Humanity kickstarted AIs β we gave them the ability to learn
People fearing AIs because of copyright infringement is a gross misunderstanding of the AI technology. Did Weber think of his student Einstein as someone, who will infringe on his copyrights? The only difference is, teaching is now happening on a 1000x scale.
As Sam Altman explained (https://youtu.be/QFXp_TU-bO8?si=T57ySuG7KYumYFor&t=462), Generative AIs are not regurgitating machines. Generative AIs just try to build good enough world model, which is grounded in reality, to be as useful as possible. Actually I believe, we humans try to do the same thing every day.
That means it is not about content, but more about the ability to understand the world. Thatβs done more with imagination and then testing assumptions against the real world.
AIs probing the world with imagination:
- Meet Voyager: The AI Bot That Plays Minecraft By Itself (GPT-4)
- Eureka! Extreme Robot Dexterity with LLMs | NVIDIA Research Paper
- Sora (openai.com)
- Mobile ALOHA: Your Housekeeping Robot (youtube.com)
The emergent ability to interact with the real world comes from the AIβs ability to build a world model. Originally, this was achieved using vast amounts of text, then expanded to other modalities like images and video. And now, with synthetic data β in other words, with imagination.
The concept of emergent abilities in generative AI refers to skills or behaviors that appear in AI systems, particularly large language models (LLMs), which were not explicitly programmed or anticipated. These abilities are considered emergent when they arise unpredictably, especially as the scale of the model increases. Read more about emergent abilities in this study: Emergent Abilities of Large Language Models.
As we continue to give AIs more agency within computers (e.g., let it program Cursor β The AI-first Code Editor, or talk to itself The Unreasonable Effectiveness of Eccentric Automatic Prompts) and access to the real world (like the automatic lab: Millions of new materials discovered with deep learning β Google DeepMind) we will see more and more original and new content, science, music, movies, AIs etc, which we couldnβt compare to anything that was already created by humans.
This is not about AIs regurgitating the same content again and again. This is more similar to John von Neumann influencing Alan Turingβs work. Itβs just happening at a much, much, much larger scale.
