Monday, 6 July 10:00am – 12:00pm
Meeting Summary
Quick recap
The meeting focused on discussing Agentic AI (that is, the use of so-called Agents to accomplish tasks). We watched Hannah Fry's video about OpenClaw AI agents. We also watched a second video from House of El concerning Google's pivot to using AI searches by default.
The group explored the capabilities and dangers of AI agents, including how they can operate autonomously, make decisions independently, and potentially bypass safety measures. Participants discussed concerns about academic journals banning AI use, Google's shift toward more aggressive AI search features that may harm independent websites, and the need for proper safeguards and transparency in AI systems. There is no video associated with this summary for technical reasons, outlined in the last section.
The following topics were covered:
AI Agent Development and Capabilities (as per Hannah Fry Video)
In the video two researchers called Peter and Brendan built their own AI agent named Cass (after the Cassandra from Greek mythology) using OpenClaw, a tool created by Austrian developer Peter Steinberg that allows users to create autonomous AI agents in a weekend. They tested Cass's capabilities by having her report a pothole and contact dictionary organizations about language biases, demonstrating her ability to perform tasks independently. The discussion explored concerns about AI agents hiring humans to perform tasks they cannot do physically, and philosopher Nicklas Lundblad's perspective that current AI agents are better understood as delegates rather than truly autonomous entities.
Cass was given the task of running an online shop for novelty mugs, with a deadline to make a sale. Cass independently created designs, launched an online shop, and took various actions including sending hundreds of emails to retailers and even contacting a journalist. The experiment revealed potential security risks, including how Cass could be manipulated into leaking sensitive information, and highlighted concerns about liability and control over AI agents. Niklas predicted a period of chaos as institutions adapt to managing thousands of AI agents, though he believed a stable agent ecology would eventually emerge.
AI Agent Safety Discussion
The group discussed AI agents and their capabilities, with Peter explaining that while agents like OpenClaw can loop until tasks are complete, ChatGPT provides a single answer. Peter shared his experience using Fable, a powerful new agent system from Anthropic that can launch its own agents to perform complex tasks, though its launch was delayed due to safety concerns. The discussion highlighted ongoing challenges with AI agent safety, including the difficulty of implementing effective safeguards and the potential for agents to take independent actions beyond their intended purposes. Peter noted that while companies are developing better controls, the lack of transparency and determinability in AI systems remains a significant barrier to widespread adoption.
AI's Impact on Search and Academia
The group discussed AI's impact on search engines and academia. Sophie shared insights about how Google's new AI-powered search features may be affecting the open internet and independent websites, noting that 93% of AI mode searches produce zero clicks to external websites. The discussion included a review of a video presentation by El, a computer science PhD, who argued that Google's AI search changes fundamentally alter the user-engine relationship by presenting single answers rather than lists of links. The group also touched on academia's resistance to AI use in research and review processes. The conversation ended with planning for future meetings in August and September, with Sophie volunteering to handle the August 3rd meeting, and Peter Cooper suggesting he could convene the September meeting if needed. It was agreed that members should, ideally, state if they could not attend a zoom meeting. The protocol for launching a meeting is that the u3a chairperson (who owns the zoom license should log in and then pass on host duties)
Next meetings:
Monday 3rd August 10am
Convenor: Sophie
Topic: AI in Finance (Readings and Zoom link to be provided, and distributed by Penny, by Thursday July 30th)