David Gleich

Thu, 10 Sep 2026 10:23:04 -0400

Is anyone else surprised Open AI can’t answer about the impact of user-info for training their LLM? It seems like searching for “the forcing function is smooth option c and d in Fefferman” (or simple variations) in user-derived training data is a feasible task. I can’t imagine it is that common.

Thu, 10 Sep 2026 09:48:22 -0400

Does anyone else seem like they can feel the impact of rope in llm writing? It loves to use a lot of words to say little.

Tue, 08 Sep 2026 10:05:34 -0400

I tried to use Fable 5.1 to review a commit. It blew through my entire usage in 10 minutes… (Anthropic Max plan…) it might as well not exist because I can’t use it even to do simple stuff.

Wed, 02 Sep 2026 15:43:39 -0400

Does anyone else find LLM generated text hard / impossible to edit? It’s like it throws words and jargon into work salads that “look right” but have no flavor / meaning. Picking them apart leaves you with nothing left. It’s the opposite of “write as simply and directly as possible”

Thu, 27 Aug 2026 08:49:29 -0400

Why am I getting tons of emails from people trying to promote my github repositories?? Is there some market here now?

Tue, 25 Aug 2026 13:42:57 -0400

In the past few days I’ve gotten job offers/inquiries for… dock supervisor in North Dakota, anesthesiologist in Colorado, cloud engineer in the bay area. Sadly, none of them are in my wheelhouse.

Mon, 24 Aug 2026 12:34:35 -0400

Prediction. First llm proof/progress that p != np within 6 months!

Wed, 19 Aug 2026 09:32:18 -0400

I’ve talked to multiple people about using LLMs in reviews. If the journal is going to systematically run an LLM to provide feedback on the paper, than that LLM output needs to be available to be run before submission. Why not give authors a chance to fix it before submitting?

Tue, 18 Aug 2026 20:39:59 -0400

Did Anthropic change something with usage limits on Claude? I’ve constantly been hitting them today. But essentially never hit them before!

Tue, 18 Aug 2026 10:22:29 -0400

Sometimes Claude wants to over-engineer stuff that just needs iterative feedback. I mean, a use case of LLMs is to build it, and iterate faster with things that work, not try and spec the thing up from the ground. If it gets it wrong, it’s fine if it just has to start over.

Tue, 18 Aug 2026 09:34:47 -0400

Can we pass a law that LLM can’t use “I”? There is no “I” in an LLM. There is no person.

Wed, 12 Aug 2026 08:27:21 -0400

I love people who email a picture of a QR code instead of GIVING A LINK! I hope cyber security scanners scan all pictures for QR codes and validate those links!

Mon, 20 Jul 2026 09:02:07 -0400

Claude (really Anthropic’s designers, who I suspect have encoded lots of color schemes and designs in the “skills”) has replaced PowerPoint and Beamer templates with … a different set of annoying but “cool looking” ways to organize text on a page.

Fri, 10 Jul 2026 12:17:16 -0400

I’ve found myself wanting to denote regions of human and LLM content in documents recently. Here’s a quick proposal for an easy-to-use way of doing this for Markdown and HTML.

Tue, 07 Jul 2026 13:00:28 -0400

Just realized that the kids were on to something … 6*7 =

Thu, 02 Jul 2026 15:35:58 -0400

Congrats to my brother-in-law Rob Lea, who is the first human to have ever done all seven summits (highest peak on each continent) and all seven swims (20+ miles channel swims on each continent). Rob’s on instagram.

Wed, 01 Jul 2026 13:26:54 -0400

Anyone know why Claude often uses British spelling vs. American? I’ve seen a few words that get this treatment … recognise is the one that made me post this. Can’t remember the other one.

Thu, 25 Jun 2026 10:49:22 -0400

Weird piece of email spam I just got. “你好! 每周至少使用1次牙线，与较低的中风风险有关，特别是由血栓和心律不整引发的中风。”, which translates to “Using dental floss at least once per week is associated with a lower risk of stroke, especially strokes caused by blood clots and irregular heart rhythm….” Along with a link. I didn’t click.

Tue, 23 Jun 2026 22:10:13 -0400

Well, this picture explains a lot!

Mon, 22 Jun 2026 12:29:24 -0400

Google’s Gemma models are funny. I had to convince a local model it wasn’t running on a server before it would try and access something running at 127.0.0.1. it was ‘convinced’ that I was incorrect and kept trying to tell me things that were very untrue.

Sat, 20 Jun 2026 16:16:13 -0400

Just had to remove and reinstall Apple Maps because it was sucking down 5gb of space. No “offline maps”.

Fri, 19 Jun 2026 10:18:00 -0400

From Daring Fireball …

Austin Powers, by the way, came out in 1997. In the film, Powers was frozen since 1967. That means next year, we’ll be as far removed from the debut of the movie as unfrozen Austin Powers was from the groovy 60s in the film.

I believe I’m now old :(

Tue, 09 Jun 2026 14:29:06 -0400

First question to fable… designs new glyph for a font! … demoted to opus for cybersecurity reasons. I guess I need a font design to exfiltrate data to understand this issue.

Tue, 26 May 2026 10:41:55 -0400

Is it me or did Claude get really bad in the last few days?

Mon, 11 May 2026 13:07:55 -0400

The Midwest Machine Learning Symposium (MMLS) 2026 will happen at Purdue!

📍 West Lafayette, IN 📅 June 24–25 📌 Poster deadline: May 24 🔗 midwest-ml.org/2026/

We awesome plenary speakers: Tong Zhang (UIUC), Jennifer Neville (Purdue), Mohit Bansal (UNC), Joyce Chai (Umich).

Join us!