Incredible piece. I felt personally attacked. It’s perfect
New blog post - Chasing the Karpathy Peak (and Falling Off Slop Cliff): Confessions from Slopaholics Anonymous
My story of becoming a slopaholic, who’s most at risk, and how to avoid it (or recover)
thomasjiralerspong.substack.…
Clayton Kim retweeted
We are hiring a Senior Clinical Specialist, AI Systems at @function - jobs.gem.com/function-health…
Can confirm. A little wonky since @muse webview doesn’t have Bluetooth enabled
The sheer craft and information density of all of his posts are astounding
New post 📝
For this article, I asked myself: what happens when motion becomes an input to our shaders?
To answer it, I spent this summer experimenting with effects like motion masking, blob tracking, optical flow, and motion blur
blog.maximeheckel.com/posts/…
Holy shit he's right. Incidentally a great signal for software engineering skill too
Let's all agree that this is the correct and final eval for agi
jackhopkins.github.io/factor…
The timing and variability of the rewards basically match video poker
It also revealed that a large number of people at companies are truly useless
Been funny watching people slowly realize these last six months that other peoples jobs are actually harder than they imagined.
AI had everyone in March saying they could do it all.
Then a bunch of PM learned about productionization of software.
Engineers learned it’s nontrivial to get customers to like and use a UI.
Designers learned that both PM and Engineering jobs are even easier than they imagined, but the amount of type A lunatics you need to argue with to get anything done collaboratively is insane.
And now everyone is back in their own lane again.
Evals are all you need
The reason enterprises struggle to go beyond basic chat bots is the talent gap to build harnesses and evals.
1. Evals: do you clearly understand your use cases and can you replicate that in the form of offline and online evals. Do the evals express your ambition and do they push the jagged frontier of the models? Do they help you pick the right models on the quality-cost-latency curve?
2. Harness: do you have a system that manages routing, multiagency orchestration, context management, tool calling, memory that is independent of the models?
3. Talent: do you have the talent to build this all out on the frontier?
This is the scarcest piece.
Everyone’s talking about using Fable or Sol as a planner, then farming work out to cheaper models. But who wants to have to specify each model for each task, including the verifiers?
I wrapped the pattern into a reusable skill:
github.com/theclaymethod/min…




























