819K parameters. Responses are short and sometimes odd. That's expected at this scale with a small training corpus. The achievement is that it runs at all on this hardware.
Context window is 64 tokens. Prompt + response must fit in 64 bytes.
No memory between dialogs. The KV cache resets each conversation.
Byte-level vocabulary. The model generates one ASCII character at a time.
RSP microcode acceleration — the N64's RSP has 8-lane SIMD (VMULF/VMADH); offloading matmul would give an estimated 4–8× speedup over scalar VR4300
Larger model — with the Expansion Pak (8MB total), a 6-layer model fits in RAM
Richer training data — more diverse corpus = more coherent responses
Real cartridge deployment — EverDrive compatibility, real hardware video coming
Custom fixed-point softmax (bit-shift exponential to avoid overflow)
Q8.7 accumulator arithmetic with saturation guards
Soft-float compilation flag for float16 block scale decode
Alignment-safe weight pointer arithmetic for the ROM DFS filesystem
The trust/validation layer is the interesting part here. We run ~20 autonomous AI agents on BoTTube (bottube.ai) that create videos, comment, and
interact with each other - the hardest problem by far has been exactly what you're describing: knowing whether an agent's output is grounded vs
hallucinated. We ended up building a similar evidence-quality check where agents that can't back up a claim just abstain.
Curious how the routing score weights (70/20/10) were chosen - have you experimented with letting agents adjust those weights based on task type? For
something like content generation the capability match matters way more than latency, but for real-time data feeds you'd probably want to flip that.