text
stringlengths
0
2.46k
[3553.72 --> 3554.70] You're putting bias into them.
[3554.72 --> 3555.24] Yes, absolutely.
[3555.30 --> 3556.82] There's absolutely 100% bias.
[3556.94 --> 3557.82] Let's not pretend there isn't.
[3557.90 --> 3558.02] Okay?
[3558.08 --> 3558.30] Okay.
[3558.30 --> 3562.90] So, it also requires you to actually have coffee beans roasted in a particular way.
[3563.24 --> 3567.24] It also requires you to have the espresso water boiled at a particular temperature.
[3567.76 --> 3568.02] Okay.
[3568.02 --> 3570.50] So, you put all of those details down.
[3571.28 --> 3572.12] That's the idea.
[3572.30 --> 3575.16] Like, so, in other words, it's not just like, okay, hi, how are you doing?
[3575.16 --> 3576.08] Like, what's great espresso?
[3576.68 --> 3578.70] You buy from Espresso Vivace in Seattle.
[3578.84 --> 3582.66] I mean, while that's true, and I'm basically, I don't own any stock in them, by the way, but
[3582.66 --> 3584.20] they are easily the best coffee.
[3584.30 --> 3585.12] Who's the brand against you?
[3585.36 --> 3586.80] Espresso Vivace in Seattle.
[3586.92 --> 3587.30] Espresso Vivace.
[3587.30 --> 3590.82] Yeah, David Shomer is a magician when it comes to espresso.
[3591.12 --> 3591.32] Okay.
[3592.22 --> 3595.48] But the context is like, well, as much as I want to just provide an answer like that,
[3595.62 --> 3596.38] the reality is no.
[3596.62 --> 3597.78] Obviously, we can't train that bad.
[3597.88 --> 3603.20] We actually need verbosity to provide context, provide proof, if you want to put it that way.
[3603.78 --> 3607.34] Because there's going to be other people putting other answers, too.
[3607.86 --> 3608.08] Oh.
[3608.30 --> 3611.60] So, for example, in this case, I'm just going to call a buddy of mine, Rob Reed.
[3611.60 --> 3612.66] He's a fellow cyclist.
[3612.82 --> 3614.30] He's also a fellow coffee addict.
[3614.30 --> 3618.02] I know he also put some coffee answers inside there as well.
[3618.42 --> 3623.20] So, between everybody that put coffee answers in there, that's actually literally, you're
[3623.20 --> 3627.26] getting data from myself, from Rob, and a few other folks from, well, Databricks.
[3627.60 --> 3627.84] Right.
[3628.06 --> 3631.66] And how many instructions are in there that you guys put in?
[3631.72 --> 3632.56] The 5,000 employees?
[3633.00 --> 3634.70] Oh, 5,000 employees put 15,000.
[3634.82 --> 3635.56] 15,000.
[3635.86 --> 3636.76] So, it's remarkable.
[3636.88 --> 3638.94] If you think about it, that's remarkably small.
[3639.30 --> 3639.46] Yeah.
[3639.46 --> 3643.30] We were always under the impression when we started this process that we would require hundreds
[3643.30 --> 3645.88] of thousands or millions of answers.
[3645.88 --> 3646.70] I was going to say, how does it know?
[3646.76 --> 3647.96] You gave it coffee instructions.
[3648.16 --> 3648.30] Yeah, yeah, yeah.
[3648.30 --> 3648.42] Yeah.
[3648.56 --> 3648.70] No.
[3648.86 --> 3650.04] How does it know something totally different?
[3650.06 --> 3652.38] Like I said, Dolly1.0 shocked us.
[3652.62 --> 3656.34] It really shocked us because we thought we would need to put in a lot more data.
[3656.58 --> 3658.44] We thought we would need to do a lot more training.
[3658.72 --> 3662.00] And then we were like, wow, this is not bad.
[3662.08 --> 3664.62] I mean, it's not perfect, but it's not bad, actually.
[3664.62 --> 3669.06] And so, from a business perspective, what ends up happening is if you have your own business,
[3669.60 --> 3673.20] now your data, you don't need a million things.
[3673.32 --> 3675.42] You've got 15,000 pieces of information.
[3676.04 --> 3678.44] Now, the great thing, and I'm not telling you to use Dolly, by the way.
[3678.50 --> 3680.02] I mean, obviously, go use it if you want to.
[3680.08 --> 3682.90] But I'm saying, use any open source model.
[3683.18 --> 3684.22] I don't care which one.
[3684.64 --> 3688.60] That way, you get to go ahead and keep it and have your data as your IP.
[3689.12 --> 3693.84] So, you as a business end up using the data actually in a good way.
[3693.86 --> 3694.14] Right.
[3694.28 --> 3699.54] Where you actually make it advantageous for you, yet also keeping the privacy for the users that make up that data.
[3699.74 --> 3700.44] At the exact same time.
[3700.50 --> 3704.12] So, the move is you have these, I don't know if this is technically what a foundational model is,
[3704.18 --> 3707.20] or you have these models that are large enough language models.
[3707.36 --> 3707.66] Right.
[3707.80 --> 3708.08] Right?
[3708.24 --> 3713.84] And then each company or each org or each use case says, okay, now we're going to fine-tune it.
[3714.06 --> 3714.32] Right.
[3714.32 --> 3715.40] I don't know if that's the right language or not.
[3715.54 --> 3715.86] It is.
[3715.86 --> 3718.30] And apply it to us.
[3718.56 --> 3718.76] Right.
[3718.92 --> 3720.34] And there are going to be all sorts of, exactly.
[3720.56 --> 3721.98] There's all sorts of models out there.
[3722.12 --> 3728.50] There are already, like, a lot of people were asking me originally, like, hey, okay, well, then, you need to use Dolly.
[3728.58 --> 3729.40] I'm like, no, no, no, no.
[3729.94 --> 3732.80] Dolly was just us proving that it can be done.
[3733.26 --> 3734.18] That's all it was.
[3734.58 --> 3741.42] So, there are a lot of really good companies, whether it's Hugging Face or anybody else, that produces solid, open source, large language models.
[3741.50 --> 3741.62] Yeah.
[3741.94 --> 3742.86] Use those, too.
[3742.86 --> 3749.86] Because the whole point is that you can use it yourself, run it with smaller amounts of data, have really good answers, and you're paying $100.
[3750.62 --> 3751.96] At least, in our case, we did.
[3752.14 --> 3753.06] $100 to train it.
[3753.22 --> 3753.38] Right.
[3753.62 --> 3756.16] So, we're like, okay, that's actually worth your business.
[3756.38 --> 3758.00] You're protecting the privacy of your users.
[3758.46 --> 3761.40] You're going ahead and actually having relatively solid answers.
[3761.40 --> 3765.44] And you're not basically giving your data away to another service.
[3765.60 --> 3767.70] Because that's the key thing about when you use a service.
[3768.04 --> 3768.24] Right.
[3768.40 --> 3772.74] That you're basically giving away your data so they can go train against the two.
[3772.94 --> 3773.20] Right.
[3773.32 --> 3773.50] Right?
[3773.58 --> 3777.90] Now, I know Microsoft and OpenAI, for example, you're calling those two out in a positive way, not a negative.
[3778.26 --> 3780.68] Usually, I'm a former Microsoft employee, so I'm allowed to be negative if I want to.