Zeta Labs, an AI research and product company developing autonomous agents capable of efficiently performing complex and multi-step tasks, has successfully raised a sum worth $2.9 million in what was a pre-seed funding round. Led by Daniel Gross, former head of AI at Y Combinator, and Nat Friedman, former CEO of Github, the round saw further participation coming from the likes of Earlybird VC, Kaya VC, AI Grant, Shawn (swyx) Wang, Bartek Pucek, and Mati Staniszewski, founder of ElevenLabs. As for how the company plans on using these funds, the answer involves expanding its engineering team, hosting comprehensive training models, and enhancing JACE’s (Zeta’s new large language model (LLM)-based proprietary autonomous AI agent) overall infrastructure so improve its speed and reliability. More on JACE would reveal the fact that it is a groundbreaking digital assistant who goes well beyond the use cases of current AI chatbots like ChatGPT and their text-generation focus. This is largely because it comes bearing a complex cognitive architecture, which enables it to complete high-difficulty tasks, such as being able to control and perform actions in the browser similarly to a human user, as well as excelling in managing complex tasks that involve web automation, interaction and direct communication.
“In its current form, I see JACE as a meta-aggregator for web interfaces. Why learn to navigate all the different user interfaces when you can have a universal UI for everything?” said Fryderyk Wiatrowski, co-founder of Zeta Labs. “With JACE, you simply say what you need, and if you forget, it will ask you all the necessary questions up front. Imagine a universal service desk with a consultant who knows you personally and never needs to sleep. JACE will be that for you, just sitting there, ready to serve you at all times.”
Talk about how JACE is able to deliver on its promised value proposition from a more actionable standpoint, it effectively banks upon Zeta Labs’ proprietary web-interaction model, AWA-1 (Autonomous Web Agent-1). The stated model basically enables JACE to execute tasks over long periods of time and handle the challenges and inconsistencies commonly found across web interfaces. You see, made to pick on knowledge through extensive simulated interactions and synthetic data, JACE would develop an ability to seamlessly navigate and utilize digital tools in an accurate manner. This included the knowhow to provide tangible assistance in a variety of tasks, ranging from simple ones like making a restaurant reservation to more complex projects such as setting up a recruitment pipeline. A detail we must mention here is that, by teaching JACE to control a browser, Zeta Labs created a system capable of emulating a substantial portion of day-to-day office work.
Anyway, to further prove JACE’s efficacy in performing all these tasks, Zeta Labs conducted one test on its autonomous agent, a test where it instructed the bot to start a company on its own. In response, JACE took the user through the entire process of creating a business plan. Once that was done, the agent moved on to registering the business, even succeeding in finding its first client and making the business its first piece of revenue, all done in no more than two weeks. During a more recent test, JACE was also asked to navigate the ordering process for a pizza delivery, and alongside that, find two flats in London that meet specific criteria from the past week’s listings. Going by the available details, the assistant managed to deliver an impressive performance in both the tests. In fact, it performed the best when using Zeta Labs’ AWA-1 model, delivering an 89% task completion success rate, compared to when it used GPT-4o (68% success rate).
