I always find so much cognitive dissonance in these kinds of posts.
> My most successful work this year was hand-writing a tool, replacing ~4,000 lines of vibe-code with a few hundred. My system has proven more flexible, easier to test, and more generalizable than what it replaced.
> agents still need human taste and experience to produce good work.
Where do you think human taste and experience in software comes from?
> overstimulating & exhausting in a way that coding never was.
> for many of us, honing the code, finding the right design, seeing the parts mesh together — that was the meditative purpose we found in our jobs.
> That’s not the only time AI waylaid me. LLMs could not figure out how to annotate that model’s training data correctly. I spent days going back and forth with them
> LLMs are not yet good companions when brainstorming complicated things.
Why do we expect this to get better? The training data will degrade in quality if more people rely on LLMs so I'd expect models to eventually start getting worse even if they currently enjoy apparent improvements in output.
> Writing code is solved.
What a strange thing to say given all the above experiences, which I think we all share.
Why are we willing to give up so much in terms of purpose and understanding for perceived productivity gains?
It's obviously getting better, by the month, and the degrading of training data leading to a collapse has been a fantasy from the start.
"Perceived productivity gain" is a baseless conclusion. At this point it's abundantly clear that the productivity gain is real and massive.
Where I see the incoherence is in that people see too much value in code review and optimizations.
I believe many of the optimizations are premature. If you did not spot the problem in your QA testing, chances are it was not that important in the first place.
If the architecture wasn't perfect, it is now extremely cheap to change it later, as you realize real deficiencies. The agent will do it for you in a matter of hours, and largely on its own.
You will have to budget a lot of QA effort for verifying the significant changes before going to production, but I don't think it's a deal breaker.
> At this point it's abundantly clear that the productivity gain is real and massive.
Oh? Where is it? Why am I not seeing it as a user if it's abundantly clear? In none of the software I've used frequently in the last year have I seen any of those “real and massive” improvements. Not saying they don't improve at all, but the rate hasn't increased since the advent of LLMs.
I'm still waiting for that shiny new era of software productivity to impact any actual users.
> The training data will degrade in quality if more people rely on LLMs so I'd expect models to eventually start getting worse even if they currently enjoy apparent improvements in output.
Its not a direct analogy, but this seems to me as like saying "I already know maths and english, why do we still hire and pay teachers". Its a massive issue that just seems to be hand-waved away as "we'll figure it out later".
To compound on the problem as well, even those of us who do still write code by hand are actively discouraged from sharing it, knowing no human will ever read it and we'll never be credited for our work, it'll just go in the pockets of a few big AI companies. I suppose there will be some that take some pride in knowing their code is helping AI models advance, but thats a small pool of an already small pool, assuming most people decide to just use the AI to write the code for them. A minority of a minority.
I guess this is the chickens coming home to roost of this whole attitude towards growth now, fallout later
I am not OP, obviously, but I am writing to address the "Where do you think human taste and experience in software comes from?" point, hoping that I provide an alternative perspective.
As someone who wrote code for 27 years in all sorts of languages, not necessarily impressive code, but got paid for it, made a career, and changed life for the better multiple times, I don't think coding was ever an important part of the journey! Maybe it's cognitive dissonance, as you wrote above, but I believe taste comes from exploration, from trial and error, from being willing to be open and change old ways if new ways can improve how things are done!
Similar to taste in food or even music! You might think certain foods taste bad, but until you taste them, you won't be certain! Sometimes your senses will be right and the food will taste horrible; other times, you'll be fairly surprised. Similar to music—it's my life hobby—you might think certain genres or artists are bad, but until you sit down, give them a shot, and hear what they have to say, you won't know for certain if your judgment is correct.
If you meant by your comment that you need to code a lot to acquire taste and experience in software, I respectfully disagree. You are not completely wrong, though, but putting emphasis on code, if that's what you were going for, is an overstatement of code's importance. For that particular case, similar to the analogies mentioned above, it's about exploration and trying different languages, paradigms, solutions, and stacks, and unfortunately, we as humans have finite time, so we can't be trying and exploring fast enough to form judgments!
What I genuinely love about LLMs is mostly their exploratory powers. I try new stuff that I always wanted to try with actual products and form opinions about the whole experience.
Anyways, as I finished writing this, I realized, as always, I might be wrong! You might be right, taste is highly subjective! And acquiring it is a very personal thing! So I don't think right or wrong are important here!
> Why do we expect this to get better? The training data will degrade in quality if more people rely on LLMs so I'd expect models to eventually start getting worse even if they currently enjoy apparent improvements in output.
You are assuming that AI should be trained on speicific field to reason better but it actually doesnt need to.
> The training data will degrade in quality if more people rely on LLMs so I'd expect models to eventually start getting worse even if they currently enjoy apparent improvements in output.
RL is not dependent on training data. The mathocalypse didn't happen because they found better training data somewhere.
> The mathocalypse didn't happen because they found better training data somewhere.
There are reasons to suspect that OpenAI may have intentionally or unintentionally trained on chat logs from mathematicians--mathematicians who were actively using OpenAI to try to to solve the same problems themselves.[1] If those mathematicians were on the right track, and getting close to a solution, that could be classified as finding "better training data somewhere."
sorry but
given how post-training actually works, there's essentially no direct way in which literal PII-scrubbed data made its way into a production oai datamix for RL warmup purposes in 2026 to the extent this data has influence, it's indirect/at the level of process decisions.
what i am trying to say more specifically is, the folk model some people seem to still have about ai companies "directly training on user transcripts" in this naive fashion is firstly wrong, and secondarily, would be counterproductive as a strategy for continued progress
"transcript shows human trying to do an exotic thing evals don't currently cover -> build synth constructions that broadly cover its genre of problem class -> RL on the verifier set designed to cover the failure"
that's the loop...
the """flywheel""" is... diagnostics! wow!
and so this kind of process improvement driven use of data has an obvious diffuse causal influence on what happens during training, right?
but ~none of that process influence looks like "specific mathematician's specific intuitions leak into the training data".
It’s perfectly plausible that unscrupulous people at OpenAI (and that’s the culture right from the top) searched user transcripts for relevant approaches to problems they were working on, or searched them for problems to look at, and nudged models based on that.
I agree it’s unlikely this happened by accident, or entered training data (though not impossible), but it could certainly be within the context agents were using when tackling problems.
We have no evidence either way, but they had all the data they would need to do this.
saw an interview from the mathematician involved, Tristan Buckmaster, on the numbephile yt channel, and he said he thinks it was more a case of people at anthropic leaking it in person to openai. Apparently with a hint about the area they were looking at it could have been easy (ish) to get the same approach, and indeed the paper openai released for Navier–Stokes uses their approach
Rl works on one of two things - (1) a literal set of rules for a system or (2) training data. The former matches perfectly to things like (most modern) math and games that have perfect sets of rules. Even things like physics and robotic where our rules/models are extraordinarily good, rl barely adds value. Move one step farther to chemistry or a complicated physics scenario like weather, and you are back almost entirely dependant on (2) data. Now there might be a few more domains with rigorous rulesets that haven't been explored fully (id bet with some effort programming might make some strides here, maybe something like circuit layout, maybe even law if folks decided to make a lean-like law scheme...) but the vast majority of problems AI is targeting are dependant on data.
I’m not sure that’s a good argument that models will get good at novel brainstorming and research ideation. Where do you get scalable and high signal verifiable rewards for creativity and insight?
I see what you mean, One thing I noticed is that the suggestions it gives are always generic and reproducible. If I want unique ideas I have not had good experience withh LLM's.
First of all ppl are just processing the new enironment. So go easy on conclusions. Its all in flux.
Secondly most the posts on HN are from individuals. It doesnt represent group dynamics and how group/institutional knowledge works. For example there might be a regulation on publishing data every quarter. On the surface it looks like compliance but in reality different internal groups will hide data, delay it, fudge it etc etc. LLMs train on public data. Only after they are allowed access internally do they discover group knowledge. We are far away from seeing that happening.
This always reads like “writing english is solved” just because you collected snippets of famous works and have built a spell corrector.
Code is written to process data. The struggle was never to learn programming syntax (most grammar are only a few page long). The main struggle was that it’s hard to model a domain into an information system, identifying the components and their relationships, what is ingested, and what needs to be output.
Computers have integers and floats, but you need to implement units like distance, money, various counts,… The real world have words and sentences, but that needs encoding like ascii and unicode as well as the “string” concepts to make it work. Take HN, I’d bet that if you can model it in terms of data, writing code is not an issue.
Thank you, this is what I've been seeing. Structs and objects are easy, but the hard part is encoding the domain most accurately into given data types of the computer (or creating new data structures to solve the problem!)
If anything, I do expect the casual non-tech users to be making the most adaptable software because they have what they want in their head.
Although, part of what the dev does is figuring out the weeds, and only a part of users have the patience to sit through detail corrections. And the AI user has to verbalize their wants exactly too.
> I do expect the casual non-tech users to be making the most adaptable software because they have what they want in their head.
Those are the people who are most likely to give a monolithic prompt and then spend a week telling the machine 'I don't care about encryption or file permissions, just make it work.' Not to mention that most people can barely verbalize the problems they have with devices they use for 8hrs a day, let alone form a full, self consistant, project idea and then put it in writing.
if they're persistent and patient enough to see the result of what they think they want and iterate enough times they'll eventually learn. non-coders aren't dumb by default after all and I suspect they'll learn much faster than us because they won't be bogged down by the minutiae like reading docs.
Imagine you could build a piece of furniture without actually needing to know how to cut wood, drive nails etc. Imagine you could do it in 3d without knowing how to actually handle wood, you just have an image you can iterate on without actually costing you anything but time.
I remember the first big app I wrote in high-school. I had to learn C++, learn Qt etc. I knew what I wanted but putting that into code when I didn't know any of the libraries / tools was a huge time investment. If I had LLMs I could've done the same thing it took me 6+ months in probably less than a month.
> non-coders aren't dumb by default after all and I suspect they'll learn much faster than us because they won't be bogged down by the minutiae like reading docs.
They may have a perfectly good shot at being smart, but I still don't buy the argument. The reason they struggle to become coders (assuming they tried at all), in my experience, is a lack of problem-solving and debugging skills. There is no real shortcut here. An "image" of the chair isn't enough; when the first iteration is wrong, you have to be able to explain why it's wrong.
I think there are a lot of people out there with good problem-solving skills but for whatever reason did not get into programming or they tried and they got put off with learning the languages / libraries or got stuck and didn't find a solution easily enough and gave up etc.
With a sufficiently good LLM these people will skip the tedious part of programming that most of us had the motivation / patience to go through. All the folks saying that actually writing code wasn't the bottleneck have had a LOT of years to get to that point, years a lot of people aren't interested in going through.
I'm not saying lay people will be able to join mega projects and work at FAANG. I'm saying a sufficiently courageous lay person will be able to make an app to automate whatever tedious birocratic process their niche work hasn't seen a product yet.
just because you have an idea doesn't mean it will work. kids come up with ideas all the time in their drawings, but they disobey the laws of physics and are therefore not feasible in reality. sure, the model can guide them, but eventually they would have to pick up the necessary knowledge.
The necessary knowledge needed with LLMs goes from needing to know programming languages and libraries to having a high level view of what is possible which is far easier and less boring to get
> Although, part of what the dev does is figuring out the weeds, and only a part of users have the patience to sit through detail corrections. And the AI user has to verbalize their wants exactly too.
Even with basic pseudocode and very lenient logical rules, most people won’t bother to provide a a good definition of some process, as it’s quite unnatural to be that logical. Instead, it’s mostly just playing around. When you ask the one that is supposedly 10x-ing his productivity to explain his mental model of their own projects, it’s often a wishlist instead of a real understanding of what he has built.
> The struggle was never to learn programming syntax
For 85-90% of devs, yes this was the struggle. Creating and updating rows in a database from user input is all the “domain modeling” there is in most software. That’s what is solved by AI and it’s the majority of software dev.
Learning syntax takes a weekend. Learning to approach problems analytically and algorithmically takes much longer, and is the only "hard part" of computer programming.
Those who think the challenge is to overreach the limitations imposed by programming languages and types, etc, to do seemingly "magical" things. These are ones who would never understand why you might want higher-kinded types, the relative pros and cons of the functional vs imperative approaches, or even of type systems in general. They also tend to think the code written, in itself, is the valuable output of their work.
And then there are those who think the challenges is to model problem domain so well, using the language features, types and data structures, etc, that the solution (and further adjustments to it) becomes obvious and straightforward. They also tend to think the analysis and model that results is what is valuable.
Another way to see this difference: what impresses them? it is a hack that uses arcane knowledge to bypass software and hardware "limitation" or structures? Or is it code that is simple but still effective at solving complex problems in elegant ways, that may look simple but reflects a huge amount of knowledge, skill, and architectural intuition built up over many years?
I of course think that the second group are more effective and are doing proper programming, but many programmers and the general public think otherwise, or can't event tell the difference. I once worked with some colleagues could never see the light, despite a lot of support from me.
So here is the kicker, the dissonance in response to AI, I think, is kind of analogous to these two groups. The first groups likes that the AI will not stop them, but rather get them much quicker to their "magic", whereas the second group is more skeptical and sees many issues with AI not seeing the big picture and understanding deeper intent.
It takes a weekend for a software developer. For a totally novice it takes a lot more time
The issue here is that you are learning the syntax at the same time you are learning foreign concepts, and the syntax is often confusing and counter intuitive
Learning the syntax separately is of no use because without a mental model you can't do anything with it
I remember learning python in high school by following the tutorial in the official docs, then from a book. Up until function and OOP, it was fairly easy, but I gave up afterwards because the exercises touch on things I haven’t learned yet like trigonometric identity. I could see and understand the syntax, but I couldn’t understand the code.
This is why children can easily learn enough programming to create games, because it’s a domain they can comprehend easily. Just like they can use english to write about their daily lives. It will be simple sure, but it will work.
So learning to code is fairly easy, you can easily create simple programs after a weekend. More complex programs will still use most of the same basic concepts as far as coding goes, but you will need to master more complex concepts, and that takes time. And such concepts are not about coding, they are about the nature of computing or the domain the software is created for.
This is why experienced devs only need a few days doing kata to get up to speed with a new programming language. It’s usually the same old, with slight differences. What was always needed was an understanding of computing and the ability to break down problems and create a solution that can be computed.
Coding to me is the same as drafting an engineering schematics, or writing a sheet music. It’s just a tool, not the end result.
It certainly wasn't "approaching problems analytically and algorithmically" that prevented me to contribute to many open source projects in the past.
On the other hand, I've made 3 apps in 3-days' spare time, which I could not find in years and were not worth investing months of effort at the time.
I believe that anyone who says that writing code is no longer needed, was probably not a very good coder to begin with. I have seen what the best models can do, and I've seen what people call "good", and I am not impressed.
However, I'm still split. Because this could mean that we see a jump in code quality, due to bad coders getting someone better to code for them. But we could see a decrease in code quality, due to less people "caring".
Most people I've seen saying coding is solved are ML engineers who write throw away code for running one-off experiments. If you can effectively "archive" code after getting the results you need, I can see why you would say coding is solved. Sustainment is a non issue.
> I believe that anyone who says that writing code is no longer needed, was probably not a very good coder to begin with.
I’m probably the best coder in my org, which doesn’t really say much, but at least allows me to judge about the room temperature. It is solved. I don’t say that’s you, but the only people who say it isn’t “that good” are the same people who used to yak shave, rename and move things 10 times until it “feels right”.
First, you question the model. I have reams and reams of conversations with Opus 5.5 that shows it makes architectural messes all the time
And then you are saying you needed to have performance tests (which many devs don't even bother with) to catch bad architecture?? lol
So what if the architectural problem is not one that would be caught with a performance test? It seems you lack imagination to think of all the ton of issues that that would be very complicated to catch with tests.
Opus 5.5 is not normally copying the whole database to update a single row for no reason, and without mentioning a workaround. It's simply not my experience, and I don't think there is anything weird about me saying so.
I don't think I commented on performance tests at all.
Maybe I lack imagination here, yes, but generally I go by experience rather than talking about imagined issues. That is precisely what many here seem to do when it comes to explaining why AI couldn't work for them, and why reviewing the code remains necessary.
I haven't tried Sonnet 5.5 myself since I find Opus 5.5 essentially unlimited in the subscription, and okayish at API rates with a $500 budget.
When I looked at a benchmark, it suggested Sonnet 5.5 wasn't much cheaper than Opus 5.5 at Low to Medium effort for similar performance. That might not be representative of real world use though.
My workplace disabled Opus access in June, saying that it is basically the same performance for most coding tasks at a lower cost. That might have changed now. But the main point I'd like to make is that even with Sonnet 5.5 I have a similar experience to OP - code that makes odd performance choices, is over-engineered, and arbitrarily ignores rules we put in the context. For small self-contained apps doing basic stuff it's okay. For 100k+ line codebases running fairly complex algorithmic routines it needs heavy supervision.
My codebase has 500k lines and some 3D work that needed optimization on mobile. At no point was it needed to supervise the code in the implementation.
Optimizing performance is one of the tasks best suited to AI, since the outcome is measurable.
I can't comment on complex algorithms specifically, but it's not at all the case that this works only for basic apps or cannot handle complex, performance critical work.
I don't understand how it's possible to acknowledge the need for this steering so casually, as though it's obvious and expected, while simultaneously holding that "coding is solved". It comes across that "solved" means something very different to different people.
I've had a realization recently. I was also probably the top craftsman in the teams I worked in throughout my career.
I think what people are missing is that people often become craftsman because we want to be able to keep the code adaptive and performant so we can provide business value quickly and efficiently.
The change is with models, the benefit dynamics of immaculately structured code changes dramitically. Now the real area you need to focus on is correctness and performance. Structure doesn't really matter unless it serves those two purposes, at least from the business value side.
This is exactly right. Architecture -- in fact, using any kind of abstractions at all -- is basically a crutch that humans use to make ourselves more effective in implementing and evolving code over time, usually by either compressing the problem domain so that we can pack more of it into our limited brains, or making our own mistakes visible early. It's extremely valuable for humans who need to work on complex systems -- but it might simply be unnecessary overhead for entities with significantly more working memory, just as humans won't plan out a class hierarchy for a 5-line script, much less write a PRD.
Yes, directly emitting binaries will likely become practical at some point if LLM context lengths continue to increase. We may be there already for small programs, I don't know.
As with eliminating other forms of abstraction, there's not much gain from doing so. Writing source code and then compiling it offers many useful checks, and greatly simplifies making changes. It's just something that may no longer be necessary at some point.
I’ve been there, your “high standards” is just noise and annoyance for other people, focus on business value instead of jerking on variable names and “perfectly” aligned folders.
Ah, I think some people who say "coding is solved" implicitly also mean "you don't need to read the code that AI generates". It's not a super clear proclamation, but it is certainly provocative!
Between Opus 5.5 + GPT-6.1 Sol reviewer, most issues and unhandled edge cases are caught during implementation.
Then you QA test and provide the feedback.
I guarantee that this is enough to produce not only a working product, but is well capable of building apps much better than what is out there, without ever looking at the code.
Depending on your product and risk tolerance you may disagree. I believe that for the vast majority of use cases the approach I described, combined with extensive manual testing, is entirely adequate.
I run a solo SaaS that is closing in on $500k ARR, so I have real skin in the game here, and I agree with you. My experience of writing code with agents over the last 18 months in particular has been a steady erosion in my need to read the code. There are still a few parts of the codebase I'm particularly cautious about doing code review on, and I spot check PRs outside of that, but Opus 5.5 + Sol / Astra / Fable for reviews has been another paradigm shift. I don't think I"ve opened my IDE in weeks at this point.
It's worth noting that I also have a lot of guardrails and custom tooling around my own process to check code quality, QA functionality, limit blast radius of bugs, add observability, make deployments and rollbacks quick and low risk, etc.
Of course, if your objective is simply to make money, this might be relevant. You can also make quick money by polluting the environment, or being devious, or... there are many more direct ways to make money.
If you want to provide a good, secure, user-empowering, performant product, you cannot delegate everything agents, I assure you. And, building your foundation on un-reviewed mess of code will come back to bite you (partly depending on how big you grow), the same way it does with humans, which management always think is not an issue until it becomes one.
Now, you don't have to take my word for. Think of this: why is every change to Claude Code making it worse now? It is not almost unusable for me after many regressions over past hew years in its UX. Is it because of a lack of AI being used in its development?
Because by my estimation we did build our products in such a way and have that experience, while you presumably did not.
As for Claude Code, they recently delivered a significant upgrade to the desktop app where some commented that it felt like they upgraded Opus 5.5 to 5.6.
It's in fact working great, and the days where it was buggy are in my view in the past.
I mean...yeah, my objective here is to make money. It's a business. It provides for my family. I find it humorous that you feel the need to diminish it and compare it to polluting the environment or being devious, but you do you.
As for the rest, I assure you that the product is much more performant, secure, and delivers far, far more value to my users than I ever could have done prior to AI-assisted coding. And I'm a staff level engineer with 25 years of coding experience across all kinds of organizations. I'm not the best coder, but I'm at least decent. AI has been a massive force multiplier for me, even as I agree that it's not perfect.
Unpacking a little more, I think the area you might be running into issues with is the idea that because I rarely open my IDE, I'm "delegating everything to agents". I'm really not. I review PRs (just not in my IDE), I steer higher-level abstractions in the broader system. To me, automated coding with a frontier model is definitely good enough that it's utterly pointless of me to read the internals of 99% of functions, even classes. Assuming the inputs are correct, the agents will get those details correct. And for the tiny fraction of the time when that's not the case, that's why we have adversarial review, guardrails, tests, o11y, quick rollbacks, etc, etc. Working at the level of how the broader building blocks of the system fit together, bisecting the software into low / medium / high risk areas so I can direct more of my scarcest resource (my attention) towards the highest ROI portions, and building the infrastructure (the "software factory") that the agents work within building and improving the product is where I can add the most value in November 2026.
Have you ever worked in a bigger tech company with tens of millions of lines of code and senior staff / principal engineers? In my experience, sure they're reading and writing code here and there, but that's not the value they provide. The value they provide is the broader architectural view of the systems and how they fit together. It's no different with AI.
Code written by not very good coders has the possibility to be short and to the point. Even if I think the architectural decisions are wrong, the code still communicates a motive and an understanding of the problem at hand. This is all lost with LLMs.
In my experience, it's very good at following standards, so good in a structured project. Not so much in a freeform project where things get out of hand pretty quick.
I'm curious, what language do you code in? I use Go for a high scale backend service at this point it is better at writing code for a jira ticket than most of my team.
There is still questions about if that jira ticket should have been written but from ticket to mergable PR it's better than me.
Not to call you out specifically but I cannot believe people still use this as an argument in 2026. As if someone writing bad code while at least trying to learn how to write good code, and a prompter prompting a computer to write bad code that leads to the person learning nothing from the exercise are at all analogous.
I can virtually guarantee the bar for code can go so much lower as we rapidly deskill programmers.
There is no substitute for reading the code in many many cases. I spend more than 3k/mo on tokens. I find incorrect assumptions baked into the code (the ultimate spec) and missed by the agents and tests daily. This is more pronounced in data analysis and complex systems.
As a side effect having to read the code at least once inevitably makes it way shorter and usually much better.
That's right. And you have to read what the agent is doing as well. Once I said (Opus 5.5), given this list, find issues that are both critical and related to {X feature}.
Guess what: because I said "both", Opus decided I was looking to pick out exactly two of the issues. If I hadn't read that reasoning log, I would've remained oblivious and thought there were only two it could find.
> Guess what: because I said "both", Opus decided I was looking to pick out exactly two of the issues.
That seems like a misinterpretation of English that I'd expect from a much less capable model. Like, a complete lack of understanding of grammar. To have that meaning, you would have to have put "both" before "issues" instead.
To be fair, that's a pebkac problem. You got exactly what you requested (and the language was unambiguous).
This is like writing `if (cond_x && cond_y) {}` and then blaming the compiler because it didn't realise that actually meant to write `if (cond_x || cond_y) {}`.
That's not what they said, they said it limited its output to 2 issues because it saw the "both" in the sentence, which is clearly an incorrect reading of their instructions.
Yes, I think I misread -- looks like the pebkac problem was at my desk instead :/
But I don't think their prompt as written would have limited the responses to two, unless there really were only two matching both conditionals. I've not seen an LLM misinterpret language that way before.
You're right, but I do wonder if they're misremembering their prompt a bit. The AI's response would be much more likely to arise if the prompt was something closer to "find both issues that are critical and related to X". Based on my own experience, it would be a very surprising mistake for the AI to have made if the prompt really was as straightforward and precise as "find issues that are both critical and related to X". I know it's a bit annoying to question someone's account of their own experience, and obviously I'm not saying this couldn't have happened, but my guess would be that the actual prompt was at least somewhat ambiguous.
Exactly, I see insidious errors made by Claude Opus daily, all the time. Errors that would be very hard to catch with tests, because usually they are architectural level issues.
> The biggest change has been Google’s AI summary. Google’s first or second result is almost always an AI summary, and it’s usually decent.
Honestly this makes me skeptical of the whole piece. I get a pretty strong Gell-Mann amnesia effect from it[0]. It is not quite reliable enough that you can trust it, as when it is wrong it tends to be VERY wrong. In addition it is not like 99% accurate or something, but closer to like 80-90% or something.
[0]: The phenomenon of a person trusting newspapers for topics which that person is not knowledgeable about, despite recognizing the newspaper as being extremely inaccurate on certain topics which that person is knowledgeable about.
OP: [last year][1], I shared how I was using AI day-to-day as an ordinary engineer. This year, I did it again!
It's crazy realizing that Claude Code only started getting big in the last year or so. It feels so different now, with all the new models & tools we have to use. I hope you enjoy!
March 2025 was the start of the boom, IMO. Sonnet 3.7 and CC released feb 24th, Responses API + Agents SDK from OAI launched March 11. Two months after we got the first Terminal Bench, then a month later we got tau2 bench.
I guess, its debatable that o1 was the real start- but I think its hard to deny that Claude Code CLI was the first time we had what we now call a coding harness hit the mainstream. And- Sonnet 3.7 was kinda Anthropic's "o1 moment".
God. Opus 4.6 and 5.3-codex feel like ancient history.
> I also now reckon with the fatigue of using these agents. Managing 6 different conversations with 6 different robots about 6 different features is overstimulating & exhausting in a way that coding never was.
Anyone using a swarm of agents in November 2026 to write code is no longer at the frontier of capability.
Running multiple system prompts at once is a perfect example of anthropomorphism. We have this fantasy that a community of diversely aligned agents will somehow outperform a single well-aligned agent. There might be some cases where the random walk takes you where you want to go, but how stable and stressful do you want it to be?
Some of us are using AI as a means to make a living now. If I slip even a little bit on best practices my clients will likely turn into ghosts. Multi agent is an extremely expensive form of entertainment in this context. It's the entire basis for why we have these junkies trolling around HN complaining about how they can never get enough tokens. If I was still playing with these ideas in front of my customers I can guarantee I'd be 100% unemployed right now. They brought me in to make AI into a profit center, not a cost center.
The frontier of capability right now is writing most of the code by hand. I'm not joking.
Multiple studies have come out recently that AI use when not used as a tutor leads to loss of learning. It's a great way to ensure you end up jobless in the future when you forget even very basic syntax from no longer having to deal with hard problems. It's why pilots still fly manually by hand even though the autopilot can do practically everything.
Being charitable, perhaps they're mostly focused on efficiency, not overall capability. I can see that using swarms probably rarely results in solutions you couldn't reach with a single agent, at much higher cost. I think the tradeoff is that you get to the solution with less clock time though.
That said, the article text they quoted isn't even about swarms, it's just about the mental fatigue of juggling multiple single-agent threads. And there I agree with the article. I keep doing it because of agent speed, maybe I should just bite the bullet and get some $500 subs so I can use Astra ultrafast for everything. Only half joking...
The models can literally get any Jira ticket and one-shot, but we still have demand for SWE, so I don't get people saying it's gonna "take our jobs".
LLMs are just a tool, you can one-shot a full project, but you still need an engineer to validate, it will catch some adjustments, then more tickets will be created from that and the cycle goes on and on.
The "Art & music & video" section is really puzzling to me. On one hand, the author mentions liking some AI music, and then links to low-effort songs that at best you could call sort-of catchy... and on the other hand, the author writes about AI video: "Has it ever been used for good?"
Gen-AI can and is being used to make real, worthwhile movies. The best video model available, Seedance 2.5, can't produce clips longer than 30 seconds. The human artistry comes from which clips/shots you ask for and how you edit them together to tell a story. Plenty of people are doing good work here.
Here are some examples. Even if you don't like these personally, they aren't slop. You can feel the creator's intention behind them:
I enjoyed yours. TBH, I feel like some small thing is still missing. It's difficult to put the finger on it. Maybe the dialogue isn't hundred percent credible? Or maybe there is something subconscious where we notice the "actors" are not fully believable. But I think you've got talent.
What struck me is that it's really close to photo realistic, but I would still have noticed it's AI by the tiniest continuity mistakes and visual cues, like slightly unnatural color grading and such. However, if I had seen this five years ago, I would have taken it to be a real recording with post processing that makes it look a bit polished, like Benjamin Button or Grand Budapest Hotel.
Yeah there are definitely things that don't feel quite right. Part of that was my own shortcomings and the time limit on the project... but I think the tools are at the point now where with enough time and attention they can make something that feels totally real.
It was about a month, putting time into it almost every day. But a lot of that was learning how to use the platform/tools and learning how to write prompts for Seedance.
Watched the last one. Why are they drenched in sweat all the time? Also amusing they're still wrangling with certificates, a true dystopia!
I mean it's a video alright, if that's the point. Frame by frame imagery.
AI, like people, seems to get better as experience builds up. (thread context) But that experience is getting used to the workflow that fits that user, so it seems hard to know in advance. So, like the post said, since you only learn the better approach or structure "after doing it", this seems like a pretty common situation.
Whoever says they throw agents at work they (the human) are getting paid to do, i am very curious about how you make sure the thing that is being built is it's money worth.
Im doing this, against my will i preffered an anchored approach but well the thing is i work for a rich guy, like most of us, i dont care the thing has 'money worth' i only care the rich guy gives me a fraction of the money we produce.
But still, if the thing is really bad it just makes my work harder so im trying to keep it stable, i am still reading the core of the system as much as i can, the agentic pipeline in this case.
I have different reviewers working, custom company model, claude, codex they run in different stages of the projects like for SOLID, for linting, thats what we build the review agents for.
I check the architecture and design documents throughly before starting, then the AI creates documents that the AI will consume and i check the details didnt changed, the same if you need example proof of concepts, like some inherited approach from the anchored spec driven times.
At this point reading every line is not possible, not even worth it to discuss it at company level its not happening, so these kind of things are your safety net now so if anyone has other strats please do share them i think we all need them.
With all the time and resources Google has put into their products, you can probably conclude that either a) they want it that way, or b) no one there wants it to be otherwise enough to take resources away from whatever they were doing, and put it into that.
> I also now reckon with the fatigue of using these agents. Managing 6 different conversations with 6 different robots about 6 different features is overstimulating & exhausting in a way that coding never was.
6 used to be the max for me as well when I worked with Claude Code in terminal windows. Too little to hold on to and to discern the conversations. Just walls of text.
I can now go 20+ with fun because
1. I built a canvas for organizing the work spatially, so every feature has its place,
2. I condensed sessions to a WhatsApp style chat with detailed agent steps available but folded away by default, and
3. Each feature agent ends with a demo, usually a narrated video that I can watch and comment on by voice as it runs.
(My solution is available open source and free: obeya.si)
Nothing against the post itself, but something I find funny: the development in AI is so rapid, that everyone thinks they are ahead of everybody else. Here we have a blog post telling us "coding is solved". In Oct 2026, how many of us don't know this already?
This was Gemini the other day: https://imgur.com/a/7mdqwjk — this wasn't some errant load, even refresh, hard refresh, etc. got the same result for a while. An outage, AFAICT.
While trying to capture that bug, I (ironically) asked Gemini (but the Gemini-via-Google-search Gemini, since Gemini proper was an outage) for the package name of a screen recorder (gtkRecordMyDesktop); the command it returned referenced a non-existent package¹. I then found an alternative, and then hit a bug in that screen recording software that I had to then work around to get the video file there.
The site I uploaded the video to, Imgur, breaks completely — i.e., it cannot even so much as surface an error message — if your browser refuses to let it store god knows what on your disks. (So as it is, I have an extension that says "why yes, of course you can have storage, and essentially provides a black hole. But since it looks like a "success" to the Javascript, the JS doesn't crash!)
This isn't, to me, some one-off interaction. Every interaction with software these days feels like this, more and more. Just breakages everywhere that leaves me wondering if "coding is solved", where is all the bug free software?
(¹forcefully prompting the AI to actually query the package listing netted me a classic "You're absolutely right! The package name I referenced does not exist!")
> This isn't, to me, some one-off interaction. Every interaction with software these days feels like this, more and more. Just breakages everywhere that leaves me wondering if "coding is solved", where is all the bug free software?
My annoyance is more specific than "it breaks". I hate how all this software (especially "web apps") just assumes that it will have permission to do whatever it wants and never contemplates that the user might have something to say about it.
Yep. Everyone is saying that AI is writing all this great software. Where is all the great software? Really, where?
My experience is the same as yours. Every single service I use. Apps, websites, doesn't matter what. Every single one seems just softly degraded. Like I can't trust that features work any more. I could report the bug, but what's the point? It's going to go to some Claude triage agent who is going to "fix" it by slopping up the app even more. Do people even use their own websites and apps any more? Apparently not - the same people declaring software solved are having their AI agents doing computer use and going on these websites for them!
Claude, please read the Bible for me. Tell me what I need to do to get into Heaven. Oh, can you dictate it to me? I'm managing an agent swarm.
You mean the three new products we launched in a month instead of a year (or two..)? Sales are good, customers are happy, we are doing well.
I know, I know, we are shit and AIs are a godsend for small shitty companies that have at least one clever person that architect their way out of a paper bag. I’m sure all ya’ll highly competent, genius academics don’t need it and find it a hindrance but the devs in my bubble really, really need it. And there are a lot of those..
They just live in denial. I don’t have any proof, but I suspect those are the same people that would nitpick your MR to death until you hint to their manager that this is not ok.
b) you confuse technical quality with product incentives
If you see this, company doesn’t care about product. And Google doesn’t care about quality for at least a decade now, all they care is a share price and speed of delivery.
> Every interaction with software these days feels like this, more and more. Just breakages everywhere
I've noted the same but even before llms so I can't draw a conclusion there. I find strange issues with software I cannot explain. Not only breakages or bugs but also a strange lack of core features and UI/UX that leaves me wondering "am I dumb?".
> In Oct 2026, how many of us don't know this already?
The comments to this article suggest that it's still disputed.
Personally, I think it's a bit of a silly phrase, because it implies an absolute, and all anyone needs to do to disprove it is to show that coding was not solved for problem X. That's still the case -- it's just that the set of X is rapidly shrinking.
I think it's more indicative to say how much you trust LLM code. A little over a year ago, I scrutinised every line carefully, and it was buggy as hell and I had to fix most generated code by hand. Now, models write clean code with fewer bugs (far fewer) than I would. A year ago, I was doing a lot of steering during planning; now, it's rare that I don't agree with the model's first proposed solution, and even if I don't, it's most likely to fix a minor change in understanding, not to propose a substantial architectural change as it was before.
The one thing I'm not sure about is how much this is due to better prompting on my part. I doubt that this contributes much, but I still suspect that continual usage does make you more aware of what works and what does not: how much detail to provide, how to describe a problem or change, how to propose solutions while still leaving other options open.
Find a lot of tech adjacent jobs on reddit and see them talk about being outskilled by new hires using AI to automate sysadmin tasks. I used to a sysadmin very early in my career and given the tools today and the problems back then, a lot of these jobs are going to be obsoleted or dramatically automated to not require FTEs. This is already started in those fields. Powershell and bash scripting is very good from these things and has been for longer than a year IMO.
We're in our own bubbles where a good part of HN population woke up to coding agents being good enough ~1 year ago, there's a group of folks that found out ~6 months before that, and then the people banging their heads on these a bit earlier than that. If you aren't using a coding agent everyday most people are just using these things to apply an instagram filter, or sort a grocery list. Muse is probably the closest thing to 'full automation' a lot of people are seeing for the first time.
Unfortunately I don't think the bubble that exists at the moment is on the promise of automating so much coding as trying to penetrate consumer use cases, which I think are still lacking and not reliable enough yet. As much as I use coding agents every day I still will not let a claw-like near a bank account or a credit card.
The problem with Muse and similar "run your life" agents is that booking a dinner reservation or signing up the kids for soccer isn't anything that I need to automate. They need to find better examples of real problems for their marketing.
The structure of the post seems like a comparison of snapshots between a date this year and last. On HN, it's an odd statement. In the context the author is writing in, it seems fair.
I have tons of acquaintances like that.
They vibe-code and think they've gained an advantage over those who don't know how to vibe-code
But there's nothing to know — take a 3-hour course, pay 200$ a day for Claude, and there you are, a vibe-coder.
It's just that most skilled developers find it disgusting, because the AI spits out code that doesn't work the way it should, and it all turns into the routine of "threw in a prompt, waited, reviewed, repeated." — only those who were crappy programmers are happy about this.
Always have a good chuckle at these doomerism bullshit posts. Our VPs are so busy it is impossible to find a meeting time, but suddenly they’ll develop ADHD and start vibe coding enough to replace 100 devs. Lol!
Coding is solved in the sense that you don NEED to write code anymore. Of course you need to check and design and find better ways of doing things, but nothing of this requires you to write the code directly, you can always design, check and ask the LLM to implement a new version.
$3,000/month is nuts! For $36k (1 yr of spend) you can buy a system with dual rtx pro 5000 (144GB VRAM). or spend a little more ($44k) and get two M5 Ultras w/512GB RAM. the latter, clustered, is big enough to run the biggest open weights (quantized), or smaller MoE weights on each in parallel. what you get for these systems is way more than $3k/mo of tokens, because you can run it 24/7/365. you can have pass after pass after pass of optimization, bug fixing, orchestrated subagents on backlogs, etc. way more value than paying for managed inference
I guess humans have agreed, that art is a skill you must learn if you want to produce it, but engineering or software is not a skill worth learning and we are happy if computers can do that for us.
So it’s great and amazing if anyone can create their gizmos and software to help them but not if they create music or digital images, which apparently it is wrong to prompt, and not a sign of a great future.
> I guess humans have agreed, that art is a skill you must learn if you want to produce it, but engineering or software is not a skill worth learning and we are happy if computers can do that for us.
No; it's a conclusion about the product rather than the endeavour. People don't think of the software itself as "art", but as just a means to an end; so they don't mind the idea of any old fool being able to produce it.
Oddly enough, I've come to feel similarly about music. Most of what you hear in "pop music" is highly uninspired, including pre-AI; but it serves the purpose. And now we have AI that doesn't give any indication that it could compete with humans on mash-up ideas, experimental microtonal stuff, etc.; but executes "yet another four-on-the-floor, standard chord progression club anthem" very well. (Except the lyrics. Ugh.)
Isn't this directly from that most people can relate to music/art but not with programming?
If you specified to someone that the AI is replacing the farting-into-mic-core music genre, they would not care. Same with programming, the average person is not exposed to this
Because I’ve seen both and code created by your uncle might work and look good enough, and music and art created by your uncle is a fucking garbage that pollutes the internet.
This rings hollow to me for many reasons. When making CGI for movies quite a lot of the process has been completely automated for quite some time. So that scene in the movie is no longer art? Making video games you can use procedural generation to create a landscape, trees, rocks, etc. and most developers do at this point. Do we disregard all that when considering if the game is good or not? Or just blanket "it's bad because of that" like some people are doing with AI created stuff these days?
AI created art and photography keeps winning contests and people get all upset after the fact saying it's not real art. If it's not real art then why are the judges so impressed by it before they hear the word AI? I get the sentimental "but a soul that has experienced pain didn't create it" but if it's fun to look at and gives you joy is that not enough? And there is of course always a person in the loop creating it.
Definition of art:
The conscious use of the imagination in the production of objects intended to be contemplated or appreciated as beautiful, as in the arrangement of forms, sounds, or words.
A director imagines the movie scene and has a rough storyboard sketched, then CGI artists make it who then usually have some form of algorithm or ML help them. I'm sure most people would agree that's art even though the director only had his thoughts interpreted by many different people / machines. Why is that so different than a person asking ChatGPT or Seedance to create something they (and only they) imagined? Is it the effort involved? What if they spent 2 hours refining 1 photo to obsessively to get exactly what was in their imagination onto the screen?
I guess this is what is missing from vibed AI art. Usually it's someone with a quick prompt and nothing else. No effort or creativity from the human and that's probably what we don't like about AI art.
This is different from AI art looking all the same though
I kinda disagree with both points maybe because I've been using / seeing AI art extensively for a while. Some people put in absolutely monumental effort into their stuff, and if it all looks the same then toy story looks the same as ghibli looks the same as 1950s Disney looks the same as Wallace and grommit (all of which you can replicate fairly faithfully, combine into something that looks new, whatever else).
I think AI art is much like ai coding where people used 1 model a couple of times then wrote it off completely as garbage. Then years of very quick progress happened and all the 10 fingers on a hand and every picture looking the same disappeared.
If you can make it through the first 20 images on civit and tell me nobody is expressing themselves or their imagination in some of these pictures I guess we agree to disagree.
OP: my company pays for my work tokens, which is the source of the $3,000/mo. I don’t use substantial tokens with my personal subscriptions, so I didn’t factor that into my number. But good point, I’ll clarify.
> since agents write everything, it is simple to propose sweeping changes to my teammates during code review, or even write my own versions — both things unthinkable in the era of hand-written code.
That is both untrue, and should be something that occurs rarely.
Past Mid level, most dicussion in code review is centered around the conceptual level. Code is mostly use to highlight the point being made. I’ve never seen the need for someone to rewrite a patch proposed by another person. Add to it or correct misunderstandings, but never sweeping changes.
Controversial take: AI music is more interesting than the autotuned, quantized, mass-market-optimized slop that dominates contemporary pop "music".
I walked into a paint store recently and there was some interesting music playing on the PA. It was some sort of spanish-language 80s new wave I'd never heard before. It had a nice groove. I asked the guy at the counter about it and was surprised to find out it was AI.
So... was it as good as great music? No. But it beat the hell out of muzak. And it was genuinely novel in a way that pop music isn't. I will sacrifice a lot of pride for novelty.
Seems weird to contrast AI music with auto-tuned music. Most of the AI music I've heard has been tuned perfectly. And I don't generally have a problem with music being in tune.
For my part, I mostly dislike AI music. But if there was to be some AI music that I didn't already dislike for its content, then I would dislike it specifically for being AI. Call me close-minded if you will. Some people don't really like music at all. In general, I love music.
What about it? Microtonal music can be perfectly in tune. Further, when people talk about auto-tune, they're normally talking about vocals, notably absent from Angine de Poitrine.
> Controversial take: AI music is more interesting than the autotuned, quantized, mass-market-optimized slop that dominates contemporary pop "music".
> And it was genuinely novel in a way that pop music isn't. I will sacrifice a lot of pride for novelty.
For one thing, "interesting" and "novel" are not particularly the point of art. An infinite TikTok feed of AI slop is in some sense infinitely interesting and novel.
But also, those concepts are subjective. Pop music has a ton of evolution and innovation if you break it down. Sure, not every hit is doing something clever, but plenty of them are. You might not find those features to be interesting but it takes ingenuity to stand out in an era where the barrier to entry is so low.
I checked out from pop music roughly 30 years ago, when I moved out of my parents house. There is literally NOTHING that interests me in the pop music charts.
AI music is far less interesting than anything any person could create. Even if it's acoustically interesting, it's hollow and pointless because it contains no human expression.
AI music is a complete waste of compute. It exists only to further enrich an already wealthy stakeholder, and we should all be ashamed to have ended up in a place where it even exists.
It contains plenty human expression. It's ultimately made by people.
And I think it can be uniquely fun, because it allows for the very rare combination of truly off-beat ideas with excellent execution.
In real life, it very rarely happens that somebody will spend a lifetime learning, gather a band, and then record something extremely off-beat. Because there are risks to that. Destroying your image, having to find people that are on the same wavelength, having to justify the time and money it costs to record something weird and not marketable.
Sure, people make weird stuff once in a while, but very rarely, and generally it takes unique individuals that find some particular weird niche to work in, like Weird Al or Max Raabe. And even then they tend to pick something relatively tame and marketable and stick to that.
But AI can bypass all that, and that makes things interesting.
The vast majority of music is made with no regards to marketability, it's just people doing whatever they personally like. An incredibly small amount of artists have any sort of "image", that necessitates being popular already.
And often the weirdest artists are also the most prolific because they tend to like experimenting for experimenting's sake.
You need to listen more than the Top 10 charts my dude. The fact that you use Weird Al as an example of experimental musician says it all about your depth of knowledge.
> My jaw literally dropped. Do you really believe this?
Sure. It happens in every field. When you invest a lot into something, it becomes risky or unprofitable to do some things.
Tech often drives things in the opposite direction. As it gets easier/cheaper/faster to create something, doing something odd and risky becomes easier to justify because there's less invested into it and there's fewer people whose cooperation is needed.
I'm not saying nobody does anything risky traditionally of course. But there's barriers to it, which become lesser once tech trivializes a good part of the process.
> It makes sense that someone like you would like AI music, I guess. You are clueless.
I like music in general, and am fairly indifferent as to how the goose is cooked. I listen to almost everything. There's no reason to exclude AI music. If it sounds good, it's good.
What claims about musicians? The only claim I saw is that it's relatively rare to observe musicians doing the musician thing, and I think that's quite accurate because IMX musicians (in the sense that we mean for the current purposes) are rare individuals.
>Controversial take: AI music is more interesting than the autotuned, quantized, mass-market-optimized slop that dominates contemporary pop "music".
Controversial take: AI shows that people simply don't give a shit and just confirms that mass culture is a mistake and the death of everything that is good and sacred.
>Controversial take: AI shows that people simply don't give a shit and just confirms that mass culture is a mistake and the death of everything that is good and sacred.
Controversial take: For ages, it's been clear that most people have never given a shit, it just took AI for most to realize it.
Look, I don't know the song so I can't say how novel it is, but to me, it completely misses the point if its made by a music generation model. The point of music is not to make a bunch of novel sounds. Music is a means by which a human can express themselves, the whole _point_ of art is human expression.
Art is subjective. It sounds like AI music did evoke an emotional response from OP, which IMO is also just as important of what we define as the purpose of art: evoking emotion and awe from consuming art. A connection is being made from the creator to the consumer. I view AI as merely another tool to that means.
Human expression with art can go both ways from producing and experiencing.
I don't know what to tell you, but me and my kid are having a blast writing ridiculous songs and plopping them with song descriptions and some of those make it to our drive playlists.
Respectfully, the primary thing you are describing here is spending time with your children. Same could be said for spending time and attention discussing children's drawings.
<< you are describing here is spending time with your children.
Note: same guy; just not anywhere near my account
You are making the argument that the type of activity does not matter as long it is 'spending time', but to say that is almost reductionist to say that that it understates how ridiculously unnuanced an opinion it is. You might as well tell me that the quality of food does not matter and generic burger should suffice.
But what if I want a burger that is custom to me and not a generic slop generated by a human to appeal to the widest possible audience[1]? What if I want something that was created just for me? Good grief? Why do I have to defend having fun and why on earth would you assume that one is mutually exclusive to the other?
The problem with AI performances is that they're too perfect, which leaves them flat. The dynamics and imperfections of human performance have yet to be replicated. That affects every human performance whether or not you notice you're noticing (but you are). AI is still in the uncanny valley here.
This sounds like CD music 40 years ago. All my friends said that it was not like a record and left an empty feeling in the room. I feel like this is not talked about any more. People listen to digital music and have forgotten there was an alternative. I suspect that in the coming years no one will think twice about AI generated music.
OK, yes, I am not a pro dev... still, that guy blows more tokens a month then I have consumed since the conception of GPT3! Granted, I stopped using agents a few months ago, when I got my first "you reached 80% of your monthly limit" message. Since then, I am manually constructing the context and executing edits... which brought my token bill down at least 5x and gave me 2x more useable code. But yeah, I am sure, 3k$ tokens a month is not wasteful behaviour at all :-)
I find that AI should be used more extensively for System Design. I have seen a lot of people not use AI for that and come up with frighteningly complex designs and let their agents run loose to implement it.
Retroactively speaking, I engaged with the already agreed designs and used AI to find much better ways to do the same thing.
Overall, I wish people used AI even for System Design as that seems to have a higher impact.
> But I’m more and more excited about using LLMs as teachers to explore well-known concepts.
Me too! I've done several deep dives in various topics just in the last few weeks. I start off with some question I want answers to, and often end up with more questions along the way and I end up learning a ton.
For example I've been enthusiastic about nuclear energy for a long time and I've been wondering why don't we build more of it? It seems like such a great investment, the fuel is extremely cheap compared to the power it produces, even including disposal of waste and everything it seems to me like the best power source we have by far.
I've read things like "it's not economically viable" which I didn't really get - how can it both be cheap clean electricity and not economically viable? It turns out the problem is interest. If you spend $20B building a nuclear plant, you need to make enough money to pay the interest on that $20B. The plant itself will easily make much more than $20B over its lifetime but it won't cover the loan payments unless you can secure a contract with a customer willing to pay a very premium price for stable power for like 20 years.
On the other hand, China builds nuclear plants for like $3B which changes everything. They can do that because they've built lots of nuclear plants, so they have repeatable designs and established supply chains, experienced builders etc. We could do that too and I really think we should (we as in the rest of the world), I think we should invest in nuclear technology to build experience and supply chains etc in order to be able to make the technology economically viable for us. It seems to me that fusion will have exactly the same problem, if we ever even make it work. Good luck building "economically viable" fusion plants, they will certainly be a thousand times more complex than fission plants.
Anyway this wasn't supposed to be a comment about nuclear energy. I also did a deep dive on hydropower based on some claims I heard and learned lots of cool things about that. The LLM found lots of interesting source material for me, including a letter from our ministers of energy and finance talking about a proposed law change that might take effect next year which I would never normally read but since it basically answered many of the questions I had I ended up reading the whole thing and it was super interesting. It's also obviously a lot more credible than just the AI's claims.
Highly recommend trying it out about any topic you may be interested in. I use the free version of ChatGPT. I've also used the Google search AI a lot, I'll basically just Google a question and then proceed to ask followup questions etc.
I'm sure it's not all 100% correct but I'm also quite sure that most of it is pretty good quality, and it really helps when it includes links to sources, I often read the sources instead of the AI summary.
> It turns out the problem is interest. If you spend $20B building a nuclear plant, you need to make enough money to pay the interest on that $20B.
This is exactly the sort of investment governments are suited to make since they have access to extremely cheap debt. Or at least they did until recently, though still cheaper than private markets. China has access to debt with interest of only 1.6% which allows these projects to be very viable.
On the one hand, I like this post: Like developing a personal Mission Statement (one of the best pieces of advice a mentor gave me long, long ago, in a galaxy far, far away). But on the other, this sounds like the beginning of AI psychosis, or somebody that got really drunk one night and decided to blog about how amazing compilers are: "I love you, man!"
That being said, I agree with this statement: "I believe they are the greatest teachers humanity has ever created." LLMs are exceedingly powerful at accelerating the learning process.
It’s kinda nice seeing more nuanced takes from folks as the bubble kinda deflates and both sides realize their idealized vision isn’t coming to pass. AI isn’t going to replace all the jobs, but it’s also not going away; finally, it is merely a tool in our toolchest.
What does amuse me to no end, however, is how many craftsfolk in AI circles are loudly rejecting AI imagery and video in favor of humans. The only folks who seem fine with that are the grifters looking to slop out a bunch of adverts quickly for commercial gain rather than do actual art, or folks who want “art” but don’t want to pay a human to make it for them or learn the skills themselves. Both of those camps are viewed with disdain and loathing by both sides of the AI Boom/Doom arguments, so I’ll take whatever allies I can get on those fronts.
AI music…man, I live with artisans and musicians and we have a number of “artists” we suspect are AI but whose music is pretty solid. We’re waiting for the shoe to drop, but I’m trying to look at it as a kind of turbocharged Fruity Loops/Logic as opposed to a human replacement. I hope I’m right on that, but time will tell.
The game changer I think more doomers need to experience is agents. It sounds like a recipe for disaster (a bunch of models known to hallucinate and allowed to run roughshod over a computer unattended? What could go wrong?), and it often (hilariously) is for the scammers out there giving it full access to production and fucking off to hawk their latest Masterclass on X or LinkedIn, but that first time you give it a reasonably achievable prompt and come back to a halfway decent product you can review and refine into production in 15m or less?
Absolutely magical, doubly so when it’s just 27B Qwen via Hermes and not some souped up stack in a frontier lab costing you thousands a month in tokens. If I were to evangelize AI, that’s how I’d do it. Fuck subscriptions, fuck cloud compute, fuck frontier labs and token buckets and whatever slop Sundar/Scam/Dario/Elmo et al are hawking this week; run that locally, give it a reasonable prompt that solves a real problem, and show them the results.
Then do the right thing and go for a walk with the two hours of free time the AI just freed up for you. Don’t ask for permission to take back your time.
I’ve lived through three previous financial bubbles, and their most common characteristic was that they didn’t ‘kinda deflate’. Bubbles pop, dramatically, they don’t just deflate slowly. They don’t withstand shocks and bad news that have had the market up and down like a demented bunny over the last few months.
Market volatility "over the past few months" has been very much within normal variation. Look at even a 5-year chart of VIX if you don't believe me. For that matter, what are you even counting as the last three "financial bubbles"?
you can totally tell Claude to take some AI technical writing (which is, to be clear, writing where you explicitly need to not sound creative or interesting, and you need to maintain very strict "voice" [a term I only even know because claude explained it to me] so this is not a creative endeavor) and make it sound "less AI". At least working with opus 5.5, textual output already has a lot fewer of the colons and "load bearing" of a few months ago, and if a PR does have that kind of lingo I can just have Claude rewrite it "like a human" and it does great.
OP: I'm not surprised you/other people are able to get better output than me.
And I agree that "less AI" edit passes reduce a lot of the hallmarks, but I think these models truly struggle with expressing complicated ideas. Have Opus been good at that for you?
> I'm not surprised you/other people are able to get better output than me.
I think people should find this sort of result surprising by default. Like, presumably you don't know why your output isn't as good, or you'd fix it. So doesn't that lack of understanding concern you?
> I still believe I use AI fairly typically, but I’ve had a notable change in background. Last October, I was an engineer on Microsoft’s Windows AI team. I am now a research engineer at Salesforce AI Research.
Priest wears robes.
I guess it’s one of those nonsensical automatic rules.
> I am now a research engineer at Salesforce AI Research. I spend more time supporting scientists & building rapid prototypes, rather than working with a large team of engineers.
I appreciate the PoV of people with different challenges, but my full-time job is not about building rapid prototypes so the guidelines are worth a bit less to me.
Over the past year, we’ve been using AI to solve everyday challenges and troubleshoot issues we encounter in our jobs. Now, we’re taking the next step by leveraging AI to build solutions, automate workflows, and improve the way we work.
> My most successful work this year was hand-writing a tool, replacing ~4,000 lines of vibe-code with a few hundred. My system has proven more flexible, easier to test, and more generalizable than what it replaced.
> agents still need human taste and experience to produce good work.
Where do you think human taste and experience in software comes from?
> overstimulating & exhausting in a way that coding never was.
> for many of us, honing the code, finding the right design, seeing the parts mesh together — that was the meditative purpose we found in our jobs.
> That’s not the only time AI waylaid me. LLMs could not figure out how to annotate that model’s training data correctly. I spent days going back and forth with them
> LLMs are not yet good companions when brainstorming complicated things.
Why do we expect this to get better? The training data will degrade in quality if more people rely on LLMs so I'd expect models to eventually start getting worse even if they currently enjoy apparent improvements in output.
> Writing code is solved.
What a strange thing to say given all the above experiences, which I think we all share.
Why are we willing to give up so much in terms of purpose and understanding for perceived productivity gains?
"Perceived productivity gain" is a baseless conclusion. At this point it's abundantly clear that the productivity gain is real and massive.
Where I see the incoherence is in that people see too much value in code review and optimizations.
I believe many of the optimizations are premature. If you did not spot the problem in your QA testing, chances are it was not that important in the first place.
If the architecture wasn't perfect, it is now extremely cheap to change it later, as you realize real deficiencies. The agent will do it for you in a matter of hours, and largely on its own.
You will have to budget a lot of QA effort for verifying the significant changes before going to production, but I don't think it's a deal breaker.
Oh? Where is it? Why am I not seeing it as a user if it's abundantly clear? In none of the software I've used frequently in the last year have I seen any of those “real and massive” improvements. Not saying they don't improve at all, but the rate hasn't increased since the advent of LLMs.
I'm still waiting for that shiny new era of software productivity to impact any actual users.
Its not a direct analogy, but this seems to me as like saying "I already know maths and english, why do we still hire and pay teachers". Its a massive issue that just seems to be hand-waved away as "we'll figure it out later".
To compound on the problem as well, even those of us who do still write code by hand are actively discouraged from sharing it, knowing no human will ever read it and we'll never be credited for our work, it'll just go in the pockets of a few big AI companies. I suppose there will be some that take some pride in knowing their code is helping AI models advance, but thats a small pool of an already small pool, assuming most people decide to just use the AI to write the code for them. A minority of a minority.
I guess this is the chickens coming home to roost of this whole attitude towards growth now, fallout later
As someone who wrote code for 27 years in all sorts of languages, not necessarily impressive code, but got paid for it, made a career, and changed life for the better multiple times, I don't think coding was ever an important part of the journey! Maybe it's cognitive dissonance, as you wrote above, but I believe taste comes from exploration, from trial and error, from being willing to be open and change old ways if new ways can improve how things are done!
Similar to taste in food or even music! You might think certain foods taste bad, but until you taste them, you won't be certain! Sometimes your senses will be right and the food will taste horrible; other times, you'll be fairly surprised. Similar to music—it's my life hobby—you might think certain genres or artists are bad, but until you sit down, give them a shot, and hear what they have to say, you won't know for certain if your judgment is correct.
If you meant by your comment that you need to code a lot to acquire taste and experience in software, I respectfully disagree. You are not completely wrong, though, but putting emphasis on code, if that's what you were going for, is an overstatement of code's importance. For that particular case, similar to the analogies mentioned above, it's about exploration and trying different languages, paradigms, solutions, and stacks, and unfortunately, we as humans have finite time, so we can't be trying and exploring fast enough to form judgments!
What I genuinely love about LLMs is mostly their exploratory powers. I try new stuff that I always wanted to try with actual products and form opinions about the whole experience.
Anyways, as I finished writing this, I realized, as always, I might be wrong! You might be right, taste is highly subjective! And acquiring it is a very personal thing! So I don't think right or wrong are important here!
You are assuming that AI should be trained on speicific field to reason better but it actually doesnt need to.
RL is not dependent on training data. The mathocalypse didn't happen because they found better training data somewhere.
There are reasons to suspect that OpenAI may have intentionally or unintentionally trained on chat logs from mathematicians--mathematicians who were actively using OpenAI to try to to solve the same problems themselves.[1] If those mathematicians were on the right track, and getting close to a solution, that could be classified as finding "better training data somewhere."
[1] https://cybernews.com/ai-news/openai-accused-stealing-anothe...
I agree it’s unlikely this happened by accident, or entered training data (though not impossible), but it could certainly be within the context agents were using when tackling problems.
We have no evidence either way, but they had all the data they would need to do this.
I actually found this to be one of their strong points. Depends on the person / thinking style / communication style, I guess.
Secondly most the posts on HN are from individuals. It doesnt represent group dynamics and how group/institutional knowledge works. For example there might be a regulation on publishing data every quarter. On the surface it looks like compliance but in reality different internal groups will hide data, delay it, fudge it etc etc. LLMs train on public data. Only after they are allowed access internally do they discover group knowledge. We are far away from seeing that happening.
This always reads like “writing english is solved” just because you collected snippets of famous works and have built a spell corrector.
Code is written to process data. The struggle was never to learn programming syntax (most grammar are only a few page long). The main struggle was that it’s hard to model a domain into an information system, identifying the components and their relationships, what is ingested, and what needs to be output.
Computers have integers and floats, but you need to implement units like distance, money, various counts,… The real world have words and sentences, but that needs encoding like ascii and unicode as well as the “string” concepts to make it work. Take HN, I’d bet that if you can model it in terms of data, writing code is not an issue.
If anything, I do expect the casual non-tech users to be making the most adaptable software because they have what they want in their head. Although, part of what the dev does is figuring out the weeds, and only a part of users have the patience to sit through detail corrections. And the AI user has to verbalize their wants exactly too.
Those are the people who are most likely to give a monolithic prompt and then spend a week telling the machine 'I don't care about encryption or file permissions, just make it work.' Not to mention that most people can barely verbalize the problems they have with devices they use for 8hrs a day, let alone form a full, self consistant, project idea and then put it in writing.
Imagine you could build a piece of furniture without actually needing to know how to cut wood, drive nails etc. Imagine you could do it in 3d without knowing how to actually handle wood, you just have an image you can iterate on without actually costing you anything but time.
I remember the first big app I wrote in high-school. I had to learn C++, learn Qt etc. I knew what I wanted but putting that into code when I didn't know any of the libraries / tools was a huge time investment. If I had LLMs I could've done the same thing it took me 6+ months in probably less than a month.
They may have a perfectly good shot at being smart, but I still don't buy the argument. The reason they struggle to become coders (assuming they tried at all), in my experience, is a lack of problem-solving and debugging skills. There is no real shortcut here. An "image" of the chair isn't enough; when the first iteration is wrong, you have to be able to explain why it's wrong.
With a sufficiently good LLM these people will skip the tedious part of programming that most of us had the motivation / patience to go through. All the folks saying that actually writing code wasn't the bottleneck have had a LOT of years to get to that point, years a lot of people aren't interested in going through.
I'm not saying lay people will be able to join mega projects and work at FAANG. I'm saying a sufficiently courageous lay person will be able to make an app to automate whatever tedious birocratic process their niche work hasn't seen a product yet.
Even with basic pseudocode and very lenient logical rules, most people won’t bother to provide a a good definition of some process, as it’s quite unnatural to be that logical. Instead, it’s mostly just playing around. When you ask the one that is supposedly 10x-ing his productivity to explain his mental model of their own projects, it’s often a wishlist instead of a real understanding of what he has built.
For 85-90% of devs, yes this was the struggle. Creating and updating rows in a database from user input is all the “domain modeling” there is in most software. That’s what is solved by AI and it’s the majority of software dev.
Those who think the challenge is to overreach the limitations imposed by programming languages and types, etc, to do seemingly "magical" things. These are ones who would never understand why you might want higher-kinded types, the relative pros and cons of the functional vs imperative approaches, or even of type systems in general. They also tend to think the code written, in itself, is the valuable output of their work.
And then there are those who think the challenges is to model problem domain so well, using the language features, types and data structures, etc, that the solution (and further adjustments to it) becomes obvious and straightforward. They also tend to think the analysis and model that results is what is valuable.
Another way to see this difference: what impresses them? it is a hack that uses arcane knowledge to bypass software and hardware "limitation" or structures? Or is it code that is simple but still effective at solving complex problems in elegant ways, that may look simple but reflects a huge amount of knowledge, skill, and architectural intuition built up over many years?
I of course think that the second group are more effective and are doing proper programming, but many programmers and the general public think otherwise, or can't event tell the difference. I once worked with some colleagues could never see the light, despite a lot of support from me.
So here is the kicker, the dissonance in response to AI, I think, is kind of analogous to these two groups. The first groups likes that the AI will not stop them, but rather get them much quicker to their "magic", whereas the second group is more skeptical and sees many issues with AI not seeing the big picture and understanding deeper intent.
The issue here is that you are learning the syntax at the same time you are learning foreign concepts, and the syntax is often confusing and counter intuitive
Learning the syntax separately is of no use because without a mental model you can't do anything with it
This is why children can easily learn enough programming to create games, because it’s a domain they can comprehend easily. Just like they can use english to write about their daily lives. It will be simple sure, but it will work.
So learning to code is fairly easy, you can easily create simple programs after a weekend. More complex programs will still use most of the same basic concepts as far as coding goes, but you will need to master more complex concepts, and that takes time. And such concepts are not about coding, they are about the nature of computing or the domain the software is created for.
This is why experienced devs only need a few days doing kata to get up to speed with a new programming language. It’s usually the same old, with slight differences. What was always needed was an understanding of computing and the ability to break down problems and create a solution that can be computed.
Coding to me is the same as drafting an engineering schematics, or writing a sheet music. It’s just a tool, not the end result.
The struggle is not where to put a semicolon. The struggle is, how do I handle validation problem X with framework Y.
I see this as HN bubble. For many on here it wasn’t a struggle, but for most people I think it is. Hence the SWE salaries.
However, I'm still split. Because this could mean that we see a jump in code quality, due to bad coders getting someone better to code for them. But we could see a decrease in code quality, due to less people "caring".
I’m probably the best coder in my org, which doesn’t really say much, but at least allows me to judge about the room temperature. It is solved. I don’t say that’s you, but the only people who say it isn’t “that good” are the same people who used to yak shave, rename and move things 10 times until it “feels right”.
Spec was good, all tests green, app worked. If I didn't actually go in and try to understand the code I would have shipped it like this.
So here is my question to people saying it's solved : do you actually know that it is or does it simply appear so because all the tests are green?
First of all, it sounds unlikely to me that such an implementation was made with Opus 5.5 while it never mentioned the workaround.
Secondly, if the performance was indeed inadequate, that's something I would have caught during QA testing and simply have had the agent fix it.
First, you question the model. I have reams and reams of conversations with Opus 5.5 that shows it makes architectural messes all the time
And then you are saying you needed to have performance tests (which many devs don't even bother with) to catch bad architecture?? lol
So what if the architectural problem is not one that would be caught with a performance test? It seems you lack imagination to think of all the ton of issues that that would be very complicated to catch with tests.
I don't think I commented on performance tests at all.
Maybe I lack imagination here, yes, but generally I go by experience rather than talking about imagined issues. That is precisely what many here seem to do when it comes to explaining why AI couldn't work for them, and why reviewing the code remains necessary.
When I looked at a benchmark, it suggested Sonnet 5.5 wasn't much cheaper than Opus 5.5 at Low to Medium effort for similar performance. That might not be representative of real world use though.
Optimizing performance is one of the tasks best suited to AI, since the outcome is measurable.
I can't comment on complex algorithms specifically, but it's not at all the case that this works only for basic apps or cannot handle complex, performance critical work.
I think what people are missing is that people often become craftsman because we want to be able to keep the code adaptive and performant so we can provide business value quickly and efficiently.
The change is with models, the benefit dynamics of immaculately structured code changes dramitically. Now the real area you need to focus on is correctness and performance. Structure doesn't really matter unless it serves those two purposes, at least from the business value side.
As with eliminating other forms of abstraction, there's not much gain from doing so. Writing source code and then compiling it offers many useful checks, and greatly simplifies making changes. It's just something that may no longer be necessary at some point.
… Yes? Of course recognizing the flaws in the code correlates with having high standards. That doesn't mean the flaws aren't there.
Between Opus 5.5 + GPT-6.1 Sol reviewer, most issues and unhandled edge cases are caught during implementation.
Then you QA test and provide the feedback.
I guarantee that this is enough to produce not only a working product, but is well capable of building apps much better than what is out there, without ever looking at the code.
Depending on your product and risk tolerance you may disagree. I believe that for the vast majority of use cases the approach I described, combined with extensive manual testing, is entirely adequate.
It's worth noting that I also have a lot of guardrails and custom tooling around my own process to check code quality, QA functionality, limit blast radius of bugs, add observability, make deployments and rollbacks quick and low risk, etc.
If you want to provide a good, secure, user-empowering, performant product, you cannot delegate everything agents, I assure you. And, building your foundation on un-reviewed mess of code will come back to bite you (partly depending on how big you grow), the same way it does with humans, which management always think is not an issue until it becomes one.
Now, you don't have to take my word for. Think of this: why is every change to Claude Code making it worse now? It is not almost unusable for me after many regressions over past hew years in its UX. Is it because of a lack of AI being used in its development?
Because by my estimation we did build our products in such a way and have that experience, while you presumably did not.
As for Claude Code, they recently delivered a significant upgrade to the desktop app where some commented that it felt like they upgraded Opus 5.5 to 5.6.
It's in fact working great, and the days where it was buggy are in my view in the past.
As for the rest, I assure you that the product is much more performant, secure, and delivers far, far more value to my users than I ever could have done prior to AI-assisted coding. And I'm a staff level engineer with 25 years of coding experience across all kinds of organizations. I'm not the best coder, but I'm at least decent. AI has been a massive force multiplier for me, even as I agree that it's not perfect.
Unpacking a little more, I think the area you might be running into issues with is the idea that because I rarely open my IDE, I'm "delegating everything to agents". I'm really not. I review PRs (just not in my IDE), I steer higher-level abstractions in the broader system. To me, automated coding with a frontier model is definitely good enough that it's utterly pointless of me to read the internals of 99% of functions, even classes. Assuming the inputs are correct, the agents will get those details correct. And for the tiny fraction of the time when that's not the case, that's why we have adversarial review, guardrails, tests, o11y, quick rollbacks, etc, etc. Working at the level of how the broader building blocks of the system fit together, bisecting the software into low / medium / high risk areas so I can direct more of my scarcest resource (my attention) towards the highest ROI portions, and building the infrastructure (the "software factory") that the agents work within building and improving the product is where I can add the most value in November 2026.
Have you ever worked in a bigger tech company with tens of millions of lines of code and senior staff / principal engineers? In my experience, sure they're reading and writing code here and there, but that's not the value they provide. The value they provide is the broader architectural view of the systems and how they fit together. It's no different with AI.
This is very much not my experience.
There is still questions about if that jira ticket should have been written but from ticket to mergable PR it's better than me.
I can virtually guarantee the bar for code can go so much lower as we rapidly deskill programmers.
As a side effect having to read the code at least once inevitably makes it way shorter and usually much better.
Guess what: because I said "both", Opus decided I was looking to pick out exactly two of the issues. If I hadn't read that reasoning log, I would've remained oblivious and thought there were only two it could find.
That seems like a misinterpretation of English that I'd expect from a much less capable model. Like, a complete lack of understanding of grammar. To have that meaning, you would have to have put "both" before "issues" instead.
This is like writing `if (cond_x && cond_y) {}` and then blaming the compiler because it didn't realise that actually meant to write `if (cond_x || cond_y) {}`.
But I don't think their prompt as written would have limited the responses to two, unless there really were only two matching both conditionals. I've not seen an LLM misinterpret language that way before.
0. "find issues that are both critical and related to X"
...could be read as "issues that are X and are Y":
1. [valid, intended] find [all] issues that are[, at the same time,] 1) critical; and 2) related to X
...or as "issues, both of which issues are X and are Y":
2. [valid, your reading] find [an to-be-specified number of] issues that are, both [so I want two issues], possessing X and Y attributes.
I'd be curious to learn how you're using them.
Honestly this makes me skeptical of the whole piece. I get a pretty strong Gell-Mann amnesia effect from it[0]. It is not quite reliable enough that you can trust it, as when it is wrong it tends to be VERY wrong. In addition it is not like 99% accurate or something, but closer to like 80-90% or something.
[0]: The phenomenon of a person trusting newspapers for topics which that person is not knowledgeable about, despite recognizing the newspaper as being extremely inaccurate on certain topics which that person is knowledgeable about.
It's crazy realizing that Claude Code only started getting big in the last year or so. It feels so different now, with all the new models & tools we have to use. I hope you enjoy!
[1]: https://ben.stolovitz.com/posts/how_use_ai_oct_2025
I guess, its debatable that o1 was the real start- but I think its hard to deny that Claude Code CLI was the first time we had what we now call a coding harness hit the mainstream. And- Sonnet 3.7 was kinda Anthropic's "o1 moment".
God. Opus 4.6 and 5.3-codex feel like ancient history.
On the CLI side Aider had a good CLI product and the only relevant benchmark back in the day.
Anyone using a swarm of agents in November 2026 to write code is no longer at the frontier of capability.
Running multiple system prompts at once is a perfect example of anthropomorphism. We have this fantasy that a community of diversely aligned agents will somehow outperform a single well-aligned agent. There might be some cases where the random walk takes you where you want to go, but how stable and stressful do you want it to be?
Some of us are using AI as a means to make a living now. If I slip even a little bit on best practices my clients will likely turn into ghosts. Multi agent is an extremely expensive form of entertainment in this context. It's the entire basis for why we have these junkies trolling around HN complaining about how they can never get enough tokens. If I was still playing with these ideas in front of my customers I can guarantee I'd be 100% unemployed right now. They brought me in to make AI into a profit center, not a cost center.
Multiple studies have come out recently that AI use when not used as a tutor leads to loss of learning. It's a great way to ensure you end up jobless in the future when you forget even very basic syntax from no longer having to deal with hard problems. It's why pilots still fly manually by hand even though the autopilot can do practically everything.
Being charitable, perhaps they're mostly focused on efficiency, not overall capability. I can see that using swarms probably rarely results in solutions you couldn't reach with a single agent, at much higher cost. I think the tradeoff is that you get to the solution with less clock time though.
That said, the article text they quoted isn't even about swarms, it's just about the mental fatigue of juggling multiple single-agent threads. And there I agree with the article. I keep doing it because of agent speed, maybe I should just bite the bullet and get some $500 subs so I can use Astra ultrafast for everything. Only half joking...
LLMs are just a tool, you can one-shot a full project, but you still need an engineer to validate, it will catch some adjustments, then more tickets will be created from that and the cycle goes on and on.
Gen-AI can and is being used to make real, worthwhile movies. The best video model available, Seedance 2.5, can't produce clips longer than 30 seconds. The human artistry comes from which clips/shots you ask for and how you edit them together to tell a story. Plenty of people are doing good work here.
Here are some examples. Even if you don't like these personally, they aren't slop. You can feel the creator's intention behind them:
https://www.youtube.com/watch?v=2__VhU7_3lw
https://www.youtube.com/watch?v=gl3oob03rdE
https://higgsfield.ai/@nue_nue/projects/listen-to-meh
https://higgsfield.ai/@sssacccade/projects/hunger-ai-short-f...
(This one is mine, just FYI) https://www.youtube.com/watch?v=g8XUH2OXXfA
What struck me is that it's really close to photo realistic, but I would still have noticed it's AI by the tiniest continuity mistakes and visual cues, like slightly unnatural color grading and such. However, if I had seen this five years ago, I would have taken it to be a real recording with post processing that makes it look a bit polished, like Benjamin Button or Grand Budapest Hotel.
But still, if the thing is really bad it just makes my work harder so im trying to keep it stable, i am still reading the core of the system as much as i can, the agentic pipeline in this case. I have different reviewers working, custom company model, claude, codex they run in different stages of the projects like for SOLID, for linting, thats what we build the review agents for. I check the architecture and design documents throughly before starting, then the AI creates documents that the AI will consume and i check the details didnt changed, the same if you need example proof of concepts, like some inherited approach from the anchored spec driven times. At this point reading every line is not possible, not even worth it to discuss it at company level its not happening, so these kind of things are your safety net now so if anyone has other strats please do share them i think we all need them.
By keeping my manager, and their manager happy.
I guess he hasn't seen CRAFT (1979) yet, which would not have been possible before AI.
Not that it redeems all the other AI thingies, it's always a mixed bag, yet I can't say a single bad thing about it.
For whom thinks that there is no economic value in local AI stations beside privacy, availability, transparency, unresrtictibility etc...
Big +1 to that.
6 used to be the max for me as well when I worked with Claude Code in terminal windows. Too little to hold on to and to discern the conversations. Just walls of text.
I can now go 20+ with fun because
1. I built a canvas for organizing the work spatially, so every feature has its place,
2. I condensed sessions to a WhatsApp style chat with detailed agent steps available but folded away by default, and
3. Each feature agent ends with a demo, usually a narrated video that I can watch and comment on by voice as it runs.
(My solution is available open source and free: obeya.si)
While trying to capture that bug, I (ironically) asked Gemini (but the Gemini-via-Google-search Gemini, since Gemini proper was an outage) for the package name of a screen recorder (gtkRecordMyDesktop); the command it returned referenced a non-existent package¹. I then found an alternative, and then hit a bug in that screen recording software that I had to then work around to get the video file there.
The site I uploaded the video to, Imgur, breaks completely — i.e., it cannot even so much as surface an error message — if your browser refuses to let it store god knows what on your disks. (So as it is, I have an extension that says "why yes, of course you can have storage, and essentially provides a black hole. But since it looks like a "success" to the Javascript, the JS doesn't crash!)
This isn't, to me, some one-off interaction. Every interaction with software these days feels like this, more and more. Just breakages everywhere that leaves me wondering if "coding is solved", where is all the bug free software?
(¹forcefully prompting the AI to actually query the package listing netted me a classic "You're absolutely right! The package name I referenced does not exist!")
My annoyance is more specific than "it breaks". I hate how all this software (especially "web apps") just assumes that it will have permission to do whatever it wants and never contemplates that the user might have something to say about it.
My experience is the same as yours. Every single service I use. Apps, websites, doesn't matter what. Every single one seems just softly degraded. Like I can't trust that features work any more. I could report the bug, but what's the point? It's going to go to some Claude triage agent who is going to "fix" it by slopping up the app even more. Do people even use their own websites and apps any more? Apparently not - the same people declaring software solved are having their AI agents doing computer use and going on these websites for them!
Claude, please read the Bible for me. Tell me what I need to do to get into Heaven. Oh, can you dictate it to me? I'm managing an agent swarm.
Bleh.
I know, I know, we are shit and AIs are a godsend for small shitty companies that have at least one clever person that architect their way out of a paper bag. I’m sure all ya’ll highly competent, genius academics don’t need it and find it a hindrance but the devs in my bubble really, really need it. And there are a lot of those..
Just find it funny how vibes based it all is, capabilities, experiences, anti-ai and pro-ai.
b) you confuse technical quality with product incentives
If you see this, company doesn’t care about product. And Google doesn’t care about quality for at least a decade now, all they care is a share price and speed of delivery.
I've noted the same but even before llms so I can't draw a conclusion there. I find strange issues with software I cannot explain. Not only breakages or bugs but also a strange lack of core features and UI/UX that leaves me wondering "am I dumb?".
The comments to this article suggest that it's still disputed.
Personally, I think it's a bit of a silly phrase, because it implies an absolute, and all anyone needs to do to disprove it is to show that coding was not solved for problem X. That's still the case -- it's just that the set of X is rapidly shrinking.
I think it's more indicative to say how much you trust LLM code. A little over a year ago, I scrutinised every line carefully, and it was buggy as hell and I had to fix most generated code by hand. Now, models write clean code with fewer bugs (far fewer) than I would. A year ago, I was doing a lot of steering during planning; now, it's rare that I don't agree with the model's first proposed solution, and even if I don't, it's most likely to fix a minor change in understanding, not to propose a substantial architectural change as it was before.
The one thing I'm not sure about is how much this is due to better prompting on my part. I doubt that this contributes much, but I still suspect that continual usage does make you more aware of what works and what does not: how much detail to provide, how to describe a problem or change, how to propose solutions while still leaving other options open.
We're in our own bubbles where a good part of HN population woke up to coding agents being good enough ~1 year ago, there's a group of folks that found out ~6 months before that, and then the people banging their heads on these a bit earlier than that. If you aren't using a coding agent everyday most people are just using these things to apply an instagram filter, or sort a grocery list. Muse is probably the closest thing to 'full automation' a lot of people are seeing for the first time.
Unfortunately I don't think the bubble that exists at the moment is on the promise of automating so much coding as trying to penetrate consumer use cases, which I think are still lacking and not reliable enough yet. As much as I use coding agents every day I still will not let a claw-like near a bank account or a credit card.
LLMs still mess up a lot if you just let them loose and don't guide what they do.
To "know" this, people would have to believe it from their personal experience.
Many people have standards and experiences that differ from yours. Yes, even if they use frontier models.
It's just that most skilled developers find it disgusting, because the AI spits out code that doesn't work the way it should, and it all turns into the routine of "threw in a prompt, waited, reviewed, repeated." — only those who were crappy programmers are happy about this.
And on top of that, these characters deny the fact that agents are evolving too, and meat proxies will soon be unnecessary — or maybe already are: https://x.com/tunahorse21/status/2117353858544660585
...
> Last October, I was an engineer on Microsoft’s Windows AI team. I am now a research engineer at Salesforce AI Research.
That's generally not how "typical use" works.
> my uncle can create software
Is considered great
> my uncle can create music or art
Is not considered great.
I guess humans have agreed, that art is a skill you must learn if you want to produce it, but engineering or software is not a skill worth learning and we are happy if computers can do that for us.
So it’s great and amazing if anyone can create their gizmos and software to help them but not if they create music or digital images, which apparently it is wrong to prompt, and not a sign of a great future.
No; it's a conclusion about the product rather than the endeavour. People don't think of the software itself as "art", but as just a means to an end; so they don't mind the idea of any old fool being able to produce it.
Oddly enough, I've come to feel similarly about music. Most of what you hear in "pop music" is highly uninspired, including pre-AI; but it serves the purpose. And now we have AI that doesn't give any indication that it could compete with humans on mash-up ideas, experimental microtonal stuff, etc.; but executes "yet another four-on-the-floor, standard chord progression club anthem" very well. (Except the lyrics. Ugh.)
If you specified to someone that the AI is replacing the farting-into-mic-core music genre, they would not care. Same with programming, the average person is not exposed to this
AI created art and photography keeps winning contests and people get all upset after the fact saying it's not real art. If it's not real art then why are the judges so impressed by it before they hear the word AI? I get the sentimental "but a soul that has experienced pain didn't create it" but if it's fun to look at and gives you joy is that not enough? And there is of course always a person in the loop creating it.
Definition of art: The conscious use of the imagination in the production of objects intended to be contemplated or appreciated as beautiful, as in the arrangement of forms, sounds, or words.
A director imagines the movie scene and has a rough storyboard sketched, then CGI artists make it who then usually have some form of algorithm or ML help them. I'm sure most people would agree that's art even though the director only had his thoughts interpreted by many different people / machines. Why is that so different than a person asking ChatGPT or Seedance to create something they (and only they) imagined? Is it the effort involved? What if they spent 2 hours refining 1 photo to obsessively to get exactly what was in their imagination onto the screen?
I guess this is what is missing from vibed AI art. Usually it's someone with a quick prompt and nothing else. No effort or creativity from the human and that's probably what we don't like about AI art.
This is different from AI art looking all the same though
I think AI art is much like ai coding where people used 1 model a couple of times then wrote it off completely as garbage. Then years of very quick progress happened and all the 10 fingers on a hand and every picture looking the same disappeared.
https://civitai.com/images
If you can make it through the first 20 images on civit and tell me nobody is expressing themselves or their imagination in some of these pictures I guess we agree to disagree.
...
>In case you’re wondering, I’m subscribed to: > Claude Pro ($17/ mo) > GitHub Copilot Pro ($10/ mo)
Nothing wrong there at all, amirite?
Maybe AI scrapped basic arithmetic and now those numbers are totally fine
That is both untrue, and should be something that occurs rarely.
Past Mid level, most dicussion in code review is centered around the conceptual level. Code is mostly use to highlight the point being made. I’ve never seen the need for someone to rewrite a patch proposed by another person. Add to it or correct misunderstandings, but never sweeping changes.
I think I would rather shoot myself in the head
I walked into a paint store recently and there was some interesting music playing on the PA. It was some sort of spanish-language 80s new wave I'd never heard before. It had a nice groove. I asked the guy at the counter about it and was surprised to find out it was AI.
So... was it as good as great music? No. But it beat the hell out of muzak. And it was genuinely novel in a way that pop music isn't. I will sacrifice a lot of pride for novelty.
For my part, I mostly dislike AI music. But if there was to be some AI music that I didn't already dislike for its content, then I would dislike it specifically for being AI. Call me close-minded if you will. Some people don't really like music at all. In general, I love music.
> And it was genuinely novel in a way that pop music isn't. I will sacrifice a lot of pride for novelty.
For one thing, "interesting" and "novel" are not particularly the point of art. An infinite TikTok feed of AI slop is in some sense infinitely interesting and novel.
But also, those concepts are subjective. Pop music has a ton of evolution and innovation if you break it down. Sure, not every hit is doing something clever, but plenty of them are. You might not find those features to be interesting but it takes ingenuity to stand out in an era where the barrier to entry is so low.
My experience has been that AI music mostly is that.
AI music is a complete waste of compute. It exists only to further enrich an already wealthy stakeholder, and we should all be ashamed to have ended up in a place where it even exists.
Most of the "top 40" consistently comes across the same way to me (and pretty much always has), so.
And I think it can be uniquely fun, because it allows for the very rare combination of truly off-beat ideas with excellent execution.
In real life, it very rarely happens that somebody will spend a lifetime learning, gather a band, and then record something extremely off-beat. Because there are risks to that. Destroying your image, having to find people that are on the same wavelength, having to justify the time and money it costs to record something weird and not marketable.
Sure, people make weird stuff once in a while, but very rarely, and generally it takes unique individuals that find some particular weird niche to work in, like Weird Al or Max Raabe. And even then they tend to pick something relatively tame and marketable and stick to that.
But AI can bypass all that, and that makes things interesting.
And often the weirdest artists are also the most prolific because they tend to like experimenting for experimenting's sake.
My jaw literally dropped. Do you really believe this?
I think you just don't spend much time around people who love music.
It makes sense that someone like you would like AI music, I guess. You are clueless.
This doesn't contradict GP at all. Very few people genuinely "love music" in the way that you have in mind.
Sure. It happens in every field. When you invest a lot into something, it becomes risky or unprofitable to do some things.
Tech often drives things in the opposite direction. As it gets easier/cheaper/faster to create something, doing something odd and risky becomes easier to justify because there's less invested into it and there's fewer people whose cooperation is needed.
I'm not saying nobody does anything risky traditionally of course. But there's barriers to it, which become lesser once tech trivializes a good part of the process.
> It makes sense that someone like you would like AI music, I guess. You are clueless.
I like music in general, and am fairly indifferent as to how the goose is cooked. I listen to almost everything. There's no reason to exclude AI music. If it sounds good, it's good.
Controversial take: AI shows that people simply don't give a shit and just confirms that mass culture is a mistake and the death of everything that is good and sacred.
Controversial take: For ages, it's been clear that most people have never given a shit, it just took AI for most to realize it.
>it was genuinely novel
Look, I don't know the song so I can't say how novel it is, but to me, it completely misses the point if its made by a music generation model. The point of music is not to make a bunch of novel sounds. Music is a means by which a human can express themselves, the whole _point_ of art is human expression.
Human expression with art can go both ways from producing and experiencing.
Bird's aren't making art when they sing, I still like listening to birds. Why can't the sounds AI make be in the same category.
I saw a live band last week (daft punk meets prince), they wrote a song about Ai and used an auto tuner for it
2. Call it slop
3. Claim victory
I don't know what to tell you, but me and my kid are having a blast writing ridiculous songs and plopping them with song descriptions and some of those make it to our drive playlists.
Note: same guy; just not anywhere near my account
You are making the argument that the type of activity does not matter as long it is 'spending time', but to say that is almost reductionist to say that that it understates how ridiculously unnuanced an opinion it is. You might as well tell me that the quality of food does not matter and generic burger should suffice.
But what if I want a burger that is custom to me and not a generic slop generated by a human to appeal to the widest possible audience[1]? What if I want something that was created just for me? Good grief? Why do I have to defend having fun and why on earth would you assume that one is mutually exclusive to the other?
Geez man. [1]https://genius.com/Bo-burnham-repeat-stuff-lyrics
Retroactively speaking, I engaged with the already agreed designs and used AI to find much better ways to do the same thing.
Overall, I wish people used AI even for System Design as that seems to have a higher impact.
Me too! I've done several deep dives in various topics just in the last few weeks. I start off with some question I want answers to, and often end up with more questions along the way and I end up learning a ton.
For example I've been enthusiastic about nuclear energy for a long time and I've been wondering why don't we build more of it? It seems like such a great investment, the fuel is extremely cheap compared to the power it produces, even including disposal of waste and everything it seems to me like the best power source we have by far.
I've read things like "it's not economically viable" which I didn't really get - how can it both be cheap clean electricity and not economically viable? It turns out the problem is interest. If you spend $20B building a nuclear plant, you need to make enough money to pay the interest on that $20B. The plant itself will easily make much more than $20B over its lifetime but it won't cover the loan payments unless you can secure a contract with a customer willing to pay a very premium price for stable power for like 20 years.
On the other hand, China builds nuclear plants for like $3B which changes everything. They can do that because they've built lots of nuclear plants, so they have repeatable designs and established supply chains, experienced builders etc. We could do that too and I really think we should (we as in the rest of the world), I think we should invest in nuclear technology to build experience and supply chains etc in order to be able to make the technology economically viable for us. It seems to me that fusion will have exactly the same problem, if we ever even make it work. Good luck building "economically viable" fusion plants, they will certainly be a thousand times more complex than fission plants.
Anyway this wasn't supposed to be a comment about nuclear energy. I also did a deep dive on hydropower based on some claims I heard and learned lots of cool things about that. The LLM found lots of interesting source material for me, including a letter from our ministers of energy and finance talking about a proposed law change that might take effect next year which I would never normally read but since it basically answered many of the questions I had I ended up reading the whole thing and it was super interesting. It's also obviously a lot more credible than just the AI's claims.
Highly recommend trying it out about any topic you may be interested in. I use the free version of ChatGPT. I've also used the Google search AI a lot, I'll basically just Google a question and then proceed to ask followup questions etc.
I'm sure it's not all 100% correct but I'm also quite sure that most of it is pretty good quality, and it really helps when it includes links to sources, I often read the sources instead of the AI summary.
This is exactly the sort of investment governments are suited to make since they have access to extremely cheap debt. Or at least they did until recently, though still cheaper than private markets. China has access to debt with interest of only 1.6% which allows these projects to be very viable.
That being said, I agree with this statement: "I believe they are the greatest teachers humanity has ever created." LLMs are exceedingly powerful at accelerating the learning process.
What does amuse me to no end, however, is how many craftsfolk in AI circles are loudly rejecting AI imagery and video in favor of humans. The only folks who seem fine with that are the grifters looking to slop out a bunch of adverts quickly for commercial gain rather than do actual art, or folks who want “art” but don’t want to pay a human to make it for them or learn the skills themselves. Both of those camps are viewed with disdain and loathing by both sides of the AI Boom/Doom arguments, so I’ll take whatever allies I can get on those fronts.
AI music…man, I live with artisans and musicians and we have a number of “artists” we suspect are AI but whose music is pretty solid. We’re waiting for the shoe to drop, but I’m trying to look at it as a kind of turbocharged Fruity Loops/Logic as opposed to a human replacement. I hope I’m right on that, but time will tell.
The game changer I think more doomers need to experience is agents. It sounds like a recipe for disaster (a bunch of models known to hallucinate and allowed to run roughshod over a computer unattended? What could go wrong?), and it often (hilariously) is for the scammers out there giving it full access to production and fucking off to hawk their latest Masterclass on X or LinkedIn, but that first time you give it a reasonably achievable prompt and come back to a halfway decent product you can review and refine into production in 15m or less?
Absolutely magical, doubly so when it’s just 27B Qwen via Hermes and not some souped up stack in a frontier lab costing you thousands a month in tokens. If I were to evangelize AI, that’s how I’d do it. Fuck subscriptions, fuck cloud compute, fuck frontier labs and token buckets and whatever slop Sundar/Scam/Dario/Elmo et al are hawking this week; run that locally, give it a reasonable prompt that solves a real problem, and show them the results.
Then do the right thing and go for a walk with the two hours of free time the AI just freed up for you. Don’t ask for permission to take back your time.
And I agree that "less AI" edit passes reduce a lot of the hallmarks, but I think these models truly struggle with expressing complicated ideas. Have Opus been good at that for you?
I think people should find this sort of result surprising by default. Like, presumably you don't know why your output isn't as good, or you'd fix it. So doesn't that lack of understanding concern you?
> I still believe I use AI fairly typically, but I’ve had a notable change in background. Last October, I was an engineer on Microsoft’s Windows AI team. I am now a research engineer at Salesforce AI Research.
Priest wears robes.
I guess it’s one of those nonsensical automatic rules.
> I am now a research engineer at Salesforce AI Research. I spend more time supporting scientists & building rapid prototypes, rather than working with a large team of engineers.
I appreciate the PoV of people with different challenges, but my full-time job is not about building rapid prototypes so the guidelines are worth a bit less to me.