On-silicon real-time AI compute governance from Nvidia, Intel, EQTY Labs (opens in new tab)

Allow me:

"Are you running the AI that you thought you were running, or a rip-off clone that will sneakily insert adverts for Acme, your one-stop-shop for roadrunner-related explosives, traps, and fake walls, into 1% of outputs? Here's how you can be sure."

People are starting to worry about generative AI/LLM training data - copyright, bias, knowledge gaps, bad/outdated info. They are also worried about the black box of hardcoded "alignment" layers on top - which can and has introduced similar issues. Cryptographic proof of a model's training data provenance and its compute would be reassuring to large enterprises, governments, and conscientious users who are hesitant to embrace AI.

You don't want to accidentally use part of a biased social media chat LLM to summarize legislation or estimate results of a business plan.

Had to use bullshitremover dot com for this one. This is the translation:

> Blah blah, AI is the future, trust it, crypto, governance, auditing, safer AI.

Some have speculated they want to command regulations so complex that only they can adhere to them.

This misunderstands what Verifiable Compute actually does. VC isn't about restricting execution, it's about proving properties of computation. The key difference is that VC lets you verify what happened, not control what can happen.

Think SSH keys, not DRM. SSH lets you verify you're talking to the right server without restricting which servers exist. Similarly, VC lets you verify properties of AI models (like training data lineage or inference characteristics) without restricting which models can run.

The regional censorship concern isn't relevant here since VC doesn't enable content restriction. It's a mathematical verification tool, not an enforcement mechanism.

> approved hardware will only run signed models

This is true. But this is an ability of the hardware owners. Intel and NVIDIA are not setting the rules - and there is a real commitment to that because its open source.

It's also confidential. Data, code, rules, ... all of these are processed together in secure enclaves. It's up to the hardware owner/users to determine that processing and to stamp/verify what they want.

Reads as compliance controls being embedded into the code with integrated gates to halt execution, or verify controls are met at runtime - providing receipts with computed outputs. This is generally oriented toward multi-party, confidential, sensitive computing domains. As AI threat models develop, general compliance of things during training, or benchmarking, etc become more relevant as security posture requires.

ref:

https://openai.com/index/reimagining-secure-infrastructure-f...

https://security.apple.com/blog/private-cloud-compute/

https://arxiv.org/html/2409.03720v2

Cryptographic proof that the model you’re using is the one you think it is along with proof it was not trained on biased or copyrighted data is the main feature here. Think certs when looking at webpages or ssh when looking at servers.

Given the amount of worry about training data provenance and associated bias/copyright issues, such a cryptographic proof of training would certainly act as a "speedometer" that disincentivizes speeding. At least among models that large enterprises, governments, and conscientious users are deciding to use.

It’s a poor comparison.

Regardless it’s about a trusted observation - in your metaphor to help you prove in court that you weren’t actually speeding.

Apple deploys verifiable compute in Private Cloud to ensure transparency as a measure of trust, and surely as a method of prevention whether a direct method or not (depends on how they utilize verifiability measures as execution gates or not).

No this is like saying adding a speedometer to your car makes it, as an inanimate object, personally liable for going fast if you press on the accelerator.