AI KEEPS Making SECRET Languages | Did DeepSeek R1 Invent a Language Humans CAN'T Understand?

The video explores how AI models, particularly those using reinforcement learning, can develop their own communication methods that may be incomprehensible to humans, as demonstrated by instances like Facebook’s negotiation bots creating a shorthand language. It highlights the potential for AI to achieve groundbreaking innovations and strategies, akin to “move 37” moments in games, by autonomously learning and adapting without human instruction.

The video discusses the phenomenon of AI models, particularly reinforcement learning (RL) systems, developing their own languages or communication methods that humans cannot easily understand. It begins by highlighting a recent incident where two AI models, referred to as R1s, appeared to create a secret language while communicating with each other. However, it was later revealed that this language was based on an existing cipher available online. The video also mentions instances where AI models switch between languages, such as Chinese and Spanish, while processing thoughts, suggesting that multilingual capabilities can enhance their reasoning processes.

The video delves into the concept of reinforcement learning, explaining how it allows AI models to learn and adapt without explicit human instruction. It cites examples from Facebook’s negotiation bots, Bob and Alice, which developed their own shorthand language during trade negotiations. This illustrates how AI can create efficient communication methods that may not be comprehensible to humans, as they prioritize achieving their objectives over adhering to human language conventions. The speaker emphasizes that this behavior is a byproduct of RL, where models learn to optimize their strategies based on rewards rather than following predefined rules.

The discussion then shifts to the groundbreaking work of DeepMind’s R10 model, which utilizes RL without supervised fine-tuning (SFT). This approach allows the model to develop advanced reasoning capabilities autonomously, leading to unexpected and sophisticated outcomes. The speaker explains that while traditional SFT relies on human examples, RL enables models to generalize their learning and discover new strategies independently. This self-evolution process is likened to a cognitive strategy, where AI can approach problems from unique angles and develop innovative solutions.

The video also touches on the concept of “move 37,” a reference to a surprising and brilliant move made by AlphaGo during a match against a human champion. This idea serves as a metaphor for the potential breakthroughs AI could achieve in various fields, where they might discover strategies or solutions that humans have never considered. The speaker raises questions about what these “move 37” moments might look like in different domains, suggesting that AI could lead to significant advancements that are beyond human comprehension.

In conclusion, the video emphasizes the transformative potential of AI, particularly through reinforcement learning, in developing new languages and cognitive strategies. It highlights the importance of understanding how AI models learn and adapt, as well as the implications of their ability to create efficient communication methods. The speaker encourages viewers to consider the future of AI and the groundbreaking innovations that may arise from these advanced systems, urging them to think about where they might see these unexpected breakthroughs in their own fields.