Three Ways To Protect An AI Prompt ๐
Every AI company now has an answer for what happens to your prompt, and the answers fall into three groups that are worth telling apart.
$ARX sits in the third group, on a compute network that apps on $SOL have used since Mainnet Alpha opened on February 2.
The three approaches:
A retention policy, where the provider promises not to keep your prompt afterwards, which still means a machine reads it while the work happens
Trusted hardware, where the job runs in a sealed area on the processor and you trust that seal and the company that made it
Cryptography in the stack, where the keys that open the job are split across independent parties so less of the system has to be trusted
Arcium put Blackthorn in that last group on November 3, describing it as its own category of confidential inference. It still uses the GPUโs confidential mode, but the sealed area on the host processor is taken out of the trust chain and the GPUโs session keys are held as fragments across Arcium nodes.
My caveat is that a promise and a cryptographic design are audited in very different ways, and Blackthorn is in early access, so its design has not yet run at the scale retention policies already do.
For anything sensitive, the useful thing to ask is how many parties you are being asked to trust, and Blackthorn is built to make that number smaller.
#AI
Every AI company now has an answer for what happens to your prompt, and the answers fall into three groups that are worth telling apart.
$ARX sits in the third group, on a compute network that apps on $SOL have used since Mainnet Alpha opened on February 2.
The three approaches:
A retention policy, where the provider promises not to keep your prompt afterwards, which still means a machine reads it while the work happens
Trusted hardware, where the job runs in a sealed area on the processor and you trust that seal and the company that made it
Cryptography in the stack, where the keys that open the job are split across independent parties so less of the system has to be trusted
Arcium put Blackthorn in that last group on November 3, describing it as its own category of confidential inference. It still uses the GPUโs confidential mode, but the sealed area on the host processor is taken out of the trust chain and the GPUโs session keys are held as fragments across Arcium nodes.
My caveat is that a promise and a cryptographic design are audited in very different ways, and Blackthorn is in early access, so its design has not yet run at the scale retention policies already do.
For anything sensitive, the useful thing to ask is how many parties you are being asked to trust, and Blackthorn is built to make that number smaller.
#AI
