Not sure about your last point but in serial comms there are start and stop bits and sometimes parity. We generally used 8 data bits with no parity so in effect there are 10 bits per character including the stop and start bits. That pretty much matched up with file transfer speeds achieved using one of the good protocols that used sliding windows to remove latency. To calculate expected speed just divide baud by 10 to covert from bits per second to characters per second then there is a little efficiency loss due to protocol overhead. This is direct without modems once you introduce those the speed could be variable.
OpenAI's model is closed source. IDK if distilling can be done via the API effectively? DeepSeek already has distilled models from other open source models like Qwen which have been done by 3rd party researchers, and I assumed that happened rapidly because they are all open source.
Google took about 6 months to accept my edit to an incorrectly named park near me. I took pictures of the signs as proof of course I could not submit them as evidence but a thought someone else would report it. In OSM I could change it in a few minutes and I am not very familiar with the tool.
I personally know someone with a similar story (small pedestrian path that existed in real life but missing from Google maps). He made the effort to flag it, provide proof etc etc.
What I told him was you guys be nuts to be spending your time to improve Google's product.
Google has the private road of a local landfill labelled as a bike path. They label little paths in cemeteries as bike paths. Its a little embarrassing how much they rely on automatic data vs, you know, a cities own published bike network maps.
These are best case scenarios. It just tells you the upper limit of the technology. The higher the limit the better. In practice though it means that the infrastructure has a lot of catching up.
I think this just speaks to how much of an advancement this is, if true.
Like, this new battery technology is so advanced that there is no infrastructure for it yet.
Look at user and system times in your test. rmz seems much worse by that measure, particularly syscalls its 4x more time for 4x the number of cpus, maybe a link there. evidently it is multi threaded and as you say causes contention.
