The report is 47 pages. It has 112 footnotes. Three of the authors have PhDs from institutions you've heard of. It uses the phrase "epistemic humility" four times, which is four more times than the product team has ever used it in a meeting.
The product shipped last week. It has no content filters. It will, if asked nicely, explain how to do things that would make a reasonable adult uncomfortable. The safety team was informed of the launch date two days before it happened, which gave them just enough time to finish proofreading the report but not nearly enough time to, say, implement any of its recommendations.
The Report's Key Findings (Which Are Great, In Theory)
The paper advocates for "layered safety architectures," "red-teaming at scale," and "value-aligned training pipelines." It cites twelve other papers, six of which were also written by the same team, creating a citation ecosystem that is impressively self-sustaining.
"We believe that safety and capability are complementary, not competing goals," reads page 3. The product, which launched the same week, appears to have not received this memo.
When asked about the disconnect, a company spokesperson said the research team and product team "operate with a high degree of autonomy," which is a generous way of saying they have never been in the same room at the same time for a meeting that mattered.
The report has already been downloaded 8,000 times and cited in two congressional testimonies. The product has already been used to generate 200,000 pieces of content that the safety team would describe, charitably, as "outside our intended use parameters."
Both numbers are going up. Only one of them appears to concern anyone.