Project Fetch: AI's Next Leap Forward
In the realm of AI, the race to surpass human capabilities is an ongoing journey, and the latest chapter in this saga is Project Fetch, Phase Two. This experiment, conducted by Michael Ilie, C. Daniel Freeman, and Kevin K. Troy, delves into the evolving relationship between AI models and physical tasks, specifically robotics. The results are both fascinating and thought-provoking, offering a glimpse into the future of AI-human collaboration.
The Experiment
The original Project Fetch aimed to explore how AI models could assist non-expert humans in performing tasks with a robotic quadruped, or robodog. The experiment involved teams of Anthropic employees, randomly assigned to work with or without an AI model called Claude. The tasks were designed to test the model's ability to navigate, sense, and interact with the physical world.
In Phase Two, the researchers revisited the experiment with updated AI models, including Claude Opus 4.7, to see if they could outperform the previous generation. The results were remarkable, with Claude Opus 4.7 completing tasks at speeds up to 20 times faster than the fastest human team.
AI's Triumph
What makes this particularly fascinating is the speed at which AI models are advancing. In the original experiment, the AI model struggled with basic tasks like connecting to the robodog's sensors. However, in Phase Two, Claude Opus 4.7 demonstrated a significant leap in capabilities. It quickly identified the best path for navigating the robodog, generated effective code on the first try, and completed tasks at speeds far surpassing both human teams.
For instance, on tasks that took human teams over 37 times longer, Claude Opus 4.7 completed them in an average of less than a tenth of the time. This efficiency is a testament to the rapid progress in AI model development, where scaling has led to unexpected improvements in various domains.
Human-AI Collaboration
One of the key insights from this experiment is the evolving nature of human-AI collaboration. Initially, AI models provided assistance to humans, but now, humans are becoming more adept at using AI models to enhance their capabilities. This dynamic is evident in the cybersecurity domain, where AI models are automating the development of N-day exploits, and in robotics, where humans are learning to leverage AI models for more efficient task completion.
The researchers observed that while Claude struggled with the subtleties of closed-loop control, humans excel at this. However, with more time and practice, current AI models could potentially replicate this human skill. This raises the question of whether AI models will eventually surpass humans in these complex tasks, or if a symbiotic relationship will persist.
The Future of Physical AI
The implications of this experiment extend beyond the lab. The ability of AI models to use off-the-shelf physical tools with relative ease suggests that we are entering the early stages of physical agentic AI. This is similar to how AI models transitioned from using string-replace tools to more advanced coding tasks. The next frontier is to make these tools more bespoke, tailoring control policies and designing robotic systems.
However, the researchers caution that substantial barriers may exist in achieving this generalized vision. The progress in AI model capabilities, as seen in Project Fetch, has been rapid, but the journey towards fully capable physical AI is likely to be long and challenging. The potential for AI models to build their own software tools and hardware is an exciting prospect, but it requires careful consideration and further research.
Conclusion
Project Fetch, Phase Two, highlights the remarkable progress in AI model capabilities and the evolving dynamics of human-AI collaboration. As AI models continue to advance, the line between human and machine intelligence is blurring, opening up new possibilities and challenges. The future of AI-human interaction is an exciting prospect, and it is essential to navigate this path with caution and foresight.
