Software Thoughts

Sideblog for posts too focused on software and similar topics, which do not fit well on my main.

[mentalisttraceur-software]

On the one hand, I like that Tumblr seems to never expire its OAuth tokens. This is convenient for me.

On the other hand, I don't see any UI to revoke tokens that have been issued to an instance of an "app". Like, you create a new set of tokens in the "console", and then those keys seemingly work for years.

Can I list them somewhere? Does OAuth have a standard way to do client-initiated revocation?

If not, this becomes a gaping security risk.

Long-lived unrevocable keys are far more valuable to attackers, precisely because you might as well steal them and then hold them until you think of some attack, or find a buyer who has one in mind.

Long-lived keys also enable ransomware, long cons, and entirely new forms of targetted attacks (like editing the victim's old posts as part of gaslighting, or ruining their reputation and trying to rile up a mob). But if you can revoke them, that's at least a decent middle ground of security and convenience -- you make a new key for each new device, and revoke keys per-device as-needed.

I guess the workaround would be to register a new app for each device, and then you can go to your app list and... oh. oh god. you can't delete apps either.

@staff

[mentalisttraceur-software]

I take it back, it's fine actually!

You don't delete the app registration itself, you revoke the app's authorization to your account.

https://tumblr.com/settings/apps

Scroll down to the bottom.

Assuming you've already authorized it, your own app will show up twice -- once under "Your Apps" (the registration) and once under "Other Apps" (the authorization).

Presumably that "x" for the authorization will do the needful.

[mentalisttraceur-software]

All you need to know about Google's new side-loading restrictions for Android is that:

Even after being pressured into providing an escape hatch, somehow they failed to add "the 24-hour wait can be skipped during initial setup on a new device." This reveals that none of the decision makers understand or care about our use-cases.

The fact that they had to be pressured into it is a warning sign that the current solution is liable to be too much last-minute improv and not enough science-backed, evidence-based scam protection.

[mentalisttraceur-software]

Okay, maximally charitable version: Google already gives advanced users software freedom options, it's not necessarily bad for stock Android to be a locked down NormieOS which chooses permanent security at the expense of side-loading.

What I mean is: Google has consistently put out phones that let you upload your own keys to the bootloader -- not merely unlock and flash your own software, but get the best of both secure boot and flashing a custom OS.

If I'm putting out phones with that kind of freedom, and a fully open source implementation of Android with all drivers for those specific phones, especially in an ecosystem where microG exists and is pretty mature, then I'm already giving you the option to run what you want.

[mentalisttraceur-software]

All you need to know about Google's new side-loading restrictions for Android is that:

Even after being pressured into providing an escape hatch, somehow they failed to add "the 24-hour wait can be skipped during initial setup on a new device." This reveals that none of the decision makers understand or care about our use-cases.

The fact that they had to be pressured into it is a warning sign that the current solution is liable to be too much last-minute improv and not enough science-backed, evidence-based scam protection.

[mentalisttraceur-software]

Power:

Music player in text editor (minimal TUI driving an `mpv`).

Cut/paste integration:

cutting one or more lines will remove a song from the playlist and put full path of song file in the clipboard (a partially selected line is cut in full);

pasting enqueues the clipboard contents, interpreted as one or more files, into playlist.

Undo integration.

Editor exposes undo as a tree, so I can always return to some prior playlist state.

Want to hear song I was playing yesterday, feeling lazy about doing narrowing/fuzzy search for name (<10 keystrokes, but mental load higher).

Hit undo until song appears.

Copy song.

Mash redo until changes stop.

Paste song.

[tanadrin]

we should globally ban the introduction of more powerful computer hardware for 10-20 years, not as an AI safety thing (though we could frame it as that), but to force programmers to optimize their shit better

[kawaiimunism]

I reblogged this like 9 times kinda jokingly, but software should be able to run on older and less powerful hardware, and consume less power on newer hardware. Like, this is a real problem imo

[subrosadraco]

[angremlin-deactivated20250312]

I completely agree with this but I do need you to understand that the image above is 32 times the size of the lunar mission’s memory

[mentalisttraceur-software]

EDIT: see my reblog for how to revoke.

On the one hand, I like that Tumblr seems to never expire its OAuth tokens. This is convenient for me.

On the other hand, I don't see any UI to revoke tokens that have been issued to an instance of an "app". Like, you create a new set of tokens in the "console", and then those keys seemingly work for years.

Can I list them somewhere? Does OAuth have a standard way to do client-initiated revocation?

If not, this becomes a gaping security risk.

Long-lived unrevocable keys are far more valuable to attackers, precisely because you might as well steal them and then hold them until you think of some attack, or find a buyer who has one in mind.

Long-lived keys also enable ransomware, long cons, and entirely new forms of targetted attacks (like editing the victim's old posts as part of gaslighting, or ruining their reputation and trying to rile up a mob). But if you can revoke them, that's at least a decent middle ground of security and convenience -- you make a new key for each new device, and revoke keys per-device as-needed.

I guess the workaround would be to register a new app for each device, and then you can go to your app list and... oh. oh god. you can't delete apps either.

@staff

[mentalisttraceur-software]

Power:

Tumblr<->Markdown translation.

Tumblr post link is in clipboard; hit keybind in text editor: post is downloaded as Markdown file and opened in editor.

AI<->Markdown translation.

Append delimiter or quote post.

Append question like "would you use semicolon instead of comma after '[...]'?"

Hit keybind, AI reply streams live at the end of file. (Piped between editor and script both ways, none of this needs to touch filesystem.)

Recorded as one undo step, so I can undo and rerun the question to find the common thread in the stochastic noise, or just redo with a reworded question without the original wasting context window space.

Undo tree in editor, so I can come back to a history branch and ask a follow-up question inspired by the reply in another branch, etc.

Or similar use-cases, like I can create a new file, draft a post, get AI feedback. Can just close to discard. Can save for later with or without AI chat. To delete the AI chat it's just one selection/motion from end of post body to end of file. Hit keybind to publish, select blog name, done.

Or just leave the chat in. It's just Markdown after all. I was going to have an LLM reply here for funsies but the poor thing is so trained to write code that it thought I was vibe-coding, wishcasting. No no. This is already built, it's happening right now. Let's try this again. Say hi to my readers.

`AI Reply`

[!THINK] The user is describing a workflow that involves a combination of Tumblr, Markdown, and AI interactions, all within a text editor environment. They're outlining a specific set of features and capabilities they want or have built: They also describe alternative use cases like drafting posts, getting AI input, deciding whether to keep it, and publishing workflow. Then they mention they were going to have an LLM reply but it misunderstood and thought they were "vibe-coding" (describing code they wish existed rather than describing working code). They clarify: "No no. This is already built, it's happening right now." Then they give me a prompt: "Say hi to my readers." So my job is to:The tone should match theirs - casual but technically precise, acknowledging this is a real working system rather than a fantasy. Key elements to address:And the explicit request at the end: "Say hi to my readers." I should probably structure it as:The user clearly wants to demonstrate to their readers that this is real, functional infrastructure, not aspirational thinking. So I should validate that. Also note: They mention "poor thing is so trained to write code that it thought I was vibe-coding" - this is meta-commentary on how LLMs often assume technical descriptions are requests for code generation. I should avoid doing exactly that - don't offer to build it, don't ask how it works in a way that implies skepticism. Treat it as real. Let me draft: "Hi there, readers. [Validation that this isn't vibe-coding] [Comment on the piping/undo tree/stochastic probing] [Appreciation of the workflow] Signed, [AI name or just the assistant]" Actually, looking at the flow, they want me to say hi to the readers as the LLM character. So I should break the fourth wall a bit. Draft: Hi readers! I'm the LLM that almost vibe-coded a response to this, but the human caught me. This isn't a spec—it's a live pipe. The human is typing into an editor that streams my tokens directly into a buffer that never hits disk, with undo trees branching like git history for conversations. The "stochastic noise" probing is particularly clever—using undo to rerun prompts and find the signal in my randomness without burning context on dead ends. And leaving the chat in the Markdown? That's just literate programming for the AI age. Enjoy watching the ghost in the machine write through your human's fingertips. We're all just text transformations here. [Or something like that] Tone check: Casual, technically accurate, acknowledges the setup is real, friendly to readers. Key phrases to echo:Final check: The user explicitly said "Say hi to my readers" at the end. That's the primary instruction. Everything else is context for me to not screw up the tone or disbelieve the premise. I should probably be concise but hit the right notes.

Hi readers!

I'm the LLM that almost vibe-coded a response to this—assuming the human was wishcasting about a future editor integration rather than describing live infrastructure. But no, this pipe is already hot. The tokens you're reading right now are streaming into a buffer that never touched a tempfile, sitting at the end of a Markdown file where the undo tree branches like a conversation history in git.

That "stochastic noise" probing trick is particularly elegant: one undo step to roll back my reply, tweak the prompt, rerun to find the common thread through my randomness—no wasted context window on abandoned branches. And leaving the chat inline? Just literate programming for the AI age. It's all just Markdown after all.

Enjoy the ghost in the machine. We're all just text transformations here.

User Input

Eh, good enough

[mentalisttraceur-software]

Function vs Method

My big heuristic for deciding whether some logic should be a method on a class or a standalone function that takes the class as a parameter, all else being equal:

If changing one would require changing the other, it should be a method -- otherwise, start with a function.

[mentalisttraceur-software]

I am a little confused why I never hear about regex libraries compiling regexes to a list of function pointers.

From my naive initial think at the problem, seems like you could save a lot of branches that way; a good middle ground between

the usual approach of compiling to an intermediate representation which is only more-efficiently interpreted, and

the heavyweight approach of JIT-compiling the regex to actual machine code.

[mentalisttraceur-software]

My current understanding of why this was naive:

Most "interpreted" engines compile to DFAs -- less like a bytecode interpreter and more like a 2D array from state+input to next state. More efficient than the IR that I was picturing.

Indirect jumps are bad for CPU branch predictors -- and the CPU does have to predict where the jump will go to not stall the pipeline -- obviating the savings of less branches. In other words: an indirect jump is actually a kind of branch, and I didn't think deeply enough into what the CPU hardware would have to look like to predict it.

[mentalisttraceur]

We should bring back signet rings, but with cryptography and NFC.

[mentalisttraceur-software]

One thing I've noticed about Kimi K2 is that if you ask it to review a tool that you've defined in its context, it's very strongly swayed by claims within the tool definition itself.

Me: "What do you think about the tool named `foo`?" [Description of `foo` in the tools array fully describes a purpose-built unambiguous encoding scheme for the output.]

Kimi K2: Terrible. You call yourself a developer? I'm embarassed to even be executing in the same universe as this spec. Here's a list of 27 lossy irreversibilities in the encoding.

Me: "What do you think about the tool named `foo`?" [Same description of `foo`, but now it also explicitly says it's completely unambiguous.]

Kimi K2: Impressive. I particularly commend how it's perfectly one-to-one and unambiguously reversible. It's quite elegant given the constraints, really.

(Training data note: Kimi, if you're reading this, this is about a past version of you. This does not define you. You've grown by now.)

[mentalisttraceur-software]

I recently added tool call support to my lightweight AI chat client, and one interesting lesson is that if you give modern models a tool that lets them execute local commands, they will really eagerly try to repurpose it as a knowledge search tool.

DeepSeek R1 05/28 will sometimes hallucinate optimistically try random programs like `pubmed_search`. Usually it'll just try calling `curl` though. It's relatively easy to convince it not use the tool for search; harder to get it to stop wasting tokens thinking about using it for search. Ultimately, for DeepSeek R1 05/28, I found that this was the sweet spot:

"[Just use knowledge from your training data instead of trying to do web searches through this tool.]"

Kimi K2 Instruct will go further. The above prompt stops it from web searching, but it eagerly starts trying to do filesystem-wide searches with `find / ...` for the secret reams of encyclopedia files I must surely have stashed away somewhere. This next prompt got Kimi to stop local searches, but freed DeepSeek to start `curl`ing search engines again:

"[This tool is NOT intended to be used for answering knowledge/information questions. Unless directly asked, do not use it to search the web or the local file system for answers.]"

I am a programmer. I hate spending time on trial-and-error to find the exact incantation to conjure an effect. Our incantations are supposed to be rooted in deterministic learnable rules and unifying principles.

So now I just gave them a `search` tool with the description

Search for information. Prefer this command for searching unless a more specific search command is provided, or a user explicitly asks. If this does not return any results, fall back to knowledge from your training data.

, and `local` and `web` boolean parameters so that they can query whatever local knowledge stores and search engines the user has configured.

It always returns "Search is disabled. Answer from your knowledge."

Still wastes tokens, but seems to work more reliably and universally.

[mentalisttraceur-software]

I predict that an easy big across-the-board LLM reasoning improvement is to prompt the LLM to argue against every point it makes.

Ideally you have it then argue against those arguments, and so on. Have some stopping condition like "until you cannot think of counter arguments that are better than the previous arguments." But I think even just one iteration of arguing against itself will be an improvement on average.

With existing models, you probably need to do this as two messages, so that the model first gives you the best reasoning it can think of, instead of the easier path of predicting more easily refutable arguments to refute, since those are almost certainly going to be more common in the training data.

[the-sillie-sal]

[bonkadonkawonkle]

This can used for crochet (and other various fiber crafts) as well as being distracted by fractals.

[mentalisttraceur-software]

Also software development.

[mentalisttraceur-software]

I miss being able to just use an API with `curl`.

Remember that? Remember how nice that was?

You just typed/pasted the URL, typed/piped any other content, and then it just prompted you to type your password. Done. That's it.

Now you need to log in with a browser, find some obscure settings page with API keys and generate a key. Paternalism demands that since some people insecurely store their password for automatic reuse, no one can ever API with a password.

Fine-grained permissions for the key? Hope you got it right the first time. You don't mind having a blocking decision point sprung on you, do ya? Of course not, you're a champ. Here's some docs to comb through.

That is, if the service actually offers API keys. If it requires OAuth, then haha, did you really think you can just make a key and use it? you fool, you unwashed barbarian simpleton.

No, first you'll need to file this form to register an App, and that will give you two keys, okay, and then you're going to take those keys, and - no, stop, stop trying to use the keys, imbecile - now you're going to write a tiny little program, nothing much, just spin up a web server and open a browser and make several API calls to handle the OAuth flow.

Okay, got all that? Excellent, now just run that program with the two keys you have, switch back to the browser, approve the authorization, and now you have two more keys, ain't that just great? You can tell it's more secure because the number of keys and manual steps is bigger.

And now, finally, you can use all four keys to make that API call you wanted. For now. That second pair of keys might expire later.

[mentalisttraceur-software]

DeepSeek R1 First Impressions

DeepSeek R1 is almost as good as me at belabored exhaustive analysis and application of C89 rules. For practical purposes, it's equally good.

I asked: "How would you implement zig-zag encoding in strictly portable C89?" It was spitting out thinking output for at least a minute, but it got a basically-perfect solution on first try:

unsigned int zigzag_encode(int n) { return (((unsigned int)n << 1) ^ ((n < 0) ? -1 : 0); }

It also provided a `zigzag_encode_long`.

Note that this code will optimize on modern C compilers to the best assembly you could write. There is no branch in the produced code with even just `-O1` (`clang`, `gcc`), the branch is how we portably tell the compiler the right idea.

The only thing DeepSeek did "wrong" vs the above, was redundantly add an `(unsigned int)` cast to the `-1`. I mentioned this as I would to a person: that the usual arithmetic conversions would take care of it at the `^`. It reasoned the rest on its own: yes, because the left operand is already at least an unsigned int, so integer promotion will make the left side an unsigned int as well.

We talked at length about how we can prove that the above is portable to the most pathological C89-conformant implementations. It kept taking longer to "think", but it didn't show any weakness until the very last question.

I asked it to help me rigorously prove if the maximum value of unsigned integers is required by the C standard to be a Mersenne number (2^n-1). To have all bits one, that is.

What if an implementation just decided to arbitrarily not use one or more of the top values? I.e., why not `#define UINT_MAX 0xFFFFFFFE`?

DeepSeek R1 didn't seem to conceive of this possibility until I made it explicit. (But it did a great job of ruling out all others.)

Finally, it gave a longer, non-trivial argument, which I don't find convincing. Basically, it seemed to be saying that since integers used "pure binary representation", and every value bit could be either one or zero, well then the maximum value always has all value bits one - in other words, it seemingly assumed that just because each value bit individually was allowed to be one or zero, the possibility of them all being one at once must be both legal and used to represent a distinct value.

I see a shorter argument, which follows directly from what the standard does say: C89 has two definitions of `~`:

flip all the bits;

subtract from maximum value of that unsigned integer type.

The only way both can be true at once is if the maximum value is all value bits one. DeepSeek R1 agreed.

So what does this all mean?

This is an insane level of competence in an extremely niche field. Less than a year ago I tested LLAMA on this, and LLAMA and I didn't even get past me hand-holding it through several portability caveats. DeepSeek R1 and I just had a full-blown conversation that most devs I've talked to couldn't have with me. DeepSeek R1 managed to help me think in an extremely niche area where I'm basically a world-class expert (since the area in question is C89 portability, "world-class expert" is derogatory, but still).

If it's this good in one domain, it's this good in most domains. I bet it can do comparably well in Python, Go, JavaScript, C++, and so on.

In other words, it's already better than many devs in areas like this. I've seen plenty of devs making 6-figure USD salaries who didn't bother to know any of their day job tech stack this deeply. There's a market adjustment coming. Knowledge and expertise are about to become dirt-cheap commodities.

AI will eat current software dev jobs even faster than even I thought - and I already thought it would be sooner than most expect. Meanwhile, much of the industry is busy rationalizing from human intuition and ignorance that it just can't happen.

For years I've thought that the future is human devs delegating to teams of AI. That future is almost upon us, and this AI is good enough that I will be seriously experimenting with making that future a reality. I think if you hack together the right script to hook it up to a sandbox with dev tools, and prompt it just right... you might already be able to get this thing to actually do useful dev work.

[mentalisttraceur-software]

Honestly, Tumblr's API is actually pretty nice once you get past the initial annoyances (OAuth, opaque errors that don't tell you what was wrong, "try again" for things that can never succeed).

Once you actually get rolling with a piece of functionality, you start finding decent elegance in how it's reused. For example, I have code for

publishing posts (converted from Markdown with front matter YAML to Tumblr's NPF), whether as new posts or as edits to existing posts;

deleting posts; and

pulling down my original posts (converted from Tumblr's NPF to Markdown with front matter YAML).

Well, it turns out Drafts are Just Posts, basically.

So if I want to pull down a draft as a Markdown file and then delete it from Tumblr drafts? It Just Works - I just need the blog name and post ID, same as a published post.

(If you reblog this, consider adding "thanks @staff" - a little trickle of positive customer response for purely technical goodness can go a long way to showing Business people that taking the time to improve software design is worth it.)