Convergence secured $12M to develop personal AI assistants able to learn new skills

Convergence, a London-based startup founded by former Shopify and Cohere employees Marvin Purtorab (CEO) and Andy Toulis (CTO), has officially launched an early access alpha version of Proxy, a customizable AI assistant capable of navigating the web and completing low-level repetitive tasks on their users' behalf. Additionally, the company announced it raised $12 million in pre-seed funding from Balderton Capital, supported by Salesforce Ventures and Shopify Ventures.

A product demo showcases Proxy helping a user update a Salesforce dashboard with missing information obtained from Slack and Google Drive and assisting with writing and publishing a job opening to Glassdoor and Indeed. From what can be seen in the video, Proxy is transparent about the steps it thinks it should complete, provides screenshots and other evidence of the tasks it is performing, and asks for help when it needs additional context to proceed.

This invites an obvious question: how much time and effort do these assistants save? Everyone knows no model is completely trustworthy, and that Proxy outlines its step-by-step plan to complete a task and provides screenshots and other evidence is a testament to this. Undoubtedly, supervising Proxy as it does a task is a recommended practice. Then, any time Proxy reaches a point where it cannot proceed without help, the user must explain exactly what to do and provide the missing information so the assistant can complete the task.

In a telling demo that promises to show Proxy's potential for helping users in their daily lives, a user asks Proxy to buy a previously picked-out garment. This entails that the user already spent some time doing everything from finding the store to choosing the garment and ensuring the correct size is available. Then, Proxy reaches the checkout form and tells the user that it cannot proceed because it needs the user's personal information. Naturally, the user proceeds to type into the chat interface the exact information they could have typed directly into the checkout form.

Security concerns aside, the Proxy demos frequently make it seem like the user has spent just as much time and effort in supervising and feeding information to a middleman as it would take to perform some of the showcased tasks directly. Convergence looks at these issues more optimistically, claiming that these steps are necessary so Proxy can acquire new skills through memory and continual learning.

The company claims its key differentiator is a new class of models called Large Meta Learning Models (LMLM) which enable Proxy to adapt, learn, and remember "by integrating memory as a core component of the model architecture." Although technical details on the models are not yet available, Convergence claims that LMLMs, unlike traditional LLMs, can learn during inference time and improve over time, relying on user feedback only and eliminating the need for massive training datasets.