Yes! We need to maximize information gain from experiments per unit of time and unit of compute.

Lots of funny banter in the comments, but worth considering how frontier labs measure this. Many consider code AI-generated if a dev used autocomplete or if the AI added a test or comment. Measurement matters; and can be misleading.

In the past few months, I’ve talked to too many data scientists and engineers who use codegen tools for everything BUT wrangling and analyzing data. They opine that codegen tools just don’t know how to “look at their data.” No longer. Enter Notebook Agent.

Damn proud of the teams I back. Ari and the Datologists have integrity. That matters and will undoubtedly be key to their continued success.

I strongly believe we will see a new wave of AI apps born of technical insights (probably related to training strategies or engineering for scale) that displace incumbents by eliciting or suppressing behaviors that unlock new user experiences. Not thin wrappers; real innovation.

So many VCs talking about consensus; so few have even heard of Raft…

Never thought Die Nacht ist Leben would displace CLRS as the most popular algorithms book…but Listen Labs had other plans amazon.de/-/en/Die-Nacht…

Data quality, evals, agent UX and more. This episode is killer.

When people ask me what makes Runway different, I encourage them to visit the office…see the art and photography on the walls, listen to the communal playlist the team has curated over the years, talk to the team about their favorite shorts from the last film festival…they

Enjoying jj so far. Tired of pretending git doesn't kind of suck.

Im fairly convinced that most people on here don’t really understand what evals are…which is ok…because it is a pretty nebulous concept. But we should align on what it means and why we do it before waxing poetic on it.