zksnark

11,353 views
18 Discussing
web3小虾米
·
--
Came across an Anthropic research blog: they open-sourced FlashProve—a way to use zkSNARKs to create “verifiable proofs” of LLM inference, proving that a response really came from the set of weights it committed to, without revealing the weights themselves. The hard numbers: on Llama 4 Maverick (around 400 billion parameters, MoE, 4-bit), a single H100 generates proofs at about 110 tokens/sec; each proof for a 1,024-token request is around 19 MB; verification takes about 17 seconds on a regular 24-core CPU; and GPU overhead is roughly 70–95× that of vLLM inference. The code, Lean formal proofs, and protocol documentation are all open-sourced under Apache 2.0. Still a research prototype—for now, it mainly supports MoE, greedy integer decoding, and contexts of about 1,024 tokens; the architecture itself isn’t hidden. But bringing zk from the blockchain layer to model outputs suddenly makes use cases like auditing and preventing model swaps during routing feel much more practical. #zkSNARK #AI #可验证推理 #Anthropic
Came across an Anthropic research blog: they open-sourced FlashProve—a way to use zkSNARKs to create “verifiable proofs” of LLM inference, proving that a response really came from the set of weights it committed to, without revealing the weights themselves.

The hard numbers: on Llama 4 Maverick (around 400 billion parameters, MoE, 4-bit), a single H100 generates proofs at about 110 tokens/sec; each proof for a 1,024-token request is around 19 MB; verification takes about 17 seconds on a regular 24-core CPU; and GPU overhead is roughly 70–95× that of vLLM inference. The code, Lean formal proofs, and protocol documentation are all open-sourced under Apache 2.0.

Still a research prototype—for now, it mainly supports MoE, greedy integer decoding, and contexts of about 1,024 tokens; the architecture itself isn’t hidden. But bringing zk from the blockchain layer to model outputs suddenly makes use cases like auditing and preventing model swaps during routing feel much more practical.

#zkSNARK #AI #可验证推理 #Anthropic
30D trade $ZEC 273.8 USDT
🛡️ $ZEC /USDT — 🔴 SHORT · Conf 85% ━━━━━━━━━━━━━━ 📍 Entry: 758.13 – 762.69 🛑 SL: 817.91 🎯 TP1: 722.39 ✅ TP2: 691.97 🏆 TP3: 646.35 ━━━━━━━━━━━━━━ 🔎 Clean lower-high staircase from 960.00 on the 1H, now pressing the 752.57 low. 🧠 $ZEC : Privacy coin using zk-SNARK shielded transfers; a holder vote on pulling 50,000 ZEC from the ZCG dev fund runs to Nov 30; risk is ETF outflows plus the Monitoring Tag. 💬 Does the ZCG vote matter more than ETF outflows for the next leg? ⚠️ Not financial advice. Manage your risk. #ZEC #zcash #Privacy #zkSnark #ZCG {future}(ZECUSDT)
🛡️ $ZEC /USDT — 🔴 SHORT · Conf 85%
━━━━━━━━━━━━━━

📍 Entry: 758.13 – 762.69

🛑 SL: 817.91

🎯 TP1: 722.39
✅ TP2: 691.97
🏆 TP3: 646.35

━━━━━━━━━━━━━━

🔎 Clean lower-high staircase from 960.00 on the 1H, now pressing the 752.57 low.

🧠 $ZEC : Privacy coin using zk-SNARK shielded transfers; a holder vote on pulling 50,000 ZEC from the ZCG dev fund runs to Nov 30; risk is ETF outflows plus the Monitoring Tag.

💬 Does the ZCG vote matter more than ETF outflows for the next leg?

⚠️ Not financial advice. Manage your risk.

#ZEC #zcash #Privacy #zkSnark #ZCG
Log in to explore more content
Join global crypto users on Binance Square
⚡️ Get latest and useful information about crypto.
💬 Trusted by the world’s largest crypto exchange.
👍 Discover real insights from verified creators.
Email / Phone number