Why Decentralized AI Compute Needs Two Assets, Not One
Bittensor pays eight dollars in token emissions for every dollar of real AI revenue. The cause is mechanism design: one token cannot do both jobs the network needs done.
Rob··11 min read
Field notes for deploy-safety teams.
Practical writing on deploy risk, runbooks, incident response, dependency changes, and the habits that keep small engineering teams out of avoidable fire drills.
Written for
Lean engineering teams
Focus
Deploys, incidents, runbooks
Bias
Useful over theoretical
Brute-Force Retrieval Holds Through 5,000 Memories. Then It Doesn't.
I measured where the linear cosine scan in a local AI memory store stops being free. Through 5,000 entries it's negligible. By 10,000 the tail widens visibly. The threshold is sharper than I expected.
Rob··10 min read
Cross-Machine Memory Query: About 20 Milliseconds, Most Days
I built a peer-memory endpoint so my Mac can query my Linux box's memory store over a WireGuard mesh. The first measurement felt suspiciously clean. I measured three times before publishing.
Rob··12 min read
An AMD GPU Beat My Mac on Llama 8B. The Same GPU Lost on Phi-3.
Same hardware. Same benchmark. Opposite winner depending on model size. Cross-platform inference numbers on Mac M2 Pro Metal, Linux RTX 2080 Ti CUDA, and Windows RX 6600 XT Vulkan, plus the hardware tier ceiling none of them clear.
Rob··6 min read
Your GPU Probably Isn't Helping Your Retrieval System
I benchmarked a small embedding model across five hardware backends. DirectML was break-even with CPU. CUDA only won by 20 percent. The reason is the part that generalizes.
Rob··5 min read