TensorFlow

And One of the first tools that I tried on it was essentially like mapping out like an entity mapping thing.

There's a lot of like TensorFlow has a tool that you can use to sort of check things out a little bit.

But it's very techie, like it's very ugly.

Like UX designers have not been there.

Um, but you need to make sure that your entity is clear and that your entity makes sense.

So have, have a look at kinds of entity associations that you have that are, they're related to your brand.

TensorFlow again, as I said, has, has that.

Um, you can also ask the LLM so you can say like game 10 entities associated with this brand.

I mean, way back when I had a...

I was using PyTorch and TensorFlow.

I don't know if anybody here knows that kind of technology.

Using PyTorch and TensorFlow, I taught my computer how to play Mario Kart.

I trained it on being able to run the Mario Royal Raceway, which is basically an oval track.

But that took...

You know, historically, so Cohere is sort of an interesting company in that one of the co-founders was one of the people who created Attention Is All You Need.

And, you know, there was Faster Transformer as a library that they had that was, you know, a little bit better than just, you know, pip import transformers, right? And then later, TensorFlow TLM, and then later VLLM.

And I've been looking at the page attention implementation as somebody who came from sort of the background of virtualization.

So I've been looking at GPU virtualization and thinking...

They behave slightly differently.

And, you know, from our perspective, the applications there, the attention algorithms themselves, as far as we're concerned, we have like one top or several top level applications.

I, you know, you have SG Lang, TensorFlow, TLLM, LMCache, VLLM.

And from the point of view of what user model is being loaded, that's really what we're optimizing for.

And they started releasing these as public papers.

So go back and look at TFX from Google about, uh, using TensorFlow, a way to string together all the steps of a machine learning pipeline.

You wanna shape your data, you wanna control it, you want to, um, normalize it, you wanna split it into training and inferent or training and test hold back.

You wanna tune your hyper parameters, et cetera, et cetera.

But I still felt there was a self-hosted thing.

I talked to a bunch of customers, and they're like, "Yeah, we actually have some self-hosted problems.

We would love this." And so I ended up finding, uh, Jeremy Louis, who's now at OpenAI, who had created a what's called a CRD, that's a custom resource definition for Kubernetes, that allowed you to spin up a small TensorFlow component.

And I was like, "Oh, what would this look like if we expanded this a little bit?" Where it wasn't just that one component, I think it was just training.

He's probably one of the most famous and world-renowned engineers in the world.

You could probably make a fair argument that Google might not exist, [chuckles] at least in the form that it is today, if Jeff Dean hadn't been there with all the contributions he made to, you know, MapReduce, Bigtable, Spanner, TensorFlow, Google Brain.

Like, he's just this godlike figure.

The fun Jeff Dean facts and, that are like these like Chet Norris the- memes and stuff like that, that I remember it was like, uh, only Jeff Ne- Dean knows the final number of pi and-

i the point is the part that we're always i'm bad at anyway is why how did you make it why was it unsafe before and is safe now and i'm sure that that's the answer that you just need to read this thing but now it's the end of the show and i'm tired yeah i'm not gonna read yeah

i'm not gonna read this whole thing but What I wonder, because there's models out there, not large language models, right? Like TensorFlow models, like image recognition.

3D models.

Hot dog, not hot dog.