Convergence, and the end of dumb tools

Humans are a tool-using species. But that's only part of the hack that led to our evolutionary speed run. More important than using tools is our ability to teach others how to use them too. From person to person, generation to generation, tribe to tribe, ideas spread as each of us become both pupil and, in time, teacher—innovating as we go.

To date, however, most of the software tools we have built have been unable to learn. In fact they have been remarkably, disappointingly dumb. Just think how many times a day you click the same buttons and enter data into the same fields, without the applications behind them ever working out how to improve this process for you. I have never seen a calendar app realise I’m running late and send a message ahead to the person waiting, or have my CRM realise I never use certain buttons and so try and streamline itself. The most dynamic software most of us use that reacts in real time to your interests is social media, and that’s hardly because they want to make you more productive.

Even GPT models have proven themselves to be pretty limited in the classroom. They are incredible tools for generating and manipulating content. You can create data, edit essays, and reformat videos with a text prompt. Amazing. But they are lousy learners. If you want to teach a foundational model new tricks, you need access to huge GPU clusters, data centres, teams of annotators, and the top 0.1% of engineering talent in the world.

If we want transformational tools, they need to learn and adapt. And Convergence is allowing AI models to do just that.

Convergence has built a unique agent platform for performing web-based tasks, Proxy, which they have launched in beta today. You can ask Proxy to perform any task, from "add this website to my CRM" to "book me a table for lunch in King's Cross," and like any of us, it can navigate websites, fill in forms, and book something for you.

I've had many moments over the last few years where I've used a tool that felt like magic. Stable Diffusion was one of those, as was GPT3. The first time I used Proxy blew me away, and the release is just the first step, for the first product, of Convergence’s ambitious roadmap. It’s one of the reasons I am delighted to say we have led their first funding round also announced today.

Autonomous web browsing while impressive is not unique - what’s truly impressive is that Proxy can learn from you— literally from your actions, and personalise it’s response to your requests going forward. As well as building a powerful workflow to navigate the web on your behalf, the Convergence team has created a unique approach to enable models to learn at inference time—a setup they're describing as Large Meta Learning Models. In the background, this LMLM (pronounced "LimLim") approach is proving to scale much more efficiently on data-to-first-success benchmarks. On the frontend, a user just needs to annotate a webpage whenever Proxy gets stuck—by drawing a box around the error it has made—and it will learn, creating a personalised model - an ‘approxymation’ of you as it goes.

This sounds simple, but it's a huge leap in user experience and, hopefully, productivity. As Convergence scales, it will mean that AI agents become useful in many millions or more ways, as millions of us train them. Agents won't just be interns; more importantly, we will be their mentors. If that sounds like how tools should be to you, sign-up to try the beta at www.convergence.ai.