Research · Satire
Anthropic Releases 84-Page Safety Paper That Only Anthropic Employees Will Read
The paper, titled 'Towards Provably Beneficial Artificial General Intelligence,' outlines a rigorous theoretical framework for ensuring AI systems are aligned with human values.
SAN FRANCISCO — AI safety company Anthropic has published a new research paper detailing its approach to ensuring that advanced AI systems remain beneficial to humanity, the company announced on its blog alongside a brief summary written for people who will not read the paper.
The paper is 84 pages long and contains 312 footnotes.
“We believe transparency is essential,” said a company spokesperson, standing next to a stack of printed copies that would never be touched.
The research introduces a new concept called “Constitutional Recursive Reward Modeling with Adversarial Preference Elicitation,” which the authors describe as “a step toward solving the alignment problem” and which three people in the world are qualified to evaluate.
On social media, the paper received approximately 4,000 likes and 23 quote posts from people who had read the abstract.
“Great paper,” wrote one machine learning researcher, who had read the title.
“Important work,” wrote a second researcher, who had read the abstract and one figure caption.
The paper concludes that more research is needed.
At press time, Anthropic had announced it was hiring.
