MRiabov's AI short execution thread

[MRiabov]

Damn, just gotten a message on MJ's email list that inactive accounts may have their grandfathered status revoked..

Ah, I have to distract myself from actually executing to retain the status.
So let's make a little execution thread right here, something I planned to expand on about 1-5 months later:

So I was creating my invention - the machine that cooks anything you want with a press of a button (10 months of work in, 3 iterations later), and thought... Damn, is saving one hour a day for likely millions, perhaps billions (yes for real - think about it), people the best that I got? I want to do more and I can do more.

After debating it for 4 days, I came to conclusion that "the man is measured by the amount of value that he brings into the world", as usual, and made the switch.

So:
I'm creating an AI that should be able to do mechanical and electronics machinery engineering, and I've been at it for 5 months. Why? Because:
1. Engineering takes hell of a lot of time. As with my last product, I had to work for 10 months, and I haven't even accomplished everything. The AI? It could do that work in hours - modern computational machinery can do computation on 2.5 petaflop/s (2.5*10^15), and that's only one GPU, compared to humans which can - what - 1-2 good thoughts per second maybe amounting to a few million operations per second? The AI is also much more competent. I believe we will be able to design inventions in hours, not years.
2. Machines can automate things. For most people, work is suffering. I dedicate my own self to creating a piece of software that would automate, I hope, 25% of labor in the world. Big goal, but actually most of it can be done via a few robotic joints and PCBs. Farming can be fully done by machines. Factories can be entirely replaced with machines, some are already.
So why aren't ALL factories fully automated yet? Simple. We only have 25 million engineers in the world. There is simply not enough inventors to fulfil all demands of society. This will do it. And also much, much more..

Insofar, I have (in order):
June 24 - a full month of research, basically invalidated all incorrect assumptions I had about what can be built and how can be built
July-October 24 - let's say I have moved in to live with very wrong and disingenuous people here, and execution was slowed down by 70% during this whole time, I literally couldn't concentrate on shit. Out of 3 months, I've spent 1.5mo bootstrapping and trying to move tf out, but I only succeeded to the end of October. However, in the other 1.5mo I've written code. Code I will most definitely use, however later. The kind of code that will cut my costs (which are big due to amount of compute that has to be done, by 5-7x (yes).
Oct-Nov - I've moved out and ......... ... Damn, I work 3-5x faster that I did there! Hooray! Big gains on product progress and thousands of lines of code are written.

Just to be clear - this is perhaps one of the biggest attempts in my field. The practices that I employ to achieve this goal - I believe nobody has ever applied them, not in my scale. I have to pass about 70 different tensors to the model to totaling about 9gb per training loop to make it happen. Amount of compute will be immense for this project, although I've prepared and found ways which can reduce the costs before I've even written the code.

I actually intended to start training this Sunday, but I found a few bugs, and I needed to rewrite something. (it's Wednesday now). I actually have no clue how many bugs I will find out before I train, and whether the training will converge at all.
It most obviously isn't your average ChatGPT-LLM project, it's something that does much, much more.

My backup plan is to use Innovation Vouchers Innovation Voucher | Business Support | Enterprise Ireland, which is essentially government giving 10000€ of free time from researchers - that's 300 hours from smart people who do it for career. Quite sure I'll need it during the bootstrap phase.

Right, I'll update when I'll have the first training results, should be soon.
To be fair: this will not be a fast project. Up until I will train enough to create something useful as, say, an injection molding machine, it will take months. Because to construct an injection molding machine the AI needs to know a computational fluid dynamics, melting and solidifying dynamics (for plastics), control, and general physics sim; also how to interact with the code that I've built. And that's a lot of data, and helluvalot of compute. No doubt it is possible, but I don't know how much I need to spend to train it; trial runs will tell.

Good news is that the best ML algorithm to date has already beaten minecraft diamond collection without configuration. It only took a single GPU and 9 days, which is a few hundreds of dollars, not millions.

If I'd have some x*10^4 value-vouchers for research purposes, would be easier of course, but I'll use what I have to use. To have more, I also asked my college for free GPUs which they will give me in 3-4 weeks. I also will acquire free startup credits from AWS, Google Cloud, and other "hyperscalers". A thousand dollars from each will go a long way.

P.S. If anybody is willing to be negative for whatever reason, I won't join the forum in the next few weeks for sure, so may as well not try.

Cheers everybody, and create great things,
Maksym Riabov

(shit, I didn't intend to spend an hour writing this! (what's the point if I don't even have the model trained?))

 

Dislike ads? Become a Fastlane member: Subscribe today and surround yourself with winners and millionaire mentors, not those broke friends who only want to drink beer and play video games. :-)