OpenAI's Swarm Framework and Google's Project Jarvis

Is Agentic AI really going to match the hype? We haven't even found a good framework yet.

Hey Everyone,

Lately I’ve been thinking about Agentic AI and if it’s more bark than bite. Chatbots certainly didn’t exactly manifest the utility I was expecting in the last two years and while Anthropic’s Computer Use is a fascinating idea, the idea and the reality seem fairly distant.

Claude can code and write, but even Google will have trouble making Chrome click and push a button. Project Jarvis to me will enable Google’s SGE to be more actionable like help us automate our shopping and travel experiences. Something by the way, Google has been promising for a very long while.

The problem is a lot of the AI agent frameworks aren’t very good or scaleable to more complex tasks. In mid October, 2024 OpenAI has unveiled “Swarm,” an experimental framework designed to orchestrate networks of AI agents.

Google’s “computer-using agent” under the codename Project Jarvis seems to be mostly a Chrome tool that enables you to do some rudimentary things by taking over your screen and let’s say completing travel bookings or helping you find a certain E-commerce shopping item you might need.

Meanwhile people are taking what OpenAI’s Orion model can do with o1 or Claude’s Computer Use, way too literally. Swarm as a framework however is worth taking a closer look into.

Apple’s AI is right to question OpenAI’s claims about the reasoning capabilities of their o1 model. There’s been a lot of academic papers about AI’s reasoning capabilities and there’s a global consensus on where we stand today, and it’s not in OpenAI’s favor. Their layers of AGI doesn’t make a whole lot of sense.

Still, OpenAI’s experimental framework, Swarm, has revived discussions about the societal impacts of AI-driven automation. While RPA stalled there’s no reason to assume Generative AI’s Agentic capabilities won’t stall as well.

Swarm is not an official OpenAI product. Think of it more like a cookbook. It’s experimental code for building simple agents. It’s not meant for production and won’t be maintained by us,” Shyamal Anadkat, a researcher at OpenAI, wrote in a post on X. This as Meta Platforms calls NotebookLlama like a recipe in a cookbook as well. But do we really want to DIY the future with open-source models?