OpenAI and Nvidia are trying to need each other less

OpenAI and Nvidia are being less codependent.

OpenAI has built its first custom chip. Whilst Nvidia is reportedly spending $12.9 billion to buy Hugging Face, the main shop window for open AI models.

Preparation for a breakup? Or at least an amicable separation.

That is the fight underneath this week's AI news. OpenAI wants Nvidia out of the middle. Nvidia wants a bigger stake in the models people run without OpenAI.

WATCH THE FULL SHOW →

OpenAI is coming for Nvidia

Jalapeño is OpenAI's first serious move into Nvidia's territory.

OpenAI says the inference chip produces 1.5 to 1.9 times more AI work per watt than Nvidia's GB200 and GB300 systems. The chip exists. It works. And on OpenAI's numbers, it beats Nvidia's current kit.

Should we trust the test though? Ehhh. OpenAI ran the tests itself (bias) and left Nvidia's next Vera Rubin platform out of the comparison (handy that). Independent results still need to land before we know how it stacks up.

BUT the direction is important. OpenAI now has hardware, models, ChatGPT/Codex products and a billion users.

They are putting the full stack in place. This is one of Google’s main advantages - Google has owned that lot for years. OpenAI wants the same grip on its stack, including the ability to set its own costs and move without waiting for Nvidia.

Every Jalapeño deployment is future money Nvidia does not get.

(Update: Anthropic are looking to buy a chip design company - suggesting they too are moving towards non reliance on Nvidia)

Watch from 00:38 →

Nvidia wants open AI

And what about the other side?

Nvidia's answer may be to buy Hugging Face for $12.9 billion. The reported deal is still unconfirmed by both companies.

Whats the big deal?

Hugging Face is where millions of people find and download open models.

Crudely: it’s the app store of AIs.

Those models need hardware to run. Nvidia sells the hardware. The more people and companies running local open weight models the more GPUs Nvidia can sell.

If Nvidia owns Hugging Face, it owns the shop window for open AI. That is uncomfortable for those who want neutrality. It is also commercially a very smart play.

OpenAI, Google, Amazon and Anthropic are building their own chips. These are Nvidia’s main customers - which is a risk. Nvidia needs open models to keep spreading because more models running locally means more Nvidia GPUs in homes and businesses.

I've been banging on about local AI because I think this is where we end up: private, offline and running on a machine you own. My local AI guide will get you started without anything technical.

Watch from 04:28 →

ChatGPT Work can get on with it

Three quick updates from ChatGPT last week:

You can finally connect more than one Gmail account and more than one Google Calendar to ChatGPT. This sounds tiny. I have wanted it for years!!

ChatGPT Work can sign into websites through your password manager, then continue after you approve the login. Its scheduled tasks can now react to changes in Gmail, Slack and GitHub.

A client email, Slack request or GitHub change can now start the next job by itself. Basically scheduled tasks triggered by events instead of timers - allowing you to set up much more powerful workflows.

 My ChatGPT Work guide shows you how to set ip ChatGPT work if you haven’t already.

Watch from 10:24 →

I am going to buy this robot duck

Microduck is probably the most important story this week. Well not really but it’s cool!

It is a 25cm open-source robot with 15 motors, a camera, depth sensor and a beak that can pick things up. It walks, sits, recovers from a fall and roller-skates.

$399 buys you a small robotics lab with roller-skates. You can train behaviours in a physics simulator, then move them onto the physical duck.

I want to learn more about robotics and physical AI, and this duck looks like my way in. Reply and tell me if you want me to share the process. Your replies will also help me justify buying it!!

Watch from 14:18 →

To the Task,

Kyle