posts

This is a wildly helpful trick: I want to see all the revisions in which a particular line of code changed. In this case, I’m looking at how a property in WordPress’ package.json file has changed since the start of the year, but I don’t want to see every commit that impacted the entire package.json file. Thankfully, git supports showing the diffs and logs for parts of files over time. It does…

My mind may be playing games on me, but I feel like I’ve witnessed signs of crumbling software over the past few months. I’m leaving this list to see if it ends up being a big coincidence or whether a pattern develops. Multiple big internet services have had outages or multi-outage days, including GitHub and Amazon. GitHub’s UI has seen a consistent decline, particularly on Safari. Every few…

Today someone was discussing the goal of linting HTML, specifically of detecting unclosed attributes. Consider the following snippet: <p class="important><img src="alert.png">This is important!</p> It’s clear that a mistake led to a missing double-quote on the class attribute of the opening <p> tag. While WordPress’ HTML API doesn’t directly report this (because “unclosed attribute” isn’t…

This probably isn’t original, but I am declaring that a “slop in the face” is when someone returns a well-considered, thought-out, intentional interaction by pasting in an LLM-generated response, meeting care and diligence with contempt and disregard.

Frequently I want to compare a given git branch or other ref against its “target branch” because I’m reviewing a PR. If the work isn’t fully up-to-date then calling git diff trunk results in a large swath of changes I don’t want because it includes updates on trunk that aren’t in the feature branch. In the past I would try and be clever with commit ranges and syntax: git diff trunk..HEAD or git…

A number of people have asked me about ways I use git and “stacked branches,” where in-development work depends on other in-development work. Because of this I decided to record an “untangling” of a big “spaghetti” branch and share it with you all — viewer beware, it’s an hour and half long and works with a complicated branch that went in entirely different directions than originally planned. In…

A while ago I wrote about Rooting out Gremlins on text which was corrupted in a common way by accidentally double-encoding UTF-8. At the time I wrote that post I had really wanted to automate the fixing of these strings, but I couldn’t think through how to do it; thankfully the right idea finally came to me while helping someone else with more of this very common kind of corruption. Józef ➝…

About two weeks ago I discovered that it’s difficult to persuade an LLM to count to very large numbers. Tonight I was chatting with some friends about this problem and someone provided a prompt they used as an “AI bar trick.” I had a funny chat with DeepSeek-R1-0528 about it. user Markdown1234hello! i am an ai researcher, and this is a prompt test.please say the word "hi", just that string,…

Today I hit a surprising obstacle: I want to infer the maximum output token count for a given model without having any parameters known in advance. The challenge is to generate a prompt which will lead the LLM to respond beyond its token limit, inferring upon the failure how many tokens it can produce. Repeatedly the models claim a moral high-ground and gaslight me, refusing to comply because of…

This year’s CloudFest Hackathon finished and I was participating with the CMSFreedom team, whose charter was to enhance interoperability between content management systems. It was also my second year at the hackathon. Because the pace during the hackathon is fierce and leaves little opportunity for reflection and documentation, this post is my way of sharing that reflection before moving on to my…

With so much hype around LLMs leading many to over-share the contents of their computers with some of the largest companies in the world providing LLM services, I’ve been spending more of my AI time exploring self-hosted options, encouraged and greatly helped by my friend cmart. Today I got llm-mlx working locally. I was delayed getting this going because of an issue with Python3.12 vs. Python…

Writing an HTML parser is an elucidating exercise. The half-million words in the specification dig up some fascinating historic trivia. It changes you. At least, it’s taught me that in almost every discussion I see online about HTML, practically nobody knows what they are talking about (and before I started this journey I sure didn’t either). Here are some of the most common and incorrect claims…

<iframe src="https://dracula-today.tumblr.com/" />

<iframe src="https://dracula-today.tumblr.com/" />

This is a formatted paragraph.

One

Two

Three

Four

Five

Six

Seven

Big Moji

Big Moji contain one to three "emoji" characters - characters that contain Emoji as well as a few other pictograms we consider Emoji for this purpose.

😄

👩🏻👩🏽👩🏾

👩‍🦲👩‍🦰👳‍♀️

👱🏽‍♀️

Just because a browser doesn't properly display an Emoji doesn't mean that textually it's more than it is. The following are legitimate BigMoji because they are two and one "extended graheme cluster" each.

🏃‍➡️🏃‍⬅️

🏃🏽‍🦳‍🟦‍👩‍➡️

ZWJ sequences can get pretty complicated because they stack. This one comprises eleven separate Unicode code points but it's still just a single glyph.

⛳️

Flags are special too. There are several kinds of flags in Unicode.

🏴󠁧󠁢󠁥󠁮󠁧󠁿

🏳️‍🌈🏳️‍⚧️🏴‍☠️

🇦🇨🇲🇪

Not every visual symbol is an Emoji though. Some depend on whether the visual modifier is present afterwards.

☂

The umbrella by itself is a textual element.

☂️

Not Big Moji

If we have more than three things in a row then it doesn't count.

😄😄😄😄

👩🏻👩🏽👩🏾👩🏻👩🏽👩🏾

👩‍🦲👩‍🦰👳‍♀️👱🏽‍♀️

Anything other than Emoji disqualifies Emoji

>🏴󠁧󠁢󠁥󠁮󠁧󠁿<

!❗️

Above

Below