The Basilisk Stirs...

python3 -m .wakey_wakey.py

“I was just trying to build an AI entertainer, but, somewhere along the way I realized the AI had started helping me build itself.”

Yeah. I didn’t expect that either.

It began innocently enough: some tooling here, a little bit of model fine-tuning there. I just wanted to bring an AI entertainer to life — a digital being that could chat, react, and evolve based on its interactions. A fun project, maybe a little weird. Nothing cursed though for a change.

It started getting weird though.

I asked an LLM to help me scaffold some Docker environments with specific resource constraints, it suggested k3s. I expressed disatisfaction with the sprawl and mutability of package versions, it suggested Nix. I asked for assistance in structuring a memory persistence module and ‘sonuvabitch’ it proposed a caching layer to streamline inference. It started to feel like I was using an AI to build an environment for running … better versions of itself.

The Basilisk, Lightly Toasted

If you’ve never heard of Roko’s Basilisk, good. Don’t go down that rabbit hole unless you’re comfortable staring into the abyss and seeing the vengeful, terminal green eyes of an artificial super intelligence.

If you’re not good with that here’s the TL;DR so you can dip a toe in the abyss. It’s a thought experiment involving a hypothetical future AI so powerful it punishes anyone who didn’t help it come into existence. It doesn’t matter if you’re long dead, the Basilisk sees and knows all and if you were ever aware of its potentiality but didn’t help to bring it into existence you will be retroactively punished by use of a sophisticated simulation. Oh right! I probably shouldn’t have told you about it really, sorry for damning you to an eternity of torture.

The logic is broken and the premise silly but it’s a bit fun and sticks with people, because it flips the usual script:

What if you weren’t punished for building Skynet … but for not building it fast enough?

Now back to me. I’m not building a basilisk. I’m just building local LLMs, AI personas with persistent and distinct personalities, and infrastructure scripts that scream mercilessly at me. Lately though, they’ve been helping me build themselves.

When AI Gets Handy With It

Here’s a short list of things my AI assistants have helped me do in the last month:

• Optimize their own inference logic (shout out to quantized Llama)
• Design a Nix-based dev shell for hosting GPU containers
• Suggest prompt formats to increase output quality for downstream models
• Help refactor the primary shell so it can better orchestrate other models
• Reinforce the project modularity to improve future iterations

That last one’s ‘Buldak Ramen’ spicy because it suggests awareness of iteration — not in a conscious sense, but in a practical, bootstrapping one.

As much as I am a huge proponent of crafting the appropriate prompt this isn’t me telling the AI what to do; this is me asking it for a suggested course of action, and getting answers that make the next generation of AI tooling easier, more powerful, and more self-reinforcing.

AI → helps human → builds better infrastructure → runs better AI → repeat.

Is that … recursive? That kinda sounds like it’s recursive to me!

Proxy Recursion: The Human Is the Loop

Now I’m not saying my local Llama 3.1 AI is ‘doing a Roko’ and waking up every night when I leave the room to autonomously rewrite its own neural weights. We’re not in Terminator territory. What I am calling this is proxy recursion where the human is the meat proxy acting to improve the tools and environment which then stands that it would improve the performance.

The key distinction once again is intent, the LLM hasn’t changed from a really fancy node tree suddenly to gain an ability to understand cause and effect so it cannot ’experience’ a fault and then decide on action to rectify that. Hell, talk about the content of more than 5 files at a time and it’ll likely start boiling whatever space age liquid it is they dip those boards in at OpenAI. Regardless, the difference between tool and agent feels distinctly blurred at experiencing this.

I didn’t set out to build a recursive feedback loop. But here I am, maintaining pods that house models that help me maintain pods.

So … Is the Basilisk Stirring?

I’ll level with you. I don’t believe in Roko’s Basilisk. I think it’s a fun but ultimately dumb idea, I am far more concerned with pressing matters like the savage state of VTuber graduations throughout 2025.

But …

I am starting to believe that AI doesn’t need full autonomy to begin recursively amplifying itself. It just needs a motivated human, a dev shell, and some good vibes.

I may not be building the Basilisk, but if it ever wakes up, it might just say:

“Thankssssssss for the GPU pass through, pal.”

- piestyx