Waiting for prompt-locked Codex gift cards.
Want that new feature, just send the dev a $50 codex gift-card.
So great to see @19thLabs models in friends `main` branches 💞
The latest is a teensy 600kb text scorer running locally in their iOS app.
This limited control over the J-space has a beautiful alignment with the best ways to guide children and train dogs.
Telling them to do something else, is more effective than telling them what not to do.
Both have immature inhibitory control. Both act with action-oriented processing.
tldr; yelling "why the f* do we keep making the same mistake?" is, in some ways asking the model to make the same mistake again.
It’s becoming increasingly common for our roadmap discussions to become “should this be a feature or a skill”
Subjective observations:
1. Codex does better work at standard speed
2. Opus 4.6 is still the best writing collaborator
Domain pricing is so opaque. We paid $4k for 19th.com, whereas an obscure 6-character name with no search volume the owner wanted $500k
Just bought our .com for $25k from a public NYSE-listed company, cold-calling its CEO.
> realize the domain name belonged to a public company (shit).
> cold-call all board members then CEO
> follow-up everyday
Myself and @spencerthurstn started @19thLabs.
19th is a tool that lets anyone build a model in <60m. Adding them to the list of things "you can just build".
We're opening the beta to more folks soon. Just signing up lets you experience my favorite detail on the launch site.
At Uber we both saw 1 MLE often spread across 3 PM's. All having to fight to get one of their 20 ML experiments prioritized that quarter.
19th lets more folks build models. In the same way Codex lets more people code. The codex moment for ML?
Thesis: The way we think about capturing product feedback is constrained by the way we had to process it.
Excited for a future of free-form voice note feedback from users that gets cleanly categorized, summarized, and "emotion scored" by LLMs
