Saturday, March 8, 2014
Oh man. Ohhhh man. It's coming. That macro AI...it's coming
Today I had a massive revelation concerning macro AI and how to think of corporations. The key is a very simple conceptual change: don't think of large-scale AI in terms of discrete actions, think of it in terms of processes
. Same for a company - don't think of it in terms of the actions that it performs, think of it in terms of the processes
that it sets up.
Let's have a look at the implications of this. If you want to get really cute, you might look at this as a shift from 'time domain' planning to 'frequency domain' planning, in a similar way that the scanner shifted
Instead of planning discrete actions, for the macro AI we want to plan periodic processes. Instead of 'planning trees,' 'process trees.' This way, we require no effort to continuously do something, which is a good thing.
Instead of thinking about preconditions, action, and postconditions, when an AI is designing a long-term plan (especially for a company's business strategy, for example), the AI will think in terms of input, process, and output. It will connect these together like a circuit to form a business operation. For example, suppose we want to build a business to produce product P1 and sell it to market. We can have a process 'produce P1 at L1,' where L1 is our factory site. As an input, the process requires the raw materials that are needed to produce P1 - call them R1, R2, R3. Let's say we know where to find R1 and R2, because they are raw ores that can be extracted from nearby asteroid fields. We can set up processes 'mine R1 at M1' and 'mine R2 at M2', two different mining sites that will provide the raw materials. These processes happen to take no inputs. However, the outputs are R1 @ M1 and R2 @ M2 - location matters, so we must set up two additional processes - 'transport R1 from M1 to L1' and 'transport R2 from M2 to L2', so that we now have R1 @ L1 and R2 @ L1, which is exactly what our production process requires. But, alas, we don't know where to find R3, and we still need it to produce. Luckily, OmegaCorp happens to sell R3 in bulk. We create a contract process with OmegaCorp to provide R3 @ L1. As an input, the process takes R3 @ L1 (which belongs to OmegaCorp), and credits (which belong to us), and outputs R3 @ L1 (which belongs to us) and credits (which belong to OmegaCorp). So a contract process basically just changes the ownership of the inputs / outputs. Contract processes are the boundaries that designate where one corporation ends and another begins. And voila, now we just need to create transportation processes that take P1 to markets, and we've got a working business! Once we have the structure of the process tree set up, it's very simple to quantitatively evaluate the the net input / output, which can tell us whether we've created a profitable business or not (in this case, it depends mostly on how cheaply we can buy R3 from OmegaCorp, and how good of a price we can get at the markets where we choose to ship P1). Furthermore, from the process tree we can ascertain things like 'startup cost' to set up the processes (we will need to purchase mining equipment and production units). To satisfy startup costs we can try to start small and fund the venture out-of-pocket, or we can try to start medium-to-large by IPOing and leveraging the shareholder capital to purchase our equipment. In the latter case we will probably need to divulge the process tree to prove to other NPCs that it is a profitable business model.
Already, with this conceptual model, you can see how an AI can quickly learn to build massive
businesses. Once the process management algorithm is in place, I see no reason why an NPC couldn't outperform any real-life operations manager
A really intelligent NPC could, perhaps, create operations involving thousands of processes occurring all over a region, perhaps involving hundreds of contracts with third-parties
For the AI it is no problem. And the process model is so conceptually elegant that...ah. You know. It's just great
Now, it gets even more exciting when we think about 'meta' processes. This is something that I was never really able to understand in time-domain planning: how to get 'better' at something. But in terms of processes, I think I really understand it. You can understand that every process has factors that influence throughput. In general, 'quantity' is always an implicit factor - you can always assign more 'processors' to increase the throughput of the task (be it production units, subordinates mining, etc.) However, consider that the process of mining can be increased in throughput by increasing the transfer unit efficiency. You can understand this to be a 'meta' input to the mining process - although it's not actually an input, the transfer unit efficiency influences the way the process works. Hence, if you want to invest in increasing the efficiency of your processes, you will want to achieve higher transfer unit efficiency. Suddenly, we see exactly where research becomes a demand in the economy! A big mining corporation that wants to focus on increasing operational efficiency will have a demand for new transfer unit technologies, which can lead to the genesis of a new company dedicated to transfer unit research, for example. One minute you're trying to build up a great mining company, the next minute you're seeing the IPO of a new research firm with a mission statement that is suspiciously similar to...well, exactly what you need! Dynamism, supply and demand, the birth and death of companies, cooperation and competition, ...just a few of the buzz words that will need to be used to describe how amazing this macro game is going to be
I know I've been fairly incoherent today and yet have only just touched on about 10% of this new macro-AI model...but...can you see how powerful this is? Can you see how ridiculously cool this is going to be for macro AI? Gah. So. Awesome!!!
I haven't even mentioned things like the demand for protection on trade routes (creating opportunity for merc corporations), the demand for corporate espionage and outright attacks (creating demand for destruction of the competition's assets), etc. There's so much to talk about. But let's save some
of it for another day
PS ~ Yes, today's revelation was also inspired by textbooks...today I read some 'operational management' and 'strategic management' books...man, I never realized that stuff actually means something
PPS ~ In case you haven't noticed, I'm on a roll. This month is going to be amazing
PPPS ~ Oh, also forgot to mention - process trees? LOD simulation? Do I even need to say it? The net effect of a process tree can be expressed as a simple list of time-dependent changes in the world. Simulate a year coarsely? Sure, just multiply the effects by a year. Boom.