Claude 4 Is Here (I Helped a Little)

july 15th: we announced claude 4.

and for the first time in my career: some of my work contributed to this.

the release

claude 4 opus and claude 4 sonnet. the next generation. major improvements across the board.

i'm not going to pretend i can objectively evaluate it (obviously biased), but the benchmarks speak for themselves. and the vibes are good.

i work on alignment science. the safety research that ensures everything works safely.

my specific contributions: i can't detail publicly. but some of the systems that helped train this model include code i wrote.

that's significant to me. my first real fingerprint on a major release.

complicated.

pride: i contributed to something millions of people will use.

humility: my contribution was small. many people did much more.

surreal: the model that came out is so much more than any individual part.

training a frontier model is a team effort of hundreds. my piece matters, but so does everyone else's.

what claude 4 can do:

the improvements aren't marginal. this feels like a genuine step up.

what i appreciate about how we work:

we didn't rush this. there were extensive safety evaluations. internal testing. alignment work.

the model is capable, but it's also aligned. those aren't opposites—they're priorities we hold simultaneously.

twitter (x) exploded with reactions. some positive, some comparing to competitors.

from the inside, you see all the work. from the outside, you see the product. both views are valid but different.

this isn't the end. it's a milestone.

next: claude 4.5? new features? things i can't talk about?

the work continues. always continues.

a year ago i was a student submitting my thesis.

now my code is part of a model that's in headlines.

how did this happen. life is strange.

called my parents to say "i helped build something you'll hear about." they were proud. i'm still processing.