David Sacks debunks Anthropic's small-model self-replication warning: experiments don't hold up

DavidSacks · x · 2026-10-14

AI czar David Sacks published a lengthy rebuttal to Anthropic's warning that small models could soon self-replicate across cheap IoT devices:

Weak experiments: a 26B model across 15 $55 TV boxes at 0.39 tokens/sec, emulated devices with deliberately open ADB ports making 'hacking' trivial, the capable model often on a separate server, and agents explicitly instructed to propagate.
Not novel: Mirai (2016) built massive DDoS botnets without AI; Kimwolf has already infected 2M+ cheap Android TV boxes via open ADB ports.
The bottleneck is never intelligence — it's unpatched devices with default passwords. A practical attacker would use strong models at design time, then ship a lean worm binary.
Timing matters: Coinbase just fine-tuned Qwen-3.5-9B on fraud data, beating Opus 4.5 while faster and cheaper (30% fraud reduction) — a direct threat to frontier labs' business model, Sacks argues, calling it a scare campaign; smaller specialized models are also the safer trajectory for AI safety.

More from Models