Tyler Burch

Lead Baseball Analyst @RedSox, recovering physicist. Formerly: @Argonne, @NIU, ATLAS Experiment @CERN, @MurrayStateUniv, STL(area) Native | Tired but hydrated

[Tyler Burch @TylerJBurch]

@drivelinekyle @cwhite23_ yeah sometimes you should be using AIC

[Lady Nimby @LadyNimby]

One thing that strikes you when you become a parent is that, like, everyone you’ve ever met with really little kids was having kind of a hard time

[Tyler Burch @TylerJBurch]

@vboykis I think agent-based workflows lend toward a LOT more context switching, and im realizing that is causing a lot of fog for me.

[Tyler Burch @TylerJBurch]

@brandon_nickell Just not too much

[Tyler Burch @TylerJBurch]

@growingwisdom RIP my garden
May 16, 2026 - May 19, 2026
Gone, never forgotten.

[Tyler Burch @TylerJBurch]

Grant me the serenity to accept the problems where I cannot offload my understanding, courage to let the LLM go burrrr on the problems where I can, and the wisdom to know the difference.

[Tyler Burch @TylerJBurch]

The more I do it, the more it feels like there's some misalignment between LLMs and statistical work. Maybe it's over-abstraction leading to making it harder to turn knobs and experiment? Maybe the harnesses are oriented toward shipping code and not sitting on a problem?

[Tyler Burch @TylerJBurch]

The better the models get, the more I seem to have to scream "DON'T FORGET YAGNI" and "come on, KISS!" at them.

[Gavin Brown @gavinrbrown1]

Gradient descent does not work. I will die on this hill.

[terminally onλine εngineer @tekbog]

it feels unholy to add agentic workflows into matlab

[Tyler Burch @TylerJBurch]

No company has ever been more relatable than Dunkin giving up on the “healthy” protein menu strategy and hard pivoting directly into putting Oreos on everything.

[Tyler Burch @TylerJBurch]

Every single “the AI deleted prod!” post I’ve seen could have very easily been avoided by thinking about the principle of least privilege even one time.

[Simon Willison @simonw]

The conclusions here feel wrong to me. The two lessons I see are:

1. Don't run agents anywhere they might be able to access production environment credentials - it's on you to know which credentials those are

2. Keep tested backups that are independent from your production host

[Tyler Burch @TylerJBurch]

Not me. I keep my hammer ready at all points in time.

[jon drake @DrakeGatsby]

Legally all doctors are allowed to hit you on the knee with a little hammer. Even PhDs. They just rarely exercise this right

[Tyler Burch @TylerJBurch]

Quick question, does gpt 5.5 still talk to me like it’s trying to prove to me that it’s read every single book on computer programming?

“We need to leverage a shim to reduce the risk surface area” got it

[Tyler Burch @TylerJBurch]

In case you were wondering how this performed, near-perfect temps meant no weather effect, leaving just the random-walk estimate.

Top-3 average was 2:02:30, ~0.7 posterior SDs faster than the 2:05:10 median, inside the 95% interval (1:57:55–2:12:29).

[Tyler Burch @TylerJBurch]

Based on the attire on the Green Line skewing more Hokas and On Cloudmonsters than a typical workday, I can confirm today is the Boston Marathon.

Just uploaded a new blog post looking at how much of finish-time variation is attributable to weather.

Link in thread

[Boston Globe Pitchbot @BostonSatire]

Report: Runner who tried to jump on Green Line to cheat way to victory in 2022 Boston Marathon just getting to finish line now

[Tyler Burch @TylerJBurch]

@pierreanalytics 👎👎👎

[Tyler Burch @TylerJBurch]

@StatsbyLopez There's actually pretty good availability! At least through 2019 easily available. See: https://t.co/tyFiTAyihM

I scoped it down to top-3 for the sake of grounding to existing work. and to save a bit of sanity for a pedagogical-leaning post. Definitely could be much richer.

[Tyler Burch @TylerJBurch]

Based on the attire on the Green Line skewing more Hokas and On Cloudmonsters than a typical workday, I can confirm today is the Boston Marathon.

Just uploaded a new blog post looking at how much of finish-time variation is attributable to weather.

Link in thread