3 ms·
Reminds me of when I first encountered the extraordinarily compressed 12 bit "words" in a ARINC717 data stream, where they're conserving bits that are written a
by MilStdJunkie 3y ago
Reminds me of when I first encountered the extraordinarily compressed 12 bit "words" in a ARINC717 data stream, where they're conserving bits that are written as-is, serially. Like three decades of computing passed it by. If something has four distinct values, it gets two bits. Then, right on top of that: is it a syncro response curve? Output the binary for the curve function. On and on, for thousands of "words", all jammed together. I used to imagine being a future researcher, trying to decode these big blocks of undifferentiated binary, the amount of legwork I'd have to accomplish to get just a few layers deep.
- userbinator 3y agoASN.1 PER - used widely in telco, notably GSM standards, is also extremely dense. Given what they charged for data, it's not surprising they'd want an encoding that's as efficient as possible for the overhead.
- Sai_ 3y agoWonder if this can be used to communicate with LLMs to reduce token costs.
- sitkack 3y agoThere was some chatter about 9 months ago I think when folks were using compression, unicode and all sorts of weird hacks to increase the context window. This is rather recent, but in the same vein. https://www.microsoft.com/en-us/research/blog/llmlingua-innovating-llm-efficiency-with-prompt-compression/ https://www.microsoft.com/en-us/research/blog/llmlingua-inno... https://twitter.com/VictorTaelin/status/1642664054912155648 https://twitter.com/VictorTaelin/status/1642664054912155648