lucid

first principle thinker, ai-generalist, e/acc, space industry optimist, ex-finance, ex-psychology

[lucid @lucidpaths33]

Ontological Refinement is a structured, continuous process that treats identity, habits, and beliefs as evolving—constantly improved through deliberate, self-guided cycles rather than static attributes.

[lucid @lucidpaths33]

@e0syn i like all the carzens dw

[lucid @lucidpaths33]

@Teknium having both astra and claude in cli and hermes using gpt vis oauth, a little bit of hoop jumping but fable is a great orchestrator

[lucid @lucidpaths33]

@vincit_amore @boneGPT vin ong just get bumble or hinge and have me or bone look trough your profile description (or pics if u fine with doxxing)

but fr it does not need to be like that 💀💀💀

[lucid @lucidpaths33]

@boneGPT @vincit_amore

[lucid @lucidpaths33]

@graphtheory oki that kinda aligns with what @vincit_amore said in a way

im inbetween the thing, i think its good, but i also think its not good enough yet lmao

[lucid @lucidpaths33]

at what point does one release something into opensource ?

is there like a rule of thumb for it or is it just gutfeeling?

[lucid @lucidpaths33]

@vincit_amore i told it to score. but you are right

i will tell it to create a determinsitic scoring script

[lucid @lucidpaths33]

@vincit_amore soon then, ssp judged by fable is already decent
but i want to be at least #2 everywhere >:3

(and thy vin, thats a good rule of thumb <3)

[lucid @lucidpaths33]

@e0syn assbench

Agi if the model can computer use this to oneshot

[lucid @lucidpaths33]

Sovereign Software Production

A control plane for machine-authored software change, reviewed as a product against the tools it competes with and borrows from.

What it is, in plain language

An operator initializes a program with a consequence level, then resumes it. A single controller process, woken by one durable event, performs at most one state transition: it publishes a prepared receipt, may call one bounded local model and one closed tool operation, publishes a terminal receipt, compare-and-swaps CONTROL.yaml to the next generation, and exits. Nothing in model prose can widen the tool set, the filesystem roots, or the external targets.

Worker output is untrusted data that the parent re-hashes before accepting.

The unit of value is one authorized, receipted transition that survives process, session, and model replacement.

I had Fable measure it against others
below are the stats

still tinkering with it but im already using it to improve itself lol

[lucid @lucidpaths33]

@tunahorse21 im tinkering on something you might like in that regard
nothing like the first stuff I got you to check out months lmao

this time its some proper code XD

[lucid @lucidpaths33]

@graphtheory the goat

[lucid @lucidpaths33]

@IterIntellectus thenonly map that matters

[lucid @lucidpaths33]

@zekramu @tunahorse21 i can literally use hermes for 30hours and PR anything i ran into that broke and see it merged to main within a week or two

where as the openclaw of my company is by no so far beyond homebrewed because nobody gives a shit about community prs over there

[lucid @lucidpaths33]

@forloopcodes bet

[Aryan Sawhney @aryansawhney17]

Everytime, I think AI models are expensive, I look back at this screenshot and thank capitalism

[lucid @lucidpaths33]

skill im currently often using:

-----
from your point of view understanding what he goal must entail, can you write a full feature update list, a roadmap. then for each feature you fan out subagents, each agent gets a pedantically specific and precisely scoped PRD of what it needs to do. a clear goal for each sub agent, when it is done it opens one PR with proper documentation and reports back to you here - you keep the pipes open for agents, you judge critic and give feedback if the pr is faulty or isnt as desired and send that feedback back to rhat specific agents work so it can finalize the PR it initially thought was done. you do this loop until the PR matched expectations of you pre-designed roadmap. the fullfil criteria are in your judgment, the work is done by a subagent of your choice (depending on complexity of that single work)

i want the design to stay slim if possible but the main goal is to get every done as proper as possible
-----

its some variant of this, more or less what i use for hermes and claude but im pretty sure by now they have improve it into a specific version for themselves (havnt checked tbh)

but key is to have one place for documentation that your orchestrator keeps track off - and each task is done by a worker that reports back

[lucid @lucidpaths33]

one of the things i learned the hardway in the recent days is that if you have automated software production and dont gate a rule around when to test or when not tl test, my system will overengineer test even around prose

bad design

Changed that today

[lucid @lucidpaths33]

@CryptoJelleNL Amen Bro 🙏🏻❤️

[lucid @lucidpaths33]

@guitaripod lets see what the numbers are if i max our every 5h limit then

Technically i have the rest of sunday and all of monday before my weekly limit resets automatically

didnt expect a usage limit yday(??) but i woke up to a reset so here we are kek

[lucid @lucidpaths33]

lets see if I can burn trough the entire claude 20x usage limits in one day for both fable and normal usage