Introducing `Agent-1`: a breakthrough foundation model that can operate software like a human. This is the brain powering Personal Assistant. We’re already well above previous state-of-the-art, and we’re improving massively each week. More details:

First, why are we building this? Current hosted APIs are amazing — but operating software isn’t a task today’s models can handle reliably. Even the next generation of unreleased closed models aren’t up to the task (and trust me, we’ve tried).

And with the complexity that comes with this type of task, costs are through the roof, and speed is an issue. So, we decided to build our own suite of models, with one purpose: to operate software reliably, quickly and cheaply.

This requires a new type of AI model — the goal is to use its parameters ~very effectively~. Current models store lots of knowledge, leaving fewer parameters for reasoning. Instead, we aim to put all of the model's horsepower to work on dynamic reasoning.

This approach enables our model to handle situations it was never trained for. Instead of relying on its knowledge of a particular site, it just figures out how to use it! And our software built on top of the model allows it to learn over time, without wasting model parameters.

This is just the beginning. As the model rapidly improves, so will its reliability on more complex software. Our goal is to surpass human ability — an assistant that can operate any software and reliably accomplish complex goals on a user’s behalf.

We’ll be deploying our newest iterations of Agent-1 into Personal Assistant in the coming weeks and months. We’re excited to see what you can accomplish!

What's special about this again? @MultiON_AI already surpasses this out of the box with ~1 sec per action speed 🥱 & can self-learn to improve itself automatically. Full post on it tmrw 😎

Oh man, Google admin console isn’t easy to figure out, even for a human… (or is it just me?)

Same here. This is why I chose it for a demo. If Agent-1 can use GCP, the sky is the limit :)

I use hyperwrite quite a bit. Right now the biggest problem with assistant is the speed. I understand the technical limitations having worked on agentic stuff as well but hoping you guys have figured out how to solve this :)

The speed is my biggest pet peeve... a big incentive for us building these models was to make it faster. Our next goal is 1 second per 'action' (right now, it's 5-10s per). We think we'll be there within a month or two.