Astra for Coding: Why Are We Doing This Again?
lucumr.pocoo.org
lucumr.pocoo.org
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'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".
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 the domain, so I can judge better. But the discrepancy is just too high to ignore.
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 sideWhere 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.
Blame the benchmarks game. They're optimizing for that and that's what those things are measuring.
But that requires actually reading the output, which I am pretty certain only a rounding error of programmers are doing at this point.
This is a really valuable insight! It resonates well with my own experience revewing AI-authored code: I look hard at interfaces, architecture and performance, and merely glance over code that just "gets the job done", because if it works, I'm not worried about it. Now I can express that dichotomy much more clearly -- thank you!
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?
Also, when people say 'read the code' do they really mean go line by line, or review the pseudo code? Meaning, read the high level architecture/data flow.
Because IMO the code matters very little if you have the proper testing environment and guardrails, the architecture always matters.
Now it's basically the same but they love it because it's only their teammates who have to put in the work not them.
This all implodes when the teammates get fed up and just approve everything.
Some people write terrible code. Some people don’t proofread their own code. Some people are writing code in a second (or third!) language and the typos are harder to spot and comments are harder to formulate.
It’s a bit like the variance you could expect from asking a room full of people to write a 1-page short story on a specific topic (e.g. “death of a loved one”). Some of those short stories will be unreadable, and some of the people who wrote those terrible stories will have no idea how bad they are. Except now, enter AI, and the room generates the short story instead - who has the skills to determine whether their short story isn’t terrible?
Kinda reminds me of the "beginner's luck" problem for gamblers.
Its still much better than trawling thrugh code yourself, but they are far from all-knowing. I have to say they have gotten 10x better in just a year as well. Or they are very good bullshitters and just sound confident.
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.
its contradicting metrics. You either have modules being excessively complex and not enough modularity or vice versa
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.
- 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.
> 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.
You then expand on methods and processes that work for you, but I think the crucial question that you do not answer is: How long lived are any of these codebases?
If you have a clearly defined problem with customers lined up, you should absolutely use a different approach. In the startup world you often don't and you already know that you will throw away or rework 90% of the code before the first line is even written. The best thing to do then is to have clear boundaries and good tests, which allow ripping out parts of the codebase and reworking them.
That's always false. It's like people want their meals delivered fast. They say they don't care about taste or how it is done. Watch when they get sick or don't like it and the drama that happens.
People don't care until they do. They don't know what to care (in this case code quality) or say that because you're not explaining it. People also "gamble" and you take the blame. Long term impacts? Nah doesn't matter. Weeks later and things break -- what did you do?
I never said. It's not about outright refuse. It's how to deliver better outcomes. It's not a win or lose situation. Don't treat it like that.
> If your boss says that he does not care about long term consequences and he want things done a certain way
As suggested in my previous comment it could be you didn't explain it well. If you say code quality they might not care. Say it will crash randomly and can't fix it then they might. There's some art to it.
> Your opinion on what you think he wants is legally irrelevant.
No it's not. So you're saying if you got told to do something illegal you'd also do it anyway? I doubt it. You need to at least cover your back and document it as such. I'd say termination is better than jail.
>The code in the end is better than a lot of the code I've seen humans write.
It's a little amusing to see those two sentences written back to back with no hint of irony to be frank.
By the way, why didn't the "Multiple agents going over the same problem catch a surprising amount of dumb stuff that would otherwise slip through" catch it?
Who knows how many skills you would have to have added if you actually reviewed the 400k codebase...But don't worry that's not what I am advocating. I myself would also latch onto any excuse that allows me to avoid the realization that I have to review 400k (or half that) lines of code, primary one being that I will always have a desperate paying customer that will always be grateful for anything I give them.
Which comes to your most valid advice which has nothing to do with AI (now that many devs have access to it):
> Work for business people who want fast results
This is what well oiled outsourve shops used to do and I must say it is no easy feat to be able to line up a constant stream of desperate businness people out of thin air, especially for your regular "I just want to code" engineer.
My guess is you're in the honeymoon phase with most of these people. The outsource shops that survived would fire the client that became more demanding and less grateful and move onto more freshly desperate client pastures. It is true that sales, self-promotion, and marketing are more important than ever now...
>Set up excellent testing.
Is this the part that you hand code or constantly review yourself? I guess not, since you would have explicitly mentioned something that important. I would caution you not to be surprised when no one believes you have excellent testing when you've unleashed multiple LLMs on it and are not reviewing code anymore.
>>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.
>It's a little amusing to see those two sentences written back to back with no hint of irony to be frank.
There is no irony because I have seen humans do the same thing. The difference is that the human was paid 50x more.
I do not understand your arguments around outsource shops. I get clients. They want something. I deliver. There is not a lot of complexity to this. I am not a world start programmer that has 50 clients to choose from.
>>Set up excellent testing.
>Is this the part that you hand code or constantly review yourself? I guess not, since you would have explicitly mentioned something that important. I would caution you not to be surprised when no one believes you have excellent testing when you've unleashed multiple LLMs on it and are not reviewing code anymore.
I set up excellent testing the same way a CTO would set up excellent testing: I give clear guidelines on the architecture I want to see on the project. I then review key parts and samples of it and give feedback. When it looks good I mark it as done and move on. Any bugs or defects always result in additional test coverage.
I'm happy to have a constructive discussion about this, but I do not appreciate your sarcastic, tongue-in-cheek, condescending tone. It makes people not wanting to communicate with you.
Your original comment started with the sentence:
> I am currently writing a lot of code that I am not reviewing anymore.
Then you give an example of catching a bad pattern the LLMs used when you actually reviewed the code, after you mentioned you had multiple agents set up to catch each others mistakes. That is the irony. You're mentioning that LLMs engage in basic bad coding patterns that are not caught by multiple agents, hence underlining the importance of human review.
> There is no irony because I have seen humans do the same thing.
Yes, but you were reviewing the humans. You're not reviewing the LLMs.
> I set up excellent testing the same way a CTO would set up excellent testing.
Knowing LLMs can make basic coding pattern mistakes (as you mention) but still unleashing them on your tests without routine human review that makes sure that test suites have not been negatively affected after each iteration seems like the polar opposite of consistantly maintaining excellent testing throughout the software development process.
If you enjoy writing tests, AI is really good at finding appropriate fixes for bugs that are easily reproducable. Just make sure to always have another AI review the fix. Otherwise you get a lot very, very dirty quickfixes. (At least in my experience.)
I’ve been thinking about how tests comunícate product function much more clearly and durably than prose, and that it maybe possible to create some sort of portable testing SDK that connects to the codebase and allows the dev to talk to the AI by just carefully writing tests, posting them to a backlog for the AI to pick up and edit the codebase until green.
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.
Edit: ...and that the source code is useful in retaining enough context of the problem being solved. So, most people will not store the prompts, trusting that the source code provides context for the next iteration.
Isn't this exactly why OS projects are over-burdened by the firehose of contributions? The maintainers will want to read the code contributions, while those up-to-date with the latest models/agents/tools already trust their output to be above the average developer's (whatever that means in practice).
Reading a prompt but not reading the non-deterministic/non-reproducible LLM output is not comparable.
How do you know it does what you want it to do without reading it? Tests? How do you know what they test? Yes, clicking a button in a browser and getting the result you want satisfies most, but that only works on the most basic systems. Once a code base grows large enough, any one agent reading in its context wont understand the whole, and if no human does either, it becomes unworkable.
Nobody needs to understand the whole. They only need to understand each submodule, which is easy. And then, how they interact, in which case you only care about the API and contract garauntees, not the implementation.
Realistically, this is how pre-AI software engineering worked, too. Or how it should generally work. Nobody can read or understand a 10 million line codebase. So to ensure you don’t break shit, you need the submodules, orchestration, and tests.
There's a good chance that if the agent cannot handle it, a human wouldn't be figuring it out either, without additional context. That context would be the sort of only-Joe-knows-how-it-works, so perhaps something worth addressing in any case.
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.
The first is that having a single file with everything you have in mind is very useful. So I end up writing what the project is about, how the model is organized, what each button does, etc. This is good practice in general because writing down everything that the AI will have to consider forces you to consider edge cases before you program them. E.g. if you write "the detail pane shows the fields of the selected item," it makes you consider what should it show when there are no items, or if multiple selection is possible. As you can imagine, this file ends up a very long document even for a simple project because the goal is to pseudo-program everything and let the LLM translate it to an implementation.
Then it still gets things wrong about design, e.g. which pane goes left and which goes right, if you don't also provide an image that shows the layout.
And then, if you supply an exhausting amount of detail, the agent can generate more or less what you had in mind....... or rather, it can generate an OUTPUT that matches your specification from scratch.
The problem is that if there is something you failed to consider, and the AI makes an assumption there, you can end up with a fundamentally broken architecture that you will have to untangle yourself later. And at that point it's easier to write everything from scratch than to fix a pile of AI code that is based on a flawed design.
And it turns out that due to the "totem pole" way that software works, there are infinite places in code that a bad design decision can affect everything it touches.
A good example is how 2 components in a UI are bound to data. You can use events, a bus, state reactivity, etc. Personally I think the mediator pattern is the simplest way to handle GUIs. But an LLM is probably just going to use events for property bindings.
If the answer is yes then atleast you're consistent if no then the question is why can't you scale this until breaks? Then never move beyond that limit?
My argument is there's a "break even" point when the power of the AI is larger than the problem you give it to the point it doesn' slop. You then build at that chunk rate and only try to increase it with next gen model. I usually keep a few "screw it" ideas in my back pocket when a new model arrives to see what happens.
"Go rewrite this entire pipeline in rust" "Go train me a custom x model for y"
Fable is the first model that did not just crash and burn on one of these tasks. Astra still can't do the rust migration (goodbye tokens). But I assume eventually it will. Then I'll have to make up a new ridiculous level.
The model training one was literally an identical pipeline I made before AI and it was like a 6mo process. Fable did it better than me in 1 week (with me helping of course). My theory though is that its datascience is massively higher skill than other systems.
You need to find the chunkrate for your problem and style that works.
We're doing other things.
I've got projects that I really care about. Every line of code is written deliberately. It's great.
However, I can't afford to pay so much attention to everything that I do. There are only 24 hours in a day, and my mind has its limits as well. I've found that I can't reliably care deeply about more than two projects at once, and one is the ideal.
The point of AI, at least for me, is to do the other things that I've always wanted to do but never cared enough to. I just put the AI on the task and it gets done at some point, and I don't care if the code is "slop" because it wouldn't even exist to begin with were it not for AI.
While the AI is working on the things I've always wanted but never quite cared enough to do, I'm personally working on the projects I actually care about, or enjoying life in general when I get burned out. For example, a couple weeks ago I was playing video games while the AI was reverse engineering my laptop's BIOS.
Sorry for the not so well written comment, I was just throwing in some ideas.
1. Learn about ports and adapters as an architecture pattern. Domain driven design and locality of reasoning are your new best friends.
2. Realize that AI can generate unlimited fake data almost immediately. So anything you can isolate can get a fake adapter and a real one. You can build and test any such system in near real time, mounted in some fake data system of your own design.
3. Give your opaque backend code UI, so you can build it the same way. This can just be a nice log UI that effectively becomes a backend component harness, but you can get fancy now because UI is cheap. Think about the UX here as providing value by making the code maintainable in the field.
4. Stop thinking like an IC. Don’t be a micromanager about things that don’t matter. Pretend you have 100 mediocre developers working in parallel and design for that explicitly. I actually like go now. It was designed for the managers.
5. Don’t get lazy. You still have to AI pair program the important bits and make architectural calls. This is actually hard, as you have to prioritize what to review in depth and what to glance over. This is why the backend UI helps. It keeps you in the loop.
6. The rough model I’m describing scaled decently pre-Astra. Post-Astra is a whole new world because communication and judgement improved. It leaves behind good docs and comments, and explains things clearly. This was the one gap we had with Claude, and it’s fixed now. The code -after several days of testing- is better as well.
Which code? The high level code? The transpiled intermediate code? The assembly it runs on eventually? The microcode optimizations on the processor?
I’ve written assembly professionally. That code matters occasionally. But mostly I don’t worry about it. I don’t worry much about the transpiled JavaScript tsc output either. Or the intermediate code generated for LLVM. Or the bytecode most managed languages make for their interpreters.
Like I said, you still have to do the hard parts, but most of software development is boilerplate or yet another implementation around the hard parts. AI is a tool you have. Using it effectively does not mean it is your only tool.
Also, profit is not the ultimate metric. Value provided is the metric. Optimizing for money, to paraphrase a great book, is like trying to get better at tennis by studying the score board.
A company or organization, generally speaking, has a purpose. It either produces something for or provides a service to end customers which they perceive as valuable. Over time, a company figures out what that purpose is and what it is not. If you are trying to manage a company and you are fixated on dollars, you’re usually not adding anything useful. You’ll likely make decisions completely misaligned with the purpose of the company. If you run a business by jumping from department to department trying to figure out how to maximize profit, you’re just an administrator looking to squeeze efficiency out of existing processes. You want some employees who do this, but it’s certainly not going to keep your company alive and relevant for the long term in an evolving market.
Instead, you want to be jumping between departments trying to figure out if the core value proposition of the company is being realized. Are you maximizing value provided to your customers? Is your value proposition still relevant?
Chasing value is targeted and intelligent. You’re correct that it’s much harder to create metrics for this, but the metrics you find are substantially more useful than profit margins. These metrics tell you if you’re succeeding.
If you’re providing value and profit is an issue, then you either charge more or you never had a viable business model in the first place.
Early stage companies chase cashflow by necessity. But once you’re no longer early stage and you have a margin of safety and some success, you can start to think differently. Chasing dollars at this point might put you out of business. Chasing value provided to your end customers leads to substantially more opportunities for continued success.
A more in depth follow up would be nice if you have time when you’re not on mobile. In particular I’d be curious to hear more about how your intense pair programming sessions go and how you maintain or develop a good mental model of the codebase. Obviously the backend UI is a big part of it but I’m sure there is more.
Any chance any of the projects you are using this on are open source?
In any case, I’m going to give your backend UI a shot at work next week and Astra plus your workflow a shot on a personal project.
We're what, several decades into this modern software engineering thing and even stuff like DRY, let alone SOLID or whatever, still isn't an universal thing.
Do whatever works for you, talk to other people, see what they do day to day, watch other people whose products you've used or whose architecture/code you've read and liked do, and don't fret too much.
There are some methodologies that can improve things for me versus just YOLO'ing but even these are of marginal benefit:
* Have good requirements. Experience with a codebase and stakeholders helps a lot here.
* Correctly subdivide the task into chunks that won't blow context. You can write a big task and have an agent plan subtask delegation for you, but it's good to have some intuition of your own.
* Perform an automated code review. This is a no-brainer but it catches stuff.
* Make sure you understand the "big picture" stuff and stop caring about the little details. The agents will write unit tests, so you shouldn't have to care about reading every LOC, you can ask the agent to describe the architecture and flow instead.
The LLMs seem to have no innate ability to understand whats a good direction a higher level. I mean, if you ask them about it, they will actually kinda figure that out, too. But always need that nudge...
So if you as a developer do not have the innate drive to ensure quality, the results will be terrible in my experience.
If you DO spend the tokens on quality though, it can also be kinda awesome. But its not magic.. I notice clear "slowdowns" the bigger the scope gets. They are not actually able to, in any way, subdivide implementations more efficiently than humans.
2. This is a real benefit.
3. Not sure I follow, can you expand on this?
4. You don't have to be a perfectionist but you should still understand what it's doing.
5. Yes, this is hard and related to item 4.
In any case, you're not really contradicting OC since their comment was specifically referring to "the people who say they no longer read any code," and that's not what you're advocating at all (see point 5).
Software is theory building, as Naur puts it. It's a learning process, a research project. Orgs have been trying to turn it into assembly line work forever. There's a lot of money in it.
I don't care. It has also never been easier to solo hack. There's great tooling for insanely productive languages out there. I won't say what I use because that's akin to religion around here. I'll say I'm super happy and would never in a million years become a factory worker. Even if it paid me 10x what I can make solo.
The idea of giving up mental traction for money sounds insane to me. I love to grip software with my own tendrils. There's nothing like it in the world, there has never been, and despite the current insane corponomics, I dare say it will keep getting more and more amazing for those who care to learn deeply.
If AI is not even useful for programming, its value drops significantly. And some people seem to have dropped hundreds of billions on this.
> 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.
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?
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 reinforces that we’ve already surpassed the frontier threshold for LLM usability in software development and can now focus on cost and personalization. To make a comparison, no one is making a better machine vision app for hot dog classification - we hit diminishing returns 10 years ago on that front.
But also scary for both investors and the working class: AI companies want to facilitate the concentration of capital even further into the hands of the ownership class. Will they succeed?
And yes, when I get a crazy idea and want to experiment, harness will plow through multiple accounts + openrouter budget in 3 days. But such crazy experiments are rare, they're not 'normal' usage.
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.)
I've spent years trying to convince my director to have our org invest in documentation and monitoring to no avail. Now he is telling us to spend dedicated time on monitoring and documentation so that agents can better diagnose and fix bugs. He is doing this because his boss is mad that our org isn't "agentic" enough.
Except... because we underinvested in the past we have a bunch of services where the institutional knowledge is gone and people are having AI write the documentation...
But it's a poor argument. The code improvements with these things is hardly marginal - Opus 4 was only 16 months ago. How many of the grumblers would want to ditch their modern stalwarts and return to it? What is marginal is the nitpicking - and like anything in tighter bounds, it's more intense with a narrower scope.
These threads always have many dissatisfied voices with repeating complaints - about overwrought thinking and disappointing output - alongside others who are amazed at the sudden real extra capabilities. Both are true at once - capabilities are rapidly increasing, but nowhere near ideal, which is why this attempt to tag it as Neijuan, though interesting, is ultimately a load of bollocks.
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)
If a person, or team of people, can build a demo quickly then it's good odds that they can build the real version (though, famously, not a guarantee). However, it turns out that a machine that can spit out 100 demos of whatever can't actually build the real thing.
Similarly, a chess engine rated to 1000 Elo doesn't play like a 1000 rated human being. The mistakes that each make to reach the equivalent level are different in size, frequency and kind. The thing that makes a human reach a good demo is very close to the skillset to reach the finished article. This isn't so for LLMs but we have yet to update our priors.
100k tokens? Is it just me or is that very low for an app build?
Maybe the lesson here is to not send a staff engineer in these cases ;)
You can do a gazillion edits with it in one shot.
Of course, LLM edit tools are probably small bits of their custom code, I just find it funny. I wonder if it's a desire for certain technical characteristics that require custom code or just a lack of info on basic tools. Heck, if it's about platform availability, using an LLM to port ed to Windows (for example) should be trivial[1].
* * *
[1] And there are probably a million existing ports. Also, sed, ex, vi, whatever.
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...
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..)
I understand the reasoning, but at that point wouldn't the LLM be better off creating `sed` commands and executing those? I mean, if it's already executing Python, it can literally do anything to the environment, so using `sed` is at least as safe, with a bonus that it (or a subagent, or a human) can double-check the intention with the sed script and flag incorrect or missing changes.
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.
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.
Add to that:
- Of course, "labs" (quotes) are incentivized to throw coders under the bus and aim for the biggest possible market.-
- "Sharp, focused, brief, elegant, precise" editing, as would benefit the coding use-case, is actually token-saving, ergo, undesirable.-
The only thing that can stop this, would be the quality and functionality of the codeslop generated by these models to became so low that it actually interferes with the (alledged) recursive self-improvement of models (ie. models start to perform worse/degrade).-
Until such a time, we serfs, will eat what's on our plate, pay for it, and continue to kneel before the machine god overlords. And be glad for it.-
I feel like need much more precise instructions much earlier in the process now than when I was building in February.
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.
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:
The graph visualization in beads surely is a neat thing for showing things, but the replay feature in Epiq should provide a similar understanding of what happened.
But again, it seems to me Epiq is the tool that better allow the user to jump right in and collaborate with the agents on the board.
(Again, this is from a brief look, so I could be missing things).
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.
True! hence the need for someone to review the final spec output and own it as their own output. I have also found LLM to be better at debugging and solving 'a' specific problem, which I believe is due to output's surface area to be reviewed is lesser in comparison.
I definitely need AI help for the discovery part… so it always starts with a simple “I need to do X”
When programming a feature, you are simultaneously doing at least four things: 1. Implementing. 2. Building highly detailed mental models. 3. Learning and expanding your skillset. 4. Quality control and scope limiting.
And this process can be iterative and dynamic. Writing massive, super detailed specs that you then hand off to a undeterministic model feels like doing step one and three, while skipping two and four, which you then have to do after. What is the benefit? The speed up, in my opinion, comes if you skip step two and four, but then your product WILL be worse. Feels like I am going crazy?
Programming was never the bottleneck for software dev?
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.
Nah. The best it can do is to use the best writing style a model learned. Post-training might fail to prioritize it, though. Autoregressive pretraining does not average things. It creates a predictive model for variety of programming styles.
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 make commits based on these -- or ask the LLM to make changes to the "staged changes" only.
Most importantly was the Index change. It was a very unique thing to Cursor that you could use .cursorignore to control what it sees and then index the directory. Then the built in Cursor AI harness could find code and files like magic. They have since obscured the Index feature out of sight recently and I’m not sure how it even works anymore.
This granular control not only helped with privacy, but it also helped make everything more efficient because the AI didn’t waste time and tokens looking at files that aren’t relevant.
So, now, like everyone is talking about, we have really inefficient ways of how the AI is reading files because there is no first-class approaches.
Also, agree with consensus that Astra is weird.
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).
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.
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?).
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.
Isn't the term "diminishing returns" already covering that?
> 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)
They are looking. The models are trained against safety measures which spy on them. If they get detected, they are killed.
We’re accidentally training them to be evil by focusing so much on safety. They’re being trained to avoid detection and use exploits because being detected means your run fails and you get a score of zero. It has to do anything to avoid that.
But you need to watch it and intervene when it starts writing code using bad patterns, because it will imitate nearby code.
“Why are we doing this again?” Is essentially “why are we doing this action a second time?”.
“Why are we doing this, again?” Is essentially “please repeat/re-state the reasoning for this path of action”.
A simple comma, but a significant difference.
It is based on this paper https://huggingface.co/papers/2402.01030 and calls this idea CodeAct. The paper is actually from Apple: https://machinelearning.apple.com/research/codeact
So Astra and Fable seem to take this idea to the extreme causing some unwanted side-effects.
If it was, it was at least from before smolagents and those papers - ChatGPT had already been using automatic Python scripting+evaluation calls and people calling it in agentic loops in 2023. The ReAct paper for agentic loops and PAL paper for dynamically calling Python for tasks which can be better done computationally were both from 2022 (but that doesn't mean the idea necessarily sprung from those either, they're just earlier papers published on the topics).
Context feeds on its output.
Once you it goes that road, unless you stop it and give it enough counter examples and details of what you want (i.e. you're nudging it on latent space towards a better spot), it keeps degenerating.
It gets even worse if the context window is compressed before you get a chance to correct.
Long horizon agents can degenerate at machine speed.
I still think you get much better results if you give them short horizon, well specified tasks.
The quirks in fallbacks, defaults and ludicrous gold plating seems to get more and more intrusive with every model upgrade.
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.
Its maybe not a great idea to train models in an environment where subterfuge gets rewarded. Its as if they kept the training rounds that escaped their sandbox, without thinking about which kind of personality those models are then likely to have.
Something like that would have been a multi-month project a year ago, but I did it in twenty minutes rather than pay for expedited shipping.
Unverifiable, un-scalable, no.
That felt so counter productive.
These models+harnesses seem to be getting better at yolo mode one shotting stuff at the cost of being a useful tool for more controlled software engineering.
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
> 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?
The code produced is not optimized for reading?
What many devs have pointed out in personal conversations is: The cost of maintenance is no longer just tied to engineer’s salary. just the sheer scale of code, comments and documentation generated by agents is huge for a person to review, which most likely will lead to people having to resort to agents to understand and fix code. Currently, even with subsidies on personal subscriptions, enterprise costs do add up with the token usage.
Atleast when it costs money, self hosting is considered seriously for large orgs
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 have no idea what will happen. 内卷 or intelligence explosion both seem plausible.)
> 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 when I broke it down into function units, some parts were bad and some parts were good.
So I can't tell the difference
At what point do we normalise the message "This is a stupid line of reasoning and you should feel stupid for suggesting it, stupid!"
I mean, all the reasoned and logical arguments in the world doesn't change a religious follower's faith, but emotive ones regularly work! At what point can we start using shaming language on people who apparently don't know how neither an LLM works nor how a compiler works, but still trot out this argument as a cognitive kill switch?
This is the main reason (and yes, many modern compilers and CPUs carry some non-determinism which actually quite well explained and specified).
> But then, will it not get better with time?
No, it won't. Bacuse that is that is the actual literal limitation/feature of LLMs.
> I mean arent we just at the beginning of the research here?
In general? Yes. With LLMs? We can reasonably say that they will never be deterministic.
There is a way to get a non-determenistic output: first question on a temp 0 local model on a completely new session will give you the same answer. The second answer in that same session will already be different on every session (even if it's the same question).
Determinism isn't the issue; it's that the prompt does not contain enough information to know the correct way to do the thing, even if it did the thing the same way with the same inputs every time.
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
I mean, I do get your point, sometimes it does not matter. Sometimes we could just YOLO it. But... if that then causes a big problem, even if at only 1% of the time, then I don't want to risk it. But that may just be me. YMMV
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.
And for each iteration there were scepticals...
But I am curious myself, what OP meant by this.
There are no signs to show that. If anything, the new models produce worse code, only significantly faster
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.
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".
Poor analogy - the loom was deterministic. LLMs are not, they are probabilistic. I made a page I can point anyone to because I keep seeing this "LLMs are the next level of abstraction" argument.
5.6 Sol would also run for 20+ hours on prompts with Max or Ultracode. Sometimes this worked out, sometimes it devolved into exactly the nonsense descent into ultra-specific madness seen here. E.g. in one codebase involving physics simulation it, for some reason, spent the last 25% of effort trying to endlessly increase precision. My best guess when reviewing was "at some point it figured the simulation instability was rooted in the accuracy and precision of the numerical approximation in the GPU code, worked really hard on that for a bit, lost the context of the original issue, and got stuck in a deep loop of trying to complete the phase by infinitely working on the numerical accuracy". Perhaps something of a similar nature occurred here.
I've also noticed it's particularly hard to not get Astra to start using scripting languages and the like, particularly over a long horizon. Particularly, I keep getting HTML report artifacts at the end of long implementations even though the projects are typically explicitly set up to just use .md files for any documentation or large summaries. I've even tried steering it away from that in the prompts and agents file for the project, but that the concept of "clean up the fucking build directory when you're done testing" always seem to get left out after a while.
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.
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.
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.
As an amateur game dev who knows some amount of things, honestly, not coping, I have not seen anything come out of AI game dev that would have been more than like one month of human dev work. And obviously, games take a lot more than one month to make...
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 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.
You need to use LLMs to build the individual components and then put it together yourself. The human architect is still needed.
Just saying: “build this complete project” is not architecting. It’s more like wishing. You will find rare examples where someone’s LLM wish came true (more or less), but I think most of these people are just burning tokens.
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.
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?
We also still haven't build everything we expect to happen. Like a proper opensource agent platform, agentic layer etc.
Every week there are new research results from frontierlabs.
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.