Michael Horrell

PhD in Statistics
Views are my own

Working on

[Michael Horrell @horrellmt]

As a dad of a 4 week old and a 4 year old, codex remote is very cool.

Hear about a new feature during some down time. Dash off a few paragraphs of text to codex from my phone. Come back later to a completed experiment.

[Michael Horrell @horrellmt]

XGBoost's new vector leaves seem pretty great. When working on Wideboost (succeeded by GBNet), I had devised a test to explicitly eval models in a multi-output-with-correlated-outputs scenario. Vector leaves do great here.

Re-doing that test, vector leaves performed best with a much smaller model. A great addition to XGB codebase!

[Michael Horrell @horrellmt]

Linking the XGBoost post on this: https://t.co/eMSX59B2mi

[Michael Horrell @horrellmt]

XGBoost's new vector leaves seem pretty great. When working on Wideboost (succeeded by GBNet), I had devised a test to explicitly eval models in a multi-output-with-correlated-outputs scenario. Vector leaves do great here.

Re-doing that test, vector leaves performed best with a much smaller model. A great addition to XGB codebase!

[Michael Horrell @horrellmt]

@GergelyOrosz Before LLMs, thinking, coding and verifying happened simultaneously. Different types of feedback would reverberate across all three processes.

Now, LLMs have removed coding. Thinking to verifying has a larger gap with less reverberation.

[Michael Horrell @horrellmt]

@NielsRogge Codex remote is so good

[Michael Horrell @horrellmt]

Claude: I only work in this folder

Also Claude: you mind if I `cd` over here?

[Michael Horrell @horrellmt]

Ordered a mac mini for phone + remote codex. If this really works I will spend less time on X.

[Michael Horrell @horrellmt]

@mitchellh For stuff I'm not familiar with, getting the median result from an llm is a real super power.

But it's still a median result.

[Michael Horrell @horrellmt]

My quibble with the "Bitter Lesson" is that it isn't very prescriptive.

I may as well say something like -- LLMs will get replaced by something that scales better.

Exactly what that thing is is left as an exercise for the reader.

[Michael Horrell @horrellmt]

I've been learning more about the Cox PH Partial Likelihood. It is really really clever.

It avoids usual survival-data pitfalls. It asks: given this event and the candidate units, which unit most probably saw an event?

It's a multinomial calculation. It uses softmax.

[Michael Horrell @horrellmt]

@Noahpinion Driving costs down creates abundance

[Michael Horrell @horrellmt]

@mattyglesias Vibecoding kind of gives just another version of this xkcd
https://t.co/uzLncabuja

[Michael Horrell @horrellmt]

Purely as a reflex, I turned down the "Upgrade to Opus 4.7 or whatever" pop-up today.

The good: 4.6 or whatever must be good enough for me.

The bad: (a) too many new models (b) I guess I don't expect any step changes (eg I bet it still try-excepts imports) so why upgrade?

[Michael Horrell @horrellmt]

BTW, the big delete here was just removing catboost from gbnet 😿. There's something I don't understand about their algorithm. It seemingly does NOT admit performant, single round manual boosting updates.

[Michael Horrell @horrellmt]

Over the weekend, I updated GBNet to correctly work with the latest versions of XGBoost, etc.

I did a side-by-side manual fix vs. LLM fix.

Manual fix all in: +28 (lines added) -244 (lines removed)
LLM fix part 1: +130 (added) -17 (removed)

Yikes!

[Michael Horrell @horrellmt]

@JFPuget Almost all my prompts to the LLMs tell them to avoid try excepts now.

It's orders of magnitude harder to figure out if they've done something silly (or if I have) if the code never fails.

[Michael Horrell @horrellmt]

@patio11 I once advised someone to let something fail to get the right attention. Instead, said individual put the team on their back and delivered a major net new revenue source for the company.

I don't think I was necessarily wrong, but you can't argue with results.

[Michael Horrell @horrellmt]

Exactly what the LLMs need to understand also. Let it break. Don't mask failures.

[Yaroslav Bulatov @yaroslavvb]

Google has an internal "let it break" essay about a hero engineer whose hard work ends up being a net negative (by masking the underlying issues). My manager sent me that essay when I was trying too hard to get the collective TensorFlow unit test suite green.

[Michael Horrell @horrellmt]

Over the weekend, I updated GBNet to correctly work with the latest versions of XGBoost, etc.

I did a side-by-side manual fix vs. LLM fix.

Manual fix all in: +28 (lines added) -244 (lines removed)
LLM fix part 1: +130 (added) -17 (removed)

Yikes!

[Michael Horrell @horrellmt]

@staysaasy Getting the median solution to something new to you is a great thing.
https://t.co/Esa6hE6bcS

[Michael Horrell @horrellmt]

LLMs are criticized for not getting beyond conventional responses, but conventional responses are underrated.

- What does a typical X team look like?
- What’s a standard approach to this problem?
- Make this email/tweet sound normal.

You get baseline quality for free.

[Michael Horrell @horrellmt]

What if platforms were not end-to-end, but instead stopped rather abruptly? 🤔