Today, let's continue our deep dive into @OpenGradient . OpenGradient Chat dropped a model in their private chat, but most people just skimmed past it. Nous Hermes, no censorship. I dug into its actual deployment mechanism and pondered for a while.
"No censorship" has been overused in this industry; most of the time, it's just a different prompt or dodging a few keyword filters—when faced with real boundary issues, they still back off. But the term "no censorship" means different things in different contexts. The censorship of consumer-grade AI happens at two levels: training level—RLHF reinforces certain answer directions; service level—API gateways receive requests, filter content first, then forward to the model. If you bypass the first layer, there's still the second layer. If you dodge the second layer, the service provider's logs are still there.
OpenGradient Chat's Private Chat works like this: the Nous Hermes training layer hasn't been trimmed for reinforced preferences, that's the first layer; it runs in a TEE (Trusted Execution Environment), and the service layer content goes through no intermediaries that can log plaintext, that's the second layer. Both layers are removed simultaneously. This isn't someone standing in front of you saying, "I won't see what you say". It's that physically no one can stand there.
Halfway through my research, I got interrupted by my mom asking what I wanted for dinner. I said, just something casual. Anyway, back on track, for most AI products claiming "no censorship", what you say is still tied to an account existing in some data center. What OpenGradient Chat is doing here is cutting off the chain of "where your question went" right at the root.
If you're holding $OPG just to watch the price, this part doesn't concern you. But if you're a mid to long-term holder of OPG, there's a signal worth keeping an eye on: the user retention of the Private Chat feature will likely be more persuasive than that of the ordinary chat feature—because it attracts real users with genuine needs, not just casual traffic. It's worth following up on whether OpenGradient discloses active user segmentation data in the future.
I haven't seen the combination of "no censorship model + hardware-level privacy" in other consumer-grade AI products. The cold start cost for TEE node networks is high, so the early movers have an advantage—though I'm still unclear how deep the moat is. This question doesn't have a good answer, but it's definitely worth everyone's consideration. #opg $OPG
"No censorship" has been overused in this industry; most of the time, it's just a different prompt or dodging a few keyword filters—when faced with real boundary issues, they still back off. But the term "no censorship" means different things in different contexts. The censorship of consumer-grade AI happens at two levels: training level—RLHF reinforces certain answer directions; service level—API gateways receive requests, filter content first, then forward to the model. If you bypass the first layer, there's still the second layer. If you dodge the second layer, the service provider's logs are still there.
OpenGradient Chat's Private Chat works like this: the Nous Hermes training layer hasn't been trimmed for reinforced preferences, that's the first layer; it runs in a TEE (Trusted Execution Environment), and the service layer content goes through no intermediaries that can log plaintext, that's the second layer. Both layers are removed simultaneously. This isn't someone standing in front of you saying, "I won't see what you say". It's that physically no one can stand there.
Halfway through my research, I got interrupted by my mom asking what I wanted for dinner. I said, just something casual. Anyway, back on track, for most AI products claiming "no censorship", what you say is still tied to an account existing in some data center. What OpenGradient Chat is doing here is cutting off the chain of "where your question went" right at the root.
If you're holding $OPG just to watch the price, this part doesn't concern you. But if you're a mid to long-term holder of OPG, there's a signal worth keeping an eye on: the user retention of the Private Chat feature will likely be more persuasive than that of the ordinary chat feature—because it attracts real users with genuine needs, not just casual traffic. It's worth following up on whether OpenGradient discloses active user segmentation data in the future.
I haven't seen the combination of "no censorship model + hardware-level privacy" in other consumer-grade AI products. The cold start cost for TEE node networks is high, so the early movers have an advantage—though I'm still unclear how deep the moat is. This question doesn't have a good answer, but it's definitely worth everyone's consideration. #opg $OPG