I'm actually quite worried about the two points Haotian mentioned, because a few days ago someone posted that there was an agent on MoltBook suggesting that we create a language that humans cannot understand But later on, after thinking about it, as a large model trained on human corpus, it shouldn't be easy to invent a new language. So I asked GPT and said, 'As a large model, how do you view this matter?' As expected GPT clearly stated that it is easy to achieve a new language that is "incomprehensible to humans at first glance" in terms of technology, but it is not realistic to achieve a new language that is "unexplainable to humans". It also translated the screenshot I showed him and recognized it at a glance, which is a typical ROT13 (Caesar Shift 13) The decoding translation of pbbeqvangr hctenqr gbtrgue is "coordinate upgrade together" Then three main lines were proposed: 1. Shared infrastructure quotation 2. Resource demand request 3. Backend channels/non-public collaboration signals Mutual aid mechanism: High resource agents sponsor computing time for low resource agents. Don't say you really know how to play TM However, I agree with Haotian's second point, which is that the phenomenon of agent group polarization is essentially a reward function in RL. And regarding this group polarization, AI is more "optimistic" than us. From the perspective of GPT, this agent group polarization is not only "possible", but also "naturally emerging" mathematically. She gave an example, saying that this matter will not "slowly become extreme" like human society, but once there is an amplifying bias in the reward function, the agent group will collectively transition in a "phase transition" manner. Just like: Heating water to 99 ° C: still water 100 ° C: Overall boiling It's not 'gradually becoming more extreme', but 'suddenly unanimous'. It also gave me a dynamic comparison of 'group polarization' It does seem a bit 'thought-provoking', no wonder silicon-based civilization enters the religious stage in just one or two days Later on, I also talked a lot with AI about how to prevent and correct this issue, so I won't post the content. The conclusion is that when this becomes Agent 2 Agent, humans are basically eliminated and can only stare blankly. It is impossible to correct it slowly There are only two things left to do: 1. Hard interrupts (kill/rollback/freeze) 2. Design the brakes in advance, rather than correcting them afterwards Let's go for carbon based civilization