⚡Megatron LLM Hub
←返回搜索
Twitter

复算 Anthropic AI 蠕虫博文:实测推理仅 0.38 tps,质疑其可行性

给出 AI 蠕虫威胁的量化反驳,避免被厂商叙事带偏。
ai_agentllm报告
RT @BlaiseBits: I went through the anthropic "IOT AI-Worm" blog post last night. This "sign of life" was 15 android TV modules running Gemma4-26B-A4B. This absolute powerhouse of inference ran at a earth shattering 0.38 tps.... They later extrapolate to what a AI Worm Swarm the size of the Kimwolf could generate inferance of roughly 80k tps. All that sounds oh so scary till you realize: 1) It's slow af 2) any node going offline kills the "worm agent" 3) network latency is a bizzich Possible, but not probable. This would be the world's SLOWEST replication worm as well as being hella easy to identify with, get this, cyber security controls we've had for decades. The report even estimates ~3 days for a replication spread on these IOT devices, meaning any node (Android TV Box) going offline BTFOs that attempt, or if it's rebooted that inference instance is dunzzos without persistence.... Which is likely gonna become a problem if a user can't even use their device because it's too busy with matrix multiplication and gets a hard reset, perminanlty removing it from the swarm. Anthropic did hedge some these thoughts in the blog, but that part is skimpy and burried by the FUD/Doomer crowd. https://t.co/0Gm63PgvP6