Vitalik: 'CROPS Ethereum Access Layer' and 'CROPS AI' Have Significant Overlap

Foresight News reports that Ethereum co-founder Vitalik Buterin has provided updates on the progress of projects related to CROPS AI. He mentioned that Deepseek v4 has been released, and there is now a 2-bit quantized version that can run within 90GB of memory. It performs faster on Apple hardware (approximately 35 tok/s) but only achieves about 7 tok/s on AMD hardware. He emphasized that truly supporting multiple hardware vendors rather than relying on a single platform is a key distinction between "decentralized AI" and "true CROPS AI." Additionally, he shared updates on several tools: the messaging-daemon now supports Telegram (though configuration remains complex); Lucebox-hub performs excellently when running intensive models (such as Qwen 27B), achieving speeds roughly twice that of llama.cpp on his 5090 laptop; and VoxTerm (a local AI recording tool) is also under continuous development. Vitalik specifically pointed out that there is significant overlap between the 'CROPS Ethereum Access Layer' and 'CROPS AI.' For instance, using ZK to enable paid calls to remote large models can also be applied to Ethereum's private RPC reads. Meanwhile, small models fine-tuned for specific scenarios (such as Leanstral) perform exceptionally well in code generation. In the future, more specialized fine-tuned models for the Ethereum ecosystem should be developed to enhance code security.