Datasets:
willtheorangeguy
/
2023-Changelog-Interviews-Transcripts

Dataset card Data Studio Files
xet
 Community
text
stringlengths
0
2.46k
[4559.80 --> 4560.68] What's the model?
[4560.68 --> 4562.06] It's called Pythia.
[4562.52 --> 4562.88] Pythia.
[4563.20 --> 4568.62] It's a suite of language models, actually, that we put out a couple of months ago.
[4568.84 --> 4569.04] Okay.
[4569.34 --> 4573.56] But in general, Eleuther.i has trained several of the largest open source language models
[4573.56 --> 4575.12] in the world in the past three years.
[4575.72 --> 4576.12] Okay.
[4576.68 --> 4577.56] Very nice.
[4578.04 --> 4579.20] So what do you want to tell the world then?
[4579.52 --> 4580.76] What do I want to tell the world?
[4582.10 --> 4583.70] Honestly, didn't think that far in advance.
[4583.90 --> 4584.22] Okay.
[4585.74 --> 4586.38] All right.
[4586.64 --> 4588.00] Well, what should the world know?
[4588.36 --> 4589.38] What should the world know?
[4589.38 --> 4593.38] About what you do in terms of training models that Databricks uses, that's open source, etc.?
[4593.90 --> 4599.46] Honestly, especially like the open source world, should really know that the AI world really
[4599.46 --> 4601.78] needs help from the open source community writ large.
[4602.32 --> 4606.66] That's actually, broadly speaking, why I'm here at the Linux Open Source Summit.
[4607.22 --> 4607.48] Okay.
[4607.48 --> 4614.28] You know, we're struggling with a lot of issues about maintainability, issues about licensing,
[4614.68 --> 4621.48] issues about regulation, issues about building sustainable ecosystems that the open source
[4621.48 --> 4625.52] community writ large has been working on for years, if not decades.
[4625.76 --> 4626.00] Yeah.
[4626.00 --> 4631.08] And a lot of people in the AI world are a little too proud to ask for help from non-AI people,
[4631.96 --> 4635.26] which is definitely a real systemic problem.
[4635.26 --> 4642.80] But there's, I think, a lot of, if people are excited about foundation models, large language models,
[4642.90 --> 4647.42] whatever you want to call them, and want to get involved and don't know, or want to help
[4647.42 --> 4653.88] and don't know that much about AI, there's a ton of open source work that needs to be done
[4653.88 --> 4659.32] that we need help with to build a robust and enduring ecosystem.
[4660.10 --> 4661.56] Where is the money coming from?
[4661.56 --> 4663.06] Where is the money coming from?
[4663.14 --> 4663.82] Great question.
[4664.14 --> 4668.46] So, at Eleuther AI, we recently formed a non-profit.
[4670.18 --> 4678.82] And we have donations from a number of companies, most prominently Google, Stability AI, and Hugging Face.
[4679.10 --> 4679.34] Okay.
[4679.88 --> 4682.70] And CoreWeave are among our biggest sponsors.
[4682.70 --> 4692.16] We have also been applying for grants from mostly the U.S. government to pay for our, I guess, forthcoming research and work.
[4692.62 --> 4699.18] In terms of, like, computing resources, it's actually, like, training these really large language models is not that expensive.
[4700.00 --> 4701.28] Which is, like...
[4701.28 --> 4701.96] Is that a secret?
[4702.78 --> 4706.66] I don't know if it's a secret or what.
[4706.66 --> 4715.90] But, like, I think that the CS world kind of got used to the idea that anything can be done on, like, a personal laptop.
[4716.44 --> 4722.16] And that that's kind of what constitutes a reasonable amount of money to spend on a paper.
[4722.50 --> 4723.78] And, like, that's great.
[4724.12 --> 4726.00] There's a huge accessibility boon for doing that.
[4726.00 --> 4726.18] Yeah.
[4726.18 --> 4728.66] But training these large language models, it is pricey.
[4729.62 --> 4732.76] You know, it's not something that anyone can do on their own.
[4733.32 --> 4735.36] But it's not ruinously expensive.
[4735.72 --> 4741.08] There are thousands of companies around the world that can afford to do this.
[4741.14 --> 4744.02] There are dozens of universities that can afford to do this.
[4744.08 --> 4745.54] And by and large, they just haven't been.
[4746.10 --> 4746.36] Okay.
[4747.36 --> 4749.22] So there's Pythia model that you trained.
[4749.62 --> 4749.82] Yeah.
[4749.88 --> 4750.84] How much did that cost?
[4750.84 --> 4753.32] Uh, so we trained...
[4753.32 --> 4758.34] So it's part of a suite of models that had, like, 28 in it total.
[4758.84 --> 4761.50] But altogether, that was, like, less than $800,000.
[4762.02 --> 4767.12] The largest model one training run would probably be, like, $200,000.
[4767.88 --> 4768.56] Not bad.
[4768.66 --> 4769.10] Which...
[4769.10 --> 4769.92] That's more than a laptop.
[4770.08 --> 4771.06] Which is more than a laptop.
[4771.20 --> 4771.72] But it's less than...
[4771.72 --> 4774.36] It's not, like, a mind-boggling amount of money.
[4774.62 --> 4775.78] It's less than a Super Bowl commercial.
[4776.06 --> 4776.44] It's true.
[4776.76 --> 4776.98] Yeah.
[4776.98 --> 4777.02] Yeah.
[4777.42 --> 4781.98] So right now, the largest open source...
[4782.76 --> 4783.26] Well, okay.
[4783.34 --> 4787.74] The second largest open source English language model in the world is called GTP NeoX.
[4788.00 --> 4788.70] We train that.
[4788.78 --> 4789.24] I train that.
[4789.58 --> 4790.36] My organization.
[4791.04 --> 4795.72] And that cost us about $350,000.
[4796.02 --> 4798.30] Or what if we weren't given the compute for free?
[4798.58 --> 4802.94] But, like, $350,000 for the second largest open source language model in the world.
[4803.02 --> 4804.72] And at the time we released it, it was the largest.
[4804.72 --> 4809.60] Later, someone else trained a bigger model with sponsorship from the Russian government.
[4810.56 --> 4812.92] But it's for...
[4812.92 --> 4816.16] So, GTP3 came out in 2020.
[4816.86 --> 4822.98] And for about two years, almost nobody was training in open sourcing language models.
[4822.98 --> 4829.50] Google was doing it with similar models, but not, like, the same kinds of models that GTP3 is.
[4830.12 --> 4831.02] And we were doing it.
[4831.46 --> 4833.22] It was really not that expensive.
[4833.82 --> 4842.30] We got into it on compute that we got for free through a Google research computing program called the TensorFlow Research Cloud.
[4842.30 --> 4851.72] And, you know, with that, we trained a 6 billion perimeter language model, the one that underpins the first version of DALI that he was talking about.
[4851.98 --> 4859.06] That's been extremely widely used, deployed in a whole bunch of different industry and research contexts, and been hugely successful.
[4859.52 --> 4862.00] And it was literally just like Google gave us for free.
[4862.44 --> 4862.52] Yeah.
[4862.52 --> 4865.18] It ran preemptively on their research...
[4865.18 --> 4870.96] Basically, the idea of TRC is that they have a research cluster that they don't always use all of.
[4871.52 --> 4882.64] And so other researchers, independent researchers, academics, nonprofits, can apply to be able to run preemptible jobs on their research cluster and just use the compute that they're not using at the time.
[4882.64 --> 4887.18] And using that, we trained this model in, like, two and a half months.
[4887.52 --> 4887.92] Wow.
[4888.16 --> 4890.46] And it was a really big deal when it came out.