Spent time with chat.opengradient.ai during a CreatorPad task today. @OpenGradient $OPG #OPG . The OG Portal is sitting at 895.47K inference transactions on mainnet right now — live counter, verifiable at portal.opengradient.ai — and I kept thinking about what "user-controlled" actually means inside that number.
The chat product lets you pick the model. Claude, GPT, Gemini, Grok. Your prompt goes through an OHTTP relay, identity stripped before it touches the TEE-isolated enclave. That architecture is real and it holds up. But hold up — the "control" is sitting in the delivery layer, not the intelligence layer. You're choosing which private pipe your request travels through. You're not choosing model weights, system prompts the underlying provider enforces, or behavioral guardrails baked into whatever model version is running at the other end.
I opened it expecting something else and had to adjust. The actual experience is closer to: a more private front-end to models you'd use anyway. The TEE attestation proves the enclave ran correctly. It says nothing about what the model decided to do once it received your prompt.
Still useful. Still a real product. Different from the framing, though.
The question I'm sitting with... is whether users arriving through chat.opengradient.ai understand that distinction, or whether "user-controlled AI experiences" is doing narrative work that the architecture — however well-built — can't quite fully back up.
The chat product lets you pick the model. Claude, GPT, Gemini, Grok. Your prompt goes through an OHTTP relay, identity stripped before it touches the TEE-isolated enclave. That architecture is real and it holds up. But hold up — the "control" is sitting in the delivery layer, not the intelligence layer. You're choosing which private pipe your request travels through. You're not choosing model weights, system prompts the underlying provider enforces, or behavioral guardrails baked into whatever model version is running at the other end.
I opened it expecting something else and had to adjust. The actual experience is closer to: a more private front-end to models you'd use anyway. The TEE attestation proves the enclave ran correctly. It says nothing about what the model decided to do once it received your prompt.
Still useful. Still a real product. Different from the framing, though.
The question I'm sitting with... is whether users arriving through chat.opengradient.ai understand that distinction, or whether "user-controlled AI experiences" is doing narrative work that the architecture — however well-built — can't quite fully back up.