Learning how to navigate these hitches is a really important process

If we streamline every bit of difficulty or complexity out of our lives, it seems trivially obvious that we will soon have no idea what to do when we encounter difficulty or complexity. Is that just me thinking that?

I can make the problem input bigger as I want.

Each LLM have a different thresholf for each problem, when crossed the performance of the LLM collapse.

One feature others have noted is that the Opus 4 context buffer rarely "wears out" in a work session. It can, and one needs to recognize this and start over. With other agents, it was my routine experience that I'd be lucky to get an hour before having to restart my agent. A reliable way to induce this "cliff" is to let AI take on a much too hard problem in one step, then flail helplessly trying to fix their mess. Vibe-coding an unsuitable problem. One can even kill Opus 4 this way, but that's no way to run a race horse.

Some "persistence of memory" harness is as important as one's testing harness, for effective AI coding. With the right care having AI edit its own context prompts for orienting new sessions, this all matters less. AI is spectacularly bad at breaking problems into small steps without our guidance, and small steps done right can be different sessions. I'll regularly start new sessions when I have a hunch that this will get me better focus for the next step. So the cliff isn't so important. But Opus 4 is smarter in other ways.

And one of the things with current generators is that they tend to make things more complex over time, rather than less. It's always me prompting the LLM to refactor things to make it simpler, or doing the refactoring once it's gotten to complex for the LLM to deal with.

So at least with the current generation of LLMs, it seems rather inevitable that if you just "give LLMs their head" and let them do what they want, eventually they'll create a giant Rube Goldberg mess that you'll have to try to clean up.

ETA: And to the point of the article -- if you're an old salt, you'll be able to recognize when the LLM is taking you out to sea early, and be able to navigate your way back into shallower waters even if you go out a bit too far. If you're a new hand, you'll be out of your depth and lost at sea before you know it's happened.

Imagine that you have your input to the context, 10000 tokens that are 99% correct. Each time the LLM replies it adds 1000 tokens that are 90% correct.

After some back-and-forth of you correcting the LLM, its context window is mostly its own backwash^Woutput. Worse, the error compounds because the 90% that is correct is just correct extrapolation of an argument about incorrect code, and because the LLM ranks more recent tokens as more important.

The same problem also shows up in prose.

This is also can be made much worse by thinking models, as their CoT is all in context, and if there thoughts really wander it just plants seeds of poison feeding the rot. I really wish they can implement some form of context pruning, so you can nip irrelevant context when it forms.

In the meantime, I make summaries and carry it to a fresh instance when I notice the rot forming.

The solve was to define several Cursor rules files for different views of the codebase - here's the structure, here's the validation logic, etc. That and using o3 has at least gotten me to the next level.

I suspect it has something more to do with the model producing too many tokens and becoming fixated on what it said before. You'll often see this in long conversations. The only way to fix it is to start a new conversation.

> The article opens with a statement saying the author isn't going to reword what others are writing, but the article reads as that and only that.

Hmm, I was just saying I hadn't seen much literature or discussion on trust dynamics in teams with LLMs. Maybe I'm just in the wrong spaces for such discussions but I haven't really come across it.

It might not solve every problem, but it solves enough of them better enough it belongs in the tool kit.

3 have obviously only read the title, and 3 comments how the article require JS.

Well played HN.

Most of the current discourse on AI coding assistants sounds either breathlessly optimistic or catastrophically alarmist. What’s missing is a more surgical observation: the disruptive effect of LLMs is not evenly distributed. In fact, the clash between how open source and industry teams establish trust reveals a fault line that’s been papered over with hype and metrics.

FOSS project work on a trust basis - but industry standard is automated testing, pair programming, and development speed. That CRUD app for finding out if a rental car is available? Not exactly in need for a hand-crafted piece of code, and no-one cares if Junior Dev #18493 is trusted within the software dev organization.

If the LLM-generated code breaks, blame gets passed, retros are held, Jira tickets multiply — the world keeps spinning, and a team fixes it. If a junior doesn’t understand their own patch, the senior rewrites it under deadline. It’s not pretty, but it works. And when it doesn’t, nobody loses “reputation” - they lose time, money, maybe sleep. But not identity.

LLMs challenge open source where it’s most vulnerable - in its culture. Meanwhile, industry just treats them like the next Jenkins: mildly annoying at first, but soon part of the stack.

The author loves the old ways, for many valid reasons: Gabled houses are beautiful, but outside of architectural circles, prefab is what scaled the suburbs, not timber joints and romanticism.

The problem is often really one of miscommunication, the task may be clear to the person working on it, but with frequent context resets it's hard to make sure the LLM also knows what the whole picture is and they tend to make dumb assumptions when there's ambiguity.

The thing that 4o does with deep research where it asks for additional info before it does anything should be standard for any code generation too tbh, it would prevent a mountain of issues.

It’s like if someone started bricking up tunnel entrances and painting ultra realistic versions of the classic Road Runner tunnel painting on them, all over the place. You’d have to stop and poke every underpass with a stick just to be sure.

There's so much more than "works well". There are many cues that exist close to code, but are not code:

I trust more if the contributor explains their change well.

I trust more if the contributor did great things in the past.

I trust more if the contributor manages granularity well (reasonable commits, not huge changes).

I trust more if the contributor picks the right problems to work on (fixing bugs before adding new features, etc).

I trust more if the contributor proves being able to maintain existing code, not just add on top of it.

I trust more if the contributor makes regular contributions.

And so on...

A good rule of thumb is to simply reject any work that has had involvement of an LLM, and ignore any communication written by an LLM (even for EFL speakers, I'd much rather have your "bad" English than whatever ChatGPT says for you).

I suspect that as the serious problems with LLMs become ever more apparent, this will become standard policy across the board. Certainly I hope so.

What you're seeing now is people who once thought and proclaimed these tools as useless now have to start to walk back their claims with stuff like this.

It does amaze me that the people who don't use these tools seem to have the most to say about them.

Essentially the premise is that in medium trust environments like very large teams or low trust environments like an open source project.

LLMs make it very difficult to make an immediate snap judgement about the quality of the dev that submitted the patch based solely on the code itself.

In the absence of being able to ascertain the type of person you are dealing with you have to fall back too "no trust" and review everything with a very fine tooth comb. Essentially there are no longer any safe "review shortcuts" and that can be painful in places that relied on those markers to grease the wheels so to speak.

Obviously if you are in an existing competent high trust team then this problem does not apply and most likely seems completely foreign as a concept.

Only because you already trust them to know that the code is indeed bug-free. Some cases are simple and straightforward -- this routine returns a desired value or it doesn't. Other situations are much more complex in anticipating the ways in which it might interact with other parts of the system, edge cases that are not obvious, etc. Writing code that is "bug free" in that situation requires the writer of the code to understand the implications of the code, and if the dev doesn't understand exactly what the code does because it was written by an LLM, then they won't be able to understand the implications of the code. It then falls to the reviewer to understand the implications of the code -- increasing their workload. That was the premise.

If A = B, higher B means that A is higher too.

Now it's A + Ai = B. Now Higher B doesn't necessarily mean higher A.

Especially since the current state of Ai is pretty much stochastic, and sometimes is worse than nothing at all

That being said, the prediction engine still can't do any real engineering. If you don't specifically task them with using things like Python generators, you're very likely to have a piece of code that eats up a gazillion memory. Which unfortunately don't set them appart from a lot of Python programmers I know, but it is an example of how the LLM's are exactly as bad as you mention. On the positive side, it helps with people actually writing the specification tasks in more detail than just "add feature".

Where AI-agents are the most useful for us is with legacy code that nobody prioritise. We have a data extractor which was written in the previous millennium. It basically uses around two hunded hard-coded coordinates to extact data from a specific type of documents which arrive by fax. It's worked for 30ish years because the documents haven't changed... but it recently did, and it took co-pilot like 30 seconds to correct the coordinates. Something that would've likely taken a human a full day of excruciating boredom.

I have no idea how our industry expect anyone to become experts in the age of vibe coding though.

Not they don't. This is 100% a made up statistic.

> but rather if it proves it can do what it intends to do after extensive testing.

Author here: Here I was less talking about the effectiveness of the output of a given tool and more so about the tool itself.

To take your garbage collection example, sure perhaps an agentic system at some point can spin some stuff up and beat it into submission with test harnesses, bug fixes etc.

But, imagine you used the model AS the garbage collector/tool, in that say every sweep you simply dumped the memory of the program into the model and told it to release the unneeded blocks. You would NEVER be able to trust that the model itself correctly identifies the correct memory blocks and no amount of "patching" or "fine tuning" would ever get you there.

With other historical abstractions like say jvm, if the deterministic output, in this case the assembly the jit emits is incorrect that bug is patched and the abstraction will never have that same fault again. not so with LLMs.

To me that distinction is very important when trying to point out previous developer tooling that changed the entire nature of the industry. It's not to say I do not think LLMs will have a profound impact on the way things work in the future. But I do think we are in completely uncharted territory with limited historical precedence to guide us.

An even more important question: who tests the tests themselves? In traditional development, every piece of logic is implemented twice: once in the code and once in the tests. The tests checks the code, and in turn, the code implicitly checks the tests. It's quite common to find that a bug was actually in the tests, not the app code. You can't just blindly trust the tests, and wait until your agent finds a way to replicate a test bug in the code.

That is quite a statement! You're talking about systems that are essentially entropy-machines somehow creating order?

> with the result being that test-driven development gains even more momentum

Why is it that TDD is always put forward as the silver bullet that fixes all issues with building software

The number of times I've seen TDD build the wrong software after starting with the wrong tests is actually embarassing

This new world of having to verify every single thing at all points is quite exhausting and frankly pretty slow.

You'll have to elaborate on that. How much trust was there in electricity, flight and radioactivity when we discovered them?

In science, you build trust as you go.

It's just too useful to ignore, and trust is always built, brick by brick. Let's not forget humans are far from reliable anyway. Just like with driving cars, I imagine producing less buggy code (along predefined roads) will soon outpace humans. Then it is just blocking and tackling to improve complexity.

I’m someone who put in my "+10,000 hours" programming complex applications, before useful LLMs were released. I spent years diving into documentation and other people's source code every night, completely focused on full-stack mastery. Eventually, that commitment led to severe burnout. My health was bad, my marriage was suffering. I released my application and then I immediately had to walk away from it for three years just to recover. I was convinced I’d never pick it up again.

It was hearing many reports that LLMs had gotten good at code that cautiously brought me back to my computer. That’s where my experience diverges so strongly from your concerns. You say, “No one who uses AI can claim ‘this is my work.’” I have to disagree. When I use an LLM, I am the architect and the final inspector. I direct the vision, design the system, and use a diff tool to review every single line of code it produces. Just recently, I used it as a partner to build a complex optimization model for my business's quote engine. Using a true optimization model was always the "right" way to do it but would have taken me months of grueling work before, learning all details of the library, reading other people’s code, etc. We got it done in a week. Do I feel like it’s my work? Absolutely. I just had a tireless and brilliant, if sometimes flawed, assistant.

You also claim the user won't "thoroughly understand it." I’ve found the opposite. To use an LLM effectively for anything non-trivial, you need a deeper understanding of the fundamentals to guide it and to catch its frequent, subtle mistakes. Without my years of experience, I would be unable to steer it for complex multi-module development, debug its output, or know that the "plausibly good work" it produced was actually wrong in some ways (like N+1 problems).

I can sympathize with your experience as a teacher. The problem of students using these tools to fake comprehension is real and difficult. In academia, the process of learning, getting some real fraction of the +10,000hrs is the goal. But in the professional world, the result is the goal, and this is a new, powerful tool to achieve better results. I’m not sure how a teacher should instruct students in this new reality, but demonizing LLM use is probably not the best approach.

For me, it didn't make bad work look good. It made great work possible again, all while allowing me to have my life back. It brought the joy back to my software development craft without killing me or my family to do it. My life is a lot more balanced now and for that, I’m thankful.

Even now I have this refusal to use GPT where as my coworkers lately have been saying "ChatGPT says" or this code was created by chatGPT idk, for me I take pride writing code myself/not using GPT but I also still use google/stackoverflow which you could say is a slower version of GPT.

Think this strongly depends on the developer and what they're attempting to accomplish.

In my experience, most people who swear LLMs make them 10x more productive are relatively junior front-end developers or serial startup devs who are constantly greenfielding new apps. These are totally valid use cases, to be clear, but it means a junior front-end dev and a senior embedded C dev tend to talk past each other when they're discussing AI productivity gains.

> Working with the technology and mitigating it's weaknesses is the only rational path forward.

Or just using it more sensibly. As an example: is the idea of an AI "agent" even a good one? The recent incident with Copilot[0] made MS and AI look like a laughingstock. It's possible that trying to let AI autonomously do work just isn't very smart.

As a recent analogy, we can look at blockchain and cryptocurrency. Love it or hate it, it's clear from the success of Coinbase and others that blockchain has found some real, if niche, use cases. But during peak crypto hype, you had people saying stuff like "we're going to track the coffee bean supply chain using blockchain". In 2025 that sounds like an exaggerated joke from Twitter, but in 2020 it was IBM legitimately trying to sell this stuff[1].

It's possible we'll look back and see AI agents, or other current applications of generative AI, as the coffee blockchain of this bubble.

[0] https://www.reddit.com/r/ExperiencedDevs/comments/1krttqo/my_new_hobby_watching_ai_slowly_drive_microsoft/

[1] https://www.forbes.com/sites/robertanzalone/2020/07/15/big-coffee-sellers-use-blockchain-to-connect-farmers-and-customers/

But this doesn't mean that the model/human combo is more effective at serving the needs of users! It means "producing more code."

There are no LLMs shipping changesets that delete 2000 lines of code -- that's how you know "making engineers more productive" is a way of talking about how much code is being created...

You seem to be claiming that this is a binary, either we will or won’t use llms, but the author is mostly talking about risk mitigation.

By analogy it seems like you’re saying the author is fundamentally against the development of the motor car because they’ve pointed out that some have exploded whereas before, we had horses which didn’t explode, and maybe we should work on making them explode less before we fire up the glue factories.

But I think that everyone is lossing trust not because there is no potential that LLMs could write good code or not, it's the trust to the user who uses LLMs to uncontrollable-ly generate those patches without any knowledge, fact checks, and verifications. (many of them may not even know how to test it.)

In another word, while LLMs is potentially capable of being a good SWE, but the human behind it right now, is spamming, and doing non-sense works, and let the unpaid open source maintainers to review and feedback them (most of the time, manually).

No, and there’s no reason to think cars will stop improving either, but that doesn’t mean they will start flying.

The first error is in thinking that AI is converging towards a human brain. To treat this as a null hypothesis is incongruent both wrt the functional differences between the two and crucially empirical observations of the current trajectory of LLMs. We have seen rapid increases in ability, yes, but those abilities are very asymmetrical by domain. Pattern matching and shitposting? Absolutely crushing humans already. Novel conceptual ideas and consistency checked reasoning? Not so much, eg all that hype around PhD-level novel math problems died down as quickly as it had been manufactured. If they were converging on human brain function, why this vastly uneven ability increases?

The second error is to assume a superlinear ability improvement when the data has more or less run out and has to be slowly replenished over time, while avoiding the AI pollution in public sources. It’s like assuming oil will accelerate if it had run out and we needed to wait for more bio-matter to decompose for every new drop of crude. Can we improve engine design and make ICEs more efficient? Yes, but it’s a diminishing returns game. The scaling hypothesis was not exponential but sigmoid, which is in line with most paradigm shifts and novel discoveries.

> Making these sort of blanket assessments of AI, as if it were a singular, static phenomena is bad thinking.

I agree, but do you agree with yourself here? Ie:

> no reason to think that these tools won't vastly outperform us in the very near future

.. so back to single axis again? How is this different from saying calculators outperform humans?

I made myself known to be a big fool ~4 years ago. A local newspaper published an article on a particular person with outrageous claims primarily using photographs as proof. I challenged the editor directly via email, laying out my reasoning for why I was sure the images were manipulated. My arguments relied on misunderstandings on my part and the person claims were levied against showing zero deviation in position and stance while posing with multiple people during a meet-and-greet. The editor was offended and trolled me in response. I didn't let up, and he realized I was an idiot, not an agitator, and shared the full unpublished video from where the photos were taken with me, at which point I apologized deeply and made a donation. My ego was appropriately small for the following year.

Before emailing him, I shared the photos with some level-headed friends for their opinion, specifically because I didn't want to make a false accusation. They came to the same conclusion that the images were most likely manipulated, so I was very confident going in.

Now I trust this paper and people involved implicitly, but this was a lot of work to convince just one person.