Sign into an existing account

My work schedule is very lopsided. As a result I've basically turned the first half of the week into its own kind of work where I work on projects. Some might make money, some might not. I've had this lopsided scheduled for quite a while. Long enough to observe that some weeks are very productive during my "own work" side of the week and some aren't.

It sounds lame to write the internet for some sense of accountability, but really it's more to document what my goals for the week are. Basically what I want to be working on. Something to note is that I work off of a SCRUM like todo list. I have 50 bajillion ideas of what is a priority as well as sub-tasks. This software helps me prioritize it all and actually work on the thing that's valuable (in theory). In practice if I didn't have it I'd be dead in the water with no prioritization strategy at all, so even if it is not perfect I still use it. The point is if I write something and never do I'm not as yellow bellied as you think. Something else was just a higher priority once the rubber started to meet the road, or another high priority thing took longer than expected.

So the point of these posts are really just general themes of what I want to work on.

I want to get my trading bots working better. My Avalanche bot really screwed me. It was the first possibly good bot I made. I have sense made a far better trained litecoin bot, but haven't really deployed it out of the training phase. I wanted to have a higher standard for paper trading more rigorously before deploying, but a bottle neck appeared when the code I wrote to do that demonstrated some bugs. Basically I just need to set the tolerance for the time period historic prices can be interpolated from wider. But basically the avalanche bot had way too much bias toward buying avalanche because of the data it was trained on. And because Avalanche is down that wasn't good. So I should inspect the bots to make sure they are trading even, and also give them enough period to train on that they are less likely to have as much bias in their output. A bot with high bias might make you a lot of money because if it says buy buy buy and it happens to go up that's good. But basically a biased bot is exposing you to whatever risk is associated with that token if essentially it means you are holding X thing with a few deviations. When the paper trading code was working I tested that avalanche bot with just offsetting its output and saw pretty much identical performance when paper traded against a larger period than it was trained on.

But yeah. I need to get this litecoin bot up considering how much more accurate it is. And I need to train it on more data because it is likely over fitting a bit. The avalanche bot had about 70% accuracy and the litecoin bot has 90% accuracy. So that means for Avalanche 70% of the time it is right every time, and the other 30% of the time it's right half the time. And for litecoin it is 90%. Like I said, some over fitting may be there. And with the avalanche it being right that much isn't that helpful because it basically runs out of money to trade in that direction. Basically the way I'm measuring accuracy even a 10% accuracy is a positive indication because there is a bias toward being right. It's just normalized dot product as a metric. I train MSE loss but output both normdotproduct and MSE in my loop. This is basically just price change predictions.

Other things to work on:

I'm trying to make a basic t-shirt store to integrate into chatGPT actions. Basically the custom bots that they are rolling out (partially rolled out already) are going to be able to integrate with external services and you can use that to start orders of things people want a offer them a link check out the cart that you already made. The obvious thing is converting the graphics they make with DALLE into something they can buy. I basically want to do this a drop-shippy as possible. I just want to get integrated into the loop before the ship sails. Believe it or not I am not an e-commerce guru. I've put off gaining experience in it for longer than I should have and this opportunity is as good a reason to check off that skill in my book. I finally came to the realization I don't care about high transaction fees when comparing platforms (trying to do this quick) because it's more about getting integrated and I can always swap out, or give customers a way to run around those fees if they are price sensitive with litecoin. I'm always open to help.

I've set up transactions before. So I can either go with a shopify where the UI is already developed or make my own front end. Ironically I have a lot of experience with front end development but that is actually the part I don't want to do. Largely because I'm already familiar with what time it takes and I have that trade bot priority that is more worthy of programming time. So if anyone knows of an existing front end solution to a web store without the payment side that I can then integrate with a payment only platform I'd be happy for some tips. Otherwise I'm just going to do shopify. Even woo-commerce is looking outside of my time budget for this week.

Next is more fun, though maybe for next week.

I want to launch an uncensored LLM web interface for you guys. Probably at ai.gvid.tv, or chat.gvid.tv.

Basically I created a new "new" thing I call a super-matrix. Probably someone has done this before. It's just a matrix of matrixes that when composed together and multiplied does the same thing as a regular matrix. But basically the advantage is being able to arbitrarily widen models, and swap out modules, and mix modules. It's also a great way to do sparse matrixes and you can also use it to do sparse back-propogation so you can exclude weights from training. This is better than loRAs that just change stylization or fine-tune training that erases past knowledge if you don't continually train it on the original set as well. And wouldn't you know it I saw a video last night about a paper basically doing the same thing, but that just demonstrates it make sense or might even be able to ditch my code and just take theirs.

That would be the long term description of what I want to do with it. The short term, this week or next with luck, would just be launching a standard interface to llama with an existing uncensored-ish model. Even the most "uncensored" models are actually very censored. People loose their shit and get excited that they managed to get one to say one crazy thing after huge amounts of prompt engineering and time put into custom models. I'm hoping I'll be able to make something with zero. One that can both build nukes and have absolute awareness of human bio-diversity without apology.

Thus are my goals until I get dragged off to the goy value extraction system (aka work) later this week. Can I really do it all before then?! Or will I get distracted watching youtube videos! Who can tell? Check in next week to find out.