It's actually really good for targeted debugging. Not the tool itself - still goes into lala land if you let it (arguably even more than trying to build out something). But I get to deeply think about code, then tell the agent to go make a small prototype or explain some arcane thing to me. Then I reason about it some more, etc.
The other times, when I let it generate 3k LOC with minimal supervision, yeah, that's when I die inside.
I have a feeling this will lead to huge interoperability and ecosystem fragmentation issues.
Well, there is one way... You can have a government steal all open source code and force its citizens to only use proprietary hardware and proprietary code, all government sanctioned btw. I wonder if we're headed this way.
LLM's have 100% gotten better, but it's hard to say if it's "intrinsically better", if that makes sense.
> OpenAI’s leading researchers have not completed a successful full-scale pre-training run that was broadly deployed for a new frontier model since GPT-4o in May 2024 [1]
That's evidence against "intrinsically better". They've also trained on the entire internet - we only have 1 internet, so.
However, late 2024 was the introduction of o1 and early 2025 was Deepseek R1 and o3. These were definitely significant reasoning models - the introduction of test time compute and significant RL pipelines were here.
Mid 2025 was when they really started getting integrated with tool calling.
Late 2025 is when they really started to become agentic and integrate with the CLI pretty well (at least for me). For example, codex would at least try and run some smoke tests for itself to test its code.
In early 2026, the trend now appears to be harness engineering - as opposed to "context engineering" in 2025, where we had to preciously babysit 1 model's context, we make it both easier to rebuild context (classic CS trick btw: rebooting is easier than restoring stale state [2]) and really lean into raw cli tool calling, subagents, etc.
FWIW, AI programming has still been as frustrating as it was when it was just TTC in 2025. Maybe because I don't have the "full harness" but it still has programming styles embedded such as silent fallback values, overly defensive programming, etc. which are obvoiusly gleaned from the desire to just pass all tests, rather than truly good programming design. I've been able to do more, but I have to review more slop... also the agents are really unpleasant to work with, if you're trying to have any reasonable conversation with them and not just delegate to them. It's as if they think the entire world revolves around them, and all information from the operator is BS, if you try and open a proper 2-way channel.
It seems like 2026 will go full zoom with AI tooling because the goal is to replace devs, but hopefully AI agents become actually nice to work with. Not sycophantic, but not passively aggressively arrogant either.
I don't know why this got downvoted. I've already been so frustrated by HN LIDAR mindsets but holy shit.
Human society exists because we value humans, full stop. The easiest way to "solve" all of humanity's problems is to simply say that humans aren't valuable. Sometimes it feels like we're conceding a ridiculous amount of ground on that basic principle every year - one more human value gone because it "doesn't matter", so hey, we've obviously made progress!
Agreed. I think that sometimes people on HN lose sight of what is actually important, which is human flourishing. The other day there was someone arguing that the best thing to do to fix loneliness problems in society is to remove the human need for socializing. Which... is certainly one way to fix the problem, I guess, but completely missed the point. The point is not to fix a mismatch between essential human desires and what we can attain, the point is to work on fulfilling those desires! Just something goes with nerd autism, I guess.
> I don't know why this got downvoted. I've already been so frustrated by HN LIDAR mindsets but holy shit
The extreme sides (proponents, opponents) are clear, opposites, and fight each other. More nuanced takes get buried as droplets in a bucket. Likely a goal.
> Human society exists because we value humans, full stop.
Call me cynic, but I do not believe every human being agrees with this sentiment. From HR acting as if humans are resources, to human beings being dehumanized as workers, civilians, cannon fodder, and... well, the product. Every time human rights are violated, and we do not stand up to it, we lose.
I have a very simple question as human right: the right for a human being to know the other side is a human being yes or no, and if not: to speak gratis (no additional fee allowed) to a human being instead. Futhermore, ML must always cite the used sources, and ML programmer is responsible for mistake. This would increase insurance costs so much, that LLM's in public would die, but SLM's could thrive.
>Human society exists because we value humans, full stop.
Eh, human society exists because it is an emergent behavior of the evolutionary advantage afforded at the time of adoption by the human species. There is on iron rule stating that it must continue into the future, or even that it can exist into the future.
More so, the value of a human has wildly fluctuated over history and culture. The village chief, nobles, the king were all high value humans. The villagers would be middle to low value, and others may be considered no value.
The industrial age began to change this some as value started to move from the merchant class to the villager class as many high production jobs needed less and less training to complete. With industrialization businesses running machines and production lines needed as many people as they could get. Still human rights were hard fought in places like America where labor wars broke out.
In the modern US we've setup a dangerous set of idealism that will most likely end in disaster because they are in conflict with general human values. That is the "pull yourself up by your bootstraps", "Any collective action is communism and communism will turn you into a pillar of salt if you dare look at it", and "greed is good". Couple that with TV media and social media owned by rich billionaires you're not going to see much serious opposition to these ideals.
But if/as labor loses it's values, so will the humans that performed that labor. After decades of optimizing human society for maximal capital extraction, values are dead, and the ever present thought police owned by the rich will make sure you don't cause too much trouble by resurrecting them.
I feel so illiterate when it comes to AST editing sometimes. I understand what an AST is from a computer science perspective. But I've never worked on a huge software refactor before that required direct AST textobject editing. Maybe an indication of my skill level...
The extent of my usage is having nice textobjects to easily interact with arglists and functions which aren't native to (neo)vim. Very cute and nice to just write "daf" somewhere in a function and just have it "just delete". Or hook it up with basic macros: search for regex, "daf".
I guess it's hard for me to edit things that I don't see right in front of me or aren't super simple changes (like name changes). Or at least, basic things I can reason about (such as finding by regex then deleting by textobject or something).
As for LSP's, I do use go to definition and rename all references, which is nice. But the huge structural refactoring part I have never really done. I don't really use many LSP features besides those two either...
> I guess it's hard for me to edit things that I don't see right in front of me or aren't super simple changes (like name changes). Or at least, basic things I can reason about (such as finding by regex then deleting by textobject or something).
This is actually what's nice about tools like ast-grep. The pattern language reads almost like the code itself so you can see the transformation right in front of you (at least for small-scale cases) and reason about it. TypeScript examples:
The $X, $A etc. are metavariables that match any AST node and if the same metavariable appears twice (e.g. $X = $X ?? $Y), it requires both occurrences to bind to the same code so `x = x ?? y` will match but `x = y ?? z` won't. You can do way more sophisticated stuff via yaml rules but those are less visually intuitive.
Sadly coding agents are still pretty bad at writing ast-grep patterns probably due to sparse training data. Hopefully that improves. The tool itself is solid!
If your editor of choice supports an extension (vscode does for example) it's a very easy on-ramp for a better search/replace than regex offers. It's syntax-aware so you don't need care about whitespace, indentation etc. Very easy to dip your toes in where a regex would get complex fast, or require multiple passes.
I converted a codebase from commonjs to esm trivially with a few commands right after first installing it. Super useful.
I hope LLMs eventually start operating at this level rather than raw text. And likewise for them to leverage the language server to take advantage of built in refactorings etc
I don't know which editors support this but there's also, for lack of a better word, context aware grep. For example, search and replace foo with bar but only inside strings, or only inside comments, or only variables, or only methods, not class names (maybe ast-grep does this).
TBH, it's actually not as hard as you think, most of the time, what I do is just select the whole syntax node and delete it, copy it, or replace it, and only 20% of the time would actually require deliberate understanding of how the AST is structured in the current language I'm coding in.
Maybe, or maybe not. I think there are many different kinds of work all lumped together under the "software engineering" umbrella, and tools which are indispensable for some people can seem unimpressive or unnecessary to others. We're not all spending our time in the same way, working on the same kinds of tasks.
I don't find myself doing huge structural refactoring often enough that it would be worth my while to learn a specialized tool for doing it; I'm a quick typist, and it's easy enough to just blat through the changes by hand on the rare occasion such a thing is necessary. But I don't work on giant line-of-business apps or sprawling web services, and I can see someone who did a lot of that sort of work having a different take on the matter.
I feel the same way. Learning Elixir and getting into macro writing was really helpful for me. Lisp is also the same way though for me the syntax of Elixir resonated a lot better for me. Many paths to the destination and all that, but figured I'd share in case it's helpful to someone
Oh wow. I have noticed the GPT series was far more arrogant than its results showed sometimes (and unironically it digs in its heels even further when questioned on it). Opus rarely has this problem - but it goes a little too far in the opposite direction. Not totally sycophantic, but sometimes it can't differentiate genuine technical pushback because something is impossible, from suggestions or exploration.
Opus has a different sort of arrogance. It readily admits fault, but at the same time is quick to declare its new code as the greatest thing since sliced bread. If you let it write commit messages itself, it's almost comical how much it toots its own horn.
Yep. There was something outside of coding that gpt was plain wrong about (had to do with setting up an electric guitar) and I couldn't convince it that it was wrong.
Well needed if the changes work as advertised. I realized from talking with 5.2 that the issue is not about being a yapper, or speaking too much about random factual tangents or your own opinions. That's easy to tune out, and sometimes it's helpful even.
What's extremely frustrating is the subtle framings and assumptions about the user that is then treated as implicit truth and smuggled in. It's plain and simple, narcissistic frame control. Obviously I don't think GPT has a "desire" to be narcissistic or whatever, but it's genuinely exhausting talking to GPT because of this. You have to restart the conversation immediately if you get into this loop. I've never been able to dig myself out of this state.
I feel like I've dealt with that kind of thing all my life, so I'm pretty sensitive to it.
You're one of the forefront experts in terminal protocol parsing. Do you have opinions on "interceptor" applications like tmux or mosh, for example? These applications technically need to do extra upfront work (especially mosh as it transforms the entire protocol) and it's not a transparent "I treat vt100 as a black box, I put bytes in, I get well structured, standardized events out". Does libghostty-vt support that currently, does it intend to support these kinds of protocols in the feature, or is this kind of thing outside the scope of the project?
reply