I'm tired of the ritual where we assume AI needs a GPU cluster and a prayer to avoid crashing.
I built BESTBRAIN Core to prove that intelligence—real, deterministic, safe intelligence—can live in the "room at the bottom." This is a 12KB kernel designed for robotics where failure isn't an option.
What it is: - 12KB decision kernel (10.7KB Python + 1.8KB JS wrapper) - 100% deterministic (same input → same output, always) - 0 crashes across 10,260+ tests - <1ms latency (edge-deployable) - No GPU, no neural nets, no randomness
What it's NOT: - Not a trajectory planner - Not a perception system - Not a learning algorithm - Those belong in user space, not the kernel
Why I built this:
Everyone's racing toward bigger models. I went the other direction: smaller, simpler, provable.
The robotics industry has this ritual where we throw NumPy (15MB), SciPy (31MB), and TensorFlow at every problem. I asked: "What if we just... didn't?"
Result: A kernel that fits in L1 cache and makes decisions you can actually audit.
Try it now: git clone https://codeberg.org/ishrikantbhosale/bestbrain-core.git cd bestbrain-core && python3 -m http.server 8000 Open http://localhost:8000/demo.html in your browser
Why local? Codeberg doesn't render HTML directly (security). This setup runs the actual Python kernel, not a hosted simulation. You can inspect the code yourself.
Technical receipts: https://codeberg.org/ishrikantbhosale/bestbrain-core/src/bra...
Philosophy: https://codeberg.org/ishrikantbhosale/room-at-the-bottom
zahlman•1mo ago
In what sense is it AI?
And why not write the promotional material by hand?
setmd•1mo ago
Fair question. I’m using “AI” in the older, broader sense: automated intelligence, not specifically neural networks.
This kernel makes autonomous decisions under constraints, evaluates feasibility, rejects invalid states, and guarantees deterministic outcomes. That puts it closer to classical AI (rule-based systems, constraint solvers, control theory) than modern statistical ML.
If someone prefers to call it a deterministic decision kernel or constraint-based controller, I’m fine with that. The label matters less to me than the guarantees.
And why not write the promotional material by hand?
The core ideas, architecture, and code are mine. I used tools to help polish and structure the explanation, the same way people use spell-checkers, linters, or diagram tools.
If the project doesn’t stand on its technical merits, no amount of wording would save it. I’m happy to have it judged on the code and the claims instead.