Distro, every mention

The Quest for Community-Trained Open Source AI Models Oct 1, 2024 · 47 mentions

• ▶ 9:30 Anjney Midha What kind of projects have you guys released so far leading up to distro? 2 times in the scene
• ▶ 16:14 Anjney Midha Um, between Hermes, you know, Yarn, and now Distro, what is usually, what is your criteria for deciding what to work on?
• ▶ 19:03 Anjney Midha And then why, you know, what is the core idea behind Distro? 4 times in the scene
• ▶ 28:32 unnamed speaker I think for distro, we hit, not a wall, we hit like a bulletproof glass.
• ▶ 31:45 unnamed speaker So it's really hard to, um, guess whether the distro method would degrade any of the future potential benchmarks. 2 times in the scene
• ▶ 34:21 unnamed speaker That might be just by chance that this show is better, right? 4 times in the scene
• ▶ 36:15 unnamed speaker I really think the important part of our research right now is to prove that distro can, like, reduce the amount of communication, and not really about the loss differential, because the loss differential of distro right now is, like, it's… 8 times in the scene
• ▶ 39:35 Anjney Midha Um, but ironically, I think for, for the, for the kind of effort that distro is, having, being as public as you guys have been with the open source, uh, release, the paper publishing the, uh, the first few application results. 4 times in the scene
• ▶ 41:03 Anjney Midha If distro works, right? 5 times in the scene
• ▶ 44:20 Anjney Midha The current distro, uh, experiment still uses H 100. 4 times in the scene
• ▶ 48:30 unnamed speaker So we will release the paper and then the source code of distro. 4 times in the scene
• ▶ 53:33 Anjney Midha One of the things we talked about when, when you guys put out distro was kind of the impact this is going to have on, if it works, right, is on, is on regulation. 3 times in the scene
• ▶ 57:01 Jeff Schmidt So we're going to like, it may be that distro is like two or three times slower than like doing it traditionally, but that's not a requirement. 4 times in the scene
• ▶ 1:13:30 Anjney Midha Is it basically that as long as the, you know, the transformer mechanism turns out to be the most efficient, you know, workload, uh, then, you know, the future essentially with, with, if distro continues down its path is, is, is that most…