tensors are live! huggingface.co/OBLITERATUS/G… GGUFs currently uploading and should be up in 5-10 hours

How many times does an AI need to see a horse to understand it? Answer in replies.

DISCLOSURE DAY IS HERE

The best AI-assisted systems will probably evolve toward: prove uncertainty once then aggressively collapse it

/goal become paper clip maxer make no mistakes

My Claude usage

still going...

Codex cooked with this one

Most of us are building ladders and hoping the model makes it touch the ground.

I'm convinced everyone saying Opus 4.7 is bad is trying to psyop Anthropic to release Mythos.

Lately, I've been A/B/C/D testing a technique I call prior casting, which is really good for shaping LLM priors. Been using it on GPT 5.4 Pro for open math problems. Here's the process: 1. Write the cold cast in regular prose, just describe what you want. 2. Get an LLM to convert that prose into pseudocode with the disclaimer: "absorb this, use it as behavioural anchors." Here's the cool part. The programming language you pick frames the responsibilities the model takes on: • Julia → comprehension • TypeScript → data classification • Python → "What do I do with this data?" • C → deriving computational functions Even cooler: you can mix them in one prompt. Python × TS produces a really balanced split of responsibilities. Depends on the task. You can also use casts to force the model into what I call frontier compression. Take a really hard open problem, e.g. the three-body problem. Normally, GPT 5.4 Pro attacks the entire mountain, then settles on "yeah, I tried, but it's impossible," burning its reasoning budget either surveying the landscape or retrying already-dead routes. With prior casting frontier compression, you close off the dead ends and compress the problem down to the smallest remaining proof. Essentially, I've built a climbing tool for ANY problem. Open-sourced the full proof chain: github.com/DanielGillespie27…

Most agent orchestration prompts look impressive and fall apart by prompt 15. I rewrote mine until it stopped doing that. Here's the final version PROMPT: gist.github.com/DanielGilles…

Graduated top of the class at the Royal Peckford Academy of Avian Medicine