Posted by CoffeeOnWrite 5 days ago
(Works on older browsers and doesn't require JavaScript except to get past CloudSnare).
Verification has a high cost and trust is the main way to lower that cost. I don't see how one can build trust in LLMs. While they are extremely articulate in both code and natural language, they will also happily go down fractal rabbit holes and show behavior I would consider malicious in a person.
This new world of having to verify every single thing at all points is quite exhausting and frankly pretty slow.
The classic on the subject.
I don't agree with this at all. Writing new code is trivially easy, to do a full in depth review takes significantly more brain power. You have to fully ascertain and insert yourself into someone elses thought process. Thats way more work than utilizing your own thought process.
They basically achieve over 80% agreement with human evaluators [1]. This level of agreement is similar to the consensus rate between two human evaluators, making LLM-as-a-judge a scalable and reliable proxy for human judgment.
[1] https://arxiv.org/abs/2306.05685 (2023)
It sounds nice but it means at least 1 in 5 are bad. That's worse odds than rolling 1 on a d6. You'll be tripping over mistakes constantly.
So if 80% is abysmal, you are getting the same abysmal level from human code reviewers.
The whole point of a tool is to reduce workload. If you have to frequently correct it, then it's not worthwhile to use.
> 80% is a pretty abysmal success rate and means it's very unreliable.
Sure there's no bug with how the logic is defined in the CR or even in the context of the project, ti maybe won't throw an exception.
But the LLM won't know that the query is iterating over an unindexed field in the DB with the table in prod having 10s of millions of rows. The LLM won't know that even though the code says the button should be red and the comments say the button should be red, the corporate style guide says red should be a very specific hex code that it isn't.
Oh goodness that's like trusting one kid to tell you whether or not his friend lied.
In matters where trust matters, it's a recipe for disaster.
Give it another year and HN comments will be very different.
Writing tests already works now. It's usually easier to read tests than to read convoluted logic.
Mmmhmm. And you think this "growing up" doesn't have biases to lie in circumstances where it matters? Consider politics. Politics matter. It's inconceivable that a magic algorithm would lie to us about various political concerns, right? Right...?
A magic algorithm lying to us about anything would be extremely valuable to liars. Do you think it's possible that liars are guiding the direction of these magic algorithms?
I notice a distinct lack of blockchain hegemony.
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.
Can we really do this reliably? LLMs are non-deterministic, right, so how do we validate the output in a deterministic way?
We can validate things like shape of data being returned, but how do we validate correctness without an independent human in the loop to verify?
If no? Then congrats, you are now in a position where your software development lifecycle needs to handle non-determinism.
This fanatical vibing movement is ridicolous, but this luddite stance that LLMs cannot contribute to software dev because they are «non deterministic» is almost as ludicrus.
Sure, except the Juniors producing wrong solutions is the Juniors problem, not mine
If I give four LLM agents tasks and they call come back with slightly wrong solutions, that's me adding four problems to my own workload
I'm not sure how I'm supposed to keep up with that. I'm definitely not sure it makes me overall more productive
I check AI output for hallucinations and issues as I don’t fully trust it to work, but we also do PRs with humans to have another set of eyes check because humans also make mistakes.
For the soft sciences and arts I’m not sure how to validate anything from AI but for software and hard sciences I don’t see why test suites wouldn’t continue serving their same purpose
If we need a human in the loop to check every row of code for the deep logic errors... then we could just get the human to write it no?
I find its a slightly improved google for vague questions. Or a doxygen writer.
Its all use I've found for any ai model since i first started playing with github copilot beta.
Ive been trying the newer models as they arrived, and found they're getting more verbose, more prone to hallucinating functions that dont exist, and more prone to praise me as a god when trying to ask about basic assumptions. (you're cutting to the heart of the matter)
What kind of code do you write where its somehow replacing coding itself? I spent 30 minutes trying to get mistral to write a basic bash script yesterday.
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.
> As the use of AC spread rapidly with other companies deploying their own systems, the Edison Electric Light Company claimed in early 1888 that high voltages used in an alternating current system were hazardous, and that the design was inferior to, and infringed on the patents behind, their direct current system.
> In the spring of 1888, a media furor arose over electrical fatalities caused by pole-mounted high-voltage AC lines, attributed to the greed and callousness of the arc lighting companies that operated them.
Most of what people think they know about Tesla is not actually true if you examine the historical record. But software engineering as a discipline demands business villains and craftsman heroes, and so Edison and Tesla were warped to fit those roles even though in real life there is only evidence of cordial interactions.
Not much.
Plenty of people were against electricity when it started becoming common. They were terrified of lamps, doorbells, telephones, or anything else with an electric wire. If they were compelled to use these things (like for their job) they would often wear heavy gloves to protect themselves. It is very occasionally mentioned in novels from the late 1800's.
(Edit: If you'd like to see this played out visually, watch the early episodes of Miss Fisher's Murder Mysteries on ABC [.oz])
There are still people afraid of electricity today. There is no shortage of information on the (ironically enough) internet about how to shield your home from the harmful effects of electrical wires, both in the house and utility lines.
Flight? I dunno about back then, but today there's plenty of people who are afraid to fly. If you live in Las Vegas for a while, you start to notice private train cars occasionally parked on the siding near the north outlet mall. These belong to celebrities who are afraid to fly, but have to go to Vegas for work.
Radioactivity? There was a plethora of radioactive hysteria in books, magazines, comics, television, movies, and radio. It's not hard to find.
Humans communicate using symbols. That could be patterns of sound waves, gestures, or written characters.
If we can't trust that the communicated symbols match their agreed meaning, what is the use of them? Communication breaks down quickly, and being social individuals which inter-depend on others, we can't live without communication. Nobody likes a liar, and the reason is, the words they give us do not match what they mean. (Perhaps not only humans. I read that dolphins readily come and rescue individuals which are drowning - including humans - but they punish individuals which fake drowning.)
And that goes from every stratum of social interaction, from big treaties to selling a bagel. Would you sell a bagel for a fake dollar bill? The number on it is a symbol as well, it has a meaning.
So it is right down one of the very bases of human cooperation.
True not only in innovation, but in business settings.
I don't think there's anyone who works in any business long enough who doesn't have problems getting their job done simply because someone else with a key part of the project doesn't trust that you know what you're doing.
The moment a technology reaches trust at scale, it becomes a non-innovation in people's mind.
Happened for TVs, electrical light in homes, AI for chess, and Google. Will happen with LLM-based assistants.
LLM leads to distrust between people. From TFA, That concept is Trust - It underpins everything about how a group of engineers function and interact with each other in all technical contexts. When you discuss a project architecture you are trusting your team has experience and viewpoints to back up their assertions.
As we were debugging, my colleague revealed his assumption that I'd used AI to write it, and expressed frustration at trying to understand something AI generated after the fact.
But I hadn't used AI for this. Sure, yes I do use AI to write code. But this code I'd written by hand and with careful deliberate thought to the overall design. The bugs didn't stem from some fundamental flaw in the refactor, they were little oversights in adjusting existing code to a modified API.
This actually ended up being a trust building experience over all because my colleague and I got to talk about the tension explicitly. It ended up being a pretty gentle encounter with the power of what's happening right now. In hindsight I'm glad it worked out this way, I could imagine in a different work environment, something like this could have been more messy.
Be careful out there.
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.
If someone uses an LLM and produces bug-free code, I'll trust them. If someone uses an LLM and produces buggy code, I won't trust them. How is this different from when they were only using their brain to produce the code?
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.
That's the core of the issue. It's time to say goodbye to heuristics like "the blog post is written in eloquent, grammatical English, hence the point its author is trying to make must be true" or "the code is idiomatic and following all code styles, hence it must be modeling the world with high fidelity".
Maybe that's not the worst thing in the world. I feel like it often made people complacent.
For sure, in some ways perhaps reverting to a low trust environment might improve quality in that it now forces harsher/more in depth reviews.
That however doesn't make the requirement less exhausting for people previously relying heavily on those markers to speed things up.
Will be very interesting to see how the industry standardizes around this. Right now it's a bit of the wild west. Maybe people in ten years will look back at this post and think "what do you mean you judged people based on the code itself that's ridiculous"
You said "hence the point its author is trying to make must be true" and "hence it must be modeling the world with high fidelity".
But it's more like "hence the author is likely competent and likely put in a reasonable effort."
When those assumptions hold, putting in a very deep review is less likely to pay off. Maybe you are right that people have been too complacent to begin with, I don't know, but I don't think you've framed it fairly.
And isn't dyslexic, and is a native speaker etc. Some will gain from this shift, some will lose.
The heuristic is "this submission doesn't even follow the basic laws of grammar, therefore I can safely assume incompetence and ignore it entirely."
You still have to do verification for what passes the heuristic, but it keeps 90% of the crap away.
When someone follows standard conventions it means that they A) have a baseline level of knowledge to know about them, and B) care to write the code in a clear and approachable way for others.
“unconventional” or “fancy” is in the eye of the beholder. Whose conventions are we talking about? Code is bad when it doesn't look the way you want it to? How convenient. I may find code hard to read because it's formatted “conventionally”, but I wouldn't be so entitled as to call it bad just because of that.
Literally not: a language defines its own conventions, they're not defined in terms of individual users/readers/maintainers subjective opinions.
> Whose conventions are we talking about?
The conventions defined by the language.
> Code is bad when it doesn't look the way you want it to?
No -- when it doesn't satisfy the conventions established by the language.
> I may find code hard to read because it's formatted “conventionally”,
If you did this then you'd be wrong, and that'd be a problem with your personal evaluation process/criteria, that you would need to fix.
Where are these mythical languages? I think the word you're looking for is syntax, which is entirely different. Conventions are how code is structured and expected to be read. Very few languages actually enforce or even suggest conventions, hence the many style guides. It's a standout feature of Go to have a format style, and people still don't agree with it.
And it's kinda moot when you can always override conventions. It's more accurate to say a team decides on the conventions of a language.
If every rando hire goes in and has a completely different style and formatting -- and then other people come in and rewrite parts in their own style -- code rapidly goes to shit.
It doesn't matter what the style is, as long as there is one and it's enforced.
What you're saying is reasonable, but that's not what they said at all. They said there's one way to write cleanly and that's "Standard conventions", whatever that means. Yes, conventions so standard that I've read 10 conflicting books on what they are.
There is no agreed upon definition of "readable code". A team can have a style guide, which is great to follow, but that is just formalizing the personal preference of the people working on a project. It's not anymore divine than the opinion of a "rando."
While it's true that in principle it doesn't matter what style you choose as long as there is one, in practice languages are just communities of people, and every community develops norms and standards. More recent languages often just pick a style and bake it in.
This is a good thing, because again, code is read 1000x more times than it's written. It saves everyone time and effort to just develop a typical style.
And yeah, the code might run no matter how you indent it, but it's not correct, any more than you going to a restaurant and licking the plates.
Again, there's a couple examples of languages doing this, and everything else is a free for all.
> No, you misunderstood what they said.
Agree to disagree. Nothing in that comment talks about the conventions of a language, only the conventions of code. Again, I don't disagree with what you say, but the person you replied to was in a completely different argument.
Is that not how you review all code? I don't care who wrote the code, just because certain person wrote the code doesn't give them an instant pass to skip my review process.
a dev can write piece of good, and piece of bad code. so per code, review the code. not the dev!
I could not disagree more. The quality of the dev will always matter, and has as much to do with what code makes it into a project as the LLM that generated it.
An experienced dev will have more finely tuned evaluation skills and will accept code from an LLM accordingly.
An inexperienced or “low quality” dev may not even know what the ideal/correct solution looks like, and may be submitting code that they do not fully understand. This is especially tricky because they may still end up submitting high quality code, but not because they were capable of evaluating it as such.
You could make the argument that it shouldn’t matter who submits the code if the code is evaluated purely on its quality/correctness, but I’ve never worked in a team that doesn’t account for who the person is behind the code. If its the grizzled veteran known for rarely making mistakes, the review might look a bit different from a review for the intern’s code.
That may be true, but the proxy for assessing the quality of the dev is the code. No one is standing over you as you code your contribution to ensure you are making the correct, pragmatic decisions. They are assessing the code you produce to determine the quality of your decisions, and over time, your reputation as a dev is made up of the assessments of the code you produced.
The point is that an LLM in no way changes this. If a dev uses an LLM in a non-pragmatic way that produces bad code, it will erode trust in them. The LLM is a tool, but trust still factors in to how the dev uses the tool.
Yes, the quality of the dev is a measure of the quality of the code they produce, but once a certain baseline has been established, the quality of the dev is now known independent of the code they may yet produce. i.e. if you were to make a prediction about the quality of code produced by a "high quality" dev vs. a "low quality" dev, you'd likely find that the high quality dev tends to produce high quality code more often.
So now you have a certain degree of knowledge even before you've seen the code. In practice, this becomes a factor on every dev team I've worked around.
Adding an LLM to the mix changes that assessment fundamentally.
> The point is that an LLM in no way changes this.
I think the LLM by definition changes this in numerous ways that can't be avoided. i.e. the code that was previously a proxy for "dev quality" could now fall into multiple categories:
1. Good code written by the dev (a good indicator of dev quality if they're consistently good over time)
2. Good code written by the LLM and accepted by the dev because they are experienced and recognize the code to be good
3. Good code written by the LLM and accepted by the dev because it works, but not necessarily because the dev knew it was good (no longer a good indicator of dev quality)
4. Bad code written by the LLM
5. Bad code written by the dev
#2 and #3 is where things get messy. Good code may now come into existence without it being an indicator of dev quality. It is now necessary to assess whether or not the LLM code was accepted because the dev recognized it was good code, or because the dev got things to work and essentially got lucky.
It may be true that you're still evaluating the code at the end of the day, but what you learn from that evaluation has changed. You can no longer evaluate the quality of a dev by the quality of the code they commit unless you have other ways to independently assess them beyond the code itself.
If you continued to assess dev quality without taking this into consideration, it seems likely that those assessments would become less accurate over time as more "low quality" devs produce high quality code - not because of their own skills, but because of the ongoing improvements to LLMs. That high quality code is no longer a trustworthy indicator of dev quality.
> If a dev uses an LLM in a non-pragmatic way that produces bad code, it will erode trust in them. The LLM is a tool, but trust still factors in to how the dev uses the tool.
Yes, of course. But the issue is not that a good dev might erode trust by using the LLM poorly. The issue is that inexperienced devs will make it increasingly difficult to use the same heuristics to assess dev quality across the board.
In my experience they very much are related. High quality devs are far more likely to output high quality working code. They test, they validate, they think, ultimately they care.
In that case that you are reviewing a patch from someone you have limited experience with, it previously was feasible to infer the quality of the dev from the context of the patch itself and the surrounding context by which it was submitted.
LLMs make that judgement far far more difficult and when you can not make a snap judgement you have to revert your review style to very low trust in depth review.
No more greasing the wheels to expedite a process.
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...
Spot on, there are so many little things that we as humans use as subtle verification steps to decide how much scrutiny various things require. LLMs are not necessarily the death of that concept but they do make it far far harder.
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.
All of these make mistakes (there are documented incidents).
And yes, we can counter with "the journalists are dumb for not verifying", "the lawyers are dumb for not checking", etc., but we should also be open for the fact that these are intelligent and professional people who make mistakes because they were mislead by those who sell LLMs.
In the past someone might have been physically healthy and strong enough to physically shovel dirt all day long
Nowadays this is rarer because we use an excavator instead. Yes, a professional dirt mover is more productive with an excavator than a shovel, but is likely not as physically fit as someone spending their days moving dirt with a shovel
I think it will be similar with AI. It is absolutely going to offload a lot of people's thinking into the LLMs and their "do it by hand" muscles will atrophy. For knowledge workers, that's our brain
I know this was a similar concern with search engines and Stack Overflow, so I am trying to temper my concern here as best I can. But I can't shake the feeling that LLMs provide a way for people to offload their thinking and go on autopilot a lot more easily than Search ever did
I'm not saying that we were better off when we had to move dirt by hand either. I'm just saying there was a physical tradeoff when people moved out of the fields and into offices. I suspect there will be a cognitive tradeoff now that we are moving away from researching solutions to problems and towards asking the AI to give us solutions to problems
As if that is a somehow exonerating sentence.
LLMs are a tool, just like any number of tools that are used by developers in modern software development. If a dev doesn’t use the tool properly, don’t trust them. If they do, trust them. The way to assess if they use it properly is in the code they produce.
Your premise is just fundamentally flawed. Before LLMs, the proof of a quality dev was in the pudding. After LLMs, the proof of a quality dev remains in the pudding.
Indeed it does, however what the "proof" is has changed. In terms of sitting down and doing a full, deep review, tracing every path validating every line etc. Then for sure, nothing has changed.
However, at least in my experience, pre LLM those reviews were not EVERY CASE there were many times I elided parts of a deep review because i saw markers in the code that to me showed competency, care etc. With those markers there are certain failure conditions that can be deemed very unlikely to exist and therefore the checks can be skipped. Is that ALWAYS the correct assumption? Absolutely not but the more experienced you are the less false positives you get.
LLMs make those markers MUCH harder to spot, so you have to fall back to doing a FULL indepth review no matter what. You have to eat ALL the pudding so to speak.
For people that relied on maybe tasting a bit of the pudding then assuming based on the taste the rest of the pudding probably tastes the same its rather jarring and exhausting to now have to eat all of it all the time.
That was never proof in the first place.
If anything, someone basing their trust in a submission on anything other than the code itself is far more concerning and trust-damaging to me than if the submitter has used an LLM.
I mean, it's not necessarily HARD proof but it has been a reliable enough way to figure out which corners to cut. You can of course say that no corners should ever be cut and while that is true in an ideal sense. In the real world things always get fuzzy.
Maybe the death of cutting corners is a good thing overall for output quality. Its certainly exhausting on the people tasked with doing the reviews however.
Ultimately I don't think the heuristics would change all that much, though. If every time you review a person's PR, almost everything is great, they are either not using AI or they are vetting what the AI writes themselves, so you can trust them as you did before. It may just take some more PRs until that's apparent. Those who submit unvetted slop will have to fix a lot of things, and you can crank up the heat on them until they do better, if they can. (The "if they can" is what I'm most worried about.)
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.
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.
This is standard for any activity where accuracy / safety is paramount - you validate the process. Hence things like maintenance logs for airplanes.
Precisely this, and this is hardly a unique to software requirement. Process audits are everywhere in engineering. Previously you could infer the process of producing some code by simply reading the patch and that generally would tell you quite a bit about the author itself. Using advanced and niche concepts with imply a solid process with experience backing it. Which would then imply that certain contextual bugs are unlikely so you skip looking for them.
My premise in the blog is basically that "Well now I have go do a full review no matter what the code itself tells me about the author."
Which IMO is the correct approach - or alternatively, if you do actually trust the author, you shouldn't care if they used LLMs or not because you'd trust them to check the LLM output too.
You can say that for pretty much any sort of automation or anything that makes things easier for humans. I'm pretty sure people were saying that about doing math by hand around when calculators became mainstream too.
There's nothing wrong with using LLMs to save time doing trivial stuff you know how to do yourself and can check very easily. The problem is that (very lazy) people are using them to do stuff they are themselves not competent at. They can't check, they won't learn, and the LLM is essentially their skill ceiling. This is very bad: what plus-value are you supposed to bring over something you don't understand? AGI won't have to improve from the current baseline to surpass humans if we're just going to drag ourselves down to its level.
What? How on god's green earth could you even pretend to know how all people are using these tools?
> LLMs are not calculators, nor are they the internet.
Umm, okay? How does that make them less useful?
I'm going to give you a concrete example of something I just did and let you try and do whatever mental gymnastics you have to do to tell me it wasn't useful:
Medicare requires all new patients receiving home health treatment go through a 100+ question long form. This form changes yearly, and it's my job to implement the form into our existing EMR. Well, part of that is creating a printable version. Guess what I did? I uploaded the entire pdf to Claude and asked it to create a print-friendly template using Cottle as the templating language in C#. It generated the 30 page print preview in a minute. And it took me about 10 more minutes to clean up.
> 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.
The irony is that they're getting better by the day. That's not to say people don't use them for the wrong applications, but the idea that this tech is going to be banned is absurd.
> A good rule of thumb is to simply reject any work that has had involvement of an LLM
Do you have any idea how ridiculous this sounds to people who actually use the tools? Are you going to be able to hunt down the single React component in which I asked it to convert the MUI styles to tailwind? How could you possibly know? You can't.
This is okay for platitudes, but for emails that really matter, having this messy watercolor kind of writing totally destroys the clarity of the text and confuses everyone.
To your point, I’ve asked everyone on my team to refrain from writing words (not code) with ChatGPT or other tools, because the LLM invariably leads to more complicated output than the author just badly, but authentically, trying to express themselves in the text.
Surely it's less work to put the words you want to say into an email, rather than craft a prompt to get the LLM to say what you want to say, and iterate until the LLM actually says it?
`ask when this will be done` -> ChatGPT -> paste answer into email
vs
type: "when will this be done?" Send.
I trust this message finds you well.
I am writing to inquire about the projected completion timeline for the HackerNews initiative. In order to optimize our downstream workflows and ensure all dependencies are properly aligned, an estimated delivery date would be highly valuable.
Could you please provide an updated forecast on when we might anticipate the project's conclusion? This data will assist in calibrating our subsequent operational parameters.
Thank you for your continued focus and effort on this task. Please advise if any additional resources or support from my end could help expedite the process.
Best regards, Sebmellen
The LLMs struggle with both but REALLY struggle with figuring out what NOT to say.
I agree that LLMs often fall short when it comes to capturing the nuanced reasoning behind implementations—and when used in an autopilot fashion, things can easily go off the rails. Documentation isn't just about what is said, but also what’s not said, and that kind of judgment is something LLMs do struggle with.
That said, when there's sufficient context and structure, I think LLMs can still provide a solid starting point. It’s not about replacing careful documentation but about lowering the barrier to getting something down—especially in environments where documentation tends to be neglected.
In my experience, that neglect can stem from a few things: personal preference, time pressure, or more commonly, language barriers. For non-native speakers, even when they fully understand the material, writing clear and fluent documentation can be a daunting and time-consuming task. That alone can push it to the back burner. Add in the fact that docs need to evolve alongside the code, and it becomes a compounding issue.
So yes, if someone treats LLM output as the final product and walks away, that’s a real problem. And honestly, this ties into my broader skepticism around the “vibe coding” trend—it often feels more like “fire and forget” than responsible tool usage.
But when approached thoughtfully, even a 60–90% draft from an LLM can be incredibly useful—especially in situations where the alternative is having no documentation at all. It’s not perfect, but it can help teams get unstuck and move forward with something workable.
How are you going to know?
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.
Precisely, in the age where it is very difficult to ascertain the type or quality of skills you are interacting with say in a patch review or otherwise you frankly have to "judge" someone and fallback to suspicion and full verification.
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
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.
For what it's worth I do actually use the tools albeit incredibly intentionally and sparingly.
I see quite a few workflows and tasks that they can be a value add on, mostly outside of the hotpath of actual code generation but still quite enticing. So much so in fact I'm working on my own local agentic tool with some self hosted ollama models. I like to think that i am at least somewhat in the know on the capabilities and failure points of the latest LLM tooling.
That however doesn't change my thoughts on trying to ascertain if code submitted to me deserves a full indepth review or if I can maybe cut a few corners here and there.
How would you even know? Seriously, if I use Chatgpt to generate a one-off function for a feature I'm working on that searches all classes for one that inherits a specific interface and attribute, are you saying you'd be able to spot the difference?
And what does it even matter it works?
What if I use Bolt to generate a quick screen for a PoC? Or use Claude to create a print-preview with CSS of a 30 page Medicare form? Or converting a component's styles MUI to tailwind? What if all these things are correct?
This whole OS repos will ban LLM-generated code is a bit absurd.
> or what it's worth I do actually use the tools albeit incredibly intentionally and sparingly.
How sparingly? Enough to see how it's constantly improving?
I don't know, thats the problem. As a result, because I can't know I have to now do full in depth reviews no matter what. Which is the "judging" I tongue in cheek talk about in the blog.
> How sparingly? Enough to see how it's constantly improving?
Nearly daily, to be honest I have not noticed too much improvement year over year in regards to how they fail. They still break in the exact same dumb ways now as they did before. Sure they might generate correct syntactic code reliably now and it might even work. But they still consistently fail to grok the underlying reasoning for things existing.
But I am writing my own versions of these agentic systems to use for some rote menial stuff.
You're kidding, right? Most people who don't use the tools and write about it are responding to the ongoing hype train -- a specific article, a specific claim, or an idea that seems to be gaining acceptance or to have gone unquestioned among LLM boosters.
I recently watched a talk by Andrei Karpathy. So much in it begged for a response. Google Glass was "all the rage" in 2013? Please. "Reading text is laborious and not fun. Looking at images is fun." You can't be serious.
Someone recently shared on HN a blog post explaining why the author doesn't use LLMs. The justification for the post? "People keep asking me."
And the "I don't use these tools and never will" sentiment is rampant in the tech community right now. So yes, I am serious.
Youre not talking about the blog post that completely ignored agentless uses are you? The one that came to the conclusion LLMs arent useful despite only using a subset of its features?
So is the "These tools are game changers and are going to make all work obsolete soon" sentiment
Don't start pretending that AI boosters aren't everywhere in tech right now
I think the major difference I'm noticing is that many of the Boosters are not people who write any code. They are executives, managers, product owners, team leads, etc. Former Engineers maybe but very often not actively writing software daily
Plenty of current, working engineers who frequent and comment on Hacker News say they use LLMs and find them useful/'game changers,' I think.
Regardless, I think I agree overall: the key distinction I see is between people who like to read and write programs and people who just want to make some specific product. The former group generally treat LLMs as an unwelcome intrusion into the work they love and value. The latter generally welcome LLMs because the people selling them promise, in essence, that with LLMs you can skip the engineering and just make the product.
I'm part of the former group. I love reading code, thinking about it, and working with it. Meeting-based programming (my term for LLM-assisted programming) sounds like hell on earth to me. I'd rather blow my brains out than continue to work as a software engineer in a world where the LLM-booster dream comes true.
I feel the same way
But please don't. I promise I won't either. There is still a place for people like you and me in this world, it's just gonna take a bit more work to find it
Deal? :)
Except we aren't talking about those people, are we? The blog post wans't about that.
> Don't start pretending that AI boosters aren't everywhere in tech right now
PLEASE tell me what I said that made you feel like you need to put words in my mouth. Seriously.
> I think the major difference I'm noticing is that many of the Boosters are not people who write any code
I write code every day. I just asked Claude to convert a Medicare mandated 30 page assessment to a printable version with CSS using Cottle in C# and it did it. I'd love to know why that sort of thing isn't useful.
I didn't draw the comparison. Karpathy, one of the most prominent LLM proponents on the planet -- the guy who invented the term 'vibe-coding' -- drew the comparison.[1]
> And the "I don't use these tools and never will" sentiment is rampant in the tech community right now. So yes, I am serious.
I think you misunderstood my comment -- or my comment just wasn't clear enough: I quoted the line "It does amaze me that the people who don't use these tools seem to have the most to say about them." and then I asked "You're kidding, right?" In other words, "you can't seriously believe that the nay-sayers 'always have the most to say.'" It's a ridiculous claim. Just about every naysayer 'think piece' -- whether or not it's garbage -- is responding to an overwhelming tidal wave of pro-LLM commentary and press coverage.
> Youre not talking about the blog post that completely ignored agentless uses are you? The one that came to the conclusion LLMs arent useful despite only using a subset of its features?
I'm referring to this one[2]. It's awful, smug, self-important, sanctimonious nonsense.
[1] https://www.youtube.com/watch?si=xF5rqWueWDQsW3FC&v=LCEmiRjP...
I once had a member of my extended family who turned out to be a con artist. After she was caught, I cut off contact, saying I didn’t know her. She said “I am the same person you’ve known for ten years.” And I replied “I suppose so. And now I realized I have never known who that is, and that I never can know.”
We all assume the people in our lives are not actively trying to hurt us. When that trust breaks, it breaks hard.
No one who uses AI can claim “this is my work.” I don’t know that it is your work.
No one who uses AI can claim that it is good work, unless they thoroughly understand it, which they probably don’t.
A great many students of mine have claimed to have read and understand articles I have written, yet I discovered they didn’t. What if I were AI and they received my work and put their name on it as author? They’d be unable to explain, defend, or follow up on anything.
This kind of problem is not new to AI. But it has become ten times worse.
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.
I do not lightly say that I don't trust the work of someone who uses AI. I'm required to practice with LLMs as part of my job. I've developed things with the help of AI. Small things, because the amount of vigilance necessary to do big things is prohibitive.
Fools rush in, they say. I'm not a fool, and I'm not claiming that you are either. What I know is that there is a huge burden of proof on the shoulders of people who claim that AI is NOT problematic-- given the substantial evidence that it behaves recklessly. This burden is not satisfied by people who say "well, I'm experienced and I trust it."
You're right to call out the need for vigilance and to place the burden of proof on those of us who advocate for this tool. That burden is not met by simply trusting the AI, you're right, that would be foolish. The burden is met by changing our craft to incorporate the necessary oversight to not be reckless in our use of this new tool.
Coming from the manufacturing world, I think of it like the transition in metalwork industry from hand tools to advanced CNC machines and robotics. A master craftsman with a set of metal working files has total, intimate control. When a CNC machine is introduced, it brings incredible speed and capability, but also a new kind of danger. It has no judgment. It will execute a flawed design with perfect, precision.
An amateur using the CNC machine will trust it blindly and create "plausibly good" work that doesn’t meet the specifications. A master, however, learns a new set of skills: CAD design, calibrating the machine, and, most importantly, inspecting the output. Their vigilance is what turns reckless use of a new tool into an asset that allows them to create things they couldn't before. They don't trust the tool, they trust their process for using it.
My experience with LLM use has been the same. The "vigilance" I practice is my new craft. I spend less time on the manual labor of coding and more time on architecture, design, and critical review. That's the only way to manage the risks.
So I agree with your premise, with one key distinction: I don’t believe tools themselves can be reckless, only their users can. Ultimately, like any powerful tool, its value is unlocked not by the tool itself, but by the disciplined, expert process used to control it.
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.
Every time I tell claude code something it did is wrong, or might be wrong, or even just ask a leading question about a potential bug it just wrote, it leads with "You're absolutely correct!" before even invoking any tools.
Maybe you've just become used to ignoring this. I mostly ignore it but it is a bit annoying when I'm trying to use the agent to help me figure out if the code it wrote is correct, so I ask it some question it should be capable of helping with and it leads with "you're absolutely correct".
I didn't make a proposition that can be correct or not, and it didn't do any work yet to to investigate my question - it feels like it has poisoned its own context by leading with this.
I did get curious though. So I decided to look up some of the times where I did correct it after I dismissed a change. I only looked at a couple of prompts but most of the AI responses looked like this:
"There are two issues...", "The error is because...", "The error persists because...", "A new route, /class_ids/fully_owned, has been added...".
I was feeling confident that it wasn't bullshitting me at that point, but then I get to this one:
"Thank you for the details. The error response..."
Now, that is the AI agent. If I use the browser or one of their "apps" the LLM politeness encouragement bullshit alone will often be longer than the entire chat response in an agent. Like this is the entire response to what it was tasked with in my example:
"To add this route, I'll implement a new endpoint that queries all unique class_id values and checks if all items with that class_id have is_owned == True. The route will return a list of such class_id values.
I'll add this as a new GET route, e.g., /class_ids/fully_owned."
ChatGPT (which is supposedly the same model) will spend those lines telling me what a great question it was.
I’d love to hear more about your workflow and the code base you’re working in. I have access to Amazon Q (which it looks like is using Claude Sonnet 4 behind the scenes) through work, and while I found it very useful for Greenfield projects, I’ve really struggled using it to work on our older code bases. These are all single file 20,000 to 100,000 line C modules with lots of global variables and most of the logic plus 25 years of changes dumped into a few long functions. It’s hard to navigate for a human, but seems to completely overwhelm Q’s context window.
Do other Agents handle this sort of scenario better, or are there tricks to making things more manageable? Obviously re-factoring to break everything up into smaller files and smaller functions would be great, but that’s just the sort of project that I want to be able to use the AI for.
Context size matters a lot in my experience, but I'm not sure if it matters whether your 100k lines are in a single or multiple files. I tend to cut down what I feed the agent to the actual context, so if I have a 100k line file, but only 3000 lines matter, then I'll only feed those 3000 lines to the AI. Even in a couple of small files with maybe 200 lines of code in total, I'll only give the AI access to the 40 line which is the context it needs to work on.
English isn't my first language, so when I say context, what I mean is everything which is related to the change I want the agent to do. I will use SQLC as an example. Even though I feed the AI the Go model generated, I'll also give it access to the raw SQL file.
> Obviously re-factoring to break everything up into smaller files and smaller functions would be great, but that’s just the sort of project that I want to be able to use the AI for.
I'm guessing here, but I think part of our success is also our YAGNI approach. AI seem to have an easier time with something like Go where everything is explicit, everything is functions and Go modules live in isolation. Similarily AI will do much better with Python that is build with dataclasses and functions, and struggle with Python that is build upon more traditional OOP hierarchies. We've also had very little success with agents on C#. I have no idea whether that is because of C#'s inherrent implicity and "black magic" or because of the .net > .net core > .net framework > .net + whatever I forgot journey confusing the prediction engine.
> Do other Agents handle this sort of scenario better
I don't know. I've only used the sanctioned co-pilot agent professionally. I believe that is a GPT-4 model, but I'm not exactly sure on the details. For personal projects I use both the free version of GPT-4 in co-pilot and Claude Sonnet 4, and I haven't noticed much of a difference, but I have no hobby projects which are compareable.
So they’re even more confident in their wrongness
Not they don't. This is 100% a made up statistic.
It seems that LLMs, as they work today, make developers more productive. It is possible that they benefit less experienced developers even more than experienced developers.
More productivity, and perhaps very large multiples of productivity, will not be abandoned due roadblocks constructed by those who oppose the technology due to some reason.
Examples of the new productivity tool causing enormous harm (eg: bug that brings down some large service for a considerable amount of time) will not stop the technology if it being considerable productivity.
Working with the technology and mitigating it's weaknesses is the only rational path forward. And those mitigation can't be a set of rules that completely strip the new technology of it's productivity gains. The mitigations have to work with the technology to increase its adoption or they will be worked around.
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...
[1] https://www.forbes.com/sites/robertanzalone/2020/07/15/big-c...
I agree with this quite a lot. I also think that those greenfield apps quickly become unmanageable by AI as you need to start applying solutions that are unique/tailored for your objective or you want to start abstracting some functionality into building components and base classes that the AI hasn't seen before.
I find AI very useful to get me to a from beginner to intermediate in codebases and domains that I'm not familiar with but, once I get the familiarity, the next steps I take mostly without AI because I want to do novel things it's never seen before.
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.
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.
But anyway I'm at the point in my career where I am not learning to code/can already do it. Sure languages are new/can help there for syntax
edit: other thing I'll add, I can see the throughput thing, it's like a person has never used opensearch before and it's a rabbithole, anything new there's that wall you have to overcome, but it's like we'll get the feature done, but did we really understand how it works... do we need to? Idk. I know this person can barely code but because they use something like chatGPT they're able to crap out walls of code and with tweaking it will work eventually -- I am aware this sounds like gatekeeping from my part
Ultimately personally I don't want to do software professionaly/trying to save/invest enough then get out just because the job part sucks the fun out of development. I've been in it for about 10 years now, should have been plenty of time to save but I'm dumb/too generous.
I think there is healthy skepticism too vs. just jumping on the bandwagon that everyone else is doing and really my problem is just I'm insecure/indecisive, I don't need everyone to accept me especially if I don't need money
Last rant, I will be experimenting with agentic stuff as I do like Jarvis, make my own voice rec model/locally runs.
I think what the author misses here is that imperfect, probabilistic agents can build reliable, deterministic systems. No one would trust a garbage collection tool based on how reliable the author was, but rather if it proves it can do what it intends to do after extensive testing.
I can certainly see an erosion of trust in the future, with the result being that test-driven development gains even more momentum. Don't trust, and verify.
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.
> 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.
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
> require them to be majority hand written.
We should specify the outcome not the process. Expecting the contributor to understand the patch is a good idea.
> Juniors may be encouraged/required to elide LLM-assisted tooling for a period of time during their onboarding.
This is a terrible idea. Onboarding is a lot of random environment setup hitches that LLMs are often really good at. It's also getting up to speed on code and docs and I've got some great text search/summarizing tools to share.
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?
Some people ask for a coworker to do it for them, c/p stackoverflw, etc and never learn.
AI makes the first group that much more effective. Setup should be a learning process not a hazing ritual.
To add to this, a barrier to contribution can reduce low quality/spam contributions. The downside is that a barrier to contribution that's too high reduces all contributions.
This is not a given!
If we automated all accounting, why would anyone still take the time to learn to become an accountant?
Yes, there are sometimes people who are just invested in learning traditional stuff for the sake of it, but is that really what we want to rely on as the fallback when AI fails?
I’ve never heard of this cliff before. Has anyone else experienced this?
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 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.
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.
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.
Snipping out the flailing in this way seems to help.
People love to justify big expenses as necessary.
The online dialog about AI is mostly noise, and even at HN it is badly distorted by people who wince at $20 a month, and complain AI isn't that smart.