Posted by dnw 8 hours ago
Pretty much. I've been advocating this for a while. For automation you need intent, and for comparison you need measurement. Blast radius/risk profile is also important to understand how much you need to cover upfront.
The Author mentions evaluations, which in this context are often called AI evals [1] and one thing I'd love to see is those evals become a common language of actually provable user stories instead of there being a disconnect between different types of roles, e.g. a scientist, a business guy and a software developer.
The more we can speak a common language and easily write and maintain these no matter which background we have, the easier it'll be to collaborate and empower people and to move fast without losing control.
- [1] https://ai-evals.io/ (or the practical repo: https://github.com/Alexhans/eval-ception )
If they implement something with a not-so-great approach, they'll keep adding workarounds or redundant code every time they run into limitations later.
If you tell them the code is slow, they'll try to add optimized fast paths (more code), specialized routines (more code), custom data structures (even more code). And then add fractally more code to patch up all the problems that code has created.
If you complain it's buggy, you can have 10 bespoke tests for every bug. Plus a new mocking framework created every time the last one turns out to be unfit for purpose.
If you ask to unify the duplication, it'll say "No problem, here's a brand new metamock abstract adapter framework that has a superset of all feature sets, plus two new metamock drivers for the older and the newer code! Let me know if you want me to write tests for the new adapters."
But I can see the carnage with offshoring+LLM, or "most employees", including so call software engineer + LLM.
> LLM code is higher quality than any codes I have seen in my 20 years in F500.
And in my French brain, code or codebase is countable and not uncountable.
There is a countable "code" (just like "un café" is either a place, or a cup of coffee, or a type of coffee), and "un code" would be the one used as a password or secret, as in "j'ai utilisé tous les codes de récupération et perdu mon accès Gmail" (I used all the recovery codes and lost Gmail access).
But what set me off is an universal qualifier: there was no code seen by you that is of equal quality or better that what LLMs generate.
https://www.neatorama.com/2007/01/22/a-mathematical-cow-joke...
If that's obvious to you than you're just being rude. If it's not obvious to you, then you'll also find this is a common deviance (plural 'code') from those who come from a particular primary language's region.
Edit; This got me thinking - what is the grammar/rule around what gets pluralized and what doesn't? How does one know that "code" can refer to a single line of code, a whole file of code, a project, or even the entirety of all code your eyes have ever seen without having to have an s tacked on to the end of it?
As for the grammar rule, it's the question of whether a word is countable or uncountable. In common industry usage, "code" is an uncountable noun, just like "flour" in cooking (you say 2 lines of code, 1 pound of flour).
It's actually pretty common for the same word to have both countable and uncountable versions, with different, though related, meanings. Typically the uncountable version is used with a measure of quantity, while the countable version denotes different kinds (flours - different types of flour; peoples - different groups of people).
This was very helpful, thank you! (I had just gotten off the phone with Claude learning about countable and uncountable nouns but those additional details you provided should prove quite valuable)
As if author of the comment had not seen any code that is better or of equal quality of code generated by LLMs.
Well, the grammar is that English has two different classes of noun, and any given noun belongs to one class or the other. Standard terminology calls them "mass nouns" and "count nouns".
The distinction is so deeply embedded in the language that it requires agreement from surrounding words; you might compare many [which can only apply to count nouns] vs much [only to mass nouns], or observe that there are separate generic nouns for each class [thing is the generic count noun; stuff is the generic mass noun].
For "how does one know", the general concept is that count nouns refer to things that occur discretely, and mass nouns refer to things that are indivisible or continuous, most prototypically materials like water, mud, paper, or steel.
Where the class of a noun is not fixed by common use (for example, if you're making it up, or if it's very rare), a speaker will assign it to one class or the other based on how they internally conceive of whatever they're referring to.
You need code to get it to generate proper code.
I certainly read it as one and found it funny.
Nevermind the fact that it only migrated 3 out of 5 duplicated sections, and hasn’t deleted any now-dead code.
Can you imagine working with someone who produces 100k lines of unmaintainable code in a single sprint?
This is your future.
I admire your experience with people.
I had a coworker that more or less exactly did that. You left a comment in a ticket about something extra to be done, he answered "yes sure" and after a few days proceeded to close the ticket without doing the thing you asked. Depending on the quantity of work you had at the moment, you might not notice that until after a few months, when the missing thing would bite you back in bitter revenge.
Tool works as expected? It's superintelligence. Programming is dead.
Tool makes dumb mistake? So do humans.
You need to do this when coding manually as well, but the speed at which AI tools can output bad code means it's so much more important.
I think programming is giving people a false impression on how intelligent the models are, programmers are meant to be smart right so being able to code means the AI must be super smart. But programmers also put a huge amount of their output online for free, unlike most disciplines, and it's all text based. When it comes to problem solving I still see them regularly confused by simple stuff, having to reset context to try and straighten it out. It's not a general purpose human replacement just yet.
Set the boundaries and guidelines before it starts working. Don't leave it space to do things you don't understand.
ie: enforce conventions, set specific and measurable/verifiable goals, define skeletons of the resulting solutions if you want/can.
To give an example. I do a lot of image similarity stuff and I wanted to test the Redis VectorSet stuff when it was still in beta and the PHP extension for redis (the fastest one, which is written in C and is a proper language extension not a runtime lib) didn't support the new commands. I cloned the repo, fired up claude code and pointed it to a local copy of the Redis VectorSet documentation I put in the directory root telling it I wanted it to update the extension to provide support for the new commands I would want/need to handle VectorSets. This was, idk, maybe a year ago. So not even Opus. It nailed it. But I chickened out about pushing that into a production environment, so I then told it to just write me a PHP run time client that mirrors the functionality of Predis (pure-php implementation of redis client) but does so via shell commands executed by php (lmao, I know).
Define the boundaries, give it guard rails, use design patterns and examples (where possible) that can be used as reference.
Perform regular sessions dedicated to cleaning up tech debt (including docs).
I've definitely built the same thing a few times, getting incrementally better designs each time.
It's a tool. It's a wildly effective and capable tool. I don't know how or why I have such a wildly different experience than so many that describe their experiences in a similar manner... but... nearly every time I come to the same conclusion that the input determines the output.
> If they implement something with a not-so-great approach, they'll keep adding workarounds or redundant code every time they run into limitations later.
Yes, when the prompt/instructions are overly broad and there's no set of guardrails or guidelines that indicate how things should be done... this will happen. If you're not using planning mode, skill issue. You have to get all this stuff wrapped up and sorted before the implementation begins. If the implementation ends up being done in a "not-so-great" approach - that's on you.
> If you tell them the code is slow
Whew. Ok. You don't tell it the code is slow. Do you tell your coworker "Hey, your code is slow" and expect great results? You ask it to benchmark the code and then you ask it how it might be optimized. Then you discuss those options with it (this is where you do the part from the previous paragraph, where you direct the approach so it doesn't do "no-so-great approach") until you get to a point where you like the approach and the model has shown it understands what's going on.
Then you accept the plan and let the model start work. At this point you should have essentially directed the approach and ensured that it's not doing anything stupid. It will then just execute, it'll stay within the parameters/bounds of the plan you established (unless you take it off the rails with a bunch of open ended feedback like telling it that it's buggy instead of being specific about bugs and how you expect them to be resolved).
> you can have 10 bespoke tests for every bug. Plus a new mocking framework created every time the last one turns out to be unfit for purpose.
This is an area I will agree that the models are wildly inept. Someone needs to study what it is about tests and testing environments and mocking things that just makes these things go off the rails. The solution to this is the same as the solution to the issue of it keeping digging or chasing it's tail in circles... Early in the prompt/conversation/message that sets the approach/intent/task you state your expectations for the final result. Define the output early, then describe/provide context/etc. The earlier in the prompt/conversation the "requirements" are set the more sticky they'll be.
And this is exactly the same for the tests. Either write your own tests and have the models build the feature from the test or have the model build the tests first as part of the planned output and then fill in the functionality from the pre-defined test. Be very specific about how your testing system/environment is setup and any time you run into an issue testing related have the model make a note about that and the solution in a TESTING.md document. In your AGENTS.md or CLAUDE.md or whatever indicate that if the model is working with tests it should refer to the TESTING.md document for notes about the testing setup.
Personally, I focus on the functionality, get things integrated and working to the point I'm ready to push it to a staging or production (yolo) environment and _then_ have the model analyze that working system/solution/feature/whatever and write tests. Generally my notes on the testing environment to the model are something along the lines of a paragraph describing the basic testing flow/process/framework in use and how I'd like things to work.
The more you stick to convention the better off you'll be. And use planning mode.
Yes? Why don't you?
They are capable people that just didn't notice something, id I notice some telemetry and tell them "hey this is slow" they are expected to understand the reason(s).
"Hey, I saw that metric A was reporting 40% slower, are you aware already or have any ideas as to what might be causing that?"
Those two approaches are going to produce rather distinctly different results whether you're speaking to a human or typing to a GPU.
The suggestion to tell the agent to do performance analysis of the part of the code you think is problematic, and offer suggestions for improvements seems like the proper way to talk to a machine, whereas "hey your code is slow" feels like the proper way to talk to a human.
It can be a tool, for specific niche problems: summarization, extraction, source-to-source translation -- if post-trained properly.
But that isn't what y'all are doing, you're engaging in "replace all the meatsacks AGI ftw" nonsense.
It's a tool. It's good for some things, not for others. Use the right tool for the job and know the job well enough to know which tools apply to which tasks.
More than anything it's a learning tool. It's also wildly effective at writing code, too. But, man... the things that it makes available to the curious mind are rather unreal.
I used it to help me turn a cat exercise wheel (think huge hamster wheel) into a generator that produces enough power to charge a battery that powers an ESP32 powered "CYD" touchscreen LCD that also utilizes a hall effect sensor to monitor, log and display the RPMs and "speed" (given we know the wheel circumference) in real time as well as historically.
I didn't know anything about all this stuff before I started. I didn't AGI myself here. I used a learning tool.
But keep up with your schtick if that's what you want to do.
So what? That's honestly amateur hour. And the LLM derived all of it from things that have been done and posted about a thousand times before.
You could have achieved the same thing with a few google searches 15 years ago (obviously not with ESP32, but other microcontrollers).
I wouldn't be surprised if over half my prompts start with "Why ...?", usually followed by "Nope, ... instead”
Maybe the occasional "Fuck that you idiot, throw the whole thing out"
Are you using plan mode? I used to experience the do a poor approach and dig issue, but with planning that seems to have gone away?
I find LLMs at present work best as autocomplete -
The chunks of code are small and can be carefully reviewed at the point of writing
Claude normally gets it right (though sometimes horribly wrong) - this is easier to catch in autocomplete
That way they mostly work as designed and the burden on humans is completely manageable, plus you end up with a good understanding of the code generated. They make mistakes I'd say 30% of the time or so when autocompleting, which is significant (mistakes not necessarily being bugs but ugly code, slow code, duplicate code or incorrect code.
Having the AI produce the majority of the code (in chats or with agents) takes lots of time to plan and babysit, and is harder to review, maintain and diagnose; it doesn't seem like much of a performance boost, unless you're producing code that is already in the training data and just want to ignore the licensing of the original code.
They just write code that is (semantically) similar to code (clusters) seen in its training data, and which haven't been fenced off by RLHF / RLVR.
This isn't that hard to remember, and is a correct enough simplification of what generative LLMs actually do, without resorting to simplistic or incorrect metaphors.
We stubbornly use the same language to refer to all software development, regardless of the task being solved. This lets us all be a part of the same community, but is also a source of misunderstanding.
Some of us are prone to not thinking about things in terms of what they are, and taking the shortcut of looking at industry leaders to tell us what we should think.
These guys consistently, in lockstep, talk about intelligent agents solving development tasks. Predominately using the same abstract language that gives us an illusion of unity. This is bound to make those of us solving the common problems believe that the industry is done.
If you are using an LLM via a harness like claude.ai, chatgpt.com, Claude Code, Windsurf, Cursor, Excel Claude plug-in, etc... then you are not using an LLM, you are using something more, correct?
An example I keep hearing is "LLMs have no memory/understanding of time so ___" - but, agents have various levels of memory.
I keep trying to explain this in meetings, and in rando comments. If I am not way off-base here, then what should be the term, or terms, be? LLM-based agents?
You always need a harness of some kind to interact with an LLM. Normal web APIs (especially for hosted commercial systems) wrapped around LLMs are non-minimal harnesses, that have built in tools, interpretation of tool calls, application of what is exposed in local toolchains as “prompt templates” to transform the context structure in the API call into a prompt (in some cases even supporting managing some of the conversation state that is used to construct the prompt on the backend.)
> If you are using an LLM via a harness like claude.ai, chatgpt.com, Claude Code, Windsurf, Cursor, Excel Claude plug-in, etc... then you are not using an LLM, you are using something more, correct?
You are essentially always using something more than an LLM (unless “you” are the person writing the whole software stack, and the only thing you are consuming is the model weights, or arguably a truly minimal harness that just takes setting and a prompt that is not transformed in any way before tokenization, and returns the result after no transformations or filtering other than mapping back from tokens to text.)
But, yes, if you are using an elaborate frontend of the type you enumerate (whether web or CLI or something else), you are probably using substantially more stuff on top of the LLM than if you are using the providers web API.
However, they just look at the whole thing as "the LLM," which carries specific baggage. If we could all spread the knowledge of what is actually going on to the wider public, it would make my meetings easier, and prevent many very smart folks who are not practitioners from saying dumb-sounding stuff.
If we could all spread the knowledge of what is actually going on to the wider public, it would make my meetings easier, and prevent very smart folks from outside the field from saying dumb-sounding stuff.
- LLM = the model itself (stateless, no tools, just text in/text out) - LLM + system prompt + conversation history = chatbot (what most people interact with via ChatGPT, Claude, etc.) - LLM + tools + memory + orchestration = agent (can take actions, persist state, use APIs)
When someone says "LLMs have no memory" they're correct about the raw model, but Claude Code or Cursor are agents - they have context, tool access, and can maintain state across interactions.
The industry seems to be settling on "agentic system" or just "agent" for that last category, and "chatbot" or "assistant" for the middle one. The confusion comes from product names (ChatGPT, Claude) blurring these boundaries - people say "LLM" when they mean the whole stack.
My own experience using Claude Code and similar tools tells me that "hidden requirements" could include:
* Make sure DESIGN.md is up to date
* Write/update tests after changing source, and make sure they pass
* Add integration test, not only unit tests that mock everything
* Don't refactor code that is unrelated to the current task
...
These are not even project/language specific instructions. They are usually considered common sense/good practice in software engineering, yet I sometimes had to almost beg coding agents to follow them. (You want to know how many times I have to emphasize don't use "any" in a TypeScript codebase?)
People should just admit it's a limitation of these coding tools, and we can still have a meaningful discussion.
Anything they happen to get "correct" is the result of probability applied to their large training database.
Being wrong will always be not only possible but also likely any time you ask for something that is not well represented in it's training data. The user has no way to know if this is the case so they are basically flying blind and hoping for the best.
Relying on an LLM for anything "serious" is a liability issue waiting to happen.
For example, let's try a simple experiment. I'll generate a random UUID:
> uuidgen 44cac250-2a76-41d2-bbed-f0513f2cbece
Now it is extremely unlikely that such a UUID is in the training set.
Now I'll use OpenCode with "Qwen3 Coder 480B A35B Instruct" with this prompt: "Generate a single Python file that prints out the following UUID: "44cac250-2a76-41d2-bbed-f0513f2cbece". Just generate one file."
It generates a Python file containing 'print("44cac250-2a76-41d2-bbed-f0513f2cbece")'. Now this is a very simple task (with a 480B model), but it solves a problem that is not in the training data, because it is a generalisation over similar but different problems in the training data.
Almost every programming task is, at some level of abstraction, and with different levels of complexity, an instance of solving a more general type of problem, where there will be multiple examples of different solutions to that same general type of problem in the training set. So you can get a very long way with Transformer model generalisations.
If you've made a significant investment in human capital, you're even more likely to protect it now and prevent posting valuable stuff on the web.
This means an LLM can autogenerated millions of code problem prompts, attempt millions of solutions (both working and non-working), and from the working solutions, penalize answers that have poor performance. The resulting synthetic dataset can then be used as a finetuning dataset.
There are now reinforcement finetuning techniques that have not been incorporated into the existing slate of LLMs that will enable finetuning them for both plausibility AND performance with a lot of gray area (like readability, conciseness, etc) in between.
What we are observing now is just the tip of a very large iceberg.
If Im the govt, Id be foaming at the mouth - those projects that used to require enormous funding now will supposedly require much less.
Hmmm, what to do? Oh I know. Lets invest in Digital ID-like projects. Fun.
I don't think you grasp my statement. LLMs will exceed humans greatly for any domain that is easy to computationally verify such as math and code. For areas not amenable to deterministic computations such as human biology, or experimental particle physics, progress will be slower
I don't use a planner though, I have my own workflow setup to do this (since it requires context isolated agents to fix tests and fix code during differential testing). If the planner somehow added broad test coverage and a performance feedback loop (or even just very aggressive well known optimizations), it might work.