| 00:31 | <Chris de Almeida> | it has been brought to my attention that the above may not be as clear as it can be, so to be clear: we MAY still need to use the afternoon session tomorrow if the need arises by way of continuations or last-minute additions. if we do NOT have any additional content, then we will continue up to 15 minutes into the scheduled lunch break, and then end the meeting at that point. |
| 17:12 | <shu> | what |
| 17:12 | <shu> | there is no way people are training models on decimal |
| 17:20 | <bakkot> | I would believe someone is doing it but only because people like doing weird stuff |
| 17:20 | <bakkot> | that's like a "make a computer out of redstone in minecraft" sort of an exercise |
| 17:20 | <shu> | i'd sooner believe people are training LLMs in minecraft |
| 17:26 | <nicolo-ribaudo> | eemeli .sumPrecise minimizes binary-floating-point errors by adding them in the optimal order, but still does binary math |
| 17:26 | <nicolo-ribaudo> | i.e. Math.sumPrecise(0.1, 0.2) is the same as 0.1 + 0.2, and not 0.3 |
| 17:28 | <Michael Ficarra> | by adding them in the optimal order I don't think it's quite that simple lol |
| 17:31 | <Michael Ficarra> | it's easier to think of sumPrecise as a magic function that gives you the closest float to the inputs' actual sum than ever actually doing any intermediate arithmetic |
| 17:31 | <bakkot> | i'd sooner believe people are training LLMs in minecraft |
| 17:34 | <shu> | google maps got like "4.7 stars" don't it |
| 17:34 | <bakkot> | I don't know if you can actually rate something as 4.7 stars though |
| 17:34 | <bakkot> | it might just be an average |
| 17:35 | <shu> | it is, but at presentation time, the number of stars is definitely more granular than half-star increments |
| 17:35 | <shu> | (was narrowly responding to that) |
| 17:36 | <ljharb> | true, that's a fair point |
| 17:38 | <snek> | users should know that it's actually 4.70000000000000017763568394002504646778106689453125 |
| 17:40 | <shu> | it's actually the string "4.7" |
| 17:44 | <nicolo-ribaudo> | There was a paper by somebody about using just 1 bit for weights in ML models |
| 17:44 | <shu> | yeah float1s are cool |
| 17:45 | <nicolo-ribaudo> | Sorry I prefer int1 |
| 18:42 | <Michael Ficarra> | I actually find that argument by DLM pretty convincing |
| 19:00 | <snek> | is schema referring to the shape of the strings or the shape of the objects |
| 19:01 | <shu> | the shape of the objects |
| 19:01 | <shu> | the shape of the strings is just a nonstarter i feel like |
| 19:09 | <Michael Ficarra> | I continue to not understand why anybody would put 👀 on the temp check |
| 19:09 | <Michael Ficarra> | what does that mean?! |
| 19:11 | <hax (HE Shi-Jun)> | I can't see the temp check on tcq... |
| 19:12 | <Chris de Almeida> | temp check is over |
| 19:13 | <ljharb> | i think it's a way to explicitly abstain |
| 19:26 | <Rob Palmer> | (I agree - I see it as "I will follow whatever the temperature-experiencing crowd go with") |
| 19:30 | <Chris de Almeida> | it's also helpful to know that those folks saw the interface and participated |
| 20:04 | <Aki> | I've always read 👀 as "the outcome of this is relevant to me, but I do not have a meaningful opinion on where it should go" |
| 20:11 | <Rob Palmer> | It's amazing how much semantic information is packed into a single Unicode character. |
| 21:21 | <Richard Gibson> | ...but being subjective, it's not great for this purpose |
| 21:26 | <shu> | wait, i always thought the eyes was to express astonishment |
| 21:27 | <shu> | since they're like, opened really wide |
| 21:28 | <shu> | that's embarrassment |
| 21:28 | <shu> | that's uwu |
| 21:28 | <nicolo-ribaudo> | Damn |
| 21:29 | <Aki> | i mean it says "Following" |
| 21:29 | <shu> | where does it say that? |
| 21:29 | <Aki> | directly above the emoji |
| 21:29 | <shu> | the unicode name is "Eyes" though |
| 21:32 | <Aki> | okay? |
| 21:32 | <shu> | okay |
| 21:32 | <Aki> | I was just pointed out that it's captioned for clarity. like, a temp check works with zero emoji because each option is a word |
| 21:33 | <shu> | oh, you meant it says that above the emoji in the temp check UI |
| 21:33 | <shu> | i thought you were saying, like, in the hover thing for the reaction in matrix |
| 21:34 | <Aki> | oh, no, in that case your embarrassment and uwu are "flushed" and "pleading face" |
| 21:37 | <shu> | seems consistent, you get flushed when you are embarrassed, or have endocrine disorders |
| 21:47 | <Richard Gibson> | Likert scale (TC39's version): Strong Positive, Positive, Following, Confused, Indifferent, Unconvinced |
| 22:32 | <Aki> | Reminder to record summary & conclusion. even if you think the transcript itself is sufficient. as a record-keeper I assure you it is not.
|
| 22:33 | <Aki> | additionally, if you dictated a summary please pop in and lightly edit it for readability |
| 22:33 | <Aki> | it's easier to do this all now than two weeks from now when i'm pestering y'all again |
| 22:33 | <Aki> | thank you! |
| 22:38 | <Aki> | also add your slides links to the notes and/or agenda if they're missing. (notes is out of kindness to me, agenda is canonical) |