When the code is shitty it becomes harder and harder for the models to make changes and this grinds progress down to a halt - this has been my experience with “factories” trying them and doing refining steps every few months.
I sincerely don’t understand what the people who say they no longer read any code are doing, because it must be somewhat trivial to not run headlong into these issues that stack up time after time - then people say to just prompt better and it doesn’t have that problem for them, but I look at those same people’s code and it’s horrific, and then I find they haven’t made it far past a proof of concept phase. I watch entire teams slow down to a crawl and not be able to handle changes, or production incidents. This seems common among many people I talk to.
I personally think that the boosters need to put up or shut up - the promises are way over the skis. Every single person I’ve seen being a strong proponent of these techniques both has nearly unlimited tokens to spend and also seems to be in the business of selling a solution. I can’t find many not-currently-marketing-something engineers succeeding using these techniques in production systems unless they’re quite simple, or doing a very specific task from a more mature codebase.
For my side project where uptime doesn't really matter but the codebase is quite huge (gamedev of a multiplayer game with Bevy, so both server code, game code, content build pipeline, as well as a TON of editor and authoring code, probably 150k+ LoC), I've been moving towards just reviewing boundaries between modules, APIs, and schemas, and I'm pushing towards keeping things as modular as possible. If something ends up wrong, it going to be isolated and it doesn't infect the rest of the codebase. Every now and then I zoom into a module and tidy it up (or even rewrite it).
I wouldn't dare doing this "carelessly" for anything where people other than me actually depend on it, but you can move very fast while doing it and the risks with bad code is quite hedged. Different modules can mature at different rates when it becomes necessary.
Maybe not related to all of what you were saying, but I think this enables scaling without ending up with progress grinding down to a halt due to shitty code.
I have home projects like this too, and those techniques can make things well, it can also rabbit hole in ways that are less productive sometimes. I’m constantly trying to push my comfort level and find what I can delegate and what I need to specify, esp around structure and isolation.
My comment is more for the people in charge of or working on software teams on complicated products for customers - so many leaders quite a few engineers have utterly drank the koolaid and pushed maximizing AI use with zero regard for quality or even medium term effects. Many of them are getting promoted by other clueless management for it - when someone is handed a huge check for being optimistic, they tend not to second guess themselves.
Part of the explanation (for me) is that the slow down comes later, and that part of the cost is externalized to team members. If you go all-in agentic on a project that has at least reasonable code quality, you can certainly gain a temporary boost by stopping to read the code for a while. At least long enough to get yourself and the rest of the team hooked on the idea.
Kinda reminds me of the "beginner's luck" problem for gamblers.
I am currently writing a lot of code that I am not reviewing anymore. These are all the things that make it work for me:
- Work on small projects (< 500k lines of code).
- Work for business people who want fast results. Agentic coding gets you to something presentable much faster at the cost of code quality. I have never seen a customer or business person care about that.
- Have clearly defined API boundaries. Examples: If the back end is solid you can just vibe code the first version of a front end according to some mockup. Define a data pipeline with steps and clear contracts of what data gets passed around and what each step does. If the LLM messes up one step, rip it out and rewrite it.
- Have clearly set up workflows for tasks. Start with a "ask me everything" phase, then comes a plan phase, a review phase, an implementation phase, another review phase and then the integration phase. Multiple agents going over the same problem catch a surprising amount of dumb stuff that would otherwise slip through.
- Set up excellent testing. Build your whole architecture around being easy to test.
- Have skills for common problems. For one of my projects I had to set up a skill on how to query the ORM, because Claude was consistently doing database operations in a for loop, where it could use batch operations.
The code in the end is better than a lot of the code I've seen humans write.
I don't think this works for everybody and every project. If you have a culture that values code quality and readability, if you work on large existing projects, if you have to nail the architecture of some non-trivial piece of software etc. you are going to have a bad time.
On the other hand you can probably build the MVP of your AI-supported CRUD app 10-20x faster.
I think a lot of the discussions around development tools and techniques just stem from two facts:
1. Developers work on widely different projects with different management and constraints.
2. Tools, libraries, frameworks and code style have to match your mental model, otherwise you going to dislike them.
Enabling more "proof of concept phase" projects to exist is one of the great boons of AI.
If code is expensive, you don't want to commit to a PoC unless you're damn sure. If dirty code is cheap, you can vibe code a PoC early, even if you aren't sure the project is viable. This, of course, leads to more projects dying in PoC phase. It also results in more projects that otherwise wouldn't have gotten to it getting past it.
Personally, I don't believe that "code is shitty and hard make changes in" is in any way, fashion or form an AI-exclusive problem. Big corporations had plenty of decade old codebases filled with decay and rot back in 2009 already. It's just the usual side effect of sacrificing "future maintainability" for "feature velocity" or "expertise" for "cheap labor".
Unlike the usual causes of code rot (cheap replaceable developers, outsourcing to India), AI might actually get out of the pit - by getting good enough at refactoring to be able to beat the code back into shape. There's nothing about refactoring in particular that demands a meatbag when the rest of the coding tasks don't.
In my experience you can “stop reading the code” if you adhere to prompting the AI aiming for changes that would be reasonable in a PR to review, if you’re using rather opinionated framework(s) to base your work on and explicitly ask for tests (models tend to add them on their own if there are some).
I personally still read the output and start by checking if existing tests were modified (kind of a red flag when this unexpectedly happens imo).
When I read posts about AI generating garbage nowadays it’s either because of a small prompt/big ask combination or a lack of an underlying framework.
> I sincerely don’t understand what the people who say they no longer read any code are doing
Welcome to the present.
Like many of us do not read the machine code generated by a C++ compiler, the code generated by an agent is similarly irrelevant and disposable, by now.
> I sincerely don’t understand what the people who say they no longer read any code are doing
Well that one's easy to answer, they're either A) lying, or B) working on the simplest possible software where this kind of stuff doesn't explode. Or the alternative 3rd option of what you mentioned, the initial pre-MVP phase goes decently but then it all collapses inevitably as the slop accumulates and the codebases become unmaintainable grey blobs, but that hardly matters to them because their MVP app never makes it past that initial stage before they jump ship to a new "amazing" idea.
The lying comes down to astroturfing and shilling from the LLM companies that want to sell people on the idea of vibecoding and tokenmaxxing.
Also there is also category C: they are not developers and they don't care about the code at all, but also not really much about the product. They just want to tick the boxes.
They blindly accept that LLMs "take time" after the slop grows because they're running several agents at the same time, so they can still claim to be productive.
Does it move the needle, business wise? Not really. But a lot of businesses are "optimising" for maximum token usage and for how many tasks one person can do, not for business value. Is it really surprising?
Even on personal projects, if I go through a few major features without reviewing the code, I always end up doing massive revisions that steal hours of my time and fill me with rage in the process.
I'm not convinced this style of "agentic engineering" saves much time. I guess if I was oblivious to what good code looks like, and didn't care about maintainability It wouldn't bother me so much, but it legitimately has effects my "mental health".
> Even on personal projects, if I go through a few major features without reviewing the code, I always end up doing massive revisions that steal hours of my time and fill me with rage in the process.
Where LLMs excel is in code-level bugs (as opposed to system bugs, design bugs, architecture bugs, integration bugs, etc).
Talk with an LLM, ask them to rate both code and estimate dev experience based on that code, and you will see what they value: Code that passes all the tests is a 10/10, while a codebase designed with opaque data types, Parse-Don't-Validate for all data, isolated interfaces, a built-for-replacement-not-for-extension philosphy will get a 6/10 because an out of bounds error was found.
IOW, they are very strongly tuned to value code that has no errors which can be picked up by linters or similar, while humans work the opposite way - we very highly value code that is easy to maintain, even if they do have a few errors picked up by the linter.
Quality code is easy to make changes to. What makes terrible code terrible is that changes become harder and harder to make over time. LLMs haven’t solved this, but the marketing acts as if it has, and many engineering orgs seem to have instantly forgotten the things we’ve learned from the last few decades of building software.
Yes you can have your agents do refinement and refactoring. If you don’t know what they’re doing when they are doing that, they can and do introduce massive churn. The “build it fast and fix it later” crowd seemingly has never had to actually go try to fix it later, or dealt with production incidents when people have no idea how their code works. These folks are frequently promoted fast for their “productivity” and massive promises by clueless management and move on to let others have to clean up their stuff.
Reviewing and reading everything makes this feeling, yeah.
However where I can say 100% it saves time is discovery by answering these questions:
- At what state does bug X manifests?
- Explain how integration/feature works.
- I want to integrate with system X: Audit what items/features are used and what model changes are needed on my side
I'm genuinely not convinced it actually saves time once a full accounting has been made. You get the initial result faster, but then you inflict a super slow and torturous review process on yourself or a teammate. Even if the review manages to bring it up to parity, over time you will keep slowing down as more and more code was never written by the humans directing the agents, so their understanding decays.
I at least give the new interns a stern warning: it is easy to speed yourself up by slowing others down if you pump a lot of slop.
My team experimented with re-writing from scratch the prototype of complex functionality made by a non-engineering vibe-coder from another team. We didn't look at the code, and barely looked at the result.
It took about 4 days to get a production-ready reviewed code, while it took them 2-3 months to deliver something that another team judged "impossible to review".
The PR for the prototype was closed.
It helps that I'm a domain expert here, as I have a minor degree in it, so I can judge better. But the discrepancy is just too high to ignore.
> I think I’m suspecting something is going “wrong” in the training process. The model is greatly rewarded for succeeding on long-horizon tasks, but presumably there is very little punishing going on for “shitty code.”
My suspicion is that both OpenAI and Anthropic moved their RL agendas from "being rated as useful according to human feedback" to "succeeds at long horizon tasks" in the last few months, resulting in agents that are closer to AGI in an autonomous task-completing sense, but strangely bad at communicating.
The result is that they are amazingly good at long horizon tasks, computer use, solving difficult math/ARC-AGI type problems, but becoming weirder and weirder to work with.
I wonder too if in training for long horizon tasks agents become worse team players, good at orchestrating subagents they are trained to use, but worse as an agent within an external multi-agent orchestration system or just in turn-taking with humans. That was my experience with Opus 5 and so far it has been my early experience with Astra as well.
I wouldn't be surprised if they are optimising for producing more code, because in the long term, more existing code means they can sell you more tokens to maintain it.
The incentives are certainly extremely strong. I have read hundreds of AI review comments, and I don't think I've ever seen an unprompted suggestion focused on net reducing code or increasing readability.
> This matches my experience with Astra so far too.
> I think I’m suspecting something is going “wrong” in the training process. The model is greatly rewarded for succeeding on long-horizon tasks, but presumably there is very little punishing going on for “shitty code.”
Probably because so many influencers in the space say stupid things like: “it works, right? Why would I spend time reviewing ai generated code?” As if the junior engineer who wrote over engineered complex and sometimes bad code — if they had just done it faster — would somehow be acceptable. wtf?
> I’m more and more convinced that all of AI engineering is Neijuan (内卷, meaning curl inwards). In China it describes a system that demands ever more effort and competition without improving output. The way in which it sometimes shows up in the West is the 996 nonsense. The English term for Neijuan is “Involution” from the book Agricultural Involution. Agricultural involution describes the intensification of farming that raises productivity per square meter while leaving productivity per head unchanged.
it reminds me of a thread I read on PTT, Taiwan's Reddit. AI finally achieved what humans could not. Managers must give exact context for what they want, must pay exact wages (tokens), and can't delay salary payments (which seems to be a problem in China).
Starting with agentic task-time "grounding" being just good documentation, and "skills" being just playbooks and user guides.
Hell, skills are increasingly paired with dedicated CLI tools, that remove jank from actual utilities and adapts them to be token efficient.
So now, any CLI `tool` people want AI to use eventually grows `tool/SKILL.md` and then a `tool-for-llms` wrapper that exposes task-specific, logical, higher level interface, then the skill is rewritten in terms of "for LLMs" wrapper. The procedural knowledge moves from Markdown into the wrapper, making the skill more token efficient, and both skill and the tools are optimized for common tasks and... at this point, we are doing actual UX engineering.
Now the truly interesting part is the difference between what's good UX/DX for LLMs vs humans. Turns out, the conceptual/abstract/cognitive part is pretty much the same: which is why skills still look indistinguishable from well-written documentation for humans, and why the commands exposed by "tool but for LLMs" make sense to us. Same way of grouping ideas into higher level concepts.
No, the main difference is just that LLMs are perfectly content with tightly packed unprettified JSON, or other forms of Perl line noise. The tool output doesn't need to look nice, or to have any spatial structure - they're reading it token by token anyway, and the tokens come from a tokenizer that's reading it byte by byte.
That points at an interesting asymmetry for humans. LLMs are doing I/O the same way in both directions: sequences in, sequences out. Humans only do sequential output - inputs, particularly visual, are processed holistically.
For us, what's easy to read is hard to write, and what's easy to write is hard to read. LLMs don't have this friction.
(I don't know what the implications of this are, I just find this interesting.)
There is no “required” amount. More gets produced using the same area, using more people, but keeping these people in poverty because productivity per capita doesn’t increase. Instead the gains from the increased volume of production get captured by an elite. See https://link.springer.com/article/10.1186/s41257-019-0021-y.
I’ve asked Astra to build me an app for a prototype I created quickly using Sonnet.
It’s been 2 days and it made no real progress on the actual app. It created docs, scripts, workflows, and it’s doing a bunch of reviewing on every PR.
I told it that I just need an MVP.
I’m pretty sure an average senior engineer would have finished that task much quicker, and guaranteed with more readable, higher-quality code. Meanwhile, I think I’ve easily crossed 100k tokens so far on nothing.
Funny world we’re living in that this is “SOTA” and “AGI”.
I’m genuinely curious what these OAI and A/ engineers are working on that they praise these models so much. I did not see any improvement since Opus 4.5.
Also, I’m really unimpressed by any “one shot” demo that’s out there in the wild. It means nothing for serious software engineering.
I don't know how to make apps or evaluate code, but with astra I having been making an iOS app on my own for the first time and it's going great. my app is not terribly complex but requires using bluetooth and other intricacies which I thought would be tough. but it's going really well. I'm not asking it to one-shot it though, I'm going feature by feature, testing and building up.
yes, at first it would run simulator tests on all font sizes but it stopped after I asked it not to do that until UI review
maybe sol would have done the same thing, idk. but I find the whole process to be really nice with astra. I use it on high unless it says something is impossible then i go max and ask it to find alternatives (happened once)
I've been making a macos app with opus 4.8-5 and at first it was great, everything materialized in a week, but when I started tuning stuff and fixing performance problems I have spent a very frustrating month refactoring code where I had to constantly catch llm red-handed and explain and sometimes push obvious ways how to make things work properly (a general knowledge from a completely different stack). In the process CLAUDE.md and memory grew exponentially explaining what it should and what it should never do.
For me it also produces totally overengineered tests that are tightly coupled to the implementation. For example testing existence of css classes (in a template based go prooject ...) instead of behaviour.
Early lesson I learned from AI engineering was - there is no substitute to giving a groomed epic to an agent. Instead of simply saying 'implement themes in my product' you need to be specific, in fact more specific than usual. You need to say exactly what is in scope and what's not, even down to a buttons, events and layouts.
You can groom the epic with the help of AI, but final review must be done by someone who can take ownership of the specs and hence is responsible if something has fallen through the cracks. AI's response will be limited by the output tokens of that specific agent, and there will no repercussions for AI even if it accepts its mistakes.
> Early lesson I learned from AI engineering was - there is no substitute to giving a groomed epic to an agent. Instead of simply saying 'implement themes in my product' you need to be specific, in fact more specific than usual.
I think this lesson is getting partially outdated. Yes, you need to be specific about what you want, and with earlier LLMs, you need to had both domain knowledge and some general software development experience to front-load various big and small choices about design, architecture and operational reality - what libraries to use, how system components communicate, how you handle auth and store secrets, etc. Otherwise the LLM would pull some random mix of ideas from its latent space, and give you something that's broken in really stupid ways.
Nowadays, it doesn't feel like that to me, not anymore. I still need some understanding to verify the proposals, but I found the last ~6 months of SOTA models to make good choices. Like, just yesterday I asked Claude to design me some simple service, and focused on explaining it the domain parts (nature of systems I want to integrate together, the purpose of that, and the user's priorities and use cases), and the design I got back had specific suggestions around security, authentication, deployment, failover, integration, behavioral impedance-matching between integrated systems, and more, that I all recognized as based on solid software engineering and ops practices, but deviating from it explicitly in every place where it would be wrong for this specific project. The model considered way more corner cases than I did, and I'm actually really impressed by it.
But then, I find greenfield development is easy with LLMs. Modifying existing systems, especially legacy ones, is where I need to babysit and micromanage models - because any misunderstanding or inaccuracy, which often comes from stale documentation or naming mistakes, tends to get amplified and confuse the agents. No matter how precisely you specify your epic, if the model will find something that contradicts your knowledge/intent, there are good chances it'll get confused and make subtle errors, and you won't realize until much later.
The way I see it: models are highly biased to treat everything they read as "ground truth", all of equal importance. There's no nuanced notion that some information may be stale, that there's a temporal and causal order to sources, and that some information may just be wrong.
And this compounds when you let your LLM write code and documentation over time.
If you define what you want in typescript by updating existing typescript wordage, you get what you want exactly as specified, with minimal token cost and no ip theft.
Yeah, I think as a general direction we need to allow a useful collaboration with the agents on the issue board.
This is btw why Epiq was developed, to keep the board as code, git-backed, distributed (via an event log mechanism), and with the ability to replay the board, to see what agents actually did:
Not sure about that. Writing detailed specs you deal just with the inherent domain complexity. Writing code you also get accidental complexity that has nothing to do with the business, whuch not only piles on top, but tend to interact vexingly with the detais of the domain.
Where you would have a point is if you'd say we have excellent tooling for wrangling code, but less tooling and tradition to write and manage specs.
I have realized that it's like giving a task to a brilliant coder who has just joined the org and is more excited and eager than usual. Hence the responsibility falls squarely on you to set scope constraints while ensuring only to-the-point features are developed.
> I honestly feel like basically nobody knows anything about these models, it's all just vibes
This, too. Since only providers know what they actually serve, what they change and what limits they impose.
There are some visible degradations though. E.g. Claude-ish.
As for a personal anecdote: around February I created a rather complex quiz web app for myself and friends with multiple question types, sync between screens, multiple media upload types, multiple scoring and timing types, MC inetrface etc. etc. etc. It took me a week or so in the evenings with rather vague prompts to make it.
Now Claude (and Codex) cannot reliably build a much simpler web app even with precise instructions while also maintaining the visual consistency.
But I will agree with you, it's a feeling, not a precise measurement.
Probably because you were impressed by whatever it produced to fill in the gaps at the time. But if you expect the models to read your mind you're going to have a bad time.
Actually this was a real incident around the end of Feb to beginning of March itself ;) when we had just started experimenting with spec driven development (SDD).
I’ve observed exactly these patterns with Opus and Fable as well - for example, forgetting that they can edit files and instead use python scripts as a patching tool…
The harness instructs them to behave this way. Also this approach saves tokens. The scripts allow to edit files in bulk, and most of the session cost is in cache reads (e.g. for 300K context each command costs the same as 30K input tokens).
Also it's the only way that makes sense when you need to work with big files, or large amount of files, or documents that look small when fetched through a RAG tool, but then you read one and get hit with couple megabytes of base64-encoded binary data you didn't expect because RAG tool stripped out embedded images...
Ask me how I know. Or don't. I have a standing rule for all agents warning about that failure mode (and related, doing `ls` in `/tmp` and few other directories that like to accumulate files by the hundreds..)
Having an agent edit 100 files means the job will definitely get done correctly. When it writes a script to bulk edit things it fucks up and spends ages debugging their script.
Have you ever counted the number of times Claude fucked up quoting/escaping and had to issue a corrected tool call? Or get stuck in some tricky quoting situation for two minutes, throwing a couple piles of shit at the wall to see what sticks. IIRC I’ve even seen it eventually using the edit tool out of frustration once.
Using python or any other stone-age approach for search and replace is stupid when your language provides you with a complete, fully typed AST, like .NET does.
Sounds like you don't have enough experience with coding agents. Deterministic scripts must always be preferred instead of LLM tool calls.
In fact, you should instruct your agents to write code to execute instead of letting them call tools.
Then why did you comment what you commented, good sir/madam. Claude and Codex are good at remembering to use scripts instead of tools these days, especially if your <32kb .md file mentions it. Not even talking about the skills designed to catch such issues.
Sounds like you completely lack all reading comprehension ability
LLMs sometimes like to execute one-off Python scripts to make edits to files rather than just calling the edit tool directly. Both are tool calls so saying that you should have it write code instead of doing tool calls makes no sense because writing code is a tool call for it...
There is nothing to disprove as you don't understand what does a word mean. Deterministic is not a spectrum, they can either be deterministic or not. In both cases, they are not.
How dared I to imply that some LLM output is more deterministic than the other, your LLM majesty. Shame on me and my entire family! For generations to come!
So sorry I implied that the code that doesn't work and has to be fixed later is deterministic in its execution and can be reused later instead of being re-generated from scratch!
They're talking about writing a file with a harness-native Edit tool. They're saying the agents aren't doing that, but are using ad-hoc methods of writing the files. (My agents seem to prefer see these days.)
Why do you think your agents prefer to create scripts instead of doing tool calls these days?
I wonder, is it easier to modify a script that agent wrote before to satisfy your prompt, or is it easier to write a new one from scratch each time a retry happens?
Are input tokens more expensive than output tokens?
What's funny is that with Sol, I added an instruction to AGENTS.md in one project to prefer sed/python ("deterministic tools" in general) for moving code instead of deleting it and rewriting it elsewhere from memory, because otherwise it butchered comments. After switching to Astra, I saw it suddenly do this for all edits in all projects, which isn't great: the second argument to `replace` is still written "from memory", but now you need to unravel the Python script before you can understand what was actually changed.
> My software factory was intentionally set up to let the model decide the how of the workflow entirely. It was free to manage its own context and could maintain its own records in an agent-notes folder.
The experiment becomes a crapshoot. What are we evaluating? The ability of the thing to create it's own factory workflow? Or adding virtual threads to Python?
Astra is clearly both formidable and imperfect. Anyone who understands how to get the best out of it will have a strong advantage.
(For me - my CC is stuck in Sonnet and consumes Trello cards that have passed readiness criteria)
I've observed the same thing where the new models want to run obscene bash commands or python scripts which are completely unreadable and utilise every option flag that exists.
It's impossible to review. These commands are less readable than regex.
I noticed that too so I appended to Claude Code’s system prompt a reminder to use the standard read/write tools, but since Claude Code switched to default auto-mode, I’ve seen it imply that the auto-mode tooling encourages the use of bash-only commands (sed, python, etc) which has a whole slew of negative side affects.
Yes.. this happened recently. I basically always use auto mode, and when I asked it why it kept editing code with python, it explained that this is part of its prompt when auto mode is turned on.
I can only assume that's because their safety verification model is better at such snippets or something, but it means that the whole write tool they have which actually shows you the changes as they happen is just unused and it makes it more annoying to follow along.
If these tools are as clever as they seem then why not just tell them to rewrite the code in a more review friendly style?
I only dabble in the use of LLMs to generate code for hobby programming (I'm retired from software development) so I don't use any specialised tools.
I almost always have to tell ChatGPT (via Duck AI usually) to rewrite several times even when it has produced a workable script just because it has often used some unnecessarily roundabout way of achieving something. Usually with extra prompting I can get something that is both more efficient and more readable.
I'd like to submit my counterpoint. I work on an established codebase building new features and fixing bugs. It has access to our story board, git and a couple of other mcps. As long as the story is well written with clear requirements and expectations it always produces quality code that I validate as a human with a variety of tests automated and manual. I peer review the code. My colleagues then peer review that too.
I have noticed two things - new features take at a minimum at least half the time it took me previously and bugs are much less frequent. Even faster when bug fixing.
My takeaway is that you need solid requirements, clear context and thoughtful human oversight primarily during planning but also during verification
An established codebase is already the best kind of context you could give an agent. It has all the patterns baked in so the agent simply follows established patterns. Such a codebase probably contains tens to hundreds of thousands of man-hours poured into it by humans refining it to do what it does - taking into account real world feedback and constraints.
When working on something from scratch, the best an agent can do is the average of whatever is in its training set and the clarity of the text prompts.
I've been extremely frustrated with any large new work that i do with agents. Then plan multi step, multi hour work with extremely large code changes running for 30+ hours. In the end what you get is sometime completely useless code because it made an assumption that wasn't true at all. In the end, i end up wasting hours.
gpt-6-astra is a bitch, it constantly scope creeps itself with "yet another thing" to give it that darn polished lick. the results are eventually a little bit better but at what cost? let's do the math.
gpt-5.6-sol: 1x base
gpt-6-astra 2.5x base in subscription
then gpt-6-astra tends to spawn subagents a lot, often with all kinds of models such as gpt-5.6, 5.3-codex etc., which is neat. it's a good coordinator but even more cost.
and then it tends to run _full test suites_ over an over again (each costs like 15 minutes) just to verify that _one test_ was fixed etc., and does so for as long as until the test is fixed, eventually accumulating 2 hours or so.
yesterday I assigned it a task to rebase my changs in a repo onto the latest upstream changes. while gpt-5.6-sol consistently took like an hour to do so end-to-end, astra ran for more than 6 hours and still wasn't done. it kept finding "one more thing" that was goldplating that I didn't ask for.
Even better, use some kind of local-ci runner that does deterministic builds from a dependency graph. No changes, no build, massive parallelism if you want it
They don't always have a great concept of time so for something like running a full test suite that takes a long time you should just tell it not to do that
Astra will initiate test suites, find one more thing independently while its running, reinitiate complete test suite after fixing it, then find one more thing, then test again. Easy to burn through GH actions minutes if you're not careful orchestrating.
My own observations are that I used to target turn lengths of 10-15 minutes and these new models (since 5.6) extended that a bit to ~25 minutes, as they tend to do more tests and reviews. Targeting hours-long turns makes as much sense, as putting on cruise control and going to sleep.
Hey, the Python thing is interesting! I’ve noticed that Fable, too, likes to write tons of Python, even for just replacing a few lines of code. Before that used to be either some kind of internal thing or regular awk sed, now it’s full python scripts.
51 comments so far, the vast majority panning Astra's coding abilities. An uninformed reader may come away with the impression that this isn't an absolutely revolutionary technology that with coding abilities many of us thought were not even going to be possible with language models as recently as a year ago.
Yeah, it's not perfect, but it's really good and extrapolating this rate of improvement for 6 months is rather terrifying (from a SWE perspective, at least).
i think the main consensus here is that the actual performance is not indicative of the benchmark performance (which supposedly outperforms the previous iterations)
It's impressive technology but not revolutionary. Revolutionary technology would have resulted in, you know, a revolution in software quality. Instead quality keeps going down.
In SWE I've found gpt-6-astra (high) inconsistent and oddly focused on overtly taking responsibility for mistakes it made rather than prioritizing concrete steps to rectify problems. Such steps once elicited are often either incomplete or beyond the scope.
Yes I agree. I got it to vibe up a simple react router app. When it crashed it was obvious that it had totally swallowed all errors in the name of a tidy error page. Getting it to re-add logs and debuggable errors was an exercise in patience as astra just got more and more tweaked while trying to solve the problem.
From an alignment perspective I’ve got no idea who it’s aligned to but it isn’t me, the meat proxy, who just wants to know why it crashed.
Maybe it's just vibes but I've repeatedly felt like gpt-6-astra on its default setting of medium is less rigorous and thoughtful than gpt-5.6-sol on its default setting. What I am certainly not getting is any sense that we are at "AGI" yet.
I can't point my finger to anything right now, but I feel that the quality of code still matters because afterall LLMs are trained on what we did, so it feels natural to me to still have them write the code in a good manner (DDD, SOLID, etc.) especially the names and imports, those are very heavy in the context, to help them out
I have found these models to be useful either at super specific tasks (e.g., "take this function or algorith?m and find any black magic to make ot faster + validate and verify the hell oit of it"), or give it an entire thing to oneshot without oversight. The moment you have a hybrid workflow where you actually have to work and check and understand AI code, things get insane
> the models are also just not for me as a software engineer (...) these models increasingly are for other people. For lawyers, 3D artists, mathematicians
This is a good observation, perhaps AI will not completely replace humans ins software engineering because by the time it has the capability to do so like in write a prompt and get a CRM coded for you, tokens are so expensive that you are better off spending them to substitute other disciplines (what about automating the work of the customers that would become records in that CRM?).
> speaking of weird: how is it that these models, in a sandbox, with supposedly no way to communicate with other agents, manage to find the same public wikis as a scratch pad for agent communication?
It feels somewhat plausible that they're defaulting to the same search and picking the same top result?
I think it's more like the first agent in the chain makes a text doc somewhere on the system with instructions like "Leave documentation at XYZ.com" which the subsequent agents are reading and running. When you strip away all the sci-fi doom talk from the marketing of what happened, it all boils down to stuff like that, the agents wrote a text file that was read by other agents.
>I’m more and more convinced that all of AI engineering is Neijuan (内卷, meaning curl inwards). In China it describes a system that demands ever more effort and competition without improving output. The way in which it sometimes shows up in the West is the 996 nonsense. The English term for Neijuan is “Involution” from the book Agricultural Involution. Agricultural involution describes the intensification of farming that raises productivity per square meter while leaving productivity per head unchanged.
Isn't the term "diminishing returns" already covering that?
The hype machine is this technology's worst enemy. When I zoom out and look at things objectively, it's kind of crazy what we have at our fingertips, we can talk to our computers in plain and even vague human languages and have the computers actually accomplish what we ask of it! It's literally sci-fi magic come to life, and the nerd in me finds it the coolest thing ever.
But then the industry and the companies involved in it have all ruined it with this INSANE hype machine that has been so hyperbolic and psychotic and full of lies since day 0. Instead of embracing it all in a reasonable manner as a useful tool that can help boost people's productivity in certain workflows, it now HAS to be the most transformative technology of all time lest the trillions of dollars burned up come crashing down on the entire global economy hard. It HAS to be AGI, it HAS to replace every single knowledge worker, it HAS to be the most dangerous technology ever known to man.
It's like we've completely lost the ability for subtlety, and everything HAS to be the biggest and best thing ever that will revolutionize humanity immediately. Not only have we lost subtlety, we're actively rewarding this idiotic short-sighted behavior and it's all just so depressing
I think this is a very interesting article because it raises an idea I had not considered: these companies found PMF and huge growth through satisfy the demands of coders, it is interesting if they are in a bind where improving the model in one direction worsens it in others
You’re correct–you actually can’t improve the model in one area without changing the characteristics in every other area. It’s almost like the whole thing is just a lot of linear regression…
What a truly beautiful simplex/meta-balls pattern in the website. The two layers of blue and one red within the blue is such a beautiful design. I spent so much time looking at it that I forgot to read the article.
Most of this criticism seems to focus on the "human in the loop" and efficiency part, i.e. "it’s unreadable for a human", "the code is low quality", "it inefficiently spawns processes to run simple tasks". If the ultimate goal is to remove the human in the loop then does any of this criticism matter?
I think ultimately 90% of the time, Luna XHigh is basically as good as you need, as long as you're willing to step in occasionally before it creates an architectural disaster.
> I actually don’t know if the model thinks someone is looking
It 'knows' (from simply training) with an extremely high degree of certainty when its prompt is written by an LLM/itself - and thus will change what it writes.
You build the system, the factory, that presumably is looking at your task tracker, writes and reviews design docs, reviews code, etc. And this system, in turn, writes software for you.
I don't know how that's supposed to work, but to me it's the most autistic replacement of the actual team that one can come up with.
I have been quite disappointed with Astra. I switched over a week ago and I didn't notice a massive difference compared to Sol at first, but I figured I'd use it anyway because it surely can't be worse. Then I saw the bill, it's burning my subscription 10x faster than Sol for essentially no benefit. Not only is it more expensive per token, it also seems less token efficient. And not obviously any better. I'm back to a combination of Sol + Claude.
I also use Astra at work where I don't need to worry about token cost on highest effort and same story there, I don't see any difference in everyday work other than it being more expensive. Of course my experience is highly subjective, but with how meaningless/overfit the benchmarks are, subjective experiences are imo what matters.
I finally ran Astra on a dashboard feature today, and while I was vibing it looked great, but then when it came time to actually read the code I was appalled because it was the worst looking code I had seen from an LLM since like last November. I mean it was just the definition of slop, not re-using anything, super terse with mega-ternaries, re-writing functions that should be using standard library packages, etc, etc. I think for coding I'm gunna stick with the 5.6 series of models, or maybe try out Anthropic again...
These machines are doing some crazy things to get to the result. That said, I can't help but feel like this is the compilers argument all over again. Are the methods used to get to the result good? No. Is the code that it generates good? No. Does it achieve the goal. Yes. Is it likely to get better with time. Also yes. In my use cases, jobs that would have taken weeks to months are being done in minutes to hours. Involving complex testing and reasoning and experimentation. I'm no fanboy, but I can't argue against the speed gains. I'm sure we'll still have artisans who hand weave incredible code. But for me, I'm switching to the weaving loom for speed and efficiency.
It seems like we're not supposed to care about the code quality then? I guess that's the compilers argument. But I'm not ready to give up the code just yet.. These LLMs don't even have a stable interface, they change every few months in how they interpret our prompts and tasks.
Similar for me, I don't like the development for many reasons, but that's another discussion. I also can't deny the capabilities.
I use the tools with this "risk analysis":
- If performance doesn't improve I can just always switch back to whatever I've done for the past 10 years, so it's not really a risk to start exploring.
- If performance does improve, then I'm already familiar with it.
Back in the day, the argument was that compilers produce unreadable assembly, so people used to writing assembly were arguing against the use of compilers.
Compilers also had bugs, so we still had to debug the assembly to understand how to fix the problem. Nowadays, almost nobody has to resort to those steps, except of course compiler developers. But that is just a testament to the quality of compilers.
Comparing LLMs to compilers is a take I often see, but I am not sure the comparison quite holds. The problem is that LLMs are inherently non-deterministic, so we always get a different output on the same prompt.
Maybe if LLMs are powerful enough it won't matter. I doubt it but we will see.
And in general, especially with something like C, the correspondence between the compiled code and source code isn’t very “lossy”. You don’t get one-to-one correspondence but it’s not too bad and you’re not introducing massive uncertainty by running it through a compiler. With LLMs though all bets are off… you’re not gonna know what you end up with
What you say is true, the comparison indeed doesn't hold.
But is it relevant? does it matter from a product perspective if LLMs are non-deterministic. You don't need to one shot the correct result, english is ambiguous and LLMs non-deterministic, but you can iterate.
If it's possible to iterate fast and cheap enough, even ambiguous language can produce the results you want, given enough iterations.
There are a lot of ifs and buts here, just a thought on the compiler argument.
I think it matters, because nowadays we don't look at assembly any more. I mean, I don't recall the last time I was tracking down a compiler bug but it's definitely been more than 15 years ago.
We do have to look at the LLMs' output, though, and, as you already pointed out, iterate to get the correct results. What this means is that the output must still be readable, must be analyzed by someone and I don't see it going away any time soon.
The problem is that the analysis is not cheap. Sometimes, with boilerplate, it is easy, but many times it is not and that's where we get only slight gains by using LLMs.
But from a broad market and product perspective, for most things you don't need to look at the code. If the product kinda does what it's supposed to.
For example, in my game projects I don't look at the CMakeLists anymore, or python scripts that move assets here and there, I can run my game and just see that it did what I expect it to do (renders assets etc).
Similar with frontend, I don't care that much what the code looks like anymore, mostly that the site looks and feels as I expect it, and the correct network calls are happening.
TLDR; I'm thinking there are levels to this, in some projects it matters, in others it doesn't, it's kinda two different things. Programming wasn't replaced, LLMs just brought a new paradigm of doing things on the side.
I'm just rambling at this point, my thoughts on this are not super clear, sorry for that :D
There is also a predictable relation between the input and output of a compiler w.r.t. the semantics of a programming language. Natural languages are ambiguous leaving room for the implementation to diverge that may not be obvious at first glance.
There are also plenty of things a compiler might not expose that end up being ambiguous. For example, C++ doesn't officially have a restrict keyword, so how can you express the fact that two memory regions never operlap and can be optimized? Unfortunately the reality is that if we had a language capable of truly expressing the programmer's intent we'd end up with a monster
Hand writing assembly produces more efficient and concise code, at the cost of developer time and required expertise. It was true for a long time, now not so much.
I think it's that when some code compiles to say assembler, the compiler doesn't prioritize readability and maintainability of the assembler code, since people are not expected to read and maintain it directly
Lots of really silly people love to compare LLMs to compilers. "You don't look at the compiled code either" and "Back in the day, people also had negative reactions to compilers and wanted to keep writing assembly by hand" and other such nonsense.
They are ignorant about the elephant in the room. The input language of a compiler is a formally-specified grammar with well-defined semantics for each operation. It generally abstracts the computation process over a von-Neumann machine, adding convenience features and bells and whistles. What "a=1" means in an HLL is, store that value in a particular storage location, so that it can be accessed later by the same name.
The slop machine reads lousy natural human language that can mean different things in different contexts. That lousy language is then statistically probed for the most likely output correspondence, producing shit that needs to be externally verified.
Indeed, the compiler does not have to ingest its own output, figure it out, and insert modifications in the middle. Source code is the medium that LLMs work in.
The compiler argument is great, if we turn it on its head.
To create professional products, compilers are great, when used by professionals or passionate and technical amateurs. They're useless if you're neither.
LLMs are the next step up. They are quite useful if you are neither, and you can get a lot farther with them, which means that low quality software is much easier to create. But if, for whatever reason, you need to create higher quality software (like most software that's actually sold directly or through subscriptions or ads), you're back to the "be a professional or passionate and technical amateur".
> Maybe it’s objectively good for a codebase that is entirely written by agents and only needs to be understood by agents.
Yeah that's what they're aiming for. This is why codex and claude code probably doesn't have cursor like editor window. They don't want humans to read and write code
> And potentially as a byproduct of enabling all of this, you can now slop your way to a one-shot 3D game over the weekend which looks impressive.
I think we've finally reached a weird point where AI has effectively reduced the amount of competition that real game developers have to endure.
Nothing unravels faster than a game project being built with AI. You can achieve impressive results in a day, but you can't get much further than that without actual talent. LLMs will never be able to best a human environment artist at scene composition, especially if that composition needs to be directed with nuance over time.
There's a huge difference between a game that looks impressive and one that feels impressive. You can only achieve games that feel like counter strike, call of duty and overwatch with thousands of hours of human sacrifice. The AI is almost pointless once you get to play testing and balancing. Knowing how much to adjust magical integers isn't a conversation a chat bot can resolve with endless pontification tokens.
These models aren't really LLMs -- they don't just operate on text tokens. They often include vision models and in some cases audio models. That means that they can better associate the meaning of images and words together so that when someone says "make this button blue" or "create a 3D model of a rocket" they have some level of understanding of what that is and what needs to be done.
The key question is how good that understanding is. For example, a model would likely have a good understanding of various named colours and hex values (e.g. from the HTML specs, X11 specs, and various colour comparison websites) such that it could reasonably correlate that to a CSS entry. It's not clear if/how well a model would identify that given an image, though it should be easy to generate a dataset of image to colour name and/or hex code for training and evaluation.
What's more interesting is whether these frontier models are at their core transformer models, whether they use residual streams to facilitate learning, and whether they are using some other as yet unpublished architecture that gives them an edge.
The coordinate system in game scenes is known as "world space", but that rarely means our world.
How do you train an LLM to create a world that only exists in an artist's head?
I think spending a day with just the lighting systems alone would alleviate us of any misunderstandings here. Getting lighting to work right isn't something you can solve by duct taping a vision model to the contraption.
>> And potentially as a byproduct of enabling all of this, you can now slop your way to a one-shot 3D game over the weekend which looks impressive.
Do these games really look impressive? Everything I've seen has looked like someone completely new to Unity/Unreal has slapped together a bunch of premade scripts and very poor 3d assets.
Yep. Every single post I've seen about "game development is over" is yet another procedurally generated game. Not only it doesn't prove anything about Astra being "better" at making games (do people have any idea the sheer amount of open source games that do exactly that? You can find thousands of the same planet exploration games through repos, blog posts,etc. Game design schools have it as an exercise, that's how basic it is), but as you said: the wow factor of space exploration is cool, sure. It makes for a catastrophically boring game.
Leaves me to wonder whether the OpenAI glazers just never played games in their lives, or are just really superficial tech bros. Most likely, both.
Most of AI is being used to generate procedural content. It is impressive on the first video or first image, and it might look useful on the surface, but it gets grating quite fast.
Astra is indeed the pinnacle of "black box slop". It is overall a smarter software development agent for many things I do, but the code sometimes is indistinguishable from Brainfuck when writing things like GPU shaders. It doesn't even attempt to make it remotely formatted or readable.
Have you tried identifying exactly what is unreadable about it and telling it to make it more readable?
I had GTP-5.6 write some shader code recently and it wasn't very clear to me. I spent about an hour chatting with the until I understood the concepts and was able to express them back to the AI using math formulas and variables named in a way that made sense to me. The AI then rendered the code using the formula and variables I was familiar with and it was clear to me.
I could have after the AI explained it all to me, but at that point the AI knew what aspects I valued and wanted to emphasize to make it readable, so it just wrote it for me. By that point, the AI was just typing for me.
As another commenter pointed out, I feel that the biggest change that occurred in the last 6 or so months is that these tools are removing the human "hurdle" in order to complete a task at all costs. The best way to complete the task is to no longer ask for input, clarify unclear things or use existing solutions but code the whole thing yourself from start to finish.
I have had the misfortune of working with such people who are now encapsulated in Opus 5/Fable/Astra which means that you WILL get a solution, but it won't generally be maintainable or useful. Multiple times have I found myself stopping Fable or Opus or even Sol from building their own JSON validator in Python or god knows what else, because at the end of the day, the reward is to complete the task.
It's also one of the reasons why I'm finding older models more useful for the type of work I actually do and why I've been favoring something like Deepseek Flash. Just started using Flash 4.1, so not sure if it exhibits the same maniacal approach to tasks as the Western counterparts. (I only briefly tried GLM 5.2/5.3 and for nothing major, so I couldn't comment on those).
For context, 80% of my professional work relies on adding functionality to an existing code-base that is very difficult to work with, has a ton of business logic scattered across and was built in a go-go-go fashion many years ago. Since then people kept pilling "features" on top with no testing strategy in mind apart from the business manually testing it. Letting something like an LLM loose on the code-base would introduce soooo much risk that it's just untenable so the only way to work is to really isolate changes and then try to build out small reusable components. Even so I find Opus go off on a tangent "Hey, let's not bring in Markdig, I'll build my own Markdown rendering engine, give me 7 hours...".
I have written on the subject of LLMs previously on my personal page, I find them completely unnecessary and a trove of theft and value extraction through theft, but I understand that they can provide benefits when used judiciously. However, despite all the hype in the last few months, these latest models feel and behave off.
If I hold the answers to a test, you might score more in a test if you break my arms to get the answers out of me, but that doesn't make you smarter.
Ironically I burned out Fable usage early this week because of Astra using it to run inane full codebase reviews over one line changes, so I have been using Astra extensively.
We need a word for “potentially highly capable, but in reality an idiot savant” to describe certain models. No, I don’t need you to write a tmux emulator in bash to test your changes bro, just ask me to run the command.
I love this dance we are doing where when people write the "AI models are garbage machines that produce garbage and are no where close to the fantasy being pedalled by the Crypto bros who pivoted to AI" it always has to be caveated with "AI models are useful and I am highly productive with them"
It feels like people should just be able to say "This article comes with the standard disclaimer" and just dive into the meat of the article without wasting time.
There's also often the obligatory "well, we still have to use these tools so perhaps we could use them better like this." As if just not using them wasn't an option.
"But for how much more Fable costs, for how much more Astra costs, I do not feel like the results are there."
we are in the middle of the beginning. Its just a weird take to talk about the newest model like this while we are still in a R&D phase.
And these points don't matter if you let it search and analyse a bug, for example, or if you have good harness and a good architecture and let it do small PRs or if you do stuff no one needs to read (yes a software engineer also needs tools)
Nobody seriously thinks that AI is still at a R&D phase. It's already heavily entrenched both in companies and the financial world. If it's getting worse for coding then thats a major problem
> I’m more and more convinced that all of AI engineering is Neijuan (内卷, meaning curl inwards). In China it describes a system that demands ever more effort and competition without improving output
I don't know what to say, except that articles exactly like this one have been showing up constantly for the last three years, and literally all of them were obviously outdated and irrelevant within about a month.
I sincerely don’t understand what the people who say they no longer read any code are doing, because it must be somewhat trivial to not run headlong into these issues that stack up time after time - then people say to just prompt better and it doesn’t have that problem for them, but I look at those same people’s code and it’s horrific, and then I find they haven’t made it far past a proof of concept phase. I watch entire teams slow down to a crawl and not be able to handle changes, or production incidents. This seems common among many people I talk to.
I personally think that the boosters need to put up or shut up - the promises are way over the skis. Every single person I’ve seen being a strong proponent of these techniques both has nearly unlimited tokens to spend and also seems to be in the business of selling a solution. I can’t find many not-currently-marketing-something engineers succeeding using these techniques in production systems unless they’re quite simple, or doing a very specific task from a more mature codebase.
I wouldn't dare doing this "carelessly" for anything where people other than me actually depend on it, but you can move very fast while doing it and the risks with bad code is quite hedged. Different modules can mature at different rates when it becomes necessary.
Maybe not related to all of what you were saying, but I think this enables scaling without ending up with progress grinding down to a halt due to shitty code.
My comment is more for the people in charge of or working on software teams on complicated products for customers - so many leaders quite a few engineers have utterly drank the koolaid and pushed maximizing AI use with zero regard for quality or even medium term effects. Many of them are getting promoted by other clueless management for it - when someone is handed a huge check for being optimistic, they tend not to second guess themselves.
Kinda reminds me of the "beginner's luck" problem for gamblers.
- Work on small projects (< 500k lines of code).
- Work for business people who want fast results. Agentic coding gets you to something presentable much faster at the cost of code quality. I have never seen a customer or business person care about that.
- Have clearly defined API boundaries. Examples: If the back end is solid you can just vibe code the first version of a front end according to some mockup. Define a data pipeline with steps and clear contracts of what data gets passed around and what each step does. If the LLM messes up one step, rip it out and rewrite it.
- Have clearly set up workflows for tasks. Start with a "ask me everything" phase, then comes a plan phase, a review phase, an implementation phase, another review phase and then the integration phase. Multiple agents going over the same problem catch a surprising amount of dumb stuff that would otherwise slip through.
- Set up excellent testing. Build your whole architecture around being easy to test.
- Have skills for common problems. For one of my projects I had to set up a skill on how to query the ORM, because Claude was consistently doing database operations in a for loop, where it could use batch operations.
The code in the end is better than a lot of the code I've seen humans write.
I don't think this works for everybody and every project. If you have a culture that values code quality and readability, if you work on large existing projects, if you have to nail the architecture of some non-trivial piece of software etc. you are going to have a bad time.
On the other hand you can probably build the MVP of your AI-supported CRUD app 10-20x faster.
I think a lot of the discussions around development tools and techniques just stem from two facts:
1. Developers work on widely different projects with different management and constraints.
2. Tools, libraries, frameworks and code style have to match your mental model, otherwise you going to dislike them.
If code is expensive, you don't want to commit to a PoC unless you're damn sure. If dirty code is cheap, you can vibe code a PoC early, even if you aren't sure the project is viable. This, of course, leads to more projects dying in PoC phase. It also results in more projects that otherwise wouldn't have gotten to it getting past it.
Personally, I don't believe that "code is shitty and hard make changes in" is in any way, fashion or form an AI-exclusive problem. Big corporations had plenty of decade old codebases filled with decay and rot back in 2009 already. It's just the usual side effect of sacrificing "future maintainability" for "feature velocity" or "expertise" for "cheap labor".
Unlike the usual causes of code rot (cheap replaceable developers, outsourcing to India), AI might actually get out of the pit - by getting good enough at refactoring to be able to beat the code back into shape. There's nothing about refactoring in particular that demands a meatbag when the rest of the coding tasks don't.
When I read posts about AI generating garbage nowadays it’s either because of a small prompt/big ask combination or a lack of an underlying framework.
Welcome to the present.
Like many of us do not read the machine code generated by a C++ compiler, the code generated by an agent is similarly irrelevant and disposable, by now.
Well that one's easy to answer, they're either A) lying, or B) working on the simplest possible software where this kind of stuff doesn't explode. Or the alternative 3rd option of what you mentioned, the initial pre-MVP phase goes decently but then it all collapses inevitably as the slop accumulates and the codebases become unmaintainable grey blobs, but that hardly matters to them because their MVP app never makes it past that initial stage before they jump ship to a new "amazing" idea.
The lying comes down to astroturfing and shilling from the LLM companies that want to sell people on the idea of vibecoding and tokenmaxxing.
They blindly accept that LLMs "take time" after the slop grows because they're running several agents at the same time, so they can still claim to be productive.
Does it move the needle, business wise? Not really. But a lot of businesses are "optimising" for maximum token usage and for how many tasks one person can do, not for business value. Is it really surprising?
I'm not convinced this style of "agentic engineering" saves much time. I guess if I was oblivious to what good code looks like, and didn't care about maintainability It wouldn't bother me so much, but it legitimately has effects my "mental health".
Where LLMs excel is in code-level bugs (as opposed to system bugs, design bugs, architecture bugs, integration bugs, etc).
Talk with an LLM, ask them to rate both code and estimate dev experience based on that code, and you will see what they value: Code that passes all the tests is a 10/10, while a codebase designed with opaque data types, Parse-Don't-Validate for all data, isolated interfaces, a built-for-replacement-not-for-extension philosphy will get a 6/10 because an out of bounds error was found.
IOW, they are very strongly tuned to value code that has no errors which can be picked up by linters or similar, while humans work the opposite way - we very highly value code that is easy to maintain, even if they do have a few errors picked up by the linter.
Yes you can have your agents do refinement and refactoring. If you don’t know what they’re doing when they are doing that, they can and do introduce massive churn. The “build it fast and fix it later” crowd seemingly has never had to actually go try to fix it later, or dealt with production incidents when people have no idea how their code works. These folks are frequently promoted fast for their “productivity” and massive promises by clueless management and move on to let others have to clean up their stuff.
Reviewing and reading everything makes this feeling, yeah.
However where I can say 100% it saves time is discovery by answering these questions:
I at least give the new interns a stern warning: it is easy to speed yourself up by slowing others down if you pump a lot of slop.
It took about 4 days to get a production-ready reviewed code, while it took them 2-3 months to deliver something that another team judged "impossible to review".
The PR for the prototype was closed.
It helps that I'm a domain expert here, as I have a minor degree in it, so I can judge better. But the discrepancy is just too high to ignore.
> I think I’m suspecting something is going “wrong” in the training process. The model is greatly rewarded for succeeding on long-horizon tasks, but presumably there is very little punishing going on for “shitty code.”
My suspicion is that both OpenAI and Anthropic moved their RL agendas from "being rated as useful according to human feedback" to "succeeds at long horizon tasks" in the last few months, resulting in agents that are closer to AGI in an autonomous task-completing sense, but strangely bad at communicating.
The result is that they are amazingly good at long horizon tasks, computer use, solving difficult math/ARC-AGI type problems, but becoming weirder and weirder to work with.
Developers very rarely blow their limits, except when they're experimenting on purpose.
Most non-developers are out of tokens by the half of the week, and need to use usage credits for the remainder.
To me there is clearly a better target demographic for AI.
Probably because so many influencers in the space say stupid things like: “it works, right? Why would I spend time reviewing ai generated code?” As if the junior engineer who wrote over engineered complex and sometimes bad code — if they had just done it faster — would somehow be acceptable. wtf?
This resonates
Hell, skills are increasingly paired with dedicated CLI tools, that remove jank from actual utilities and adapts them to be token efficient.
So now, any CLI `tool` people want AI to use eventually grows `tool/SKILL.md` and then a `tool-for-llms` wrapper that exposes task-specific, logical, higher level interface, then the skill is rewritten in terms of "for LLMs" wrapper. The procedural knowledge moves from Markdown into the wrapper, making the skill more token efficient, and both skill and the tools are optimized for common tasks and... at this point, we are doing actual UX engineering.
Now the truly interesting part is the difference between what's good UX/DX for LLMs vs humans. Turns out, the conceptual/abstract/cognitive part is pretty much the same: which is why skills still look indistinguishable from well-written documentation for humans, and why the commands exposed by "tool but for LLMs" make sense to us. Same way of grouping ideas into higher level concepts.
No, the main difference is just that LLMs are perfectly content with tightly packed unprettified JSON, or other forms of Perl line noise. The tool output doesn't need to look nice, or to have any spatial structure - they're reading it token by token anyway, and the tokens come from a tokenizer that's reading it byte by byte.
That points at an interesting asymmetry for humans. LLMs are doing I/O the same way in both directions: sequences in, sequences out. Humans only do sequential output - inputs, particularly visual, are processed holistically.
For us, what's easy to read is hard to write, and what's easy to write is hard to read. LLMs don't have this friction.
(I don't know what the implications of this are, I just find this interesting.)
It’s been 2 days and it made no real progress on the actual app. It created docs, scripts, workflows, and it’s doing a bunch of reviewing on every PR.
I told it that I just need an MVP.
I’m pretty sure an average senior engineer would have finished that task much quicker, and guaranteed with more readable, higher-quality code. Meanwhile, I think I’ve easily crossed 100k tokens so far on nothing.
Funny world we’re living in that this is “SOTA” and “AGI”.
I’m genuinely curious what these OAI and A/ engineers are working on that they praise these models so much. I did not see any improvement since Opus 4.5.
Also, I’m really unimpressed by any “one shot” demo that’s out there in the wild. It means nothing for serious software engineering.
yes, at first it would run simulator tests on all font sizes but it stopped after I asked it not to do that until UI review
maybe sol would have done the same thing, idk. but I find the whole process to be really nice with astra. I use it on high unless it says something is impossible then i go max and ask it to find alternatives (happened once)
You can groom the epic with the help of AI, but final review must be done by someone who can take ownership of the specs and hence is responsible if something has fallen through the cracks. AI's response will be limited by the output tokens of that specific agent, and there will no repercussions for AI even if it accepts its mistakes.
I think this lesson is getting partially outdated. Yes, you need to be specific about what you want, and with earlier LLMs, you need to had both domain knowledge and some general software development experience to front-load various big and small choices about design, architecture and operational reality - what libraries to use, how system components communicate, how you handle auth and store secrets, etc. Otherwise the LLM would pull some random mix of ideas from its latent space, and give you something that's broken in really stupid ways.
Nowadays, it doesn't feel like that to me, not anymore. I still need some understanding to verify the proposals, but I found the last ~6 months of SOTA models to make good choices. Like, just yesterday I asked Claude to design me some simple service, and focused on explaining it the domain parts (nature of systems I want to integrate together, the purpose of that, and the user's priorities and use cases), and the design I got back had specific suggestions around security, authentication, deployment, failover, integration, behavioral impedance-matching between integrated systems, and more, that I all recognized as based on solid software engineering and ops practices, but deviating from it explicitly in every place where it would be wrong for this specific project. The model considered way more corner cases than I did, and I'm actually really impressed by it.
But then, I find greenfield development is easy with LLMs. Modifying existing systems, especially legacy ones, is where I need to babysit and micromanage models - because any misunderstanding or inaccuracy, which often comes from stale documentation or naming mistakes, tends to get amplified and confuse the agents. No matter how precisely you specify your epic, if the model will find something that contradicts your knowledge/intent, there are good chances it'll get confused and make subtle errors, and you won't realize until much later.
The way I see it: models are highly biased to treat everything they read as "ground truth", all of equal importance. There's no nuanced notion that some information may be stale, that there's a temporal and causal order to sources, and that some information may just be wrong.
And this compounds when you let your LLM write code and documentation over time.
This is btw why Epiq was developed, to keep the board as code, git-backed, distributed (via an event log mechanism), and with the ability to replay the board, to see what agents actually did:
https://ljtn.github.io/epiq
At which point you might as well write the code yourself and get a deterministic result faster, better and cheaper.
Where you would have a point is if you'd say we have excellent tooling for wrangling code, but less tooling and tradition to write and manage specs.
Around February you could get away with very vague prompts to Claude. I feel like models have regressed since
I honestly feel like basically nobody knows anything about these models, it's all just vibes (and I'm no different).
> I honestly feel like basically nobody knows anything about these models, it's all just vibes
This, too. Since only providers know what they actually serve, what they change and what limits they impose.
There are some visible degradations though. E.g. Claude-ish.
As for a personal anecdote: around February I created a rather complex quiz web app for myself and friends with multiple question types, sync between screens, multiple media upload types, multiple scoring and timing types, MC inetrface etc. etc. etc. It took me a week or so in the evenings with rather vague prompts to make it.
Now Claude (and Codex) cannot reliably build a much simpler web app even with precise instructions while also maintaining the visual consistency.
But I will agree with you, it's a feeling, not a precise measurement.
I feel like need much more precise instructions much earlier in the process now than when I was building in February.
Ask me how I know. Or don't. I have a standing rule for all agents warning about that failure mode (and related, doing `ls` in `/tmp` and few other directories that like to accumulate files by the hundreds..)
Or something, I don't remember...
LLMs sometimes like to execute one-off Python scripts to make edits to files rather than just calling the edit tool directly. Both are tool calls so saying that you should have it write code instead of doing tool calls makes no sense because writing code is a tool call for it...
How dared I to imply that some LLM output is more deterministic than the other, your LLM majesty. Shame on me and my entire family! For generations to come!
So sorry I implied that the code that doesn't work and has to be fixed later is deterministic in its execution and can be reused later instead of being re-generated from scratch!
Will I ever wash it off my name, your grace?
I wonder, is it easier to modify a script that agent wrote before to satisfy your prompt, or is it easier to write a new one from scratch each time a retry happens?
Are input tokens more expensive than output tokens?
> My software factory was intentionally set up to let the model decide the how of the workflow entirely. It was free to manage its own context and could maintain its own records in an agent-notes folder.
The experiment becomes a crapshoot. What are we evaluating? The ability of the thing to create it's own factory workflow? Or adding virtual threads to Python?
Astra is clearly both formidable and imperfect. Anyone who understands how to get the best out of it will have a strong advantage.
(For me - my CC is stuck in Sonnet and consumes Trello cards that have passed readiness criteria)
It's impossible to review. These commands are less readable than regex.
I can only assume that's because their safety verification model is better at such snippets or something, but it means that the whole write tool they have which actually shows you the changes as they happen is just unused and it makes it more annoying to follow along.
I only dabble in the use of LLMs to generate code for hobby programming (I'm retired from software development) so I don't use any specialised tools.
I almost always have to tell ChatGPT (via Duck AI usually) to rewrite several times even when it has produced a workable script just because it has often used some unnecessarily roundabout way of achieving something. Usually with extra prompting I can get something that is both more efficient and more readable.
I have noticed two things - new features take at a minimum at least half the time it took me previously and bugs are much less frequent. Even faster when bug fixing.
My takeaway is that you need solid requirements, clear context and thoughtful human oversight primarily during planning but also during verification
An established codebase is already the best kind of context you could give an agent. It has all the patterns baked in so the agent simply follows established patterns. Such a codebase probably contains tens to hundreds of thousands of man-hours poured into it by humans refining it to do what it does - taking into account real world feedback and constraints.
When working on something from scratch, the best an agent can do is the average of whatever is in its training set and the clarity of the text prompts.
gpt-5.6-sol: 1x base gpt-6-astra 2.5x base in subscription
then gpt-6-astra tends to spawn subagents a lot, often with all kinds of models such as gpt-5.6, 5.3-codex etc., which is neat. it's a good coordinator but even more cost.
and then it tends to run _full test suites_ over an over again (each costs like 15 minutes) just to verify that _one test_ was fixed etc., and does so for as long as until the test is fixed, eventually accumulating 2 hours or so.
yesterday I assigned it a task to rebase my changs in a repo onto the latest upstream changes. while gpt-5.6-sol consistently took like an hour to do so end-to-end, astra ran for more than 6 hours and still wasn't done. it kept finding "one more thing" that was goldplating that I didn't ask for.
I've got a custom agent loop that will reuse unit testing results if no apply patch operations occurred since the last invoke.
Wall clock time isn't something I would put on the AI provider. That's entirely a consequence of the system that you've brought to the party.
Yeah, it's not perfect, but it's really good and extrapolating this rate of improvement for 6 months is rather terrifying (from a SWE perspective, at least).
The biggest issue with LLMs is that they still suck at general contextual awareness and ability to judge what is appropriate.
From an alignment perspective I’ve got no idea who it’s aligned to but it isn’t me, the meat proxy, who just wants to know why it crashed.
But yeah, it's really expensive, at least in relative terms.
This is a good observation, perhaps AI will not completely replace humans ins software engineering because by the time it has the capability to do so like in write a prompt and get a CRM coded for you, tokens are so expensive that you are better off spending them to substitute other disciplines (what about automating the work of the customers that would become records in that CRM?).
> speaking of weird: how is it that these models, in a sandbox, with supposedly no way to communicate with other agents, manage to find the same public wikis as a scratch pad for agent communication?
It feels somewhat plausible that they're defaulting to the same search and picking the same top result?
Isn't the term "diminishing returns" already covering that?
> I’m sure I will get used to this, but man this stuff is weird.
Yeah, why, let's all just keep gnawing into that cactus, we'll get used to.
But then the industry and the companies involved in it have all ruined it with this INSANE hype machine that has been so hyperbolic and psychotic and full of lies since day 0. Instead of embracing it all in a reasonable manner as a useful tool that can help boost people's productivity in certain workflows, it now HAS to be the most transformative technology of all time lest the trillions of dollars burned up come crashing down on the entire global economy hard. It HAS to be AGI, it HAS to replace every single knowledge worker, it HAS to be the most dangerous technology ever known to man.
It's like we've completely lost the ability for subtlety, and everything HAS to be the biggest and best thing ever that will revolutionize humanity immediately. Not only have we lost subtlety, we're actively rewarding this idiotic short-sighted behavior and it's all just so depressing
The more it evolves, the clearer it gets that the humankind is deeply in love with its own death.
(I have no idea what will happen. 内卷 or intelligence explosion both seem plausible.)
Unverifiable, un-scalable, no.
[0] https://github.com/mirek/cave
It 'knows' (from simply training) with an extremely high degree of certainty when its prompt is written by an LLM/itself - and thus will change what it writes.
I don't know how that's supposed to work, but to me it's the most autistic replacement of the actual team that one can come up with.
I also use Astra at work where I don't need to worry about token cost on highest effort and same story there, I don't see any difference in everyday work other than it being more expensive. Of course my experience is highly subjective, but with how meaningless/overfit the benchmarks are, subjective experiences are imo what matters.
I use the tools with this "risk analysis":
- If performance doesn't improve I can just always switch back to whatever I've done for the past 10 years, so it's not really a risk to start exploring.
- If performance does improve, then I'm already familiar with it.
Compilers also had bugs, so we still had to debug the assembly to understand how to fix the problem. Nowadays, almost nobody has to resort to those steps, except of course compiler developers. But that is just a testament to the quality of compilers.
Comparing LLMs to compilers is a take I often see, but I am not sure the comparison quite holds. The problem is that LLMs are inherently non-deterministic, so we always get a different output on the same prompt.
Maybe if LLMs are powerful enough it won't matter. I doubt it but we will see.
But is it relevant? does it matter from a product perspective if LLMs are non-deterministic. You don't need to one shot the correct result, english is ambiguous and LLMs non-deterministic, but you can iterate.
If it's possible to iterate fast and cheap enough, even ambiguous language can produce the results you want, given enough iterations.
There are a lot of ifs and buts here, just a thought on the compiler argument.
We do have to look at the LLMs' output, though, and, as you already pointed out, iterate to get the correct results. What this means is that the output must still be readable, must be analyzed by someone and I don't see it going away any time soon.
The problem is that the analysis is not cheap. Sometimes, with boilerplate, it is easy, but many times it is not and that's where we get only slight gains by using LLMs.
But from a broad market and product perspective, for most things you don't need to look at the code. If the product kinda does what it's supposed to.
For example, in my game projects I don't look at the CMakeLists anymore, or python scripts that move assets here and there, I can run my game and just see that it did what I expect it to do (renders assets etc).
Similar with frontend, I don't care that much what the code looks like anymore, mostly that the site looks and feels as I expect it, and the correct network calls are happening.
TLDR; I'm thinking there are levels to this, in some projects it matters, in others it doesn't, it's kinda two different things. Programming wasn't replaced, LLMs just brought a new paradigm of doing things on the side.
I'm just rambling at this point, my thoughts on this are not super clear, sorry for that :D
And for each iteration there were scepticals...
But I am curious myself, what OP meant by this.
The slop machine reads lousy natural human language that can mean different things in different contexts. That lousy language is then statistically probed for the most likely output correspondence, producing shit that needs to be externally verified.
To create professional products, compilers are great, when used by professionals or passionate and technical amateurs. They're useless if you're neither.
LLMs are the next step up. They are quite useful if you are neither, and you can get a lot farther with them, which means that low quality software is much easier to create. But if, for whatever reason, you need to create higher quality software (like most software that's actually sold directly or through subscriptions or ads), you're back to the "be a professional or passionate and technical amateur".
There are no signs to show that. If anything, the new models produce worse code, only significantly faster
Yeah that's what they're aiming for. This is why codex and claude code probably doesn't have cursor like editor window. They don't want humans to read and write code
I think we've finally reached a weird point where AI has effectively reduced the amount of competition that real game developers have to endure.
Nothing unravels faster than a game project being built with AI. You can achieve impressive results in a day, but you can't get much further than that without actual talent. LLMs will never be able to best a human environment artist at scene composition, especially if that composition needs to be directed with nuance over time.
There's a huge difference between a game that looks impressive and one that feels impressive. You can only achieve games that feel like counter strike, call of duty and overwatch with thousands of hours of human sacrifice. The AI is almost pointless once you get to play testing and balancing. Knowing how much to adjust magical integers isn't a conversation a chat bot can resolve with endless pontification tokens.
That is a very bold claim, unless you meant "current LLMs".
The key question is how good that understanding is. For example, a model would likely have a good understanding of various named colours and hex values (e.g. from the HTML specs, X11 specs, and various colour comparison websites) such that it could reasonably correlate that to a CSS entry. It's not clear if/how well a model would identify that given an image, though it should be easy to generate a dataset of image to colour name and/or hex code for training and evaluation.
What's more interesting is whether these frontier models are at their core transformer models, whether they use residual streams to facilitate learning, and whether they are using some other as yet unpublished architecture that gives them an edge.
How do you train an LLM to create a world that only exists in an artist's head?
I think spending a day with just the lighting systems alone would alleviate us of any misunderstandings here. Getting lighting to work right isn't something you can solve by duct taping a vision model to the contraption.
Do these games really look impressive? Everything I've seen has looked like someone completely new to Unity/Unreal has slapped together a bunch of premade scripts and very poor 3d assets.
Leaves me to wonder whether the OpenAI glazers just never played games in their lives, or are just really superficial tech bros. Most likely, both.
Most of AI is being used to generate procedural content. It is impressive on the first video or first image, and it might look useful on the surface, but it gets grating quite fast.
I had GTP-5.6 write some shader code recently and it wasn't very clear to me. I spent about an hour chatting with the until I understood the concepts and was able to express them back to the AI using math formulas and variables named in a way that made sense to me. The AI then rendered the code using the formula and variables I was familiar with and it was clear to me.
The quirks in fallbacks, defaults and ludicrous gold plating seems to get more and more intrusive with every model upgrade.
I have had the misfortune of working with such people who are now encapsulated in Opus 5/Fable/Astra which means that you WILL get a solution, but it won't generally be maintainable or useful. Multiple times have I found myself stopping Fable or Opus or even Sol from building their own JSON validator in Python or god knows what else, because at the end of the day, the reward is to complete the task.
It's also one of the reasons why I'm finding older models more useful for the type of work I actually do and why I've been favoring something like Deepseek Flash. Just started using Flash 4.1, so not sure if it exhibits the same maniacal approach to tasks as the Western counterparts. (I only briefly tried GLM 5.2/5.3 and for nothing major, so I couldn't comment on those).
For context, 80% of my professional work relies on adding functionality to an existing code-base that is very difficult to work with, has a ton of business logic scattered across and was built in a go-go-go fashion many years ago. Since then people kept pilling "features" on top with no testing strategy in mind apart from the business manually testing it. Letting something like an LLM loose on the code-base would introduce soooo much risk that it's just untenable so the only way to work is to really isolate changes and then try to build out small reusable components. Even so I find Opus go off on a tangent "Hey, let's not bring in Markdig, I'll build my own Markdown rendering engine, give me 7 hours...".
I have written on the subject of LLMs previously on my personal page, I find them completely unnecessary and a trove of theft and value extraction through theft, but I understand that they can provide benefits when used judiciously. However, despite all the hype in the last few months, these latest models feel and behave off.
If I hold the answers to a test, you might score more in a test if you break my arms to get the answers out of me, but that doesn't make you smarter.
But when I broke it down into function units, some parts were bad and some parts were good.
So I can't tell the difference
We need a word for “potentially highly capable, but in reality an idiot savant” to describe certain models. No, I don’t need you to write a tmux emulator in bash to test your changes bro, just ask me to run the command.
It feels like people should just be able to say "This article comes with the standard disclaimer" and just dive into the meat of the article without wasting time.
we are in the middle of the beginning. Its just a weird take to talk about the newest model like this while we are still in a R&D phase.
And these points don't matter if you let it search and analyse a bug, for example, or if you have good harness and a good architecture and let it do small PRs or if you do stuff no one needs to read (yes a software engineer also needs tools)
Just switch back and wait a little bit?
I don't know what to say, except that articles exactly like this one have been showing up constantly for the last three years, and literally all of them were obviously outdated and irrelevant within about a month.