BREAKING
OpenAI Reveals Alignment Research
0
/53
evals improved
0
domains trained
0
%
GPT-4 disallowed cut
How the method works
1
RL on real talks
↓
2
12 domains
↓
3
Beneficial traits
Health training aids other domains
Strengths vs open questions
Reported gains
●
Resists adversarial prompts
●
Keeps helpful responsiveness
●
Early fine-tuning resistance
Still unknown
●
No model names or versions
●
No paper confirmed
●
Real-world behavior unclear
Research stage, not a product
AI NEWS BLITZ
On June 18, OpenAI disclosed new research on keeping AI beneficial beyond its training data.