclusion, which will mean that the AI controlled by a person will change accepted laws to follow. Given a stable solution.

 

Since the world is big and have different groupings, there is no need that every body needs to agree to everything. The law can be divided into jurisdictions, where different people are allowed to vote in different jurisdictions. Or it can be the same people, who just want to isolate different parts of the law. This make sense since some parts changes very slowly (freedom of the press), and other very fast (financial laws).

One question on could ask is why fundamental democracy is needed, when we already have existing democracies.

And there is really not much of an advantage compared to representative democracy. You can use an AI to evaluate if politicians make sense, and the politicians can use AI to do their daily job efficiently.

But it still make sense for existing democracies to implement to put pressure on totalitarian states.

The big argument from the Chinese government why they don’t want to implement democracy or freedom of the press, is that people not educated enough and are also easily manipulated.

So if an AI driven democracy works those excuses are removed.

 

But would such a system be implemented in current democracies. The answer is probably yes, for two reasons. The first one is that it is based on the fundamental power structures, so the best system for implementing will win. The power and needs doesn’t change, just because a different system is used.

This is in stark contrast to e-currency, where the system it self creates a limited resource. Which means that anybody can just take that code from a e-currency and create a new one.

The other reason is the system can be made in a non-enforced mode. So people can accept the law from the fundamental democracy, but it has no jurisdiction since an existing democracy is already implemented. But it can generate the laws, and if the existing democracy find that the fundamental democracy have generated sensible laws it can change.

 

Open Source renaissance

With AI software engineers open source will probably see a big surge. At the moment you need to be extremely skilled to contribute to open source. But if an AI can do all the complicated work, and individuals only have to supply ideas and money, we will probably see a big transformation in this space. The example from Devins homepage is also just how to run github code and fix github issues.

 

Iterative neural network

To have an AI that is cheap and capable enough to run all of the above, we probably need to make some improvements.

As previously written about, a smaller neural network can be made, that uses a big prompt, instead of relying on all information being in the LLM. This can be improved by having the neural network be queried many times for generating a single token. We could call this recursive neural network or repeating neural network, but both of these are too similar to recurrent neural network. So I just call it iterative neural network to differentiate.

But the idea is that it keeps some state between runs, and use that to decide which part of the neural network to run. This means that the neural network will be much better at logical thinking. What part of the neural network that will be rerun, is of course part of the end-to-end training.

So we need to balance how much to put in the prompt, how big the neural network and what it knows, is and how many times we query it.

The last piece needed is to change it into a reinforcement learning problem. This is quite easy, since we just need to be able to create training data with question/answers.

If such an AI can be made, it will outperform existing LLMs by a high margin.