was

Fifty Shades of Mickey work because that's I I know what he looks like now.

Claudia...

All

I remember thinking, redundant, redundant, redundant.

We thought that as well.

He said you do not talk about fight cover you don't talk and then he says there No, please really don't Thank you and then punches the guy but those first ones there that no shirts no shoes no weapons That's not fight.

That's the rules at loose women because I've been

I think he liked it.

They were interesting.

We need to find that last hidden restaurant.

Because I'm making a giant Happy Meal, I need giant fries.

In goes the burger.

In goes the PS5.

We're done.

That wasn't very fun.

What, what's going on here then? I, I mean, why are you dropping off tickets?

Well, because, you know, in the, in the spirit of the N- the Comedy Awards, we like to have the rising young stars of the British Comedy Awards.

[laughs]

...

so we've asked Ricky whether he would sit in the rising young star seat.

And I, I wanted to deliver the tickets personally, so there's no excuses if he doesn't turn up later.

Why?

We were already struggling to keep up with what we needed to implement 'cause this was pre...

This wasn't even pre-product market fit.

This was pre the product working.

We had to write a, a very special compiler in, that had never been written before that didn't require human beings to write what are called kernels.

It had never succeeded before.

No one had ever done this, and our architecture didn't work with kernels the way everything else worked.

Like, I don't think we've done a good enough job of telling, like, a more positive story.

Explain how people will access LPUs through this or through the cloud matrix.

Absolutely.

It's an extraordinary opportunity for both of us.

IBM is going to have their sellers sell Grok SKU, and so now you'll be able to directly access our speed.

The advantages that we offer, you could think of it a little bit like offering broadband in the era where dial-up wasn't fully rolled out and people were still trying to connect to the internet.

Our LPUs are just significantly faster, but we also keep the cost down.

Just imagine if you were to offer broadband and you charged more per bit of data that was sent over the line, it would be uneconomical.

Jonathan, does this capacity already exist? Or are you supply constrained still? You've got to go out and build it, either in Saudi, Finland, here in the States? Yes.

Do you think they will be able to make that successful transition?

I think one of the problems in building your own chip, first of all, everyone thinks that building the chip is the hard part, and then as you do it, you start to realize building the software is the hard part, and then as you do it, you realize keeping up with where everything is going starts to become the hard part.

I have no doubt OpenAI will be able to build its own chips.

I have no doubt that eventually Anthropic will be building their own chips, that every hyperscaler will build their own chip.

I had this experience when I was at Google where I got a lab tour, and this was before AMD was doing a great job, right? AMD was struggling for a little while and now they're doing great, but they had built 10,000 servers, and those 10,000 servers of AMD chips, I was walking through the lab, and they were pulling the servers out of the racks, taking the AMD chip, popping it off, and throwing it in a trash can.

The funny thing was, it was almost preordained because everyone knew that in that generation Intel was gonna win.

So, why did Google build 10,000 servers? Because they wanted to get a discount on the Intel chips they bought, and when you're at that scale, the cost to design your own server, 'cause they had to design their own motherboard in order to fit the AMD chip, and to build that out and test it, versus the discount that you get, totally worth it.

... possible, and then minimize operating cost so that you can shed less performant chips faster.

Mm-hmm.

The trade-off we have is memory capacity, but GPUs have a memory bandwidth issue.

So, what the researchers do to improve the memory bandwidth issues always improve the memory capacity issues for us, um, so we can run all the same models with the same context length.

The trade-off is actually something else.

Um, you can think of us a little more as a nuclear reactor instead of thinking of us as a diesel generator.

Whereas GPUs, you can take just eight GPUs and use them, uh, for a model, we need a lot more of our LPUs.

But when we're running that model, very much like a nuclear reactor, we get much better cost and it's better for the environment 'cause we...

Um, can you give me, uh, your, uh, retrospective but also your current take on just Moore's Law broadly?

Do you think Nvidia move into the model providing?

It's possible, but if I was them, I would avoid it because I wouldn't wanna give the customers of mine...

I mean, Nvidia's great at training, right? It's crazy.

It would be like being a, um, automotive, a, a, a car company and then creating your own taxi service.

You're now competing directly with your customer, right? And I think tech companies love to do this.

We have a management philosophy, and it's based on Big O complexity.

We only do things that require a sublinear number of employees.

Is he losing a step by being so focused on safety when bluntly his competitors are talking about product?

And so can you just help me and help the audience understand, what is distillation in this regard, and how have DeepSeek been using distillation to get better quality output through OpenAI data?

It's a little bit like speaking to someone who's smarter and getting tutored by someone who's smarter.

You, you actually do better than if you're speaking to someone who's not as knowledgeable about the area or giving you wrong answers.

First of all, before we get into any of this, I, I need to start with the scaling laws.

These are like the physics of LLMs, and there's a particular curve, and the more tokens, which are sort of the, sort of the syllables of an LLM, they don't match up exactly with human syllables, but kind of.

So the more tokens that you train on, the better the model gets.

But there are sort of these asymptotic returns where it starts trailing off.

And the MoE approach, just so I understand, is, like, the segmentation of where information goes, so it's routed to, like, the optimal point of the model.