契約したモデルが本当に動いているのか、利用者は事業者を信じるしかありませんでした。AnthropicのFlashProveは、zkSNARKで重みを明かさずに応答がコミット済みモデル由来だと証明する、Apache 2.0のプロトタイプです。Llama 4 Maverickを1基のH100で110トークン/秒、証明は19MBとのこと。個人的には、信頼を評判ではなく検証可能な形に置き換える方向性が重要だと思います。実運用にはまだ制約があるようなので、引き続き動向を追いたいと思います。 #FlashProve #zkSNARK #推論検証 #LLM
Related
New post on Fahhim: "Arabic breaks most prompt templates — here is the structure that survives"https://fahhim.moteklab.c...
New post on Fahhim: "Arabic breaks most prompt templates — here is the structure that survives"https://fahhim.moteklab.com/blog#ArabicAI #PromptEngineering #EgyptianDialect #LLM #C...
One of the product aspects I dislike about AI is the fact it seems to think I don't know how to use things.Case in point...
One of the product aspects I dislike about AI is the fact it seems to think I don't know how to use things.Case in point; Slackbot this morning "yelling" at me about how it could c...
Developers, let's settle this.What's more challenging these days?- Writing code manually- Debugging AI-generated code #P...
Developers, let's settle this.What's more challenging these days?- Writing code manually- Debugging AI-generated code #Programming #AI #SoftwareDevelopment