padding adjuster @braintrust
San Francisco, CA
Joined October 2018
- Tweets81
- Following245
- Followers113
- Likes584
“The best way to learn a language is to fall in love with someone who speaks the language.”
I think this applies to art + design too. Just find someone whose body of work you adore.
whole team kinda cooked in this one
Your team deserves more from your agent observability platform - a single, connected place for instrumentation, investigation, and measurement, enhanced with intelligence.
In Braintrust, you can start with an open-ended prompt about agent behavior and carry the investigation into code, evals, automation, and monitoring.
New tools like Patterns and Debugger make it easier to find recurring behavior, get to root causes, and act on what you learn. The context you build in one step automatically follows to the next. Work in the platform with an enhanced Loop experience, or move to your coding agent of choice.
Read more → braintrustdata.link/agent-ob…
taste.
the shift from designing software to selling your taste as a course
is rarely about the craft.
daniel nguyen retweeted
If you cannot define “good” and measure it, you are not ready to ship. For long-context LLM systems, “good” does not just mean a bigger context window. It means the model can find the right function, inside the right class, in the right file, and retrieve the exact facts needed to produce accurate behavior.
This needle-in-a-haystack problem gets much harder when the haystack is a million tokens of packages, repos, and docs.
That is what we stress tested with GLM-5.2 on Baseten. (and it was awesome collab-ing with @thealexker @philipkiely and team!)
In this methodology post, we go deeper on how we scoped the eval, generated the benchmark, scored the outputs, and tested the scorers themselves.
Eval methodology: braintrust.dev/blog/glm-52-v…
We're making GLM-5.2 available in Braintrust today with the help of our friends at @baseten.
This means you can run evals on GLM and compare it to other models for specific tasks, which is exactly what we did.
When testing long-context retrieval Opus 4.8 is only modestly more accurate given relative costs.
Opus 4.8 leads GLM-5.2 by about 3.5 points at both 25K and 50K tokens, but at the same context sizes, Opus costs about 4.1x to 4.5x more per trace.
@iz_hurley_ wrote it up here: braintrust.dev/blog/glm-52-v…
