Replying to
@ekuber@hachyderm.io @cliffle@hachyderm.io I wonder about teaching Rust as a first programming language. Would it be too much all at once, or would they have the advantage of not having to unlearn things from other languages?
Remote
Matt Campbell
@matt@toot.cafe
mastodon 4.7.3Software developer, formerly at Microsoft, now co-developer of the AccessKit open-source project (https://accesskit.dev/) and cofounder of Pneuma Solutions (https://pneumasolutions.com/). My current favorite programming language is Rust, but I don't want to make that part of my identity.
Music lover. Karaoke singer. Science fiction fan. Legally blind. Secular humanist
1448 Followers
196 Following
50 Posts
Joined May 25, 2018
Pronouns:
he/him
GitHub:
Question for people who vibe-code full apps, especially blind people, as I believe that covers most of my followers who do this: What would it take for you to be comfortable coding directly in a programming language without an LLM? A high-level programming language with better tutorials and quickstart templates than are currently available, particularly for writing real apps? Or do you believe that having an LLM generate the code is the only way you could practically do what you do?
Replying to on hachyderm.io
@tristan@hachyderm.io @prism@infosec.exchange I'm sympathetic to the people who truly believe that others continuing to use, say, generative AI, are contributing to its harms. If you truly believe something is causing harm, why wouldn't you do all you can to try to stop it, including persuading others?
Replying to
@tristan@hachyderm.io @prism@infosec.exchange As for me, I'm not sure to what extent individual users are contributing to the problems. Or maybe I've just convinced myself that we don't because I'm now benefitting from the thing myself, and I see how others are as well.
Replying to
I, for one, am not going to criticize the people who are using the most effective tool they have, or tell them (as I once did, or at least implied through a badly written toot) that they should be angry at the companies who are developing those tools.
Replying to
But for those who *do* believe that generative AI should be resisted and fought, I think there also needs to be something better to meet the need that generative AI is now meeting for some people, however badly it may be doing so.
Replying to
I'm not even sure anymore that generative AI is something to be fought and resisted in the first place. I'll be honest: I use it myself. Though I continue to refrain from using it in one specific open-source project where I agreed not to use it due to copyright concerns.
It makes me angry, perhaps unreasonably so, that teachers are now resorting to mandating the use of handwriting on exams, at least in part in response to students using LLMs to cheat.
I have enough usable vision that I *can* handwrite, but it was unusually tedious for me, and I was glad to be mostly done with it when I got a laptop. I thought the days of students like me having to handwrite, and fully blind students having to do something different from their sighted peers, were over.
1/?
RE: https://caneandable.social/@EdenLinnea/116879333978518110
I broadly agree with the sentiment. But, I think there's something wrong with the fact that people think it's just normal that computers make mistakes. Aside from occasional hardware failure, a computer reliably does exactly what the running program tells it to do. So, a computer running a properly verified program shouldn't make mistakes. I think it's a problem that people just accepted that computers make mistakes, *before* statistical machine-learning models came along.
Replying to
@glyph@mastodon.social I've also heard of developers deliberately make macro-level decisions like language and/or framework based on what's likely to be good for the LLM. For instance, using Go on the backend and React on the frontend. The latter particularly bothers me, because that's possibly hurting the user experience, especially for users on low-end devices, for the sake of the almighty "productivity".
Replying to
@Migueldeicaza@mastodon.social I respect @glyph@mastodon.social for maintaining ethical standards. It makes me wish I had not crossed the line by ever using an LLM in the first place. I can't claim that I ever did so out of *need*.
Replying to
@jonathan859@someplace.social This was standard RCS, since our family is mixed between iOS and Android, and while some of them are on Facebook Messenge r(I'm not), I don't know if any of them are on WhatsApp. And I'm not about to encourage them to get on a proprietary platform.
RE: https://mas.to/@zekjur/116684759615780829
What I take from the linked post by Andrew Tridgell, author and maintainer of rsync, is that anger about generative AI should be directed at the people who are making it and pushing it on the world, not at individual maintainers who have given us so very much over the decades and are just trying to do what they think is right in response to what the former group have done and are doing.
RE: @drkirkadams@mastodon.drkirkadams.com
We've kind of been here before. In the late 2000s through the mid 2010s, there was much hype about APIs, and big players like Facebook and Twitter offered feature-rich APIs that let you write a full alternate client for the service. And there were such clients specifically for blind people. Then the companies closed those APIs down. Some companies are opening up their services to AI agents, through MCP and/or command-line interfaces. I predict that won't last, for commercial services.
Replying to
@JonathanMosen@caneandable.social If one believes what many of the critics say, that the real purpose of creating generative models is to replace as many categories of knowledge work as possible, then the benefits to the blind community are only side effects, and they come at a cost to other groups. And possibly at a cost to blind people too; any development that reduces demand for knowledge workers seems likely to worsen the already high unemployment among blind people.
Replying to
@ricky_enger@mindly.social @alpuzz@mastodon.social Oh, my favorite is when you have a Wikipedia table of contents with a section called something like "3 Marketing", and Eloquence, at least the version still shipped by System Access, converts the "3 Mar" to "March 3", which makes a total mess of it.
The variety of weird states that a general-purpose computer can get into is truly out of control. My Windows PC occasionally gets into a state where opening Slack with Windows+5 (I have Slack pinned to the 5th slot on my taskbar) plays a Windows ding sound as Slack is opening, but only if I press that keystroke while an Edge browser window is focused. How does that happen? I don't know; I just reboot. But that shouldn't be necessary.
Question for blind programmers: Do you think you can read or write code faster, when working in a language and library ecosystem you know well? And a follow-up question: If you use LLMs to generate code, do you feel that having to read more code, as opposed to writing the code yourself from memory, is a net productivity loss?
Replying to
@JonathanMosen@caneandable.social You've certainly given me something to think about with your reply. Of course I'm aware of how generative AI has been helping blind people going back to GPT-4, and I see that an increasing number of people are using coding agents to develop their own tools, with apparently good results.
I agree with you that the future belongs to those who work to solve the problems in front of us, and I regret the times that I've strayed from that focus to join the chorus of negativity here.
Replying to
@glyph@mastodon.social @prism@infosec.exchange That bit about Anthropic putting copyright notices on their users' LLM outputs was certainly news to me. I haven't seen it.
Replying to
@glyph@mastodon.social Luckily, it looks like LLM-assisted developers don't have to pessimize on choice of language as much as we might think, at least with the big currently-subsidized models. In my small-scale uses of LLM-assisted coding at work, Claude seems to know Elixir pretty well.
Replying to
@Migueldeicaza@mastodon.social @glyph@mastodon.social I thought GLM was big enough that you can only run it on a high-end GPU or cluster of GPUs.
Replying to
@ToniBarth@troet.cafe It's easy for any of us to say that about our own personal benefit. But when you look at the benefit for a whole group of people, plus the fact that some wildly exaggerated claims about the environmental impact of using generative models have been going around, it becomes less clear.
Replying to
@TheQuinbox Have you looked at using Mozilla's uniffi to generate bindings to Swift and Kotlin? https://github.com/mozilla/uniffi-rs
Replying to
@JonathanMosen@caneandable.social I don't know what to do about any of this. There are those on this platform, the fundamentalists and zealots as you put it, who say that generative AI is completely unethical and so is using it, for any purpose. I can't say I truly believe that; I do use generative AI myself. But their certainty and vehemence do make me wonder if they're on to something.
Replying to
@JonathanMosen@caneandable.social And unlike the religious fundamentalism of my youth, which I rejected once I was convinced that the religion was wrong about facts, I haven't yet been able to refute the anti-AI fundamentalists by debunking factual claims.
Anyway, for now, I'll get on with my work, as that's the best thing I can do to make a positive difference.
Replying to
@JonathanMosen@caneandable.social Still, the fundamentalists and zealots, as you call them, point to real problems with the process of training generative AI and the effects of its use: the environmental impact of both training and using the models, the use of people's creative work as training data without consent, the resulting unclear legal status of the output of generative models (I especially care about what that means for AI-assisted coding), and the further transfer of power from labor to capital.
Replying to
@ics@tau-ceti.space For what it's worth, that post has several instances of the classic OCR error where "AI" is written as "Al". Specifically "genAl".
Replying to
@grimalkina@mastodon.social It would be sweet if the ebook could be made available DRM-free, from a store such as ebooks.com that offers DRM-free ebooks as an option at the publisher's discretion. DRM-free ebooks can be read with a wider variety of readers, particularly specialized apps and devices for blind people. But I'll accept Kindle.
Replying to
@shinmera@mastodon.tymoon.eu @TheQuinbox@dragonscave.space Is it actually necessary to pound knowledge into kids against their will to any extent? There's a famous quote by Seymour Papert (one of the original developers of the Logo language): "Children seem to be such remarkable learners on their own. But then they enter school." Was he wrong?
Replying to
@jscholes@dragonscave.space @cachondo@defcon.social @alexhall@mastodon.social I interpreted "it stops arguments" as stopping arguments specifically about the music being played.
I wish there were a way of text messaging where someone could send a message to a group but the replies would go only to the sender, not to the group. I'm glad to get my mother's announcements that she sends to multiple family members, but not everyone else's reactions.
Replying to
@TheQuinbox Fair enough. ~1.3 MB isn't bad by modern standards. I wonder how the .NET AOT footprint grows when you add an HTTP client, Windows Forms for GUI, etc. This particular console program could probably be smaller in Rust.
Replying to
@masukomi@connectified.com Damn, it ended right at "behind the scenes what we've got is...". I want to know what was in that 1.3 GB .vhdx, unless that was somehow fake.
Replying to
@TheQuinbox @JonathanMosen Back when I was working on DocuScan Plus at Serotek in 2015, I had no trouble getting a Bookshare API key. I gather they've tightened access since then.
Replying to
@glyph@mastodon.social Yes, *game* programming is clearly art, but some other kinds of programming might not be. Still, the boring, non-art programming is necessary for shipping usable software.
Replying to
@glyph@mastodon.social Re: footnote 1, about LLM-assisted development making everything into instant brownfield work: I read a post, by someone working at a coding agent startup, about how he deliberately let the LLM follow its instincts on function names and code layout, and it apparently made the agent more effective in that codebase. https://ampcode.com/notes/by-an-agent-for-an-agent I could be OK with giving up minor aesthetic preferences. *But* if the agent can code even faster, then that puts me further behind on review.
Replying to
@jaybird110127@dragonscave.space @HunterXWorld@mstdn.social I couldn't make it work with Textalker GS either. I'd load Textalker GS, then boot one of the Infocom floppies with PR#6, and it would crash. This was running under MAME, of course, with the older version of Textalker GS that worked under emulation.
Of course, it's not that big a deal now, as there are accessible ways to play Z-code games on modern PCs.
Replying to
@alpuzz@mastodon.social Somebody didn't learn the lesson of DECtalk saying things like "Sunday Microsystems" in the 90s. Granted, I only know about that example because I happened to be listening to Jonathan Mosen's old Blind Line Internet radio show in 1999 or 2000.
Replying to
@jpellis2008@dragonscave.space I don't see where you boosted something right before this post, so I'm missing the context.
Replying to
@jetlagjen@gts.phillipsuk.org @WeirdWriter@caneandable.social Wouldn't it be better still to put everyone on the same level, both in what we perceive and what we present (both deliberately and unwillingly), by communicating in text?
There's an LLM called Apertus (https://apertus-ai.org/) that has not just open weights, but open training data and a fully published training process. It's developed by public institutions in Switzerland. Has anyone tried it? Is it any good? I might play with it; I have hardware that should be capable of running a quantized version of the 70B-parameter model.
Replying to
@JStark@caneandable.social Sure, I'd be happy to join a meeting and share what I know. My personal email address is mattcampbell@pobox.com
Is there such a thing as a USB power bank / portable charger that has some way of finding out its current battery level that's accessible to blind people?
Replying to
@glyph@mastodon.social I don't want it to be, as I'm sure you don't either, but that's not the same as confidently saying that it's not.