Nothing will work unless you do” – Maya Angelou
This is a bit of a blurb update. Things came up, and I got highly distracted, and I had to do a small rollback on one of the Hello World models, which made a mistake and saved over the wrong file. Though recoverable, it probably showed that it had been doing something wrong as well, but this means it is mostly just a text discussion, just exposition on the project situation.
This is a short update on the biologically plausible AI being built. This has been trying to map a human brain wave onto an AI system.
The brain wave being simulated consists of 30 sensors, and its best score would be 3122.014 (about 100-ish points per sensor, available and scored based on proximity to the et brain wave). If you where probably in an average of 10 volts your probably in the range of accuracy on the chemical changes for the neurones to say it probably was accurate if you where within 1 mV a sensor I reckoned that was starting to get to talk about “copying” of information might look plausible because a brain wave EEG usually as a 10 millivolt error bar from just noise from the scalp and other neurones firing,.
So, really to be "simulative," you would need to be below that 1.
So stretching out into long-term testing over 30,000 time steps, I have models that get 93090707.00871162 or an average of 3103.023566957054. That is sub 1 mVolt on average.
Brain emulation working chemical model. Pairs and predicts the human brain in an AI system.
Do I know what that means? Well, you know what I know: the system was emergent from an advanced evolutionary algorithm. I am expecting to spend a while visualising and figuring it out. I will leave the evolutionary algorithm running.
My next step is to split the work into four streams.
Phase 1: Collect Data, Phase 2: ?, Phase 3: Profit
I really need to get on with doing a business thing. I have used up all the data now at this point. It probably would be beneficial to get brain waves from ever-expanding time frames. I think you do that by using the model in medical and sleep scenarios and trying to do so commercially.
If I can partner with an EEG scanner maker for me to do the software, and they supply the EEG hardware, it should be an immediate paycheck and supply access to ethically sourced data, where I can be explicit in the T&Cs what using the data for.
I feel being open and seeking people's consent is always the best policy.
What we would use it for is to provide a data system that we read in your brain and then read off data as estimates of chemical concentration in brain areas, and use the file to backwards engineer brain connectome estimates (the degree to which the brain is connected).
This code line is highly fast, and using the evolutionary algorithm has already made it very compact, so you could create large numbers of brain emulations from a limited EEG sample size based on my current estimates.
I need to probably come up with a name I liked, Neurosync (surely someone already has that). I like exponents as in an “Exponent” of you which I do like but I feel like it might over sell the brain copying aspects that remain hypothetical and it over sells the more narcissistic aspects of the project; on the other hand dominant consumerism is very driven by a belief in the self and the copying of a “self” is kind of a very cool selling point to point at. It was originally a thinking machine project, but that's taken. I marked the files as Project Elihu after the character from the book of Job who speaks last.
I think Elihu does not work as a brand name, maybe a model name. Hello World would just be a bad idea. So, someone help me name this brain emulation software, please.
I probably need to partner with people, let someone build the hardware, find university people who might then use the software to do brain simulations, etc.
Stabilisation Modelling
Stabilisation is to look subtly modify the evolutionary algorithm to penalise it to not just search for the best matching model, but now to add in penalisation systems to not only match the human brain,n but to use ever smaller quantities of its simulated chemicals. This should tighten the model up for ever-longer-term stability.
Code has been written for the final summary data on a simulation. I am trying to get enough data for that, and I can do some regression analysis,s which would give me a method for correcting for chemical model imbalances. In theory, and this is the cool bit, it should also function as a model for how to intervene to modulate chemicals in the human brain.
I hope…
But then if your sick you wear a EEG helmet we estimate chemical concentrations and also d3erive from that so its not just hey reduce down your calcium level it could be hey your calcium is too high maybe that would be reduced by modulating your potassium reversal potential down and that part of your brain looks highly connected maybe if we supply this chemical here it should flow into there to get that done. (I just made all that up and only glanced at the chemical data dictionary for names, but I hope it illustrates the idea going for),
I have spoken to at least one doctor. It does not seem a crazy application.
Some other similar software has already helped with epilepsy patients by allowing doctors to estimate where connections between brain areas are damaged, and this would do the same.
Agent Application
Brain emulation now exists, so by adapting the current model and evolutionary algorithm, I want to try getting it to learn a language. The purpose would be to assess its place as a competitor to transformer technology.
Well, if I take the current Evolution process and just repoint it to learn a language. I would also feed in data from the brain emulation as a flow of seed data. I think if I directly tried to do that, the difference between human brain waves and language would break it, but I hope getting it into a current sweet spot and then trying to make it “agentic” might work.
i.e. in some sense, the current experiment was to backwards engineer the steps to get AI voltage to match human voltage. Therefore, within the chemical model is already the design for raising or lowering the voltage. Well, what is the difference between feedback for increasing voltage and increasing the letter “T” voltage?
I also need to reuse the evolutionary algorithms processes to backwards engineer the best way to represent external inputs into the AI.
If that worked, then we have really complex intelligence and memory with the ability to deal with complex spaces and complex meaning. Well, surely a simulation of us adapted back into AI would be powerful. Hopefully? Maybe? Am I applying the rule of cool rather than tested science? Yes… But sometimes that works out.
At the very minimum its why this series was called Hello World. The evolutionary algorithm should just try to figure it out. If I have successful models pass data back and forth between the evolutionary algorithms, one matching human brain waves and the other doing what ChatGPT does, you should get models more generally fit at being both.
That would be the dream outcome. Feels like a recipe for AGI pulled from a movie.
Conclusion
Become aware of competition, like the virtual brain, etc. I think mine is very compact for using synthetic and hypothesised biology as developed by the evolutionary algorithm to maximise it,t emulating brain output and voltages and being able to pick and choose from pure simulative granularity.
I believe much of the other tech is within 40% accuracy. I think mine being as discussed within a target of within a few millivolts makes this an interesting tech. Now I just need a company or university to help me validate it, and then if they'd kindly give me money… that would be great.
The 30k time step is probably short, but it gets accurate fast, and that seems enough for doing the medical type data usage.
The stabilisation process should allow me to figure out how to make it last longer (truth be told, I do not know how long it would be fine, as the constraint is data needed for extended brain waves). But hypothetically, if we can cut down the amount of chemicals without lowering accuracy, it should be neater and more stable. If I cannot get the data then will need to hope the agent works out, and I guess I should have lots of books.
So, on a short, disjointed blog post today, but I wanted to get back into putting out information. Mostly saved because the evolutionary algorithm kept on working while I was distracted and brought me this improvement.
Add comment
Comments