Emacs in the Age of Magic
commandlinesystems.com
commandlinesystems.com
On the other hand, many people claim that machine-generated software STINKS. That it is extremely verbose, that it reimplements similar methods over and over again. That it would rather cut corners at a low level than build high-level abstractions. That it pulls in libraries and makes serious security mistakes. That it writes tests that are basically incomplete. That it is unable to synthesize and keep context. In short, it has the same flaws as humans.
I think these are binary points of view, and both are right. We have vampires that are incredibly fast, super-strong, super-intelligent and super-agile, that suck your blood (your money), and that die as soon as the sunlight of security or complexity hits them.
There has to be another way. Nobody questions the versatility LLMs provide, or the boldness they give you to take on new projects. But I am not convinced that programmers or software engineers are going to disappear. I am not convinced that a manager can ship a new feature in a couple of afternoons and then lay off, or skip hiring, a junior or senior developer. There is a line near the beginning of Dune that goes something like this: "there was a time when men handed their thinking over to machines in the hope of being freed, but that only allowed other men with machines to enslave them."
We need to find a new path between smartphone addiction and being dumbed down by LLMs.
Shills and luddites say AI is capable of replacing SWE
Haha. I imagine Security and biology are the garlic and herbs that drive away the vampires and then the vampires sends their lesser minions to do their job.
> there was a time when men handed their thinking
Reminds me of the "knowing where to tap" story. Some already know where and how to tap and only delegate the tapping part. Some start from scratch and expect the machine to tell them where to tap.
The management, like the ship owner thinks that they don't need to call anyone. That's not the main risk though, it's the majority not caring enough to learn about the ship.
only the clueless, stupid ones
I spent most of the early 2010s rewriting systems that had been outsourced to budget offshore coding shops in the 2000s. The software looked a lot like this: no code reuse, no system architecture, just copy pasta over and over.
The client didn't care about this of course, they weren't reading the code. What made them eventually fire the outsourcing shop and hire a guy like me was the bugs that wouldn't get fixed, and the request to fix a bug that caused three more bugs to get introduced.
I'd rewrite the software and it'd be functionally complete with 1/5th the code, and it'd be faster and have fewer bugs, as well as being maintainable.
I actually developed a sick enjoyment about reading through their code and understanding what they were doing. From what I can tell the guys were being put under pressure to produce code and close tickets. I would find big chunks of dead code that I think was to fool their SLOC counters.
Changing schema on one project required a change order and migrations submitted to DB folks so these guys discovered that mysql booleans were actually integers and they could stick whatever number they wanted in there, and ended up using existing boolean fields to store unrelated data! Of course this broke the original functionality in difficult to track down ways!
The way we are accepting this is extremely unsettling...
Have you ever written parsers for complex languages? It is a cumbersome slog.
It seems quite cumbersome that this is, how this is supposed to work now?
Ah right, it’s free software. There is no milkshake, only people with similar ideology contributing to a common cause.
> I believe a sufficiently interested emacs developer with aThe disappearance of bikeshedding doesn't look like an obvious one but it's a huge loss. Imagine if every scientist resolved to refer to already written books and carry on their research in closed labs with no discussion with the rest of their field.
I believe a lot of people will find out with AI that most people don't have very original ideas. And that the limitations we encounter while realizing those ideas enable us to grow it into a good art piece or product. Due to being forced to consider if the effort of working around the limitations is worth it. And then put in the effort to realize something that approaches the initial idea as close as possible. Instead of instantly being able to realize the first idea in you mind without any obstacles.
Software still feels like this grand exploration to me—what's the best way to use a computer? What's the best way to express a program?—and a hackable text editor epitomizes that ethos (though I'm partial to neovim myself). Every time I install a plugin, I feel like I'm participating in a clinical trial of UX design.
I think a lot of us feel this, and the community will have to solidify the ethos that trying out other peoples' software is noble and healthy, exposing you to their ideas, even if it's actually harder and worse in some ways than vibecoding.
(I also think, by extension, the OSS community will need a more developed ethics of project governance. This was a hot issue even before AI, e.g. the terraform license change. But when the default is a vibe-coded bubble, authors will need to communicate either
1. this is just code—use at your own risk, PR at your own risk, or
2. this is a community, here's what we want for/from users, here's what we want for/from contributors, here's how to get involved in maintainance/governance.
You can be either but you have to say which.)
With vibe coded tools you know what you got, and they're easy to adapt if you need to.
Open source projects have always had forks, public and private. The ability to fork a project and use your own version of it privately, without needing permission or buy-in from others, has always been an explicitly valued aspect of open source software. If it was always seen as a good thing before, why do we see it differently now?
I think the reason is an implicit assumption that the production of more unshared work means that less work will be shared. But I don't see why that would be true. We could just as easily, and more likely in my view, see more work shared with the world at the same time we see more work not shared with the world. The incentives for cooperation remain, even as it becomes easier to work alone.
Has anyone made a persuasive argument to the contrary, or provided evidence? It seems like the main problem AI has created for open source projects is an avalanche of bad PRs, not a drying up of good contributions.
Writing code shouldn't merely be about accomplishing a goal, but making deliberate and good decisions that lead to more good decisions with increasing ease and clarity. It's no different than any other type of planning, engineering, or architecting. FOSS demands this more than any other genre of software because it's made collaboratively.
I think AI is just too irresistible for the reckless and dumb developers who were, unfortunately, always the majority in our line of work. Now we don't have to guess at that fact anymore or be afraid of being judged as arrogant for saying it. Now we know. These developers don't bring anything to the table. The problem of rejecting garbage PRs is far from new.
But who knows, maybe AI will get good enough to meet your standards even though most humans don't.
Not every project has to be "elitist", but if the whole point is to do the best work to serve a particular use case you can call it whatever you want. It doesn't change the fact that the best work has to be done.
If an overreliance on AI is ultimately deferring decision making and responsibility, then AI is then just a threat vector for yet more ideological battles that were already muddying waters.
Ideally it doesn't matter how the best work is done, but it seems that banning AI is a very effective filter.
For all the headlines about AI finding bugs, we never celebrate all the times it didn't because a human wrote excellent code. It is human will determining how to classify and fix them. AI training data is only ever as smart and fresh as the continued human engagement with it. It's trivially easy to see now that many models are going stale and maintenance is very expensive. Even if it was cheap, the fundamental problems don't go away.
Github has the forks link for easy navigation to other spins. If core won't except contributions for whatever reason let the marketplace decide what is useful. Let a thousand flowers bloom I say.
1. There are a few, already. I haven't systematically surveyed the landscape, but here's one: https://neomacs.com/
2. AI coding is still far from mature -- with substantial improvements even in 2026. I expect the number of attempts will only burgeon as this capability matures. I can already see a proliferation of emacs packages written with AI.
3. I would expect hobbyist activity to be more correlated with the capabilities of open-weight & cheaper models, for a variety of reasons. Pure hobbyists (esp. outside Silicon Valley) don't often have the spare $$$ for a $200 subscription plan, and won't be as early to adopt AI coding as "professional" software engineers in Silicon Valley (or geographically / economically adjacent somehow).
* probably yak shaved and far too much for what I use it for
To live in ignorance, that is to be pushed, pulled or compelled by what you don't understand, for example because you've outsourced your life to the magical mystery programming box, is in his terminology bondage.
Which also explains why all these AI psychosis people sound like crack addicts. Stallman is still completely correct about what software freedom is
The modern world is too complex for anyone to understand all topics fully. So as a result of this cognitive cap, by Spinoza's definition people have grown progressively less free for centuries. AI can, for anyone who seeks it, allow easy access to fields where they previously would have to consult with a human expert in order to begin to understand. Coding agents arguably make coding more accessible then it's ever been, it's just that as a side effect they make understanding less necessary to produce software (and thus also less valuable). Any schmuck can now ask an agent "hey how does this program work" and get a competent answer in terms they can parse. Of course, most people do not do this. Human beings, as a general rule, do not do things they don't have to. But the option is there.
As a board saturated with experts, people here on HN don't tend to appreciate what kind of incredible super power vibe coding is for people who did not pursue an education and career in software engineering. Everything that a computer can do is now accessible to everyone. As the vast majority of people on earth would not and could not spend the time and effort to understand software (or simply don't have the talent to do so), this is a pure upgrade in terms of average freedom.
Something that's the output of mass-effort machine learning algorithms is categorically different to human-made-and-understandable software. Do we have terms for these different categories?
For now. But it may well become common practice to construct weights directly, or at least to directly adjust weights in trained models, rather than exlcusively train them from prior datasets.
Asbestos is an amazing fire retardant. That's why nearly every building built before we knew how cancerous it was contains it.
Actually, this one also excludes nonfree data from the training set; there are probably more capable models that include downloadable nonfree data. I'm not sure if the dataset also needs to be free in order for the model to be considered free.
I can't tell if you're being cheeky, but... you're describing human child-rearing. From food and water and two humans, substrate can be created, and parents, schools, and other community resources help to train and align.
Folks can take EMacs as is and decipher it in isolation of Stallman, or any of the other authors.
But even then, how many people understand the internals of EMacs, or the interrelationships of the zillion lines of code in hundreds of files? Yet, millions of ignorant people leverage and apply EMacs to their daily tasks.
Look at what the REA impact is having on all sorts of code bases. The ONE THING that pushed Stallman over the edge and onto his path. The lack of source code to a print driver (or something along those lines).
Well, it’s all Open Source now.
Not only is the source “open”. It can be explained.
“Here is the source tree to EMacs. ELI5 the garbage collector.” “Working…”
It goes both ways. Now everyone has access to a savant to explain software, if that’s what they like.
I would rather take the time to ask the same question to the Emacs contributors mailing-list, who have the incentive to check their posts for mistakes.
GPT-based LLMs, by design, cannot. https://arxiv.org/abs/2409.05746
Modern LLMs are very far from being next token predictors trained only on text. Text and other modalities is the base that makes them understand your language and have concepts similar to yours, but they were also subjected to RL that makes them good at reaching the goals. Two tricks in a trenchcoat instead of just one.
I've definitely seen both Labs' "best" stuff confabulate to this day.
The harnesses / RLHF do a good job of prioritizing mechanisms to check results, which works out pretty well pragmatically for programming.
Nonetheless, I see Opus 5.5 spew out false information daily (never mind its code's unreadability).
I can't speak to Fable or Astra, because my employers, while threatening to fire us if we don't use the bullshit bots, have also refused to light the barrels of cash on fire that are required for those models.
As far as I can see, their fundamental nature has not changed, and truth is not something they have access to - just the called tool's exit code.
Obviously the answer is to stop relying on trust entirely, and just proceed directly to verification.
(You need to pass a test, maintain a license from the government, in many places have proof of passing a maintenance test, pay an annual fee to the government, roads can be closed at will by the government. They can install tolls if they wish etc. In the UK, the police can stop you and search your vehicle without reason and I'm sure that's true in many other countries)
The irony is that Stallman promoted free software and allot of code got published online for free, and now LLMs have trained on this code. Without Stallman there will be no vibe coding.
Without FSF, there would have been probably somewhat less open-source data, but certainly not dramatically less.
We need to be thinking about modularity with well-defined APIs now, to isolate the chunks created by AI coding. Maintainabilty requires firewalls.
⟩ Refactor these components into standalone modules that communicate across well-defined APIs. Write the API specifications to docs/API.md. Add a note to AGENTS.md instructing agents to preserve the modular seams and focus work on one module at a time.
bitter-lesson-pilled RMS before GTA VI
> But there's one glaring headscratcher, and that's its full-blown, self-contained implementation of a terminal emulator, in Rust no less. In the before times, a developer would undoubtedly seek such a heavy lift from off-the-shelf kit, but the holmes replicated over a coffee break what a billionaire with infinite resources had been toiling over the last five years.
Is this really true, though? Did model improvement since Ghostty began simply eclipse Ghostty and anything achievable with that now "legacy" code base? Would that project be better off abandoned or started over from scratch with Opus 5.5? Or is there something more there?
These are honest questions.
It all comes down to the attitude towards hacking, or any other profession / hobby for that matter, whether one (at a specific mood) enjoys the process or the results. The latter group has something to show off in the end, the former usually doesn't. But they know what they earned.
Of course it is. The AI coded apps only (debatably) work as end-user apps. They are thin little bumps on top of the software mountain. So much, including all the AI tools, depends on efficient, perfectly deterministic tooling.
It is almost hypocritical for AI enthusiasts to say we don't need to look at code anymore, while everything they do is build on piles and piles of infrastructure and tools that are carefully hard-coded, deterministic building blocks. The whole thing would come crashing down if everyone embraced vibe coding at every level.
I am just constantly baffled by people thinking that because we can make (crappy) end-user apps and one-off tools, no one needs to code anymore. Does anyone build solid tools and libraries? Does anyone still get pleasure from building a rock solid, snappy app?
I can feel all the AI coded apps and website turning to ever buggier sludge under my fingertips, and people are rejoicing that this is the way of the future.
I refuse. I will not touch Claude.
At times it feels literally magical. But more and more I feel like my brain has slumped onto the couch, lazily barking at a butler to do my bidding. Like a weird kind of professional brainrot.
One metaphor could be a car. You have to motivate yourself to get out and walk now, since a car can take you almost everywhere.
Like the people who circle the parking lot trying to get the parking space closest to the store entrance, you can live your life around the "convenience" and not see the problems with what you're doing. Or you can be the person who intentionally parks far away from the entrance so you get a few more steps in. Or the person who cycles to the store. Many metaphors murdered malignantly.
But since this is mental rather than physical, we might end up being couch potatoes like you said, flipping through channels rather than reading, or doing a hobby. (is "doing" a hobby correct? It sounds like "doing a business")
I get where you're coming from, but this metaphor doesn't really land for me. I find driving to be far more stressful than walking, and therefore higher friction than walking.
But your point still stands.
Most of that is also written via AI these days... Even the Linux kernel got that bug.
I think this is an interesting hypothesis and want to see what happens. I'm not really willing to go all the way to vibecoding, because I want to understand the system design. But I do have bots writing all the new code.
Where were those apps even before AI came along? Every piece of software I use in my day across Windows and iOS is a slow, laggy, piece of resource hogging garbage, and that was true long before AI came along. I'm not convinced that AI makes this worse, it might make it much better.
This definitely is a new era and, personally, I love it! (Also I have broken my right arm so I can't write code even if I wanted to. I'm dictating this text.)
This, after configuring Emacs to talk to LLMs. But that's not the right angle now for me....
The README of the project itself explains; "cooked" allocates the pty and interprets dynamically whether the other side is a command-line reader or full-screen program. See: https://github.com/vodik/cooked
More discussion on it here - https://www.reddit.com/r/emacs/comments/1xren7e/emacs_in_the...
The age of LLMs is forcing me to review bigger bodies of text, and more often. Various features of Emacs, like indirect buffers, narrowing, ibuffer, a multitude of search options, etc., make it an unbeatable choice. Plain-text composition is also at its peak, and Emacs, again, is just beyond anything else with completion-preview, thesauri, etymology and definition lookup, dictionary and translation, and fantastic agentic integration - I can query a model in the middle of typing this very comment, for various tasks - I can either just ask it to look up something, or feed the selected text to correct mistakes (with various diffing options to compare the result), fact-check things, ask it to TLDR or explain something, or even change any aspect of Emacs itself, right there, without missing a beat, without ever restarting anything. The flexibility totally is at complete and utter bananas level.
LLMs are all about text. Emacs is still the uncontested king in this space. Who knew that the 50-year-old turtle, after all, would take a major lead, surpassing even the fastest rabbits? I'm so glad I spent a decade falling in love with the idea of a Lisp-based editor and ignored all the cries about Lisp's imminent demise - I'm witnessing its renaissance; it's happening right in front of my eyes.
Is the author not familiar with the ECC memory business?
Is that a common thought? I learned vim, then emacs, and various other tooling not because I wanted to increase my speed competitively, but just to reduce my frustration with the gap between thinking and doing. Using something like VSCode to code is like playing on a laggy piano (if that exists). You think about I need to quickly update the signature of this function and it turns into a mouse/keyboard chaotic dance with your eyes focusing all over the place. It's like doing a three legged race.
With emacs and vim, I think about the code, notice that I want to change to change something, and can do the necessary actions (almost) subconsciously. Which means, I never stop thinking about the code. It may be longer or faster than someone using another software, but the advantage for me is to rely mostly on muscle memory while I'm thinking deeply about some problems.
Another thing is multiple buffer/views at once. This is why I don't like editors that have vim mode, because it's a pain to browse code with them, especially when you want a global view of something. With both emacs and vim, I can have multiple related files/sections of files/output of commands visible at once. Like how you would want physical papers and lay them on a desk.
If you expect that you can complete some task just from the keyboard, then the idea of having to use the mouse interrupts what you're doing.
The original post gives "stage and commit in git" as an example. Magit is much nicer than CLI, since you can quickly invoke the actions you want; it's less effort than typing out the full command on the CLI. -- With LLM coding agents, getting an LLM agent to make a commit is slower than doing it yourself in magit (if all you're going to do is stage and commit), but it's much less effort.
I spend a lot of time thinking about AI’s effect on me, as a person. TLDR: it’s a mixed bag: for non-coding things, AI has helped me learn and understand the world better. For coding, using AI coding harnesses and agents feels like an addiction.
I will say that I have used AI coding tools to make my Emacs setup the best it has ever been in 40+ years using Emacs.
It is weird to be using tools that sometimes make me think more about myself and their effect on me than the actual technical work I am doing.
Navel gazing.
Pick MLite from https://t3x.org and you'll get it pretty fast. Or with invariants under Lisp. Or its weird cousing, CLIPS, born from Common Lisp clauses.
On code quality, Neomacs (an Emacs clone in Rust, it has Elisp support) crashes a lot. And by a lot, I mean once you hit the menu with the mouse, it segfaults.
>he only catch is you'll have no idea how it works. No one does.
Is not about the rules, but the relations. It's a network of clauses, and not something akin to if-then-else with flow controlling. There's no actual "flow".
It's a highly experimental thing someone decided to mess around with, without any prior design discussion with any Emacs core devs. It's like a forty-minute-old baby and you're asking it to parallel park without crashing. What's your point here, I don't get it?
GNU/Emacs is incredibly stable. It's one of the most reliable pieces I've ever had; I sometimes run it for months without ever restarting. Which I can't say about pretty much any app on my system. Including the system itself.
I didn't got a crash on Linux/Wayland. And I didn't tested on MacOS or Windows.
What's your device envionment?
Could you please submit a GitHub issue to neomacs? (better with a minimal reproducible script and log `RUST_LOG=INFO ./neomacs` can output the debug log)
but now everyone's just sending a drone."
Why are people so sure this will keep scaling? We're still so far away from me being able to get anything even remotely close to what I can imagine using a free model running on hardware I own.
Then you need to remember there are a lot of very powerful people currently invested in not letting us have any of this, so I really don't think it's time to call free software irrelevant yet.
The AI tooling seems too unstable still.
The friction makes it much slower than agentic coding but it helps me stay in control and it's still faster than before.
So, no.
Reach out to Qwen, he can help you.
HN avant-garde article formatting to show they are cool.
Like demanding commented machine code between 1985 and 2025.
Sure you can get it, but that's not where the software lives even if it's how the software runs.
Reverse engineering closed source software is the simplest thing in the world now. The only reason why more of it isn't done is because it's so boring even minding the AI isn't worth most people's time.
If you're in a STEM program at university, quit. By the time you graduate, you will have negative value proposition compared to Claude. Get into business. The only way you're going to make it is to join the ownership class. Or pick up a torch or pitchfork and join the revolution.
...which is rapidly running away from you. Ignore this nonsense and do useful work. It will be okay.
I hope you are not advising any young people out there.
Tools are an extension of your mind. AI is becoming like unto a mind of its own. It's taking the power to make decisions away, rather than speeding your execution on such decisions. Note that by "decisions" I don't mean just higher-level decisions, but the micro-decisions about details that go into crafting a thing also. And we're being told oh, just accept this new layer of abstraction (on top of all the others). And don't worry about the details anymore, the machine does them so much better. I mean, sure you can work at a detailed level with an LLM, but on the job they'll PIP you for not just trusting the dark factory. Judgement is the new moat. Wait, Opus 5.5 is pretty good at that too; business context is the new new moat. It's just an upward retreat into fewer and fewer, less and less concrete decisions made by people, until engineering simply isn't done by people anymore; there's only a need for human business people, product managers, and QA testers. And then AI starts coming for their jobs too, especially since the CEO has been using ChatGPT to do product-market fit analysis for what potential things we might ship to maximize profitability over the next FY...
As for university, I actually don't really know what to tell kids these days. Many of them will succumb to pressure to keep their GPA up, and use AI to cheat to ensure that they do. So whatever they choose, unless they are super disciplined, super honest, and super incentivized to come by their degree without cheating, they won't get a whole lot of learning out of it. Especially not for the money they're spending now.