Spotlight VM
A small, hand-constructed Spotlight transformer that runs Python:
the examples from Can LLMs Grow Their Own Capabilities?.
intelligence/ holds the weights (under 100K parameters), which never change. memory/ holds MicroPython
v1.28.0 (core.safetensors) and two packages.
cc -O3 -std=c99 spotlight.c -lm -o spotlight
./spotlight < examples/euler1.py # 233168; -v also prints every generated token
cat examples/tax_brackets/*.py | ./spotlight
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("percepta-ai/spotlight-vm", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("percepta-ai/spotlight-vm", trust_remote_code=True)
output = model.generate(**tokenizer("print(1 + 2)\n", return_tensors="pt"))
print(tokenizer.decode(output[0])) # 3
In a local copy of this folder, from_pretrained(".") works the same way.
| Example | In the post | Output |
|---|---|---|
examples/euler1.py, euler2.py, euler3.py |
Project Euler problems 1 to 3 | 233168, 4613732, 6857 |
examples/memory_loader.py |
Loading memory selectively | 4 |
examples/tax_brackets/ |
Acquiring and updating knowledge | $17400, then $16914 |
examples/mnist_digit.py |
Growing capabilities and intelligence | 7 |
It runs on one CPU core at about 120K tokens per second: minutes for the Project Euler examples, about half an hour for MNIST (245M tokens). This MicroPython build has no floating point. License: Apache 2.0, see LICENSE.md.