Notices by NEETzsche (neetzsche@iddqd.social), page 11

I noticed that Latin America has a fixation with Bart Simpson smoking weed and skateboarding

Don’t RT memes from instances that defederate you. Just steal them outright

One thing I noticed tech spergs are guilty of is they have all this expensive hardware and equipment and shit and then they don’t actually use it for anything. Spend a small fortune in hardware, use it to host a NAS server for pirated animes and to shitpost on fedi.

EDIT: Making this into its own OP

Yeah but for now it’s going to automate some of my elfgame bullshit.

I’m literally just hitting the GPT-4 API for this. Locally-run LLMs right now are dog shit, but because of the kind of people who work on them, they will get there… eventually. All the /g/ spergs are into them so they can have ERP LLMs and to make them say nigger. It’s just another case where ZOG will be humiliated by a food stamp recipient who wants to be mommy dommed by his computer.

No. I did Sheetposter to see if I can take GPT-4 through a complicated process. I would characterize what I did with it “prompt engineering” unironically not because anything I did is impossible or anything, but because I literally made it transform the prompt based on previous choices in the chargen process. For example, it submits to GPT-4 the list of choices you can make when you’re choosing, say, Gifts or Rotes, but it filters out the ones where the sheet so far doesn’t meet the prerequisites. It also provides simple descriptions of them in the prompt to give it context, alongside the partial sheet that’s been made.

If GPT-4 can, with this kind of guidance, navigate TTRPG chargen, it can navigate the tax code or something similar.

I wrote a separate ANSI art script that takes the JSON character sheets that Sheetposter generates by going back-and-forth with GPT-4, transforming the prompts to accommodate the range of legal choices based on previous context, and dumps it out. On the left, the code, and on the right, the code. It’s spaghetti garbage, with zero apology.

It just werks, bich niga.

WIP: New Feature / Big Fixes

Sometimes. Fine-tuning and training generative AIs is more art than science. It can make the performance worse if you do it wrong

Anybody who actually tries to build anything is in this post.

Fine-tuning isn’t even that hard or terribly expensive. And depending on your definition of fine-tuning (there are several) it can include things like using embeddings

I will also get to dunk on “clean code” faggots because it will take everybody’s batshit spaghetti code that works and return beautifully-formatted and well-organized professional code that doesn’t.

Actually, doing this might be really funny way to embarrass OpenAI if I just call it GPT-PR or something like that and let it go hogwild.

Write a Ruby/Python script that downloads random GitHub projects, tells GPT-4 to fix them, and submits pull requests with whatever garbage it returns. Make sure to use your real name on these PRs so they know you fixed their shit and are a prolific computer programming genius

What it’s going to do is automate away a very specific kind of entry level programmer from India who can’t do fizzbuzz but can pass those certification tests. Which is most people with job titles like “Software Engineer” anyway. So there will be job loss, but I doubt it’s going to get rid of people who actually build things. That requires vision, and LLMs are notoriously robotic for obvious reasons