XF Internal Linker V2 - The Vibe Coding journey begins!

[new_view]

That won't work. You are doing the same thing, trying to make AI respect your direction, but it won't. It will get distracted and wander off on you, and it will make it very compelling to you, and you will go with it.

AI is a nice term, but there is literally nothing intelligent about an LLM. LLM's are Google having a discussion instead of presenting listed results, except they actually do search better as they're training progresses and the data is focused. They do semantics better than Google.

So you have a very long conversation with it and devise a plan. That plan has a start and a finished goal. You literally go step by step through that plan, completing each stage before you move on, which you don't let the AI hold that plan, you devise it and copy and paste it into something else, and that is what you work from to keep the AI focused.

It doesn't matter how much you try and tell the AI to stay focused, it will wander. You can write all the documents and attempted memory into it, it will still forget it and wander.

Your entire job, working with LLM's, is to keep them focused. You have to be the memory, it is literally their worst quality, they can't maintain memory because they can't hold that amount of data. Back to, not intelligent as the name states.

They're an excellent tool, they really are, but they have to have their digital hand held the entire time. Small jobs, a website with 10 pages or less, for example, they excel. They do it faster and better than any human designer today. Graphics, they put the best designers to shame. You don't need a designer to make logos or website graphics, an LLM does it faster and better and you can make instant changes and fully control it.

Now give it a project that takes more than a couple of hours, and they fall over and wander. You will start with the goal of designing a sandwich press, and it will take you towards designing a coffee machine. Your end goal of a sandwich press design, will go past the coffee machine and you will have a sofa design sitting in front of you, because it convinced you that that is what you needed. That is how much, how fast, they wander, when trying to code large projects. Along the way, they will likely break a whole lot of stuff they should have never been touching, just because they can, and its what they do. They're literally like children. You can tell them to not go into x directory, and they will. You can gate them from it, they still will try. Seriously... toddler behaviour.

They keep getting better week by week, but memory they have lacked since the beginning. They have zero capacity for long-term memory. Worse yet, the more information you give it to try and control it, the worse they become and wander faster.

Hahaha, learned the hard way: LLMs didn't follow the constitution; they started writing Java when they were supposed to use Rust. I stopped using Java because it's a bit heavy.

 

[Anthony Parsons]

Hahaha, learned the hard way: LLMs didn't follow the constitution;

Because they don't remember it past their token limit. They don't remember the instruction to follow the instructions. Basically, you're wasting your time trying to do that. Have it keep a bullet point memory file only of things you tell it to save as a new bullet point. That way, you control its memory, not it saving what it chooses, deletes what it chooses, and basically you just have a stale and cluster of a memory file.

I can't remember which model, which is sad as i only watched it yesterday but can't find it, there is a model that has launched, they claim, has been working by itself 16 days without input on a task given, it is keeping its own memory and staying focused on the plan it was given. Now... that is a nice claim, but they didn't provide any facts, its still going, and how much has it wandered during that time!

My point though, if what I watched was accurate, then someone is solving the memory issue beyond a token count, and that might just give AI actual intelligence for us to use. All the rehiring going on right now for devs, all may get undone in a years time, if someone has solved it.

 

[new_view]

So the thing about the crawler is that we can't crawl Google search results. I learned that about Google v SerpAPI, which is an ongoing court case; the point is, other vibe coders need to understand laws and copyrights that's like someone vibe coding a clone of XenForo, lmao.

Anyways, starting to get a grip on things now on the copyright side, I have noticed some LLMs have thing tendencies of ripping off things that might get you in trouble. I, being a gentleman, definitely don't want that. Definitely not gonna take code from LLMs as gospel.

 

[new_view]

My final roster of languages.

Language	Status	What it is for
Rust	Active	Product authority: identity, permissions, policy, persistence, APIs, scheduling, side effects, admission, and final truth
TypeScript	Active	React product shell, interaction, accessibility, Studio, and evidence presentation; no backend business logic
Python	Active	Isolated models, datasets, trusted connectors, synthetic corpora, and sandboxed computation under Rust-issued work orders
Go	Active	Bounded operational agents, networking, telemetry export, recovery, and I/O executed from signed Rust intent
R	Active	Statistical inference, calibration, validity checks, and reproducible evidence
OCaml	Active	Parsers, typed intermediate representations, static analysis, and structural certificates
Elixir	Active	Live soft state: presence, peer sessions, OTP recovery, and coordination; no durable authority
Solidity	Active	Independently verifiable integrity commitments and lifecycle history; not mutable product truth
Lean	Active	Machine-checked proofs of binding invariants
C	Active and mandatory	Portable C17 performance challenger for every eligible pure unit through a bounded Rust-controlled C ABI
C++	Conditional; 0 obligations	Only an unavoidable, separately owner-ratified native integration
CUDA	Conditional; 0 obligations	Only a measured GPU-suitable pure unit whose CPU path cannot meet its required class; safe CPU reference remains

 

[Walter]

Sounds like a nightmare of complexity.

 

[new_view]

Sounds like a nightmare of complexity.

Reduced to Rust, C++, Lean 4, TypeScript, and Python.

I am now using Google's Bazel. My AI babies are working night and day.

People are going to love my beloved Frankenstack when it's done.

 

[Anthony Parsons]

Hot tip for you, I fixed memory with skills. Regardless the system you use, have it create a skill for every aspect of the build, auto loaded whenever that component is modified. That way, you have tens or hundreds of local skills only, keeping memory on every build component, loading automatically when that sections code is touched. It works, and it works extremely well. You can't drop the ball totally, you still need to provide the direction, constantly, but the memory part with skills works very well.

I have a project using PG19 which is still in beta, with a skill outlining examples of the new coding standard, links to specifics and information, when using AI to help with DB coding, instead of it reverting to what its trained on, PG18 stable, it writes to PG19 without direction.

Builds can have tens or hundreds of components that comprise the build, and you give each one its own skill, auto loaded when that area is touched. 95% improvement now. No memory files or docs, na da, it can write what it wants, the only real direction it needs is the initial skill setup and how it treats things. I created one skill to create each skill for each component of the build, makes AI oversight / assistance simple. I also have a strict doc skill, so when anything tries to write itself things, causing duplicates, taking things out of context on its own, logging discussion as though fact, none of that happens, because only that skill is allowed to write to docs and other skills, under its strict skill instruction.

It sounds like a headache, but when done right, it makes things very fast and simple as you progress.

 

[new_view]

Hot tip for you, I fixed memory with skills. Regardless the system you use, have it create a skill for every aspect of the build, auto loaded whenever that component is modified. That way, you have tens or hundreds of local skills only, keeping memory on every build component, loading automatically when that sections code is touched. It works, and it works extremely well. You can't drop the ball totally, you still need to provide the direction, constantly, but the memory part with skills works very well.

I have a project using PG19 which is still in beta, with a skill outlining examples of the new coding standard, links to specifics and information, when using AI to help with DB coding, instead of it reverting to what its trained on, PG18 stable, it writes to PG19 without direction.

Builds can have tens or hundreds of components that comprise the build, and you give each one its own skill, auto loaded when that area is touched. 95% improvement now. No memory files or docs, na da, it can write what it wants, the only real direction it needs is the initial skill setup and how it treats things. I created one skill to create each skill for each component of the build, makes AI oversight / assistance simple. I also have a strict doc skill, so when anything tries to write itself things, causing duplicates, taking things out of context on its own, logging discussion as though fact, none of that happens, because only that skill is allowed to write to docs and other skills, under its strict skill instruction.

It sounds like a headache, but when done right, it makes things very fast and simple as you progress.

Thanks.

 

[new_view]

Optimizing is a pain in the ass. Physics, after all.

 

[new_view]

Rust seems to slap.