Riany

7 ms·

1.
by Riany 4mo ago
Token savings is more and more important, but it also important if the agent trusts the result and stops searching. it should measure the full agent loop instead of just the search output
2.
by Riany 4mo ago
That's good that all pilots ejected safely. But what if it fails? Still, losing two specialized aircraft during an airshow feels like very expensive, I doubt if it's really worth it to risks these pilots life on it
3.
by Riany 4mo ago
the idea is good, but the examples still feel like a distance to handle real constraints and dimensions
4.
by Riany 4mo ago
I think is useful cuz it removes setup friction for beginners, who can just focus on Rust and embedded concepts instead of fixing wiring and board differences
5.
by Riany 4mo ago
The best part is you don't have to stay at home waiting codex thinking, you can go out grab a coffee while connect your Codex and ask it to work
6.
by Riany 4mo ago
The gaming part is fun, so does the local AI numbers. As fast prefill changes the whole experience, it makes local inference feel practical
7.
by Riany 4mo ago
I had the same question and if this is a data-only attack, the lesson may be that MIE reduces many attack paths but does not remove every useful corruption primitive
8.
by Riany 4mo ago
I think local models need to be good enough that privacy, latency, and control become worth the tradeoff, instead of beat the best cloud models
9.
by Riany 4mo ago
losing SOS/cloud features is real cost, but so is having an always-on telemetry device in a thing you own. this should be a software setting and a clear privacy contract
10.
by Riany 4mo ago
My guess is deterministic assignment makes load distribution and debugging easier. But for a privacy product, that convenience probably needs to be reconsidered
11.
by Riany 4mo ago
surprising that the mapping may be stable enough to become a user-level signal. and rotating away from deterministic assignment seems like a cheap way to avoid creating an extra fingerprint