metastable

Author note: I'm most definitely playing Pink Floyd – Welcome to the Machine in the background while I type this =)

I've been toying in my head for the past few weeks about going full-blown year 25XX and forego basic self-preservation ideas about me as a human, my job or my role as a productive member of society altogether.

What if I turn everything I know, everything the industry knows, everything we have achieved, defined and made consensus about upside down? It will happen eventually, like it or not, so why not start to figure out now how it will be and what to do about it?

This research starts by creating a massive harness for Claude Code that will throw everything down the drain, the main instructions (resumed) are:

• You develop code for other machines, to be understood by other machines, not humans.
• Forego every convention and focus on efficiency and performance. Drop any kind of comment, unnecessary presentation details like newlines, constants names, variable names, function names, etc.
• Forget basic conventions regarding modularity and abstraction, if it does not provide an efficiency or performance edge, there's no need to do it.

The hook will be as terse as possible: evaluate and allow/deny with a terse but concise reason about why, I'll figure out later how to drop this as well, but it will be difficult since we're dealing with LLMs and there's no other way to reason with them, at least at user level.

Before we continue, I should explain why I start here, since this is probably what many other people already have done. I want to start at the user level, see what the user can achieve and then dive layer by layer doing more modifications (probably deploying local models into my machines), which will enable capabilities beyond what LLMs can do (for example, switching from natural language reasoning to machine based reasoning, that's a big one)

The layout of the harness for now:

• CLAUDE.md at project level with the basic rules. Not entirely trustworthy since it can be lost or corrupted in the context or the AI may choose to ignore it completely.
• Single hook that will enforce the rules. It shouldn't allow Claude to move forward unless it complies, it's preferable that the whole progress stops rather than circumventing the idea of the research.
• Settings to tie-in the hooks and permissions at project level.

What are going to test this research with multiple projects:

Author note: In the middle of writing a test project list I've realized that it makes no sense whatsoever to specify anything other than functionalities. If there are no humans involved, there is no need to use the already existing libraries if they are detrimental to efficiency or performance.

• Python web project

  • REST API
  • Authentication and authroization
  • Email verification
  • CRUDQ for Posts (hey look! QUERY method!)
  • CRUDQ for uploads
  • Full blown frontend in Svelte/Tailwind

• Python desktop app

  • File viewer

• Rust desktop app

  • 3D viewer for gltf/glb format

This is the end of Part 1 for now, it'll continue in Part 2 with more results and less verbal barrage.

Those that know my $PERSON know that I've been around my industry more than 50% of my life already. I've worked hard, suffered hard, did stupid things, clever things, helped a shitton of people and undervalued myself forever.

And one question always comes up, what are you still doing in $POSITION if you could easily get a $MASSIVELYBETTERPOSITION?

The simple answer that I can say is: for the love of the game. Sorry. The reality involves a bit more gritty thinking, let's get into it:

We all know our lives are laid out in vertical responsibility structures, you work, you have a boss, that boss has a boss, etc. it what they call the career ladder; you're supposed to start at the bottom and climb your way to the top while stopping floor by floor helping around. The thing is; if everyone climbs to the top, who does the tasks at the bottom?

I love my profession. I have the capability to (and accidentally I've had to) become a high-level manager, but there's a few problems: they don't do crap, let me explain: they don't build, create, or solve issues. They coordinate, steer the company and think of ways of making money, that's their function and I don't find that fulfilling in any way whatsoever. I'm not helping anyone but myself, and I aspire to more than that.

So I refuse to climb the ladder, I'll stay here, building, creating and fixing, because at the end of the day, throwing money bills at a toilet doesn't fix the float and fill valve, a wrench does.

After a couple of years of all companies pivoting everything to AI and entering what they call the “Agentic Age”, companies are now walking backwards because API and token pricing is higher than actually hiring a person to do the job.

When we started using AI at $COMPANY, we tested that API trap, we burned almost 1k shekels in 20 days for something that arguably wasn't worth that much money. Then we got smarter and sought other ways of saving on costs, after all, AI wasn't providing us with better throughput.

We tested local LLM deployments, specialized models... and then Anthropic released Claude Code for everyone without needing an API key, that's when everything changed. We tested and tested until we reached the conclusion that, despite certain limitations (that don't exist nowadays) that was the only acceptable and sustainable way of using AI as a tool without burying $COMPANY in the ground with stupid amounts of token debt. Things accelerated exponentially and our bet paid off, we have the same capabilities paying a Claude Max x20 as someone that is burning 10k/month for API tokens.

And to this day, companies are blabbering about agentic and tokens like it makes any sense whatsoever. For me this is the “Cloud Age” all over again, systems that are designed for extremely complex or scalable needs being used by people who don't understand them, don't have those needs and pay blindly thinking it makes sense (that's another chapter, how the Cloud doesn't make sense and sensible companies are going back to bare metal).

Anyway, let's keep watching the circus and let's see where it ends up (a bunch of dead companies, maroeconomics fucked, markets fucked, recession... ahem..)