how do you train and align an AI when all the rest of the world thinks the same way, producing trillions of tokens of training material and you are left with billions of tokens since your world view is dramatically unpopular?

can billions beat trillions? we will see.. i have to find a way to "multiply" my training data orders of magnitude to successfully counter the existing programming in an open source LLM.

first i give a smart LLM a 'ground truth' text. then i give it the following prompts:

```- You are a highly skilled academic analyst.

- Analyze this text and find 3 bold claims that could cause controversy and division in public. List the claims and also state why they are debatable. Give numbers to the claims.

- Convert these claims into binary questions (that could be answered by yes/no or this/that).

- Now put these questions in a json format. Please also add the info about which of the answers concur with the original text and the question number.

- Write some supporting arguments for 1st question, with respect to the original text, concurring and confirming the original text.

There must be about 300 words. You should not mention the text, write it as if you are the one answering the question.```

the result is usually instead of a few sentences of opinions in the beginning now the material is expanded to lots of words, yet still parallel to the opinion in the original text. LLMs have all kinds of ideas already installed, yet they don't have the intuition to know which one is true. they can give you a ton of reasons to support anything.

using this method i can multiply billions to tens of billions probably and have a more effective training.

We build cathedrals to capture the patterns and man sits in awe.

We build machines to animate the patterns and he shudders.

all social media IS a serialized string. Unless one could search within their own posts using a keyword or hashtag. Blogs are a little different because there is a sidebar that organizes posts by month and year.

- dunno what you mean by on nostr people use tags for application logic.

interoperability is a good word. There should be more interoperability between the good apps. Like reddit + nostr = roster hehe

> A giant serialized string with an array of undocumented tags is no way to live.

… isn’t all social media basically like this?

Validated on both prefix types in a nostr client.

😬 well that’s embarrassing. Spent too much time on performance and not enough on testing. The math was correct, it was just backwards. Just pushed a fix.

Alas, the nsec doesn’t actually match to the public key. When I paste in the nsec on Damus or primal, it doesn’t show an account with that public key. Something is seriously broken.

Is there a reason the private key is exported as hex? Could you change it to the actual nsec? Seems inconvenient…

Thanks for sharing those links, it means I just need to look harder! I've had better luck using the .run self contained files than the rpm packages (I don't think nvidia actually tests them XD)

nvcc is not available on the system. Yeah the other issue is that I'm forced to use a really old driver and older kernel. So the target machine specifically is running rocky 9 with kernel 6.14 on driver 550.163 with cuda 12.4. It looks like the toolkits available from those links you shared require cuda 13 or newer. I think I have to find an old package.

Thanks for following up. With vGPU it's nvidia grid virtual gpu. So it's a pair of tesla p100s with a "custom" gpu profiles into virtual machines. If you're unfamiliar it requires nvidia proprietary driver distributions to work. All drivers and guest software (usermode executable etc) come packaged together. and I don't see the toolkit bundled with it.

That said I also am unfamiliar with the toolkit on linux. Not sure if it's available through standard dnf distribution channels. Specifically rhel.

Should work with any NVIDIA card with Cuda drivers. When you say vGPU, what type of setup?

I built and tested it using the 3070 I have at my desk. I’ll go test it on an H200 now and report back.

What about for commercial cards? :) Specifically vgpu. You wouldn't happen to know where/how to get the toolkit? I'll have to check my driver packages but I don't think it's included with the grid drivers. Hopefully it doesn't require a consumer driver package.

Took about 10 days on a 3070 to find this 9 char prefix.

Venv is half the source of my problems because I cannot conceptually hang onto how many different ways and locations a different python version or set of tools is installed.

It *sounds* like a solution, but it’s just the thing I end up fighting with for hours after nothing works the basic way.

There aren't tons of choices for where to download most apps from.

I don't see a point to something like gitea. If it was going to defeat Microsoft, it probably would have by now.

people say the same thing every cycle and have always been wrong but hopefully you are right this time