the OG duck... nothing is more inspiring than founding engineers who're just plain crazy

Why does DX not rise to AWS' broader bar on customer obsession? (assuming we agree that their DX leaves a lot to be desired) the only thing i could come up with is their desire to iterate fast. If customers care about it, they'll improve; but if customers can't articulate...

Anu Sharma Anyway, changing both infra and dev workflow at the same time means (imo) picking a narrower audience and having a strong point of view of what it should look like. I think that's hard if you're a very large cloud provider trying to build fairly general-purpose systems!

...and just like that we stopped talking about number of <billion> parameters in AI models except how we predict loss/pass rate at some future notional scale. does the scale "moat" hold any longer?

have to say all the huffing & puffing on alignment reduces the most valuable data point in applying models for automation... calibration/confidence. calibration drops materially from the pre-trained model. does every case need this level of alignment at the cost of confidence?

my dad once said to me, ‘how can you be depressed? you’re tall.’

‘We want AI agents that can discover like we can, not which contain what we have discovered.’ The Bitter Lesson, Rich Sutton, 2019

confession: vide coding feels a lot like drunk coding, unbridled eagerness mixed with brain farts that don't take the basic/most obvious context into account meanwhile, my husband, the sane one, is left cleaning up after me...

two years is a long time to learn a new skill, right? but it just occurred to me that ‘think step by step’ is an instruction meant for me to make the model do stuff, not for the model… 🤦‍♀️

after i got oauth working with an Streamable HTTP MCP Server and Client, i seem to have landed in a world with no dimensions, total imagination zone… like what if you had a ‘thinking’ calculator, what could you think up if you had 200 slightly smarter versions of you?