AI Just Went Sailing Past Human Intelligence

In this episode, I discuss:

🍓 How Strawberry Works

Strawberry is likely based on incremental improvements to GPT-4, not a completely new foundation model. It uses various training techniques, most of which are based on open-source research. OpenAI excels at bringing these techniques together and productizing them effectively.

🧠 AI Surpassing Human Intelligence

Strawberry demonstrates capabilities that suggest it may have surpassed human intelligence in certain areas. It's performing at a level useful to PhD researchers in specialized fields, indicating a significant leap in AI capabilities.

📊 The Importance of Benchmarks

Objective benchmarks like Simple Bench are crucial for accurately assessing AI capabilities. Current benchmarks show top models still underperforming humans, but Strawberry may change this landscape.

🤖 Embodiment and AGI

I argue that embodiment isn't necessary for achieving AGI. The real test of intelligence lies in solving complex problems, not necessarily in physical interaction with the world.

🔮 The Future of AI and Robotics

The combination of AGI data centers and AGI-controlled robots could become the two pillars of the future economy. The humanoid robotics market is projected to become enormously valuable in the coming decades.

🏴󠁧󠁢󠁳󠁣󠁴󠁿 The "No True Scotsman" Fallacy in AI

I discuss how debates about whether AI is "truly thinking" become irrelevant when the AI is practical and useful. The capabilities and real-world applications matter more than semantic arguments.

🌟 Implications for AGI

Strawberry might represent a pivotal moment in the development of AGI. While it's too early to definitively claim we've achieved AGI, the rapid progress and rumored developments (like OpenAI's Orion) suggest we may be closer than previously thought.