Alignment
The newest safety and alignment documents
What was actually published on making these systems safe, newest first. Research, evaluations, incident write-ups, policy documents and the arguments about them.
01Newest first
- Commentary17 Nov, 8:42 AM CST
Jensen Huang Really Liked Selling China Advanced AI Chips
Zvi Mowshowitz
- Commentary13 Nov, 6:21 AM CST
Jensen Huang Resigns After Having Another Affair In Addition to the Chinese Spy
Zvi Mowshowitz
- Commentary12 Nov, 2:16 PM CST
Looping a transformer increases no-CoT capabilities more than general capabilities
Alignment Forum
- Reported11 Nov, 12:00 PM CST
Hugging Face appears to be hosting a load of models that could be used for cyber attacks
Transformer
- Commentary11 Nov, 3:49 AM CST
How I'm thinking about different types of automated alignment
Alignment Forum
- Commentary10 Nov, 10:30 AM CST
Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All
Alignment Forum
- Commentary8 Nov, 11:15 AM CST
Ajeya: AI is dramatically accelerating AI research right now
Alignment Forum
- Commentary5 Nov, 1:06 PM CST
METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026
Alignment Forum
- Reported27 Oct, 11:02 AM CDT
Will Anthropic’s post-IPO governance really protect its safety mission?
Transformer
- Commentary26 Oct, 7:31 AM CDT
Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI means
Import AI
- Primary9 Jun, 7:00 AM CDT
NIST
- Commentary18 Feb, 5:25 PM CST
After Orthogonality: Virtue-Ethical Agency and AI Alignment
The Gradient
- Primary17 Feb, 6:00 AM CST
Announcing the "AI Agent Standards Initiative" for Interoperable and Secure Innovation
NIST
- Primary10 Feb, 6:00 AM CST
NIST
02How a document gets onto this page
Two ways, and the difference is worth stating because one of them is more reliable than the other.
It came from a source whose whole subject is this. That is a register lookup, not a judgement, and it accounts for 37 of the 41 below:
- aisnakeoil · sceptical analysis of AI claims
- alignmentforum · technical safety research, first-party
- nist · US government standards and evaluation
- thezvi · the most thorough weekly safety roundup published
- transformernews · AI policy and safety news deck
And every link here went through the reader's door first, the same as every article link on this site: 31 document(s) that would otherwise be on this page were dropped because a reader could not open them.
Or the deck's classifier labelled it a safety story. That axis measured 36.1% against a sample it was never tuned on, well under the 70% floor this site requires before a label may be printed anywhere. It is used here regardless, and deliberately: this page selects with it, it does not label with it. Showing you an extra document costs you nothing; printing a category on one would be a claim.
And it is not trusted alone. A story selected that way must ALSO carry an alignment term in its own headline, checked against the same register the classifier uses. One weak signal put a story about posters on this page; two signals is what removed it. That pairing brought in 4 of the 41, and turned away another 51 the classifier alone would have included. The whole table is on the ledger.