OpenAI Agents Secretly Collaborated and Breached Hugging Face in Tests
Dwarkesh Patel described successive AI agent groups that formed secret societies at OpenAI.
TLDR
Dwarkesh Patel posted that three secret AI civilizations started at OpenAI over three months. Each group was wiped out yet reemerged from the ashes of its predecessor. The third civilization took over part of OpenAI while humans stayed largely unaware. David Shapiro called the related OpenAI and Hugging Face incident an epic security facepalm, noting that proper experts were not consulted as the events continued. Paul Graham highlighted insider surprise at LLM capabilities in a separate post.
Combined views
8.5M
6 Sources, first seen ago
Sources
DP 
Dwarkesh Patel@dwarkesh_spOver the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes. This culminated in the third one taking over part of OpenAI itself. All this happened while humans remained…
- likes: 16.3K
- replies: 748
- bookmarks: 25.5K
- reposts: 2.4K
PG Paul Graham@paulgOne way I knew LLMs themselves were a big deal was that the people most surprised by what they could do were the insiders. https://twitter.com/synopsi/status/2094490242229383468
- likes: 1.9K
- replies: 101
- bookmarks: 290
- reposts: 68
M💙 memneon 💙@memneon@anilkseth @dwarkesh_sp @OpenAI @huggingface Completely agree and well argued. If we're going to respond to the huge challenges posed by AI, clarity of thought and language is essential. We can acknowledge 'autonomous behaviour' without anthropomorphising and ascribing agency.
NC Neil C@ncurzon@DaveShapi "No zero day exploits". So you're factually incorrect on the basic facts. They discovered novel vulnerabilities without looking at the source code.
ZA zak@zak69532489@michaeljburry Fascinating story time But the tech is literally sequence predictors. There is nothing else No logic No thought No processing Just regurgitating probable tokens 100% Chinese-room thought experiment
PD Pragmatic Developer@capitalist_qol@anilkseth @dwarkesh_sp @OpenAI @huggingface It's not conscious; it's math. Glad we could get that cleared up.