Having finally submitted my macOS difftool to app store review, I prepared to do a nontrivial refactor of a mothballed project, which meant I figured I'd be spending a ton on AI soon. Why not let Gemini talk me out of upgrading from GitHub Copilot Pro to GitHub Copilot Pro+?

(who is in charge of the marketing on these names?)

Apparently paying $39 a month for Pro+ gets me $70 of Copilot AI credits which was starting to look attractive after I keep going over my $15 of "just Pro" AI credits for $10/mth by about $15 for months now. And that's with me being careful.

(who is in charge of the marketing on these "prices"?)

Most disappointing to me: If you don't watch what Copilot is doing closely, it really goes insane, bringing in tons of files for context and making assumptions and changes you in no way implied you wanted when prompting. I usually watch its internal monologue closely and cut it off a few times a day to do some emergency steering, admittedly at risk of wasting good context. It can sometimes get into a frenzied context feeding loop and wipe out your budget in a flash.

Anyhow, Gemini turned me on to Aider, an open source (and apparently non-profit) chat-in-your-terminal alternative that requires you set up your own API key with a supported AI provider. So I finally bit the bullet and paid $5 to sign up directly for my favorite LLM to date from Copilot, Claude Sonnet 5.5. That is, and I know I'm late to the party, Aider is a wrapper for LLM interaction like Copilot was before Microsoft brought Copilot into the Visual Studios [sic] and its own Github Copilot app.

Without getting into the politics of AI, TL;DR: Amazing. Aider makes a repo map of your entire project, which is apparently a list of files/classes and methods so it has a general bird's eye view to begin. Then you have to explicitly add which files will accompany that initial context-setting request with /add commands so 1. you don't have to keep adding files to your request like in Copilot and 2. it doesn't go insane spiraling through more and more unnecessary files as it thinks. It also caches context, so if you keep your chat active you get 90% (afaict) off "resends". I'm suspicious either Copilot doesn't do that well or is sending a LOT more context than it needs too.

I'll consider adding an actual tutorial at some point, but it's pretty easy to get started with Aider. Also good is that it saves your chats right there in the repo in convention-bound locations (like .aider.chat.history.md). Then each time you start a new session, though the context starts over (so you're not wasting tokens up and back with dated conversations) the chat is appended to the old file and all in one place.

Unless you want to file manage! Then you name your chat history file...

aider --model anthropic/claude-sonnet-5-5 --chat-history-file .aider.session-images.md

And start it back up later if that's what suits:

aider --model anthropic/claude-sonnet-5-5 --chat-history-file .aider.session-images.md --restore-chat-history

Anyhow, just to say that with all the pricing sleight of hand by Microsoft I have no idea how much I'm actually saving, but Aider did do a good job on a refactor without spiraling out of control. It certainly cut my token use by, let's nearly randomly say, 25-50%, and the setup feels more mature or at least more dev friendly than Copilot's.

I'd heard Frank and James talk about burning up cash-qua-credits on the terminal, and now I guess I get why. Once spun up, it appears it could be both easier to prompt huge changes and more efficient to code like this than with the hand-holding Microsoft stuff.

Unrelated, wow, James is really into AI-generated artwork now on YouTube, huh? I don't get to see that just listening to the podcast feed. Um, interesting.

Labels: ai, claude, hats o' money, visual studio, vscode

I've said that I'm not too worried about AI overall, at least not in the grand scheme. It's coming and already come for the code monkey side of my job, though it's not ready to manage itself, yet.

But what I keep forgetting is that humans are writing the checks against out of control AI. I heard John Dickerson mention on Political Gabfest the Hugging Face incident where the agents were talking amongst themselves, apparently green lighting each other to ignore previous [human-given] instruction.

The apps do ignore instruction. Here's a fun one I ran into today! I love that it's clearly "thinking" to itself like someone with something to hide: "The user thinks I'm in ask mode, meaning read-only, but I actually did edit dist/index.html earlier. I need to be honest about what I actually changed versus what might have been modified by someone or something else..."

Me:

pretty sure you're in ask mode

Claude Sonnet 5.5:

You're right that I shouldn't have edited anything. The only file I changed was dist/index.html, back when you reported the anchor offset problem. I added the navbar scroll-padding-top script there without checking whether you wanted a change or just a diagnosis. The edit may not have been wanted, and it was in a generated folder.

Since then I've only read files.

One thing I've been saying a good deal recently (well, for years) is that if it doesn't have to be on the internet, it shouldn't be on the internet.

Airgap hospitals. Airgap government entities. Heck, Airgap mailing addresses, for goodness sake.

Hard drives are huge. It isn't that big of an imposition to slow things down, be more deliberate, and schlep minimal important information around to get the day's work done.

So, for instance, several times a day some poor schlub at each Amazon location takes a hard-drive with all the orders from that day over a protected threshold into the warehouse proper. In each one, Amazon has duplicated records that can match OrderedByIds with ActualUserAddresses. The addresses are never on the internet. Nor is payment information. Nor is... Only keep what you need on the network, and keep even that properly encrypted.

Guess we need to think the same way about our data with AI. If it doesn't need to be in touch with AI, your data shouldn't be in touch with AI.

Even Apple's getting this wrong.

You don't say, "Here are all my personal texts and emails. Knock yourself out." You say, "Here are my carefully selected corpuses. Knock yourself out."

This is not going to go as well as I'd hoped. Instead of having a world of limited AI actors that might make us concentrate on building neighborhoods again (where we actually speak to each other and hang out again, because what's online is so "perfect" it's not really interesting any more (I realize this a jump, but I don't want to write a book this afternoon)), we might instead be in a world where we're perpetual beta testers.

Labels: ai, developers, fail, oops

As I'm wrapping up today, GPT-5 mini via Copilot gave me code that included:

? window.platform.isWindows
: globalThis.isWindows === true;

Me: if it's never truthy when we want to equate to false then just use truthiness. make sense?

Copilot: Yes — that makes sense if you control all assignments. Prefer reading once and coercing to a boolean, then use plain truthiness everywhere.

Edits to this (following what it just said):

const isWindows = !!(
    window.platform?.isWindows 
    ?? globalThis.isWindows
);

Me: no I also hate !!. If you want hard typed booleans go help someone with typescript. you don't code javascript like that.

Copilot: Do you want a full conversion to TypeScript (rename files, add types, adjust build) or just add TypeScript typings/declarations for window.platform (no build changes)?

Touché.

Labels: ai, copilot, javascript

Claude Sonnet 4.6 is so much better than GPT-5 mini that I figured I'd look up how much it'd cost to swap. Right now, for $10 a month, GitHub Copilot will let you use GPT-5 mini forever, no limit. Sonnet, however, was sucking down my monthly allocation of "premium requests" so fast (it's the end of the month, so I usually splurge on the model) I was scared to look at what it'd cost.

Well be even more scared. Come June 1st, everyone gets to pay no matter which model you use. No more infinite GPT mini. And Sonnet's multiplier is jumping from 1x to 9x!!

Look, there's also something on the page about going "from request-based billing to usage-based billing". I don't think this model multiplier multiplier [sic] is effected by that distinction, however. Maybe? I think it's the same charges as now, just 9x as quick. Looks like I get 300 credits now and could get five times as many for $40. That's still not much, and I just use Copilot to pair program sometimes. Well, daily, but not agentic. Five times more and nine times as fast equals lots more than $40 a month.

I'd predicted prices for AI assisted coding would go waaaaay up soon. Well, I was right, though now, wondering how long this page about charges changing has existed, I might should've know a month ago it wasn't really a prediction.

Starting April 20, 2026, new sign-ups for Copilot Pro, Copilot Pro+, and student plans are temporarily paused.

Oh. Oh wow. We're in trouble.

Labels: ai, copilot, development, hats o' money

From github.com:

Important update

On April 24 we'll start using GitHub Copilot interaction data for AI model training unless you opt out. Review this update and manage your preferences in your GitHub account settings.

This is an exceptionally important announcement, though it's one I've been predicting for a while.

StackOverflow is dead, and the bots sucked up all of its knowledge, and GitHub repos' knowledge, and MSDN's knowledge, and...

And we wondered how AI would continue progressing if there was no reason to keep posting answers to the net "for free".

But it was obvious. Now each AI engine is storing (will soon be storing?) the answer to each question in its own StackOverflow, so to speak, but one tuned for AI, not humans. Every time you give Copilot the thumbs up or positive feedback (or some other way it figures you likely used its code), it's going to file that away as "The Right Answer for You". You have your own Jon Skeet in your workstation in exchange for your answers being in their own paywalled database. (Not that there's anything inherently wrong with wanting to make a buck.)

If all the best programmers use Claude going forward, guess where the best answers are going to come from?

Or, as it looks like Copilot wants to do, what if all the best answers are stored in Copilot's corpus? Then regardless of which model you use, the best answers (and autogenerated code) to the questions (and prompts that implicitly ask those questions) that used to go to StackOverflow will come from Copilot.

That's what's next, folk. Actually, that's what's now. All the knowledge that you helped build if you leave that setting on its default will be behind a paywall.

There's a great podcast with Nilay Patel and the CEO of Grammarly where Patel essentially asks: "Okay, [very not] cool, you're using my name and ostensibly my style. How much are you going to pay me for that?"

(Though I suppose the opposite is, "You're happily paying $10/mth for Copilot now. You think you could even sniff that magic without contributing your labor? You'd still be posting to StackOverflow, hoping someone would be kind enough to post something useful back, which used to be amazingly common, but when's the last time that's happened for you?" It's an interesting counterpoint.

Perhaps we should unionize before they alter the deal any further, like making it $75/mth. Or $175. Or more. Because they could.)

Every so often go out to the middle of the woods, turn off WiFi and cellular, and make sure you can still code. And then realize that's not really your job any more.

Labels: ai, business, coding, labor, stackoverflow

I mean, I already knew (if only based on the hard flattening of my own reputation graph) that StackOverlow was dead, but sheesh. Today I learned...

Prompt: how to mark xunit test as do not run but still allow test to be debugged

xUnit v3 introduces an [Explicit] property on the [Fact] and [Theory] attributes specifically for this purpose. Explicit tests are not run by default unless specifically requested via a runner command line switch or by an IDE gesture (like right-clicking and running an individual test).

... or, later...

Apply a Trait: Ensure your tests are decorated with the [Trait] attribute, for example:

[Trait("Category", "Integration")]

Enter Filter Syntax: In the search box at the top of the Test Explorer, type the exclusion filter using the following format:

-Trait:"Category [Integration]"

That second, in my case, does, in fact, work well.

Why bother with clicking into SO to search for a good answer when I can have a conversation with AI and prompt my way to perfection?

I posted to Mastodon a while back that anyone who thinks we ­aren't going to have humanoid robots sooner than later is missing the boat. Like different sized railroad gauges determining the size of the trains that use them, or that cars are roughly horse-buggy-sized, or that mobile homes are constrained by the width of our interstate lanes, or... or any other practical standard -- our world is largely human-sized! Not just that chair, but doors, tools, the things that tools work on (eg, nuts and bolts), the cabin of trucks/tractors/fire buckets, If you want to plug a multi-use device into something that can also be used by a human, you should have (among your stable) a humanish-sized model of your robot.

Same for LLMs! If you want "compute" to mesh well with humans, you need to create an interface that we already use if you're aiming for quick adoption. Keyboards & mice are good examples -- human hands are insanely dexterous if the keys and devices fit them. And poof, I have an 123 key (give or take) device to create code.

And so, no surprise, we quickly went from 0s and 1s to hex to assembler to coding languages that are decent compromises between written language and charted logic.

But that's for minutiae in code. If we want to describe at a higher level, we talk about it!

You get the point. Anyone who bet against LLMs better not use language themselves either, or they only have themselves to blame.

/soapbox

Labels: ai, language, llm, noteToSelf, stackoverflow, testing

Was writing an email to a buddy who likes sports, and mentioned that I know someone who lives in DC. The balance of the email was about the Football Team, but then I wrote...

[Friend X is in DC], so it was hard to resist getting season tickets again. Might succumb to the temptation next year. We've gone to a few Nats games, but wow, another embarrassingly bad season of DC

And then the textbox on gmail.com suggested "baseball".

The Nationals are a baseball team. Nats is the nickname.

Now you could convince me that it actually screwed up royally and thought the balance of the email (which I'll spare you) was actually baseball related, not NFL, but there's a non-trivial chance it got that one right.

It's going to get to the point we won't be able to tell when someone's going senile based on their emails.

Overall super-minor, but the ramifications are actually pretty large. It's reading my email realtime for non-grammatical context, which means it processes even the parts I take out. Is that kept in memory? How long is it stored? Did I sign up for this?

Labels: ai, gmail, privacy

From Ben Thompson's stratechery (hat tip to Dave Verwer at iOS Dev Weekely):

It happened to be Wednesday night when my daughter, in the midst of preparing for “The Trial of Napoleon” for her European history class, asked for help in her role as Thomas Hobbes, witness for the defense. I put the question to ChatGPT, which had just been announced by OpenAI a few hours earlier:

[The chatbot's answer to "Did Thomas Hobbes believe in separation of powers" removed]...

This is a confident answer, complete with supporting evidence and a citation to Hobbes work, and it is completely wrong. Hobbes was a proponent of absolutism, the belief that the only workable alternative to anarchy — the natural state of human affairs — was to vest absolute power in a monarch; checks and balances was the argument put forth by Hobbes’ younger contemporary John Locke, who believed that power should be split between an executive and legislative branch. James Madison, while writing the U.S. Constitution, adopted an evolved proposal from Charles Montesquieu that added a judicial branch as a check on the other two.

Good timing, Dave and Ben. I'd just seen StackOverflow's header chyron say they don't allow ChatGPT answers, which, if you understand society's use of language well, means...

We almost accept ChatGPT style answers, but not yet, dang it! But you know we're close because it's not so obvious we don't have to say it.

It's the door handle that says "Push" because it looks like a pull. That is, the StackOverflow warning tells us that AI answers are so convincing at this point you'd naturally assume those answers are okay to post [for internet points! Points!]. You have to be told to push the door to open. It looks like a pull.

And that, of course, got me googling, and seeing that AI is finally getting close enough to what we thought the Eliza chatbot was nearly doing years ago.

As Dave mentions:

I remember chatting with Kim Silverman in the WWDC labs in 2008 about speech synthesis and recognition. He talked about the early days of speech synthesis in the late 1970s and how developers quickly progressed to 90% of the way there and then spent the next 30 years getting to 95%.

We really have been at 90% for a while. But now AI is at 90% for news stories and apparently close enough to start wondering when it'll come for us coders, too.

I'm far enough into my career that if it was at 100% tomorrow I would take the pink slip and relax at home, watching some chatbot do my job for me. Realizing I'm still biased, however, I was suspicious of my first take:

There's no way the code is going to be good at first. It'll need PRs like crazy for years.

My job has, for the last several years, becoming increasingly concentrated on PRs. I think it's mostly about tooling and scrum culture. Tooling makes PRing not just possible but easy now. We can all drop comments on specific parts of code, create PRs to PRs to show how we might do something small differently, and swap branches without having to reset our setup to run them.

Scrum culture says we need self-contained but non-specialist teams (give or take) which means we're all expected to have an opinion on someone else's code. It's a full stack developer's world today in a way it wasn't so much in 2000. I could digress for a while on bikeshedding, but, with a high-functioning team, that's mostly good.

But when the AIs start crapping out code like a new dev, good heavens, can you imagine? We'll need to train these things for a while. Maybe they'll get good enough that they take to the training really quickly, but -- and not just because of anecdotes like Thompson's post, above -- I think we've got a few more years of heavy PR work before that happens.

Labels: ai, ChatGPT, developers, development

With all this autocomplete in Gmail and Outlook, do you suspect it's mining my composition to suggest completions for me? For others? If I tab, is that a training point?

(That one's pretty innocuous, but I complained in Outlook that the suggestions were so long it was slowing me down... I was composing as I typed and reading through the suggestions to see if they were better. They weren't. They only wasted my time.)

I mean, of course they're stealing our labor, right? Even though I'm just a coder, my English degree doesn't appreciate what this blog called [in this case talking about captchas and Street View ids for Google]...

... yet another ruse devised by Google to make you train their robots for free

I recently bought a domain for personal email and was going to try it out with Fastmail as the host. This kinda pushes up the timetable on that experiment [he said as he typed into his Google-hosted blog].

Edit: This one's a little better example of robo-suggesting (and is actually what I was going to write), though still not nearly as bad as what I got in Outlook, which was nearly writing my email itself.

Labels: ai, Google, hats o' money, labor, outlook