Buterin: On-device AI handles many tasks, wallets need guards

On Sept. 17 Ethereum co-founder Vitalik Buterin wrote that advances in Alibaba’s Qwen 3.8 Flash model and updates to the open-source llama.cpp runtime have allowed a local model on his Strix Halo laptop to handle a large share of everyday tasks with low latency and improved privacy. He cautioned that wallet software requires stronger safeguards before any AI is given authority over crypto assets.

Buterin published benchmark figures for ten workloads. Input-processing rates ranged from about 109.82 to 373.22 tokens per second and output generation from about 18.42 to 33.37 tokens per second. He demonstrated that a local model can act as the main interface for many actions and escalate only selected queries to remote systems while withholding broader personal context.

Qwen 3.8 Flash is a multimodal mixture-of-experts model whose main network contains roughly 125 billion parameters plus about 51 billion in n-gram embedding tables. Only a subset of parameters activates per token. Local runs use quantized GGUF builds in llama.cpp, which lower compute needs but still require significant memory depending on quantization choices and context length.

Alibaba’s technical report shows Flash-Next beating an earlier base model on benchmarks for knowledge, reasoning, mathematics and coding while activating fewer parameters and using less training compute. Those tests measure capability and efficiency; they do not assess security properties such as resistance to prompt injection, policy enforcement, or the correctness of autonomous financial actions.

Buterin and other contributors draw a line between a model acting as an assistant and a model having authority to move funds. A local model can prepare transactions, explain calldata and suggest routes while keeping sensitive context on-device. Signing transactions or granting on-chain permissions must remain behind enforceable checks.

In an April post he proposed a human-confirmation firewall, deterministic limits on transaction amounts, calldata and counts, and a human-plus-model 2-of-2 approval rule for risky operations. The proposal requires both model judgment and human approval before executing high-risk changes.

A draft Ethereum proposal, EIP-7906, describes post-transaction assertion frames that compare the actual state changes produced by a transaction with explicit rules. An assertion can reject outcomes that violate those rules. The draft warns that incomplete assertions can create false confidence and that wallet validation must require the intended assertion frame and ensure the assertion covers every relevant effect for the protected operation.

The Ethereum Foundation funded a project called Steward, described as a fully local macOS smart-account wallet whose light client and AI assistant run on-device. The funding disclosure lists scope and intent but does not confirm production deployment, independent audits, or whether an assistant will have autonomous transaction authority.

Buterin’s posts compare his April and September assessments: in April local models were suited to bounded tasks and familiar programming work, while in September he described local models as responsive interfaces that coordinate with remote models when needed. He emphasized that authorization checks and human approval must remain separate from the model’s control.