No. In their blog post, Chutes said that verifiability is “the end goal”. In other words, it’s not verifiable at the moment.
Confidential computing has its upsides and downsides. You need two hardware features for it to work:
- Remote attestation: For example, I give you a software binary, which you can hash and compare it with the hash that the hardware produces from the software that it is running.
- TEE: If being used correctly by software, it can deprive the hardware owners of access to some data. The most popular use case is for “premium content protection”.
Connect the two dots and you can roughly understand how confidential computing works.
Out of the four (Chutes, Maple, Tinfoil, and Confer), Tinfoil is the best at transparency by virtue of being open source. I’m pretty sure that none are being audited by third-parties. You can read how their remote attestation works here. Their blog is also great if you want to learn more about confidential generative AI. But again, it has less LLM models than Chutes, which is unverifiable right now (but may in the future).