| 20:04 | <zbraniecki> | is this the first TC39 meeting since I joined where no ECMA-402 proposals are brought up? YAY! :D |
| 20:05 | <sffc> | There's the time zone canonicalization proposal |
| 20:06 | <zbraniecki> | ah |
| 20:17 | <justingrant> | Yep! If any TG2 folks have feedback on that time zone canonicalization proposal, it'd be great to hear it here today so we can make changes or address concerns before we present to the committee tomorrow. Slides: https://docs.google.com/presentation/d/13vW8JxkbzyzGubT5ZkqUIxtpOQGNSUlguVwgcrbitog |
| 21:06 | <Richard Gibson> | sffc: what i18n data are you working with that can use binary16 values? |
| 21:07 | <sffc> | Segmentation data |
| 21:15 | <Richard Gibson> | why is that floating point at all? |
| 21:17 | <sffc> | for the LSTM model. Currently we're running f32 arithmetic but we can store in f16 with no loss in accuracy. I haven't put time into whether we would want to do arithmetic in f16 or bf16 space directly. But if your weights are going to be converted back to f32, f16 makes more sense than bf16. |
| 21:19 | <sffc> | Another place where floating point comes up in i18n is unit conversions. We could use decimals there instead, but f64 has been sufficient for formatting |