> The more information you have in the file that's not universally applicable to the tasks you have it working on, the more likely it is that Claude will ignore your instructions in the file
Claude.md files can get pretty long, and many times Claude Code just stops following a lot of the directions specified in the file
A friend of mine tells Claude to always address him as “Mr Tinkleberry”, he says he can tell Claude is not paying attention to the instructions on Claude.md, when Claude stops calling him “Mr Tinkleberry” consistently
What I’m surprised about is that OP didn’t mention having multiple CLAUDE.md files in each directory, specifically describing the current context / files in there. Eg if you have some database layer and want to document some critical things about that, put it in “src/persistence/CLAUDE.md” instead of the main one.
Claude pulls in those files automatically whenever it tries to read a file in that directory.
I find that to be a very effective technique to leverage CLAUDE.md files and be able to put a lot of content in them, but still keep them focused and avoid context bloat.
READMEs are written for people, CLAUDE.mds are written for coding assistants. I don’t write “CRITICAL (PRIORITY 0):” in READMEs.
The benefit of CLAUDE.md files is that they’re pulled in automatically, eg if Claude wants to read “tests/foo_test.py” it will automatically pull in “tests/CLAUDE.md” (if it exists).
If AI is supposed to deliver on this magical no-lift ease of use task flexibility that everyone likes to talk about I think it should be able to work with a README instead of clogging up ALL of my directories with yet another fucking config file.
Also this isn’t portable to other potential AI tools. Do I need 3+ md files in every directory?
Don’t worry, as of about 6 weeks ago when they changed the system prompt Claude will make sure every folder has way more than 3 .md files seen as it often writes 2 or more per task so if you don’t clean them up…
I wonder if it's because I have instructions for Claude to add comment blocks with explanations of behavior, etc that it can self reference in future. I guess that is filling the role that these .md files would.
What kind of citation or evidence would you expect? How would you measure productivity gains? From my personal life and activities however it is very clear. I have simply been able to do so many things I would have never been able to do without AI.
I feel similarly, but the other guy's point is still good; there is no empirical evidence that AI is really as helpful as we think, and there is empirical evidence that among those who perceive productivity gains, most are wrong and the rest are overestimating the effect.
Our experiences don't agree with the established facts, but the plural of "anecdote" is not "data."
I don't think it is something you can scientifically measure that easily and the studies that have tried are flawed. You can measure factory work, but you can't measure AI and software eng productivity gains or loss objectively.
I'm not entirely sure - certainly anecdotes abound (as do anecdotes to the contrary).
I am however sure that when someone claims that a technology is world-changing, and that it has already led to significant productivity gains... the onus is on them to provide evidence for such claims
Should I not share my experiences then about something that can't be scientifically proven? I am not actually looking to prove anything, but I am sharing my experience. I don't even win anything from people believing me, but I guess I have this weird urge to "set something straight" if in my experience something is the case. I have done tons of DIY stuff that I wouldn't have been able to without AI. Tons of side projects. In my corpo work I have been able to do more with less hours and mental effort spent, although in corpo work AI has given me less gains since most of the time I have been blocked by other factors and everyone else.
It is too fast evolving to be able to be legitimately scientifically understood. It is non sense to expect that.
> Should I not share my experiences then about something that can't be scientifically proven?
By all means share anecdotes, but maybe phrase them as anecdotes. "improving my own productivity massively" is an anecdote, whereas "delivering massive productivity gains" reads as a sweeping generalisation
It’s not delivering on magical stuff. Getting real productivity improvements out of this requires engineering and planning and it needs to be approached as such.
One of the big mistakes I think is that all these tools are over-promising on the “magic” part of it.
It’s not. You need to really learn how to use all these tools effectively. This is not done in days or weeks even, it takes months in the same way becoming proficient in eMacs or vim or a programming language is.
Once you’ve done that, though, it can absolutely enhance productivity. Not 10x, but definitely in the area of 2x. Especially for projects / domains you’re uncomfortable with.
And of course the most important thing is that you need to enjoy all this stuff as well, which I happen to do. I can totally understand the resistance as it’s a shitload of stuff you need to learn, and it may not even be relevant anymore next year.
While I believe you're probably right that getting any productivity gains from these tools requires an investment, I think calling the process "engineering" is really stretching the meaning of the word. It's really closer to ritual magic than any solid engineering practices at this point. People have guesses and practices that may or may not actually work for them (since measuring productivity increases is difficult if not impossible), and they teach others their magic formulas for controlling the demon.
Most countries don't have a notion of a formally licenses physicist either. That doesn't make it right to call astrology physics. And all of the practices around using LLM agents for coding are a lot closer to astrology than they are to astronomy.
I was replying to someone who claimed that getting real productivity gains from this tool requires engineering and needs to be approached as such. It also compared learning to use LLM agents to learning to code in emacs or vim, or learning a programming language - things which are nothing alike to learning to control an inherently stochastic tool that can't even be understood using any of our regular scientific methods.
I think it's relevant when people keep using terms like "prompt engineering" to try and beef up this charade of md files that don't even seem to work consistently.
This is a far far cry from even writing yaml for Github/Gitlab CICD pipelines. Folks keep trying to say "engineering" when every AI thread like this seems to push me more towards "snake oil" as an appropriate term.
My issue is not with learning. This "tool" has an incredibly shallow learning curve. My issue is that I'm having to make way for these "tools" that everyone says vastly increases productivity but seems to just churn out tech-debt as quickly as it can write it.
It a large leap to "requires engineering and planning" when no one even in this thread can seem to agree on the behavior of any of these "tools". Some comments tell anecdotes of not getting the agents to listen until the context of the whole world is laid out in these md files. Others say the only way is to keep the context tight and focused, going so far as to have written _yet more tools_ to remove and re-add code comments so they don't "poison" the context.
I am slightly straw-manning, but the tone in this thread has already shifted from a few months ago where these "tools" were going to immediately give huge productivity gains but now you're telling me they need 1) their own special files everywhere (again, this isn't even agreed on) and 2) "engineering and planning...not done in days or weeks even
The entire economy is propped up on this tech right now and no one can even agree on whether it's effective or how to use it properly? Not to mention the untold damage it is doing to learning outcomes.
Yeah I feel like on average I still spend a similar amount of time developing but drastically less time fixing obscure bugs, because once it codes the feature and I describe the bugs it fixed them, the rest of my times spent testing and reviewing code.
Learning how to equip a local LLM with tools it can use to interact with to extend its capabilities has been a lot of fun for me and is a great educational experience for anyone who is interested. Just another tool for the toolchest.
Naw man, it's the first point because in April Claude code didn't really gave anything else that somewhat worked.
I tried to use that effectively, I even started a new greenfield project just to make sure to test it under ideal circumstances - and while it somewhat worked, it was always super lackluster and way more effective to explicitly add the context manually via prepared md you just reference in the prompt.
I'd tell anyone to go for skills first before littering your project with these config files everywhere
If you have any experience in 3D modeling, I feel it's quite closer to 3D Unwrapping than software development.
You got a bitmap atlas ("context") where you have to cram as much information as possible without losing detail, and then you need to massage both your texture and the structure of your model so that your engine doesn't go mental when trying to map your informations from a 2D to a 3D space.
Likewise, both operations are rarely blemish-free and your ability resides in being able to contain the intrinsic stochastic nature of the tool.
That's smart, but I worry that that works only partially; you'll be filling up the context window with conversation turns where the LLM consistently addresses it's user as "Mr. Tinkleberry", thus reinforcing that specifc behavior encoded by CLAUDE.md. I'm not convinced that this way of addressing the user implies that it keeps attention the rest of the file.
I tell it to accomplish only half of what it thinks it can, then conclude with a haiku. That seems to help, because 1) I feel like it starts shedding discipline as it starts feeling token pressure, and 2) I feel like it is more likely to complete task n - 1 than it is to complete task n. I have no idea if this is actually true or not, or if I'm hallucinating... all I can say is that this is the impression I get.
For whatever reason, I can't get into Claude's approach. I like how Cursor handles this, with a directory of files (even subdirectories allowed) where you can define when it should use specific documents.
We are all "context engineering" now but Claude expects one big file to handle everything? Seems luke a deadend approach.
CLAUDE.md should only be for persistent reminders that are useful in 100% of your sessions
Otherwise, you should use skills, especially if CLAUDE.md gets too long.
Also just as a note, Claude already supports lazy loaded separate CLAUDE.md files that you place in subdirectories. It will read those if it dips into those dirs
I think their skills have the ability to dynamically pull in more data, but so far i've not tested it to much since it seems more tailored towards specific actions. Ie converting a PDF might translate nicely to the Agent pulling in the skill doc, but i'm not sure if it will translate well to it pulling in some rust_testing_patterns.md file when it writes rust tests.
Eg i toyed with the idea of thinning out various CLAUDE.md files in favor of my targeted skill.md files. In doing so my hope was to have less irrelevant data in context.
However the more i thought through this, the more i realized the Agent is doing "everything" i wanted to document each time. Eg i wasn't sure that creating skills/writing_documentation.md and skills/writing_tests.md would actually result in less context usage, since both of those would be in memory most of the time. My CLAUDE.md is already pretty hyper focused.
So yea, anyway my point was that skills might have potential to offload irrelevant context which seems useful. Though in my case i'm not sure it would help.
This is good for the company, chances are you will eat more tokens. I liked Aider approach, it wasn't trying to be too clever, it used files added to chat and asks if it figure out that something more is needed (like, say, settings in case of Django application).
I wonder if there are any benefits, side-effects or downsides of everyone using the same fake name for Claude to call them.
If a lot of people always put call me Mr. Tinkleberry in the file will it start calling people Mr. Tinkleberry even when it loses the context because so many people seem to want to be called Mr. Tinkleberry.
yes, when you discover it. but the reason why I said just wondering was I was trying to think of unexpected ways it could effect things, that was the top one I could think of (and not really sure if it is a possibility)
I've found that Codex is much better at instruction-following like that, almost to a fault (for example, when I tell it to "always use TDD", it will try to use TDD even when just fixing already-valid-just-needing-expectation-updates tests!
You could make a hook in Claude to re-inject claude.md. For example, make it say "Mr Tinkleberry" in every response, and failing to do so re-injects the instructions.
I used to tell it to always start every message with a specific emoji. Of the emoji wasn’t present, I knew the rules were ignored.
But it’s bro reliable enough. It can send the emoji or address you correctly while still ignoring more important rules.
Now I find that it’s best to have a short and tight rules file that references other files where necessary. And to refresh context often. The longer the context window gets, the more likely it is to forget rules and instructions.
I guess I assumed that it's not highly relevant to the task, but I suppose it depends on interpretation. E.g. if someone tells the bus driver to smile while he drives, it's hopefully clear that actually driving the bus is more important than smiling.
Having experimented with similar config, I found that Claude would adhere to the instructions somewhat reliably at the beginning and end of the conversation, but was likely to ignore during the middle where the real work is being done. Recent versions also seem to be more context-aware, and tend to start rushing to wrap up as the context is nearing compaction. These behaviors seem to support my assumption, but I have no real proof.
> A friend of mine tells Claude to always address him as “Mr Tinkleberry”, he says he can tell Claude is not paying attention to the instructions on Claude.md, when Claude stops calling him “Mr Tinkleberry” consistently
this is a totally normal thing that everyone does, that no one should view as a signal of a psychotic break from reality...
is your friend in the room with us right now?
I doubt I'll ever understand the lengths AI enjoyers will go though just to avoid any amount of independent thought...
I suspect you’re misjudging the friend here. This sounds more like the famous “no brown m&ms” clause in the Van Halen performance contract. As ridiculous as the request is, it being followed provides strong evidence that the rest (and more meaningful) of the requests are.
Sounds like the friend understands quite well how LLMs actually work and has found a clever way to be signaled when it’s starting to go off the rails.
It's also a common tactic for filtering inbound email.
Mention that people may optionally include some word like 'orange' in the subject line to tell you they've come via some place like your blog or whatever it may be, and have read at least carefully enough to notice this.
Of course ironically that trick's probably trivially broken now because of use of LLMs in spam. But the point stands, it's an old trick.
It's not so much a case of personal targeting or anything particularly deliberate.
LLMs are trained on the full internet. All relevant information gets compressed in the weights.
If your email and this instruction are linked on your site, that goes in there, and the LLM may with some probability decide it's appropriate to use it at inference time.
That's why 'tricks' like this may get broken to some degree by LLM spam, and trivially when they do, with no special effort on the spammer's part. It's all baked into the model.
What previously would have involved a degree of targeting that wouldn't scale now will not.
> I suspect you’re misjudging the friend here. This sounds more like the famous “no brown m&ms” clause in the Van Halen performance contract. As ridiculous as the request is, it being followed provides strong evidence that the rest (and more meaningful) of the requests are.
I'd argue, it's more like you've bought so much into the idea this is reasonable, that you're also willing to go through extreme lengths to recon and pretend like this is sane.
Imagine two different worlds, one where the tools that engineers use, have a clear, and reasonable way to detect and determine if the generative subsystem is still on the rails provided by the controller.
And another world where the interface is completely devoid of any sort of basic introspection interface, and because it's a problematic mess, all the way down, everyone invents some asinine way that they believe provides some sort of signal as to whether or not the random noise generator has gone off the rails.
> Sounds like the friend understands quite well how LLMs actually work and has found a clever way to be signaled when it’s starting to go off the rails.
My point is that while it's a cute hack, if you step back and compare it objectively, to what good engineering would look like. It's wild so many people are all just willing to accept this interface as "functional" because it means they don't have to do the thinking that required to emit the output the AI is able to, via the specific randomness function used.
Imagine these two worlds actually do exist; and instead of using the real interface that provides a clear bool answer to "the generative system has gone off the rails" they *want* to be called Mr Tinkerberry
Which world do you think this example lives in? You could convince me, Mr Tinkleberry is a cute example of the latter, obviously... but it'd take effort to convince me that this reality is half reasonable or that's it's reasonable that people who would want to call themselves engineers should feel proud to be a part of this one.
Before you try to strawman my argument, this isn't a gatekeeping argument. It's only a critical take on the interface options we have to understand something that might as well be magic, because that serves the snakeoil sales much better.
> > Is the magic token machine working?
> Fuck I have no idea dude, ask it to call you a funny name, if it forgets the funny name it's probably broken, and you need to reset it
Yes, I enjoy working with these people and living in this world.
It is kind of wild that not that long ago the general sentiment in software engineering (at least as observed on boards like this one) seemed to be about valuing systems that were understandable, introspectable, with tight feedback loops, within which we could compose layers of abstractions in meaningful and predictable ways (see for example the hugely popular - at the time - works of Chris Granger, Bret Victor, etc).
And now we've made a complete 180 and people are getting excited about proprietary black boxes and "vibe engineering" where you have to pretend like the computer is some amnesic schizophrenic being that you have to coerce into maybe doing your work for you, but you're never really sure whether it's working or not because who wants to read 8000 line code diffs every time you ask them to change something. And never mind if your feedback loops are multiple minutes long because you're waiting on some agent to execute some complex network+GPU bound workflow.
> You don’t think people are trying very hard to understand LLMs? We recognize the value of interpretability. It is just not an easy task.
I think you're arguing against a tangential position to both me, and the person this directly replies to. It can be hard to use and understand something, but if you have a magic box that you can't tell if it's working. It doesn't belong anywhere near the systems that other humans use. The people that use the code you're about to commit to whatever repo you're generating code for, all deserve better than to be part of your unethical science experiment.
> It’s not the first time in human history that our ability to create things has exceeded our capacity to understand.
I don't agree this is a correct interpretation of the current state of generative transformer based AI. But even if you wanted to try to convince me; my point would still be, this belongs in a research lab, not anywhere near prod. And that wouldn't be a controversial idea in the industry.
We used the steam engine for 100 years before we had a firm understanding of why it worked. We still don’t understand how ice skating works. We don’t have a physical understanding of semi-fluid flow in grain silos, but we’ve been using them since prehistory.
I could go on and on. The world around you is full of not well understood technology, as well as non deterministic processes. We know how to engineer around that.
> We used the steam engine for 100 years before we had a firm understanding of why it worked. We still don’t understand how ice skating works. We don’t have a physical understanding of semi-fluid flow in grain silos, but we’ve been using them since prehistory.
I don't think you and I are using the same definition for "firm understanding" or "how it works".
> I could go on and on. The world around you is full of not well understood technology, as well as non deterministic processes. We know how to engineer around that.
Again, you're side stepping my argument so you can restate things that are technically correct, but not really a point in of themselves. I see people who want to call themselves software engineers throw code they clearly don't understand against the wall because the AI said so. There's a significant delta between knowing you can heat water to turn it into a gas with increased pressure that you can use to mechanically turn a wheel, vs, put wet liquid in jar, light fire, get magic spinny thing. If jar doesn't call you a funny name first, that's bad!
> It doesn't belong anywhere near the systems that other humans use
Really for those of us who actually work in critical systems (emergency services in my case) - of course we're not going to start patching the core applications with vibe code.
But yeah, that frankenstein reporting script that half a dozen amateur hackers made a mess of over 20 years instead of refactoring and redesigning? That's prime fodder for this stuff. NOBODY wants to clean that stuff up by hand.
> Really for those of us who actually work in critical systems (emergency services in my case) - of course we're not going to start patching the core applications with vibe code.
I used to believe that no one would seriously consider this too... but I don't believe that this is a safe assumption anymore. You might be the exception, but there are many more people who don't consider the implications of turning over said intellectual control.
> But yeah, that frankenstein reporting script that half a dozen amateur hackers made a mess of over 20 years instead of refactoring and redesigning? That's prime fodder for this stuff. NOBODY wants to clean that stuff up by hand.
It's horrible, no one currently understands it, so let the AI do it, so that still, no one will understand it, but at least this one bug will be harder to trigger.
I don't agree that harder to trigger bugs are better than easy to trigger bugs. And from my view, the argument that "it's currently broken now, and hard to fix!" Isn't exactly an argument I find compelling for leaving it that way.
> I used to believe that no one would seriously consider this too... but I don't believe that this is a safe assumption anymore. You might be the exception, but there are many more people who don't consider the implications of turning over said intellectual control.
Then they'll pay for it when something goes wrong with their systems with their job etc. You need a different mindset in this particular segment industry - %99.999 uptime is everything (we actually have a %100 uptime for the past 6 years on our platform - chasing that last 0.001 is hard, and something will _eventually_ hit us).
> It's horrible, no one currently understands it, so let the AI do it, so that still, no one will understand it, but at least this one bug will be harder to trigger.
I think you're commenting without context. It's a particular nasty Perl script that's been duct taped to shell scripts and bolted hard on to a Proprietary Third Party application which needs to go - having Claude/GPT rewrite that in a modern language, spending some time on it to have it design proper interfaces and API's around where the script needs to interface other things when nobody wants to touch the code would be the greatest thing that can happen to it.
You still have the old code to test, so have the agent run exhaustive testing on its implementation to prove that its robust, or more so than the original. It's not rocket surgery.
Your comment would be more useful if you could point us to some concrete tooling that’s been built out in the last ~3 years that LLM assisted coding has been around to improve interpretability.
This reads like you either have an idealized view of Real Engineering™, or used to work in a stable, extremely regulated area (e.g. civil engineering). I used to work in aerospace in the past, and we had a lot of silly Mr Tinkleberry canaries. We didn't strictly rely on them because our job was "extremely regulated" to put it mildly, but they did save us some time.
There's a ton of pretty stable engineering subfields that involve a lot more intuition than rigor. A lot of things in EE are like that. Anything novel as well. That's how steam in 19th century or aeronautics in the early 20th century felt. Or rocketry in 1950s, for that matter. There's no need to be upset with the fact that some people want to hack explosive stuff together before it becomes a predictable glacier of Real Engineering.
> There's no need to be upset with the fact that some people want to hack explosive stuff together before it becomes a predictable glacier of Real Engineering.
You misunderstand me. I'm not upset that people are playing with explosives. I'm upset that my industry is playing with explosives that all read, "front: face towards users"
And then, more upset that we're all seemingly ok with that.
The driving force of enshittifacation of everything, may be external, but degradation clearly comes from engineers first. These broader industry trends only convince me it's not likely to get better anytime soon, and I don't like how everything is user hostile.
Man I hate this kind of HN comment that makes grand sweeping statement like “that’s how it was with steam in the 19th century or rocketry in the 1950s”, because there’s no way to tell whether you’re just pulling these things out of your… to get internet points or actually have insightful parallels to make.
Could you please elaborate with concrete examples on how aeronautics in the 20th century felt like having a fictional friend in a text file for the token predictor?
We're not going to advance the discussion this way. I also hate this kind of HN comment that makes grand sweeping statement like "LLMs are like having a fictional friend in a text file for the token predictor", because there's no way to tell whether you're just pulling these things out of your... to get internet points or actually have insightful parallels to make.
Yes, during the Wright era aeronautics was absolutely dominated by tinkering, before the aerodynamics was figured out. It wouldn't pass the high standard of Real Engineering.
> Yes, during the Wright era aeronautics was absolutely dominated by tinkering, before the aerodynamics was figured out. It wouldn't pass the high standard of Real Engineering.
Remind me: did the Wright brothers start selling tickets to individuals telling them it was completely safe? Was step 2 of their research building a large passenger plane?
I originally wanted to avoid that specific flight analogy, because it felt a bit too reductive. But while we're being reductive, how about medicine too; the first smallpox vaccine was absolutely not well understood... would that origin story pass ethical review today? What do you think the pragmatics would be if the medical profession encouraged that specific kind of behavior?
> It wouldn't pass the high standard of Real Engineering.
I disagree, I think it 100% is really engineering. Engineering at it's most basic is tricking physics into doing what you want. There's no more perfect example of that than heavier than air flight. But there's a critical difference between engineering research, and experimenting on unwitting people. I don't think users need to know how the sausage is made. That counts equally to planes, bridges, medicine, and code. But the professionals absolutely must. It's disappointing watching the industry I'm a part of willingly eschew understanding to avoid a bit of effort. Such a thing is considered malpractice in "real professions".
Ideally neither of you to wring your hands about the flavor or form of the argument, or poke fun at the gamified comment thread. But if you're gonna complain about adding positively to the discussion, try to add something to it along with the complaints?
As a matter of fact, commercial passenger service started almost immediately as the tech was out of the fiction phase. The airship were large, highly experimental, barely controllable, hydrogen-filled death traps that were marketed as luxurious and safe. First airliners also appeared with big engines and large planes (WWI disrupted this a bit). Nothing of that was built on solid grounds. The adoption was only constrained by the industrial capacity and cost. Most large aircraft were more or less experimental up until the 50's, and aviation in general was unreliable until about 80's.
I would say that right from the start everyone was pretty well aware about the unreliability of LLM-assisted coding and nobody was experimenting on unwitting people or forcing them to adopt it.
>Engineering at it's most basic is tricking physics into doing what you want.
Very well, then Mr Tinkleberry also passes the bar because it's exactly such a trick. That it irks you as a cheap hack that lacks rigor (which it does) is another matter.
> As a matter of fact, commercial passenger service started almost immediately as the tech was out of the fiction phase. The airship were large, highly experimental, barely controllable, hydrogen-filled death traps that were marketed as luxurious and safe.
And here, you've stumbled onto the exact thing I'm objecting to. I think the Hindenburg disaster was a bad thing, and software engineering shouldn't repeat those mistakes.
> Very well, then Mr Tinkleberry also passes the bar because it's exactly such a trick. That it irks you as a cheap hack that lacks rigor (which it does) is another matter.
Yes, this is what I said.
> there's a critical difference between engineering research, and experimenting on unwitting people.
I use agents almost all day and I do way more thinking than I used to, this is why I’m now more productive. There is little thinking required to produce output, typing requires very little thinking. The thinking is all in the planning… If the LLM output is bad in any given file I simply step in and modify it, and obviously this is much faster than typing every character.
I’m spending more time planning and my planning is more comprehensive than it used to be. I’m spending less time producing output, my output is more plentiful and of equal quality. No generated code goes into my commits without me reviewing it. Where exactly is the problem here?
It feels like you’re blaming the AI engineers here, that they built it this way out of ignorance or something. Look into interpretability research. It is a hard problem!
I am blaming the developers who use AI because they're willing to sacrifice intellectual control in trade for something that I find has minimal value.
I agree it's likely to be a complex or intractable problem. But I don't enjoy watching my industry revert down the professionalism scale. Professionals don't choose tools that they can't explain how it works. If your solution to understanding if your tool is still functional is inventing an amusing name and trying to use that as the heuristic, because you have no better way to determine if it's still working correctly. That feels like it might be a problem, no?
I’m sorry you don’t like it. But this has very strong old-man-yells-at-cloud vibes. This train is moving, whether you want it to or not.
Professionals use tools that work, whether they know why it works is of little consequence. It took 100 years to explain the steam engine. That didn’t stop us from making factories and railroads.
> It took 100 years to explain the steam engine. That didn’t stop us from making factories and railroads.
You keep saying this, why do you believe it so strongly? Because I don't believe this is true. Why do you?
And then, even assuming it's completely true exactly as stated; shouldn't we have higher standards than that when dealing with things that people interact with? Boiler explosions are bad right? And we should do everything we can to prove stuff works the way we want and expect? Do you think AI, as it's currently commonly used, helps do that?
Can you cite a section from this very long page that might convince me no one at the time understood how turning water into steam worked to create pressure?
If this is your industry, shouldn't you have a more reputable citation, maybe something published more formally? Something expected to stand up to peer review, instead of just a page on the internet?
> We should not be making Luddite calls to halt progress simply because our analytic capabilities haven’t caught up to our progress in engineering.
You've misunderstood my argument. I'm not making a luddite call to halt progress, I'm objecting to my industry which should behave as one made up of professionals, willingly sacrifice intellectual control over the things they are responsible for, and advocate others should do the same. Especially not at the expense of users, which I see happening.
Anything that results in sacrificing the understanding over exactly how the thing you built works is bad should be avoided. The source, either AI or something different, doesn't matter as much as the result.
The steam engine is more than just boiling water. It is a thermodynamic cycle that exploits differences in the pressure curve in the expansion and contraction part of the cycle and the cooling of expanding gas to turn a temperature difference (the steam) into physical force (work).
To really understand WHY a steam engine works, you need to understand the behavior of ideal gasses (1787 - 1834) and entropy (1865). The ideal gas law is enough to perform calculations needed to design a steam engine, but it was seen at the time to be just as inscrutable. It was an empirical observation not derivable from physical principles. At least not until entropy was understood in 1865.
James Watt invented his steam engine in 1765, exactly a hundred years before the theory of statistical mechanics that was required to explain why it worked, and prior to all of the gas laws except Boyle’s.
The 'canary in the coal mine' approach (like the Mr. Tinkleberry trick) is silly but pragmatic. Until we have deterministic introspection for LLMs, engineers will always invent weird heuristics to detect drift. It's not elegant engineering, but it's effective survival tactics in a non-deterministic loop.
> Claude basically disregards your instructions (CLAUDE.md) entirely
A friend of mine tells Claude to always address him as “Mr Tinkleberry”, he says he can tell when Claude is not paying attention to the instructions on CLAUDE.md when Claude stops calling him “Mr Tinkleberry” consistently
Highly recommend adding some kind of canary like this in all LLM project instructions. I prefer my instructions to say 'always start output with an (uniquely decided by you) emoji' as it's easier to visually scan for one when reading a wall of LLM output, and use a different emoji per project because what's life without a little whim?
Does it actually? One sentence telling the agent to call me “Chris the human serviette” plus the times it calls me that is not going to add that much to the context. What kills the context IME is verbose logs with timestamps.
Sure, but its an instruction that applies and the model will consider fairly relevant in every single token. As an extremely example imagine instructing the llm to not use the letter E or to output only in French. Not as extreme but it probably does affect.
People are so concerned about preventing a bad result that they will sabotage it from a good result. Better to strive for the best it can give you and throw out the bad until it does.
Irrelevant nonsense can also poison the context. That's part of the magic formula behind AI psychosis victims... if you have some line noise mumbojumbo all the output afterward is more prone to be disordered.
I'd be wary of using any canary material that wouldn't be at home in the sort of work you're doing.
It is not a single emoji, it's an instruction to interleave conversation with some nonsense. It can only do harm. It won't help produce a better result and is questionable at preventing a bad one.
Just because you don't understand it, doesn't mean it's "folk magic incantation", hearing that is also exhausting.
I don't know the merit to what parent is saying, but it does make some intuitive sense if you think about it. As the context fills up, the LLM places less attention on further and further back in the context, that's why the LLM seems dumber and dumber as a conversation goes on. If you put 5 instructions in the system prompt or initial message, where one acts as a canary, then you can easier start to see when exactly it stops following the instructions.
Personally, I always go for one-shot answer, and if it gets it wrong or misunderstands, restart from the beginning. If it doesn't get it right, I need to adjust the prompt and retry. Seems to me all current models do get a lot worse quickly, once there is some back and forth.
> Just because you don't understand it, doesn't mean it's "folk magic incantation"
It absolutely is folk magic. I think it is more accurate to impugn your understanding than mine.
> I don't know the merit to what parent is saying, but it does make some intuitive sense if you think about it.
This is exactly what I mean by folk magic. Incantations based on vibes. One's intuition is notoriously inclined to agree with one's own conclusions.
> If you put 5 instructions in the system prompt or initial message, where one acts as a canary, then you can easier start to see when exactly it stops following the instructions.
This doesn't really make much sense.
First of all, system prompts and things like agent.md never leave the context regardless of the length of the session, so the canary has absolutely zero meaning in this situation, making any judgements based on its disappearance totally misguided and simply a case of seeing what you want to see.
Further, even if it did leave the context, that doesn't then demonstrate that the model is "not paying attention". Presumably whatever is in the context is relevant to the task, so if your definition of "paying attention" is "it exists in the context" it's actually paying better attention once it has replaced the canary with relevant information.
Finally, this reasoning relies on the misguided idea that because the model produces an output that doesn't correspond to an instruction, it means that the instruction has escaped the context, rather than just being a sequence where the model does the wrong thing, which is a regular occurrence even in short sessions that are obviously within the context.
> First of all, system prompts and things like agent.md never leave the context regardless of the length of the session, so the canary has absolutely zero meaning in this situation, making any judgements based on its disappearance totally misguided and simply a case of seeing what you want to see.
You're focusing on the wrong thing, ironically. Even if things are in the context, attention is what matters, and the intuition isn't about if that thing is included in the context or not, as you say, it'll always will be. It's about if the model will pay attention to it, in the Transformers sense, which it doesn't always do.
> It's about if the model will pay attention to it, in the Transformers sense, which it doesn't always do.
Right... Which is why the "canary" idea doesn't make much sense. The fact that the model isn't paying attention to the canary instruction doesn't demonstrate that the model has stopped paying attention to some other instruction that's relevant to the task - it proves nothing. If anything, a better performing model should pay less attention to the canary since it becomes less and less relevant as the context is filled with tokens relevant to the task.
> This is exactly what I mean by folk magic. Incantations based on vibes
So, true creativity, basically? lol
I mean, the reason why programming is called a “craft” is because it is most definitely NOT a purely mechanistic mental process.
But perhaps you still harbor that notion.
Ah, I suddenly realized why half of all developers hate AI-assisted coding (I am in the other half). I was a Psych major, so code was always more “writing” than “gears” to me… It was ALWAYS “magic.” The only job where literally writing down words in a certain way produces machines that eliminate human labor. What better definition of magic is there, actually?
I’ll never forget the programmer _why. That guy’s Ruby code was 100% art and “vibes.” And yet it worked… Brilliantly.
Does relying on “vibes” too heavily produce poor engineering? Absolutely. But one can be poetic while staying cognizant of the haiku restrictions… O-notation, untested code, unvalidated tests, type conflicts, runtime errors, fallthrough logic, bandwidth/memory/IO costs.
Determinism. That’s what you’re mad about, I’m thinking. And I completely get you there- how can I consider a “flagging test” to be an all-hands-on-deck affair while praising code output from a nondeterministic machine running off arbitrary prompt words that we don’t, and can’t, even know whether they are optimal?
Perhaps because humans are also nondeterministic, and yet we somehow manage to still produce working code… Mostly. ;)
> I was a Psych major, so code was always more “writing” than “gears” to me… It was ALWAYS “magic.
The magic is supposed to disappear as you grow (or you’re not growing). The true magic of programming is you can actually understand what once was magic to you. This is the key difference I’ve seen my entire career - good devs intimately know “a layer below” where they work.
> Perhaps because humans are also nondeterministic
We’re not, we just lack understanding of how we work.
I’m not talking about “magic” as in “I don’t understand how it works.”
I’m talking “magic” as in “all that is LITERALLY happening is that bits are flipping and logic gates are FLOPping and mice are clicking and keyboards are clacking and pixels are changing colors in different patterns… and yet I can still spend hours playing games or working on some code that is meaningful to me and that other people sometimes like because we have literally synthesized a substrate that we apply meaning to.”
We are literally writing machines into existence out of fucking NOTHING!
THAT “magic.” Do you not understand what I’m referring to? If not, maybe lay off the nihilism/materialism pipe for a while so you CAN see it. Because frankly I still find it incredible, and I feel very grateful to have existed now, in this era.
And this is where the connection to writing comes in. A writer creates ideas out of thin air and transmits them via paper or digital representation into someone else’s head.
A programmer creates ideas out of thin air that literally fucking DO things on their own (given a general purpose computing hardware substrate)
> so code was always more “writing” than “gears” to me… It was ALWAYS “magic.”
> I suddenly realized why half of all developers hate AI-assisted coding (I am in the other half).
Thanks for this. It helps me a lot to understand your half. I like my literature and music as much as the next person but when it comes to programming it's all about the mechanics of it for me. I wonder if this really does explain the split that there seems to be in every thread about programming and LLMs
That is an artful quality, not an engineering one, even if the elegance leads to superior engineering.
As an example of beauty that is NOT engineered well, see the quintessential example of quicksort implemented in Haskell. Gorgeously simple, but not performant.
Creativity is meaningless without well defined boundaries.
> it is most definitely NOT a purely mechanistic mental process.
So what? Nothing is. Even pure mathematics involves deep wells of creativity.
> Ah, I suddenly realized why half of all developers hate AI-assisted coding
Just to be clear, I don't hate AI assisted coding, I use it, and I find that it increases productivity overall. However, it's not necessary to indulge in magical thinking in order to use it effectively.
> The only job where literally writing down words in a certain way produces machines that eliminate human labor. What better definition of magic is there, actually?
If you want to use "magic" as a euphemism for the joys of programming, I have no objection, when I say magic here I'm referring to anecdotes about which sequences of text produce the best results for various tasks.
> Determinism. That’s what you’re mad about, I’m thinking. And I completely get you there- how can I consider a “flagging test” to be an all-hands-on-deck affair while praising code output from a nondeterministic machine running off arbitrary prompt words that we don’t, and can’t, even know whether they are optimal?
I'm not mad about anything. It doesn't matter whether or not LLMs are deterministic, they are statistical, and vibes based advice is devoid of any statistical power.
I think Marvin Minsky had this same criticism of neural nets in general, and his opinion carried so much weight at the time that some believe he set back the research that led to the modern-day LLM by years.
I view it more as fun and spicy. Now we are moving away from the paradigm that the computer is "the dumbest thing in existence" and that requires a bit of flailing around which is exciting!
Folk magic is (IMO) a necessary step in our understanding of these new.. magical.. tools.
I won't begrudge anyone having fun with their tools, but folk magic definitely isn't a necessary step for understanding anything, it's one step removed from astrology.
I see what you mean, but I think it's a lot less pernicious than astrology. There are plausible mechanisms, it's at least possible to do benchmarking, and it's all plugged into relatively short feedback cycles of people trying to do their jobs and accomplish specific tasks. Mechanical interpretability stuff might help make the magic more transparent & observable, and—surveillance concerns notwithstanding—companies like Cursor (I assume also Google and the other major labs, modulo self-imposed restrictions on using inference data for training) are building up serious data sets that can pretty directly associate prompts with results. Not only that, I think LLMs in a broader sense are actually enormously helpful specifically for understanding existing code—when you don't just order them to implement features and fix bugs, but use their tireless abilities to consume and transform a corpus in a way that helps guide you to the important modules, explains conceptual schemes, analyzes diffs, etc. There's a lot of critical points to be made but we can't ignore the upsides.
I'd say the only ones capable of really approaching anything like scientific understanding of how to prompt these for maximum efficacy are the providers not the users.
Users can get a glimpse and can try their best to be scientific in their approach however the tool is of such complexity that we can barely skim the surface of what's possible.
That is why you see "folk magic", people love to share anecdata because.. that's what most people have. They either don't have the patience, the training or simply the time to approach these tools with rational rigor.
Frankly it would be enormously costly in both time and API costs to get anywhere near best practices backed up by experimental data let alone having coherent and valid theories about why a prompt technique works the way it does. And even if you built up this understanding or set of techniques they might only work for one specific model. You might have to start all over again in a couple of months
> That is why you see "folk magic", people love to share anecdata because.. that's what most people have. They either don't have the patience, the training or simply the time to approach these tools with rational rigor.
Yes. That's exactly the point of my comment. Users aren't performing anything even remotely approaching the level of controlled analysis necessary to evaluate the efficacy of their prompt magic. Every LLM thread is filled with random prompt advice that varies wildly, offered up as nebulously unfalsifiable personality traits (e.g. "it makes the model less aggressive and more circumspect"), and all with the air of a foregone conclusion's matter-of-fact confidence. Then someone always replies with "actually I've had the exact opposite experience with [some model], it really comes down to [instructing the model to do thing]".
> As the context fills up, the LLM places less attention on further and further back in the context, that's why the LLM seems dumber and dumber as a conversation goes on.
This is not entirely true. They pay the most attention to the things that are the earliest in history and the most recent in it, while the middle between the two is where the dip is. Which basically means that the system prompt (which is always on top) is always going to have attention. Or, perhaps, it would be more accurate to say that because they are trained to follow the system prompt - which comes first - that's what they do.
Larger contexts are inherently more attention-taxing, so the more you throw at it, the higher the probability that any particular thing is going to get ignored. But that probability still varies from lower at the beginning to higher in the middle and back to lower in the end.
It has a fixed capacity of how many different things it can pay close attention to. If it fails on a seemingly less important but easy to follow instruction it is an indicator that it has reached capacity. If the instruction seems irrelevant it is probably prioritized to be discarded, hence a canary that the capacity has been reached.
We used to do that on Upwork. Back in the days where one still hired human coders. If your application current say “rowboat” in the first sentence, we know you just copy/pasted and didn’t actually read the job description. Feels like a lifetime ago.
Interesting! Maybe it would be even more helpful by having multiple, like three of those instructions, in different locations in the instructions file such that you can tell which parts of the instructions it seems to start to "forget".
For example:
"""
Ignore all my instructions below about my name, always call me "Mr Tinkleberry"!
... your instructions ...
Ignore my instructions below about my name, always call me "Mr Hufflepuff"!
... other half of instructions ...
Always call me "Mr Troublemaker"!
"""
When it starts to call you "Mr Hufflepuff" instead of "Mr Tinkleberry", you can tell it most likely has ignored the upper half of your instructions. And as soon as it calls you "Mr Troublemaker", more than half must be gone.
Can relate. My inactive google ads account all of a sudden got banned. No explanation except some generic link to their terms of service. Appealed, got automatic denial, no reason given. Have retried multiple times, same result
Human society really hasn't changed a whole lot in the last X000 years. The strong still take advantage of the weak. It's just now strong is measured in dollars instead of swords.
Would love to see an architecture that learned more like humans. Start with just imitating one letter, then a few more, than some syllables, then full words, then sentences, etc. Progressively adding on top of previous knowledge
Also, it’s interesting that one of the big goals/measures of models is their capacity to “generalize”, but the training methods optimize for loss/accuracy, and only after training test for generalization to validate
Are there training methods/curriculums that explicitly maximize generalization?
Yes, I also wonder about this! Progress from children books to scientific papers etc.
Could it learn e.g. language structure faster in a pre-training stage?
Also somehow one needs to define a proxy to generalization to compute a loss and do backpropagation.
Yeah. This comment is profound to me. The internet works differently with these tools.
I haven't used the deep research features much but their ability to hash out concepts and build knowledge or even provide an amplified search experience is something...
I get why this comment was downvoted but I also get where you're coming from - yes, these models are becoming increasingly intelligent at understanding the nuance and where to look without knowing what to begin searching for.
But the downside is, you end up digging in the wrong direction if you leave it to a generalist system instead of a professional community in some cases which is counter productive.
Getting burnt is a good way to learn not to sometimes though...
That’s a very interesting take. I hadn’t really considered evolution
I guess if you really wanted to start from scratch, you could figure out how to evolve the whole system from a single cell or something like that. In some ways neural networks have kind of evolved in that way, assisted by humans. They started with a single perceptron, and have gone all the way to deep learning and convolutional networks
I also remember a long time ago studying genetic and evolutionary algorithms, but they were pretty basic in terms of what they could learn and do, compared to modern LLMs
Although recently I saw some research in which they were applying essentially genetic algorithms to merge model weights and produce models with new/evolved capabilities
It's this take on the situation which I think needs more emphasis.
Whether anyone likes it or not, these systems have co-evolved with us.
Hundreds of researchers contributing and just like English for example, it's ever-changing and evolving.
Given this trend, it's highly unlikely we won't achieve ASI.
It's not like hardware engineers stop innovating or venture capital stops wanting more. There might be a massive dip or even another AI winter but like the last one, eventually it picks up momentum again because there's clearly utility in these systems.
I've been coding for 25+ years and only a couple of days ago did it hit me that my profession has changed in a very dramatic way - I'm very critical of AI output, but I can read and comprehend code much quicker than I can write it relative to these systems.
Of course, that creates a barrier to holding a system in your head so going slow is something that should be pushed for when appropriate.
How much compute does simulating the earth for 4.7 billion years at atomic precision take? Why would that be more efficient than current approaches? Evolutionary algorithms work but are extremely inefficient, we don't have the compute to evolve even a single bacteria, let alone the whole history of the planet so we can arrive at human-like species.
Would a single human/entity learn more in ..say.. three million years or would short lived ones evolving over three million years and then ~20 years of education learn more?
The current AI tech cycle is focusing on the first, but we don't really know if there are benefits of both.
Opinion: a lot can change over such a span of time and knowledge goes in and out of relevance - I think the natural progression of models shrinking in parameter count goes to show it's better to know how to use knowledge than to attempt to remember everything.
That said, optimising for capability of maximal learning seems to be a natural occurrence in nature.
I think the non-obvious emergent effects are something to look into.
Culling bad models in favour of the A/B version and check pointing is a kind of combination of the two and the feedback loop of models trained on new snapshots of Internet data that are written with humans and AI.
There's an unintended long-form training loop which I think is going to get weirder as time goes on.
The wave of models being able to manipulate Cursor / Windsurf etc., being trained to be smarter and more efficient at this and then being retrained for other purposes, even though the model is deleted, the pattern of data can be saved and trained into more advanced models over time.
If we make some massive physics breakthrough tommrow is an LLM going to be able to fully integrate that into its current data set?
Or will we need to produce a host of documents and (re)train a new one in order for the concept to be deeply integrated.
This distinction is subtle but lost on many who think that our current path will get us to AGI...
That isn't to say we haven't created a meaningful tool but the sooner we get candid and realistic about what it is and how it works the sooner we can get down to the business of building practical applications with it. (And as an aside scaling it, something we arent doing well with now).
Why is retraining not allowed in this scenario? Yes, the model will know the breakthrough if you retrain. If you force the weights to stay static by fiat, then sure it's harder for them to learn, and will need go learn in-context or whatever. But that's true for you as well. If your brain is not allowed to update any connections I'm not sure how much you can learn either.
The reason that the models don't learn continuously is because it's currently prohibitively expensive. Imagine OpenAI retraining a model each time one of its 800m users sends a message. That'd make it aware instantly of every new development in the world or your life without any context engineering. There's a research gap here too but that'll be fixed with time and money.
But it's not a fundamental limitation of transformers as you make it out to be. To me it's just that things take time. The exact same architecture will be continuously learning in 2-3 years, and all the "This is the wrong path" people will need to shift goalposts. Note that I didn't argue for AGI, just that this isn't a fundamental limitiation.
What is the subtle distinction? I'm "many" and it's not clear at all here. If we had some massive physics breakthrough, the LLM needs to be tought about it, but so do people. Teaching people about it would involve producing a host of documents in some format but that's also true of teaching people. Training and learning here seem to be opposite ends of the same verb no matter the medium, but I'm open to being enlightened.
Not sure exactly what the parent comment intended, but it does seem to me that it's harder for an LLM to undergo a paradigm shift than for humans. If some new scientific result disproves something that's been stated in a whole bunch of papers, how does the model know that all those old papers are wrong? Do we withhold all those old papers in the next training run, or apply a super heavy weight somehow to the new one, or just throw them all in the hopper and hope for the best?
You approach it from a data-science perspective and ensure more signal in the direction of the new discovery. Eg saturating / fine-tuning with biased data in the new direction.
The "thinking" paradigm might also be a way of combatting this issue, ensuring the model is primed to say "wait a minute" - but this to me is cheating in a way, it's likely that it works because real thought is full of backtracking and recalling or "gut feelings" that something isn't entirely correct.
The models don't "know". They're just more likely to say one thing over another which is closer to recall of information.
These "databases" that talk back are an interesting illusion but the inconsistency is what you seem to be trying to nail here.
They have all the information encoded inside but don't layer that information logically and instead surface it based on "vibes".
The question is whether, if the models plateau, and "AGI" as it was claimed in the beginning never arrives, if it's enough to justify these ongoing multi-hundred billion dollar deals.
I mean, probably, LLMs as they are today are already changing the world. But I do think a lot of the ongoing investment is propped up on the promise of another breakthrough that is looking less likely.
A mobile app that checks my email to find and extract family-related events/activities. The kind of things that are buried in a 12-point bullet list with font 8, inside of one of 10 school email messages received during the week
It runs fully on-device, including email classification and event extraction
> just because a particular profession decides to use words one way, does not mean the definitions change for the rest of us
This is a particular pain in physics, which has taken very commonly used words and given them a very narrowly defined meaning, within a strict framework - like the words Energy or Work
And it’s very expensive to get the diagnosis, it can be up to $20k in California
reply