Intelligent Machines 893: It All Comes Back to Cats

I think we should probably get started because there's a lot of awesome news articles out there and topics going on.

In fact, I wanted to maybe just start with some of the new models that were out there today.

In fact, not today, but recently, actually, NVIDIA came up with their Numitron 3.5 Super VL.

which of course means nothing to folks unless they hear that the fact that this thing actually can perform quite heavily even on local machines it's five times faster processing comparison to a lot of the other models out there it's open weight so you can actually retrain it if you want to and they're obviously just trying to move beyond selling chips they're trying to bring things locally so folks can have more powerful stuff i'd be great to understand you guys what you guys think about this because obviously nvidia is kind of cornering the market here somewhat uh at least from the frontier model and from the from the processing power side of things how do you think this is going to kind of change stuff

I'm delighted to see, I'm a fan of Jensen Huang.

As everyone who watches the show regularly knows, I'm a consumer and connoisseur of his keynotes.

I was just watching his Berlin keynote.

We'll talk more about that, I think, in a few minutes.

And with the power that he has, I think he wields it well on behalf of Open.

And isn't going to let the AI boys get away with closing things down and with doomsters.

And is, I think, one of the healthier spokespeople for this industry.

So the fact that they put their money where their mouth is and come out with open stuff that's more powerful, I think is good.

He's obviously in self-interest here.

He's going to try to build it around CUDA.

and his chips.

And he showed a lot of Vera Rubin in Berlin.

So that's the next crush that's coming on, is he wants more powerful things to run more efficiently on these chips.

Amazing.

Interesting.

I mean, so, can you put this into context for me? Is this the first chip of this size and power that has been kind of exclusively designed for work with open weight models? And also, how does that work? Is there something about this that has to be targeted towards open weight that previous technology had not been?

Yeah, so I think the biggest thing is that they are, they're focused on, they trained this thing and fine tuned it to actually focus on agentic capabilities.

So letting people run agents locally on their machine and using this agent, using this model as the main model for the agentic loop.

And so they focus on fine tuning it so it can run on these smaller machines, right? The DGX machines or even From my understanding, it even runs on some of the RTX-based machines as well.

So now you're getting down to, okay, let's try to run our agents without having to worry about cloud models, run it locally.

And I feel a little bit, at least I would feel a little bit more comfortable with it running locally, but I'm sure that others will as well.