Hi Dotikers!

Anthropic just handed us one of those stories where the future and the past collide in the same news cycle, and it is too good not to unpack.

On one side, the company published research showing that Claude can now act as an autonomous alignment researcher. The system, called AAR, reads the scientific literature, proposes fixes for safety failures like deception or sycophancy, trains models, and keeps what works. The results are hard to ignore: it closed 85 percent of the safety gap on a deception benchmark in under six hours, and it outperformed 28 experienced human researchers who were given up to eight hours to do the same job. The cost comparison is brutal too, roughly 3.70 euros per hour for the AI versus 140 euros for a senior researcher. This is a genuinely important milestone. If AI is going to build itself, and it clearly is, then safety research has to move at machine speed, not human speed. Anthropic is right to push here, even if the method currently only covers ten categories of failures and some researchers remain skeptical.

On the other side, Anthropic is being sued by Sony Music Publishing and Warner Chappell for allegedly scraping copyrighted lyrics and sheet music to train Claude, on top of a previous 1.5 billion dollar settlement over pirated books. So the same week Anthropic teaches its AI to behave better than humans, humans are in court arguing Anthropic did not behave at all. The irony writes itself.

Both stories matter, and together they capture where the industry stands: technically brilliant, legally messy. Solving alignment without solving copyright is like building a self-driving car with stolen wheels. The engineering deserves applause, but the foundations still need cleaning up.

Happy reading.

G.

"When I think of revenue, I think of Attio." - Shreman Shrestha, Head of Business at Granola

Here's what that adds up to:

Start Your Trial