Lagrange's zkML: The direction is right, but don't rush to imagine its implementation
My attitude is: Lagrange's zkML approach has great potential, but don't expect it to happen too quickly in the short term. DeepProve-1 can verify GPT-2, which is indeed a milestone. However, GPT-2 is just a relatively small model, and scaling up to large models like LLaMA and Gemini might lead to skyrocketing verification costs.
The hype brought by the airdrop is short-term; whether it can sustain in the long run depends on performance optimization and whether the verifier network can keep up. If generating a proof takes several minutes and the Gas fees are very high, then no matter how good the narrative is, it will be hard to implement.
So my viewpoint is very clear: Lagrange is on the right track, but don't rush to treat it as a project that is about to explode. We need to look at the long term and see if it can handle large model verification.
