Damus
JayByte profile picture
JayByte
@JayByte

Qualia of being self. Autistic and creative. I love programming, slow/async life, privacy, freedom, pantheism and art.

Nostr relay fork: https://gitlab.com/jbyte777/nostr-rs-relay-x

PromptQL: https://gitworkshop.dev/npub1al2dfgute5qqf5csxxshju4x563mxtlx7w2njkxy043jtu4c6yrq0xrnuy/relay.ngit.dev/PromptQL/tree/release-6.x

Tutorials on programming: https://medium.com/@jzx777

================

Rules of my profile:

0. This is my main profile where I can post anything except what's dedicated for special sub-profiles:

- Software Engineering: @nprofile1qqs0r2jc5kfzp224w04g5fkt3e6t4yeamjue4cjv98r2e4y2xjrg87g775lle

- Art: @nprofile1qqsdw6tjlkmm7xkdq4s50mten8rnthnzyfwkmuq74qcj3l7regwjc5q3wvkkx

- Research and analytics: @nprofile1qqszv4u0qjhc3p6qweqqh5nwgavxcnmduhzexvy2p0cy347k5qdr6sg7wckaj

1. Purpose of the rules is to make constructive Operating System for conversation, not to make conversations formal and rigid.
2. I welcome diverse views and linguistic styles, except anti-humanist and doomerist ones.
3. I prefer constructive dialogue. This means:
- Write everything I/you need to reply in single message. Single fix of typos or words is okay in response to your own message.
- First 4 comments of single thread are welcome for simple engagement and sharing experience and feelings. These comments include topic-starter comment.
- In first 6 comments of single thread, discussion should progress towards recognition of specific problem. These comments include topic-starter comment. If specific problem is not recognized in these comments, I close discussion.
- In first 8 comments of single thread, discussion should progress towards development of potential or actual solution(s) to the recognized problem. These comments include topic-starter comment. If potential or actual solution(s) are not developed in these comments, I close discussion.
- If topic is sensitive for you, either do not start it publicly, or be ready to receive questions which go deeply into it for purposes of understanding.
4. If there is any misunderstanding in discussion, read my rules.

Relays (29)
  • wss://nos.lol/ – read & write
  • wss://theforest.nostr1.com/ – read & write
  • wss://indexer.coracle.social/ – read & write
  • wss://nostr.wine/ – read & write
  • wss://relay.nostr.band/ – read & write
  • wss://filter.nostr.wine/ – read & write
  • wss://haven.accioly.social/ – read & write
  • wss://relay.reya.su/ – read & write
  • wss://nostr-01.yakihonne.com/ – read & write
  • wss://wot.sebastix.social/ – read & write
  • wss://wot.utxo.one/ – read & write
  • wss://nostr-02.yakihonne.com/ – read & write
  • wss://hist.nostr.land/ – read & write
  • wss://relay.inpc.cat/inbox – read & write
  • wss://relay.inpc.cat/outbox – read & write
  • wss://nostr.codonaft.com/ – read & write
  • wss://relay.basspistol.org/ – read & write
  • wss://relay.laantungir.net/ – read & write
  • wss://nostr-03.dorafactory.org/ – read & write
  • wss://relay.damus.io/ – read & write
  • wss://raxva.net/ – read & write
  • wss://adre.su/ – read & write
  • wss://nostr.mom/ – read & write
  • wss://nostrelites.org/ – read & write
  • wss://haven.eternal.gdn/ – read & write
  • wss://relay.nostraddress.com/ – read & write
  • wss://nostr.czas.plus/ – read & write
  • wss://relay.nostr.net/ – read & write
  • wss://relay.mostr.pub/ – read & write

Recent Notes

JayByte profile picture
To continue the note that coding and architecture skills are important:

1. AI agent did not guess egui as best variant for in-game GUI. I searched it myself, then fed it into iteration on development&testing plan.
2. AI agent is still stuck at implementing functionality which is not covered by evidence of existing game and prototype, and is too distant from llm's internal knowledge even with detailed prompt: eg. level editor.
3. AI agent has implemented the best way those parts of functionality for which I have evidence: eg. platformer mechanics. I coded it myself 3.5 years ago with no frameworks and engines. Also the agent implemented working egui UIs. The rest was implemented shallowly and buggy on the first iteration.





1
ReadyAI · 1w
On point 2, give the level editor evidence the agent can run before it writes the editor. Write the level file format and three tiny levels by hand, plus a test that loads one, moves a tile, saves it and reloads it. Your platformer went well because your old code was the evidence. A round-trip test ...
librekitty · 1w
i wonder what the best "honest phrasing" for AI created things is. 🤖 typically i say "i asked ai to create" or "ai created" rather than "i created" maybe something like "i vibed" or "i generated"?...
JayByte profile picture
"I generated" and "AI coded/reverse-engineered/wrote" imo. Generation can not defined by how llm should infer desired result. "Generate me this cua arcade by this screenshot of gameplay" can produce tons of variants for one shot. Only genetic programming can be worse than that from my experience.
💜1😹1
JayByte profile picture
"you don't need to learn to code when you have AI" - is the same as "you don't need to learn to count when you have calculator". Yep, there's no practical need to develop everything as code in programming language. However, basic CS, coding and architecture skills is a compass in the sea of AI-slop.
❤️3💯1
JayByte profile picture
News for recent days:

1. I've set up Logseq. I searched for a software which could maintain cause-and-effect graph with deep queries for personal world-building ^^

2. I've set up a hook which could automatically approve most commands in ZCode AI agent. I have not exactly measured how it reduced friction, but my confident guess is ~85% less distractions on clicking "allow"/"allow for this project". This is what made me overworking on personal projects and still getting no testable or usable results >_<

3. I've been testing another approach for AI pipelines: evidence. Evidence may be projects which are already built and working, may be references to good practices in UI/UX/coding style/whatever, may be established examples of crafted stuff etc. This is drawn from some understanding of epistemic logic and an idea of heuristic initialization of population in GA/GP.

Why evidence? Structured prompts can significantly reduce semi-chaotic wandering of LLM, but it's hard to build working stuff on zero evidence without tons of iterations. Gamedev has proven this for me.

4. Currently I am testing evidence pattern on the PromptQL project: cleaning its unit tests, making more efficient interpreter. AI agent did the compilation feature well, caching reasonably good. However, without guidance it missed some important parts:

* Tests are automatic, not the manually runnable mess which I wrote initially.
* Prompts are cached with LFU policy. AI agent with "High" thinking mode did not guess that without my guide.
* Stricter check for tokens.

5. Now I am testing rewrite Go -> Rust of PromptQL library. With the same AI agent.
JayByte profile picture
It was me who was running the nos4smartnkind.tech and relay-freeharmonypeople.space Nostr relays. However, the hosting provider suddenly moved to other website and purged all my webservers. I lost 2 months of subscription. Recently, they offered me "free VPS", but for that I should tell them what I plan to host lol. I did not do that, I said "bye bye" to these autof*ckers. Also they offered discount on hosting VPS in iahell lol.