Everything Is a Code Problem | Uri Eliabayev

So first we have agents.

Again, it's all over the place.

I think this is for the first time we see them mature enough in order to do complex tasks.

But if we are going to talk about the next big thing, I see that everything is related to physical world, physical models, world models.

Now it's a huge trend in which models can generate world when you can interact, change, or even move around the environments.

So this is super fascinating.

Also, a very important trend for my behalf is multimodality.

For the first time, you can analyze not only text, but video, audio.

many elements of, let's say, inputs.

And last but not least, this might sound really geeky, but this is my personal preference.

For the first time, you see models that are capable to create files.

I work a lot with Claude about creating PowerPoint, Excel, Word, even much more complex formats.

And this was possible six months ago.

It wasn't possible at all.

Yeah.

And for us, very interesting to see that you are also testing new frontier models and evaluating those.

Something that we are seeing in our space, broken it down into conversational AI and agentic AI, those models have become so good, right? And many of them are good enough.

Exactly.

What's your take on that? Where's that industry heading with the new models and the new releases and everything?

Yeah, this is a good question because nowadays it's really difficult to evaluate the new capabilities of the models.

We see big announcements, but at the end of the day, people are not sure always how they can understand what were the new changes, what is the delta.

So what I say to the majority of my clients, but also in my community, that it needs to be task-specific.

You couldn't care less about the model or the benchmark because we all know that benchmark is nice for PR, but sometimes they are not telling the entire story.

So what I'm aiming for is just to have a specific task that you know what are the real benchmark.

And moreover, as a domain expert, you know how it should be completed.

So first we have agents.

Again, it's all over the place.

I think this is for the first time we see them mature enough in order to do complex tasks.

But if we are going to talk about the next big thing, I see that everything is related to physical world, physical models, world models.

Now it's a huge trend in which models can generate world when you can interact, change, or even move around the environments.

So this is super fascinating.

Also, a very important trend for my behalf is multimodality.

For the first time, you can analyze not only text, but video, audio.

many elements of, let's say, inputs.

And last but not least, this might sound really geeky, but this is my personal preference.

For the first time, you see models that are capable to create files.

I work a lot with Claude about creating PowerPoint, Excel, Word, even much more complex formats.

And this was possible six months ago.

It wasn't possible at all.

Yeah.

And for us, very interesting to see that you are also testing new frontier models and evaluating those.

Something that we are seeing in our space, broken it down into conversational AI and agentic AI, those models have become so good, right? And many of them are good enough.

Exactly.

What's your take on that? Where's that industry heading with the new models and the new releases and everything?

Yeah, this is a good question because nowadays it's really difficult to evaluate the new capabilities of the models.

We see big announcements, but at the end of the day, people are not sure always how they can understand what were the new changes, what is the delta.

So what I say to the majority of my clients, but also in my community, that it needs to be task-specific.

You couldn't care less about the model or the benchmark because we all know that benchmark is nice for PR, but sometimes they are not telling the entire story.

So what I'm aiming for is just to have a specific task that you know what are the real benchmark.

And moreover, as a domain expert, you know how it should be completed.