Shriram Krishnamurthi

One of the nice things about having a course of 20 students producing code to the same tasks means I can start to build up some "distribution" in my head of what it produces. Of course, those intuitions don't apply with "swarms" or whatever. C'est la vie.

The one thing I can say with conviction is that most of my old intuitions, carefully honed over 40 years of programming, don't really work any longer. I'm hard at work building new intuitions. But "what will Claude Code do in response to this request" is not something I can reliably say. ↵

Whereas a lot of the code we write is actually rather boring and patterned and repetitious and similar things we've written before. A machine that is good at doing that is rather well suited for what we need an awful lot of the time. ↵

Left to itself, 2+2 means you could get any number of answers. But given a checker, 2+2 won't give you anything other than 4. That makes a huge difference.

But also, I think arithmetic is a bad example: we already have great programs for doing that. ↵

Well, not like I'm any authority on this topic… The problem is it'll take far more than a few skeets. The real thing to me is not LLMs but agentic programming. LLMs have all kinds of weaknesses, but wrapping them in an agent that can react to feedback makes all the difference. ↵

Maybe this is how I'll get into using jj?

Were they also founded as a cricket club initially, the way some of the other Italian clubs were?

The AI influencer bros (mostly bros) are definitely a pox on the landscape — useful about 20% of the time and garbage about 80%, kinda' like GPT-1 or something.

You don't have to hate to tell me, I have no vested interest here. I'm not talking about what others said, I'm talking about my personal experience/view. To my mind things only changed in ~Nov 2025.

Wow — Ethan WAS the Valley Breeze, no?

For some reason, my university's Clery Act training module has the incorrect Scrabble tile values on their stock filler image. C and P are both worth *3* points, not 4!

To literally having a statue built to you in the heart of London — something that didn't happen to any of the nepos.

(The statue was total news to me — saw it last week in Leicester Square. It's a commemoration of the 30th anniversary of DDLJ, and was unveiled in Dec 2025.)

It really does!

Simply outstanding product. And when I had a question and wrote to their email, within a day the company's owner wrote me back. I wrote in English, and he replied in German. Totally excellent Germanic Mittelstand stuff (it's based in Süd Tyrol).

It does not "hallucinate the wrong answer 20% of the time", in my experience. What is happening is far more subtle than that. Of course, I don't know what you're using or how you're using it.

I don't find those kinds of threads particularly educational. They tell me more about the writer than about the topic, so unless I'm deeply invested in understanding the writer… 🤷

Until a few weeks ago, it wasn't "transformative for coding" in the slightest. What does one make of *that*?

Indeed, a question I ask myself with increasing frequency is "What would Grace Hopper say?" I think she'd actually be very positive!

I think you misunderstood what I'm trying to do.

A new way to extract detailed transcripts from Claude Code I’ve released claude-code-transcripts, a new Python CLI tool for converting Claude Code transcripts to detailed HTML pages that provide a better interface for understanding what Claude Code has done t...

We've been using
simonwillison.net/2025/Dec/25/...

One of the big problems of genAI coding is it's given Steve even more time to post!

U.S. troops who died in Iran War to be laid to rest in the Tomb of the Unknown Strategic Objectives.

Maybe just reading less Steve "me, I'm the problem, but oh, I'm just going to continue to profit from it while slamming myself in the chest in public" Yegge would be a start…

Also, that could have been written in 3 paras.

I'm very excited by them, but also feel we're creating entirely new *kinds* of messes in our codebases that won't shake out for another 2-3 years. When they do, we'll be unprepared for them because they'll be a whole new *kind* of software quality issue (not quite like those created by humans).

Product in red

I hate the janky slippers that hotels, etc. provide. They aren't comfortable, don't always work well with socks, generate a lot of waste, sometimes slip, etc. For years I've looked for something equally light+compact but w/out these issues. These are amazing!
www.orthopant.com/collections/...

I'm already doing a bit of — indeed, growing amounts of — optimization in my accelerated intro. But that's a very specialized audience of students who have a lot of background and place in. It's challenging to get true novices to that point. But maybe it will be a lot less challenging!

It's exciting to me that we can get relative novice students to appreciate the need for external solvers, and also make use of them. It really is amazing how much CS one can do when "generating code" is no longer the dominant constraint. ↵

Pascal, I updated the README to make clearer the learning objective. You may be amused/pleased to hear that we're working towards rule engines. Eventually, of course, there will be tasks that motivate combinatorial optimization solvers… This is the first step in that direction. ↵

No, I don't. Kathi ran this while I was in the UK, so I don't really have any of the resources.

Yep. I say that one should be happy to get a segfault in a C program — at least you know something went wrong, instead of just getting, oh, 17.

Thanks for the note about comments. I added that now.

CSV: yeah! I had left that as open in my review:

> There are all kinds of issues w/ spinning your own (can you list them?).