Fundamental's Shortcut builds AI Excel agent using 'Cogen' method — outperforms Microsoft Copilot

Thank you. You too. Thanks. We'll talk to you soon. Cheers. Bye. Up next, we have we have two more guests. We're going long before the 4th of July. We got to get an extra couple minutes in. Who we got? Jordy. Uh we have added a new guest. This was very last second, but hey, great. We're here. Welcome to the stream.

You're live. This is Nico. Nico. Hey, what's up, guys? How are you? Break it down. Introduce shortcut. Nico managed to go viral yesterday on a timeline that was in turmoil. That's insane. You You did You were up against some serious. I haven't I haven't uh I haven't been been able to watch the video yet.

So, you're going to have to bring uh break down the whole thing for us. But you introduced shortcut, the first superhuman Excel agent. And when I read this line, I was like, "Shots fired at co-pilot. Let's go. " Well, the the the when I read this line, I was like, "Oh, yeah.

You should just build agents on top of Excel where people already are. You don't need to reinvent Excel and build agents. You can. But so anyways, but in your words, what do you do? Yeah. What's up guys? I'm Nico from Fundamental.

I'm one of the co-founders here and we just released Shortcut yesterday and and Shortcut is a bench is essentially the first superhuman Excel agent. We actually benchmarked it against Gemini Copilot uh GenSpark. It's really not even close. What we have to do now is scale capacity.

It's about four times better than Gemini even in the limited stuff that we were like allowed to use for Gemini because we just know it couldn't do it. Um, and yeah, we're scaling up capacity now. I'm actually about to drop a million dollar check on scaling up the Infra just just to Okay, there we go. Physical check.

Do you Did you write it out? You can handle it. How much are you? I might do that just for posterity. Yeah. Yeah. Smack the thing. Yeah. You need a You need a million bucks. You need a comically large check like one of those one of the hyperscalers. Yeah. Go give it to AWS Andy Jasse. Here we go. But uh awesome.

So, so uh is this the the second, you know, talk about maybe step back fundamental is this? Yeah, sure. Uh so, Fundamentals research lab, we're in Menlo Park now, but we actually shut down our lab at MIT where my co-founder was a professor about 18 months ago.

And the mission of the company is to build digital humans back when that was like a crazy derange thing to say. But from a computational neuroscience perspective, it basically means giving AI more than just intelligence, but like fundamental understanding of life, a an ability to perceive and experience emotion.

um grounded interactions in the world, long-term autonomy and natural start with Excel, the the foundation of modern life. Well, you know, it's it's a perfect storm that we're doing Excel. Yeah. Um I cut my teeth like on Excel. I get my knuckles bruised for using the mouse.

Um and then I found my way into a research lab. And the reason it hasn't been done yet is because the people who are capable of doing it are just so out of distribution for like regular people that Excel hasn't been touched um until now. Right. Okay. Uh talk to me about actually building this thing.

I remember like Cap IQ and Bloomberg had like VBA plugins for Excel. There's some pretty crazy stuff you can do under the hood of Excel. It is very modular. I don't know if that's changed with Excel going more cloudnative, but uh you know, Microsoft talks a big game about C-pilot.

Are you worried about getting steamrolled? Is there a way that you guys work together? How does that dynamic play out? And how are you actually building this thing? Because we've seen like Cle just does screen recordings, right? And that's a different vector.

There's other LLMs that are just like using the uh like they'll integrate with any IDE by just using the uh the the ADA features like the accessibility features to just scrape the text out. Of course, but how are you working through that?

So for context, the reason we're doing what we're doing now is because we actually set the state-of-the-art in OS world which is the computer use benchmark um computer use like sweet benches to software engineering. Um but we wanted to go tackle it in a much deeper way as if we had full access to the machine interface.

Mhm. So what that means is like can you write your own code essentially to do what you want to do and to do that you actually have to own the machine interface. So we couldn't build exactly on top of Excel. What we actually wanted to do is recreate it entirely with feature parody.

And then from there the big question is how we're doing this. I don't care to give away a little bit of a secret. This is being done through codegen at runtime which is a totally new paradigm for codegen. It does codegen is not about writing your own tools to store in your codebase right.

It's about like giving the AI the ability to do what it needs to do as it has to. Interesting. So, it's writing all of it. It's essentially writing its VBA as it needs. I haven't used uh Microsoft Copilot that much, but uh I've noticed that the way Gemini is inter is uh is like surfaced in Gmail.

It's very clear that it's not doing codegen. It's actually just taking text dropping in LLM. It's a chat interface and that's it. And so I can imagine that we're in we're entering the the v2 of of like sidebar companions and this feels very logical.

Yeah, even Gemini if you actually look into the sheets integration they have now they are starting to do a little bit of Python integration. So they are doing some nump some some data frame manipulation. Okay, but they're much more hesitant to get there and we were very hesitant to even call this cursor for whatever.

Sure. Sure. cuz like it's such a caricature, but there's a truth that's much deeper about how you do that and it's actually about owning the interface the same way cursor had to fork VS Code to own it. Yeah. Um, talk about the demand.

I replied to a comment of yours 15 minutes ago and then a bunch of people are now replying to me saying, "Can I get an invite code? " So, it seems seems like All right. All right. Um, announcement. Well, we got First of all, I went viral on LinkedIn. I didn't even know that was possible. Wow.

Um, let's give it up for LinkedIn. They're always always We got about 7,000 comments now and I've only given out about 50 codes. This thing is extremely token hungry. Um, but I did just make code TBPN. No, there'll be 10 10 uses. It's probably eaten up by now. Go hunt it down.

But yeah, we are we are trying to get this out as fast as we can. That's why I'm writing these big checks for infra so that we can scale up capacity to serve. Um, the only bottleneck is GPUs. That's amazing. Incredible. Well, thank you for pinging us. This is super exciting.

Come come back on planning to so what's the the next few months looks like is scaling shortcut or are you going to Can I refine this a little bit?

I want to ask uh uh cursor very ground up adoption from what I can tell it's a developer who's at a big company and they just buy it or install it, expense it, that type of thing versus going top down to, you know, I I imagine Goldman Sachs has a lot of people working in Excel.

You could go to Mackenzie, you could go to Goldman and say, "Hey, CTO, install this and let's do an enterprise contract. " How are you thinking about go to market? Yeah. Um, I mean, the stupid answer is I'm doing a little bit of both and seeing what's working.

Um, but it's pretty obvious that the same people who adopt AI in enterprise for Excel don't have the same risk tolerance or like the appetite to try stuff at the frontier that does software engineers like for cursor. Yeah.

Um so and also my intention is that this bubbles up organically and that people bring it in as opposed to me go tell me telling some VP adopt this you can produce headcom at 20%. Because that reality also can exist if I wanted to right like makes sense. Very awesome. All right we'll come back on soon.

Thank you for the code. Hopefully hopefully we get a shot at it before the audience says I'll make you a special one. Yeah we appreciate it. Congrats on the launch. Crushed it. Thank you so much. We'll talk to you soon. Uh, and up next we have Flo Crellllo coming back on the stream. So Han's former boss. Former boss.

Former boss. So we open the show with Saomself and now we're going to a former employer to get his side of the story. And the question that I want to answer is that that point that growing Daniel pointed