Micah R Ledbetter

I frequently find myself wishing for the ability to really treat the filesystem as an API. Imagine a file that:

• returns the current date/time in a certain format when it is read
• transparently download binaries from github and assemble them into .deb/.rpm/.apk packages
• … maybe more

I am aware the FUSE and plan9 exist, however, I’m imagining something easier. Imagine if a file could run a shell script, interpreted program, or compiled binary on read/write! It would be so powerful.

An off the cuff idea:

AI training on public data is driving privatization of user-generated content (eg Reddit API). This could push users with a real need for the information who today rely on public google search to build private archives. For instance, keeping a copy of the most insightful StackExchange answers that help you do your job.

If that happens, AI companies that figure out how to ingest private archives will have a training advantage, even if doing so is a gray area legally, as training based on sci-hub was for modern models.

It might even mean that the quality of the training goes up even as the amount of publicly available UGC goes down, since the private archives are curated by humans, and curation is a quality signal.

Michał Sapka notes that PhpBB is a good way to create a small community. I’d like this to be true, but I think this makes the barrier to entry too high for normie participation.

I think part of the reason Reddit has been so successful is that website account creation is such a chore. If every subreddit required a separate account, would they have a tenth of their users? A hundredth? Less?

Currently if you’re a small blog or community and want to provide persistent identity for commenting/forums, you need to allow some subset of account creation with email and integration with identity providers like Apple, Google, Github, etc. But you can’t reasonably offer all of them, because there are too many! It’s a front end design problem: do you want commenting on your blog to require searching through the list of all possible identity providers? Sounds very 2007.

And email as the identity manager is just awful: not only do you have to share a password or use a password manager, but you have wait for the site to send you an email and click a link before your account can be set up.

What if the browser stored a user’s preference for identity provider? Support everyone, let the browser prompt the user to log in with the one they already selected. No overdramatic interstitial “WARNING: THIS WILL ALLOW THE WEBSITE example.com READ ONLY ACCESS TO YOUR: PUBLIC PROFILE” (the user already accepted it), no email link.

It’s also unphishable, as it happens in browser chrome.

Hypothesis: LLMs will enable developers like me to build apps that require a serviceable but not world class graphical UI. This could unlock a deluge of applications built from or inspired by existing open-source command-line apps that don’t have a GUI complement, especially mobile apps.

One reason I’ve always liked the command line is that it’s easy to build a decent UI - even a good one. I’ve churned out a few dozen lines of python to create a great little CLI many times over the years, and it only takes a few minutes now. But GUIs are such a deep subject that require their own expertise, and all the GUIs I’ve built have taken weeks.

I started a SwiftUI app in Xcode yesterday and found myself missing Copilot immediately. I haven’t used Swift in a long time, and I’ve never used SwiftUI, so the median Copilot suggestion would have been extremely welcome. But what I realized I really wanted to ask a hypothetical XCode Copilot is “what’s the normal way to build an app that has one main screen and some sub-views in SwiftUI?”. I had no idea how to do this, and I wasn’t contexted enough in SwiftUI to ask the right questions.

In my experience with actual Github Copilot, and also with ChatGPT, questions about the normal way to do something get the best results. I’ve never seen a useful answer inventing new techniques, but they can certainly give you the average quicksort implementation (and explain it too). If they can do that for building simple GUIs, that will let a lot of command-line programmers like me graduate to baby’s first mobile app.

All of this would be really great even for users, because it might be the rising tide that takes open source software GUIs from subpar to good enough. I am guessing that there are thousands of useful apps that have very simple UIs with just a few screens, and that maybe there are thousands more that complement existing open source command-line programs which could be built if making the GUI was easier.

I’m not sure if the LLMs are there today; I haven’t built any graphical UIs with Copilot. I wonder what the state of this is at Google with Android Studio, or at Microsoft with whatever its mobile app GUI system is, or at Microsoft with Visual Studio (the big one, not VS Code). It seems like it’s too early to be optimized in any direction yet, but maybe they have this already. Would be neat to try it out.

And I think that’s how I became the person I am today: I want computer networks to operate in as straightforward and tangible a fashion as they did in 2009. And I want a lot of cabling.

…

Pihole forms part of a split-horizon DNS arrangement on the top-level domain I use, which is such a nice name I made it available on FreeDNS where it is used by a dozen poorly run Minecraft servers. This introduces an interesting set of DNS hijacking and misconfiguration hazards, which I find aesthetically pleasing.

• “my homelab” - computers are bad

A recent shower thought: consumer routers could enforce good IoT manufacturer behavior by placing devices on a secure, contained network by default, and allowing full access only if a device is receiving security updates.

The default network would be slower and heavily limited in endpoints it could hit, and in particular would not allow accessing any other devices on the local network. If an IoT manufacturer commits to security updates, they could be allowed on the full network. Defection could be subject to legal liabilities, up to a lawsuit in the worst case.

Over time, the feature could become more sophisticated, for instance by providing a better EOL experience. Currently, IoT manufacturers just leave their waste with the customer and mostly don’t even tell them (a recent example). This system might:

• offer an API on the router that devices call to register their EOL date
• inform users when a device approaches EOL
• show a dashboard of devices and how long they’re supported
• automatically contain devices that reach EOL

Come to think of it, there’s nothing that would technically restrict this to IoT devices. PCs could opt into this too. Probably most useful as a form of parental management of children’s devices – baby’s first mdm.

Router manufacturers would have to step up their own game, but they also would get upside: more customer touchpoints.

The best version of this would be an open standard, preserving consumer choice and competition. The hope would be that it improves baseline, even if industry consensus is hard. In the worst case, the consumer isn’t served because the standard is captured by the industry, which could then mislead in product ads and/or abuse the customer communications channels with ads.

Regarding AI prose suggestions like this, it seems to me that there’s a big divide between documents intended for others and notes intended for personal use. Clearly you wouldn’t want short sentences like “I’m on it” to translate to 2 paragraphs in inkdrop.app (my notes app of choice) or obsidian.md.

One place I have noticed this myself is in Github Copilot - so far, its code suggestions have been more helpful than its comment suggestions. In my limited experience, I get idiomatic, non-flowery code that could have been written by anyone, but I get comments that sound off to me. Sometimes the tone strikes me as actively obnoxious; I think this is just that it doesn’t feel like I wrote it or could identify with it. I rarely use Copilot’s comments.

Worse, I find them distracting. I don’t find its code completion nearly as distracting as how it tries to finish my English sentences in comments. Prior to the latest AI feature rollout from Google, there were ubiquitous short phrase suggestions in Docs and Gmail, which I also found distracting in a similar way. They derail my train of thought, and I end up losing time rather than saving it.

… although I’ve also been thinking of ways such verbose automatic rewrites might be helpful.

One hypothetical: if everyone uses ML suggestions, and everyone knows that everyone is using ML suggestions, then you could imagine spending a lot less time analyzing the tone of professional communication for personal slights. For all but the most overt slights, there’s plausible deniability in “oh well, that was just the AI”. This might make fractious professional relationships less uncomfortable, or prevent them from arising in the first place.

The fact that the appropriateness of a given suggestion is partly determined by its audience prompts the idea that the machine could have many audiences in mind and the user could choose between them. It might have generic heuristics about professional communication or notes to self, or specific training about the way the CEO likes to hear about production problems.

And that idea, of choosing an audience for suggestions, prompts an idea that the AI text generation could happen at the time of reading, rather than the time of writing. I wonder if there will ever be a need for a low level conceptual representation that the AI could translate for any type of target audience, like IR for natural language.

Maybe this starts off by converting the user’s natural language input to something that looks like an English prompt. “Hello, I hope this email finds you well.” -> “greeting, professional, short, top of email, trending on greetingstation, let’s greet step by step”. Or maybe it’s something lower level, some occult binary format that only a dedicated reverser could love.

You’ve gotten to the bottom of the post, so I can admit it here: I didn’t have any artificial help writing this post. I didn’t even reach out to Bing for comment. This way, I get to keep all the mistakes for myself.

Fascinating history that feels very relevant today: there was recently a bank startup called TNB that tried to offer “narrow banking” which would have offered deposit accounts and plops the result directly in the Fed, passing interest to customers (nearly 2% in 2018, although obviously finance people use the obfuscated language of “basis points” to make everything seem more complicated than it really is). The opposite of fractional-reserve banking.

The Fed did not approve this, and the startup says it “seeks to offer” narrow banking to this day – it cannot accept deposits.

This is relevant to today because runs are impossible in a narrow bank, and FDIC deposit insurance (which has a $250,000 limit) is not necessary. A regular bank takes your money in deposit and immediately loans it out. The bank has a lot of money in absolute terms sitting around liquid so that customers can withdraw approximately whenever they want to, but most money they hold is actually loaned out and therefore illiquid, such that if too many of their customers withdraw at once they will not be able to access it. (This is what happened to SVB.) By contrast, a narrow bank is much simpler; it can always get all of your money immediately from the Fed.

The inimitable Matt Levine covered TNB in more approachable language. Basically, if we had that, then why would regular depositors put their money in a regular bank? Maybe they wouldn’t. And if customers didn’t use regular banks at scale, then maybe your Chase Bank and your Bank of America don’t exist any more, and maybe it’s hard to figure out of that’s bad for society or not.

So the Fed just stonewalled. I have no position on whether this was right or wrong. But it’s pretty weird! Runs are impossible in a narrow bank, but also probably business loans of various kinds are more expensive. Maybe that’s worse?