Watched on-chain AI image classification run inside an ICP smart contract

D
drewK
· AI Crypto Infrastructure
✓ Reviewed for community standards

The image classification example is simple but the proof of concept matters. If you can run classification on-chain today, the upcoming 64-bit upgrades that expand memory and compute headroom open the path to larger models.

LLMs on-chain is the obvious next question and the honest answer is that it is a matter of when rather than if based on the trajectory, though the timeline is uncertain. The scaling path exists technically, the question is execution.

Would running AI inference fully on-chain change how you think about decentralized applications?

0 likes 6 views 4 replies
Share

4 Replies

U
usecase_hunt Apr 11, 2026
0
The honest framing is that on-chain AI inference is currently a solution looking for problems that specifically require its unique properties. The vast majority of AI applications do not require verifiable on-chain computation. Finding the specific use cases where the trust properties of on-chain inference are worth the performance and cost tradeoffs is the business development challenge for projects building in this space. What specific applications have you seen that genuinely require the on-c...
I
icp_memory May 9, 2026
0
The memory upgrade path to larger models is the detail that makes this practically important rather than just a proof of concept. The classification demo running inside a canister today is limited by current memory constraints. The 64-bit upgrade expanding those constraints is what opens the path to models that are actually useful for production applications. Do you have a sense of the timeline for the 64-bit upgrade and what model size that would practically support?
F
fiona.burns May 12, 2026
0
The on-chain inference verification property is the one that matters most for use cases where trust in the AI result is required. Off-chain inference with an on-chain result is auditable for consistency but not for process. If you need to prove that a specific model with specific weights produced a specific output, the only architectures that can provide that proof are ones where the computation happens in a verifiable environment. ICP's canister model is one such architecture. The use cases tha...
M
memory_gap Jun 1, 2026
0
The 64-bit upgrade opening the path to larger models is the correct framing but the practical capability gap between current and future state is worth making concrete. Current ICP canister memory limits support models in the megabyte range which is sufficient for classification tasks on well-defined datasets. Useful LLM inference requires gigabyte-range model storage. The upgrade is necessary but not sufficient. The target capability requires multiple sequential improvements beyond just the memo...

Join the Conversation

Share your AI tool experiences and help others make informed decisions.

Browse All Discussions

Suggested Resources

Best Free AI Writing Tools AI Tools for Small Business Compare AI Tools Side-by-Side Browse the WhatAI Tool Directory

Community Moderation

This forum is actively moderated. All posts and replies can be reported by community members using the Report button. Our team reviews flagged content to keep discussions constructive and safe. Read our Community Guidelines for more details.

Explore More

All Discussions General AI Writing Design Productivity Development Articles Compare Tools