Hi! 👋 Welcome to The Big Y!
“Theory of mind refers to the capacity to understand other people by ascribing mental states to them. A theory of mind includes the knowledge that others' beliefs, desires, intentions, emotions, and thoughts may be different from one's own. Possessing a functional theory of mind is crucial for success in everyday human social interactions.”
As I have discussed previously, we should refrain from anthropomorphizing AI and large language models (LLMs). But as they improve and get better at interacting with us, we are going to start running into hard to answer questions about signaled intelligence and actual intelligence.
There is an interesting paper out of Stanford that has studied the idea that LLMs could have spontaneously developed theory of mind from being trained on language data that includes theories on mental states and beliefs. They found that GPT-4 was able to solve 75% of their benchmark tasks, which is the same as the average human six-year-old child.
Another paper dives into the idea that you can create a framework for LLMs that can then “direct” the LLM to display human-like tendencies through inferences, then exhibiting what we see as artificial general intelligence (AGI) style outcomes. Could the future of AGI be based less on the LLM and increasingly on the tools that govern and teach LLMs to anticipate, reason and reflect?
At the end of the day, because LLMs parrott words and thoughts back to us in a language we understand, it is hard to separate true insight from words mirrored back at us. Emergent behavior can tend to look like it was planned and intentional, but many times it doesn’t actually mean much and should be looked at with a critical eye. I think we’ll continue to see comparisons of LLMs to the human mind, but more importantly, I think we’ll see a growth in tooling and frameworks that enable LLMs to sound and reason progressively more like humans do.
Google had a big AI week: They finally, officially launched Gemini Ultra which is showing to be competitive with GPT-4 on the leaderboards, while also launching a monthly subscription one cent cheaper that ChatGPT. On the other end of the spectrum, a Waymo car was destroyed when people threw a firework into the vehicle and started a fire (nobody was in the car at the time), complicating an already complicated relationship between the autonomous vehicles and SF.
Know someone who might enjoy this newsletter? Share it with them and help spread the word!
Thanks for reading! Have a great week! 😁
🎙 The Big Y Podcast: Listen on Spotify, Apple Podcasts, Stitcher, Substack