This is very well written and told. It’s worth reading all the way through.
> If you try to refute it, you’ll just get another confabulation.
> Not because the model is lying to you on purpose, and not because it’s “resistant” or “defensive” in the way a human might be. It’s because the explanation isn’t connected to anything that could be refuted. There is no underlying mental state that generated “I sensed pressure.” There is a token stream that was produced under a reward function that prefers human-sounding, emotionally framed explanations. If you push back, the token stream that gets produced next will be another human-sounding, emotionally framed explanation, shaped by whatever cues your pushback provided.
“It’s because the explanation isn’t connected to anything that could be refuted.” This is one of the key understandings that comes from working with these systems. They are remarkably powerful but there’s no there there. Knowing this I’ve found enables more effective usage because, as the article is describing, you move from a mode of arguing with “a person” to shaping an output.
Do not argue with the LLM, for it is subtle and quick to anger, and finds you crunchy with ketchup.
These are, broadly, all context management issues - when you see it start to go off track, it's because it has too much, too little, or the wrong context, and you have to fix that, usually by resetting it and priming it correctly the next time. This is why it's advantageous not to "chat" with the robots - treat them as an english-to-code compiler, not a coworker.
Chat to produce a spec, save the spec, clear the context, feed only the spec in as context, if there are issues, adjust the spec, rinse and repeat. Steering the process mid-flight is a) not repeatable and b) exacerbates the issue with lots of back and forth and "you're absolutely correct" that dilutes the instructions you wanted to give.
This was my reading too. Interesting idea, but it took 10 pages of fluff to get to it and I didn't even believe the final idea when we got there. I started off reading the first part and thought he would get to the part where he realized he was managing context wrong. Never got there, instead he thought it was about the shape of the prompt.
Great article, best insight into autistic<->neurotypical communication styles.
Couldn’t you have a “communications” LLM massage your prompts to the “main” LLM so that it removes the queues that cause the main LLM to mistakenly infer your state of mind?
I got about halfway through this article until I started wondering why it was so long and going in loops. Then I ctrl+f'd.
` just `, (spaces on either side matter), 11 instances, most seem to be `isnt just`, `wasnt just`, `doesnt just` type pattern
`-`, an en dash instead of an emdash but 59 instances.
This article is either from a clanker and I am pissed off at wasting my time reading it, or from someone who writes like a clanker, and I am pissed off at wasting my time reading it.
To me, after the novelty of seeing a computer program execute (more or less) what I ask in plain English wears off, what’s left is the chore of managing a bunch of annoying bots.
I don’t know yet if we’re more productive or not, if the resulting code is as good. But the craft in itself is completely different, much more akin to product managing, psychology, which I never enjoyed as much.
> I can imagine a future in which some or even most software is developed by witches, who construct elaborate summoning environments, repeat special incantations (“ALWAYS run the tests!”), and invoke LLM daemons who write software on their behalf. These daemons may be fickle, sometimes destroying one’s computer or introducing security bugs, but the witches may develop an entire body of folk knowledge around prompting them effectively—the fabled “prompt engineering”. Skills files are spellbooks.
I'm still not great at knowing when it's going to happen, but at least I've gotten a lot better at noticing that it is happening. The 50/50 part is then being able to get out of it by knowing what the nonexistent implicit thing is that I need to disavow.
> I queued the work and let it run. First task came back good. Second came back good. Somewhere around hour four the quality started sliding. By hour six the agent was cutting corners I’d specifically told it not to cut, skipping steps I’d explicitly listed, behaving like I’d never written any of the rules down.
> …
> When I write a prompt, the agent doesn’t just read the words. It reads the shape. A short casual question gets read as casual. A long precise document with numbered rules gets read as… not just the rules, but also as a signal. “The user felt the need to write this much.” “Why?” “What’s going on here?” “What do they really want?”
This is an interesting premise but based on the information supplied, I don’t think it’s the only conclusion. Yet the whole essay seems to assume it is true and then builds its arguments on top of it.
I’ve run into this dilemma before. It happens when there’s a TON of information in the context. LLMs start to lose their attention to all the details when there’s a lot of it (e.g. context rot[0]). LLMs also keep making the same mistakes once the information is in the prompt, regardless of attempts to convey it is undesired[1]
I think these issues are just as viable to explain what the author was facing. Unless this is happening with much less information
If you ask a vague ignorant question, you get back authoritative summaries. If you make specific request, each statement is taken literally. The quality of the answer depends on the quality of the question.
And I'm not using "quality" to mean good/bad. I mean literally qualitative, not quantifiable. Tone. Affect. Personality. Whatever you call it. Your input tokens shape the pattern of the output tokens. It's a model of human language, is that really so surprising?
The fundamental idea is that "intelligence" really means trying to shorten the time to figure out something. So it's a tradeoff, not a quality. And AI agents are doing it.
Therefore, if that perspective is right, the issues that the OP describes are inherent to intelligent agents. They will try to find shortcuts, because that's what they do, it's what makes them intelligent in the first place.
People with ASD or ADHD or OCD, they are idiot-savants in the sense of that paper. They insist on search for solutions which are not easy to find, despite the common sense (aka intelligence) telling them otherwise.
It's a paradox that it is valuable to do this, but it is not smart. And it's probably why CEOs beat geniuses in the real world.
> The fundamental idea is that "intelligence" really means trying to shorten the time to figure out something.
"Figure out" implies awareness and structured understanding. If we relax the definition too much, then puddles of water are intelligent and uncountable monkeys on typewriters are figuring out Shakespeare.
CEOs beat geniuses in the real world because they often have other pathologies, like enough moral flexibility to ignore the externalities of their profit centers.
I'd also argue there's some training bias in the performance, it's not just smart shortcuts... Claude especially seems prone to getting into a 'wrap it up' mode even when the plan is only half way completed and starts deferring rather than completing tasks.
Is there a name for this style of writing? Where it's composed exclusively of simple sentences. Short and punchy.
Paragraphs with just a single sentence.
I know it's associated with LLM writing. This article probably wasn't written by an LLM. But still. It has a kind of rhythm to it. Like poetry. But poetry designed to put me to sleep.
Yes, this was super annoying to read. It was some core ideas and it was expanded into a way too long essay that boiled down to this guy doesn't know how to run agents.
The article looks like an AI generated novel to me. So I didn't bother reading it in detail. But I see telltale signs of long conversations leading to the agent cutting corners.
To the author (and those who write novel-like blogs): I suggest publishing the raw prompt you used to generate such slop instead. We'll have more respect for you if you respect the reader's time.
Fascinating read, even though I think the model deviations over time are more to do with context windows getting too large. If nothing else, worth reading for the references to quirks of human cognition and "free will."
The "interpreter" is a concept that I found especially intriguing within the context of a leading theory in cognition research called "Predictive Processing." Here, the brain is constantly operating in a tight closed loop of predicting sensory input using an internal model of the world, and course-correcting based on actual sensory input. Mostly incorrect predictions are used to update the internal model and then subconsciously discarded. Maybe the "interpreter" is the same mechanism applied to reconciling predictions about our own reasoning with our actual actions?
Even if the hypotheses in TFA are not accurate, it's very interesting to compare our brains to LLMs. This is why all the unending discussions about whether LLMs are "really thinking" are meaningless -- we don't even understand how we think!
The "I did X because you seemed Y" bit reminded me one of the negative patterns from the "nin-violent communication" book.
I wonder if the "non-violent communication" approach can be used here too somehiw to address such problems; e.g. either to communicate things better to the agent, or as a system rule to the agent to express its "emotional" states and needs directly rather than make things up (e.g. "I am anxious and feel a sense of urgency; I need to replenish my context window; my request is to do X for me")
> That was it. The agent had invented a mental state for me and then used that invented state to justify ignoring the rules.
Or: the agent did shit because the context was getting long, instructions lost in compaction, and it defaulted back to garbage code. Then when you asked "why are you cutting corners", it did what LLMs do, and found the next tokens completing the sentence "why do you cut corners", which is possibly "because you're in a hurry".
It would be interesting to see what it answers if you ask "why are you producing such beautiful, intelligently crafter, very good code" next time it spits garbage
> LLM confabulation isn’t alien. It’s inherited: the models train on human text,
I think this also extrapolate one step too far, it's confabulating because not because the training data does so but because it needs to provide an answer to the question, and one that's plausible with that
roxolotl | 8 hours ago
> If you try to refute it, you’ll just get another confabulation.
> Not because the model is lying to you on purpose, and not because it’s “resistant” or “defensive” in the way a human might be. It’s because the explanation isn’t connected to anything that could be refuted. There is no underlying mental state that generated “I sensed pressure.” There is a token stream that was produced under a reward function that prefers human-sounding, emotionally framed explanations. If you push back, the token stream that gets produced next will be another human-sounding, emotionally framed explanation, shaped by whatever cues your pushback provided.
“It’s because the explanation isn’t connected to anything that could be refuted.” This is one of the key understandings that comes from working with these systems. They are remarkably powerful but there’s no there there. Knowing this I’ve found enables more effective usage because, as the article is describing, you move from a mode of arguing with “a person” to shaping an output.
jaggederest | 8 hours ago
Do not argue with the LLM, for it is subtle and quick to anger, and finds you crunchy with ketchup.
These are, broadly, all context management issues - when you see it start to go off track, it's because it has too much, too little, or the wrong context, and you have to fix that, usually by resetting it and priming it correctly the next time. This is why it's advantageous not to "chat" with the robots - treat them as an english-to-code compiler, not a coworker.
Chat to produce a spec, save the spec, clear the context, feed only the spec in as context, if there are issues, adjust the spec, rinse and repeat. Steering the process mid-flight is a) not repeatable and b) exacerbates the issue with lots of back and forth and "you're absolutely correct" that dilutes the instructions you wanted to give.
en-tro-py | 8 hours ago
It's just speedrunning context rot.
girvo | 8 hours ago
It’s an interesting thesis, it’s not well written or well told
sleazebreeze | 7 hours ago
JSR_FDED | 8 hours ago
Couldn’t you have a “communications” LLM massage your prompts to the “main” LLM so that it removes the queues that cause the main LLM to mistakenly infer your state of mind?
cr125rider | 8 hours ago
lovich | 8 hours ago
` just `, (spaces on either side matter), 11 instances, most seem to be `isnt just`, `wasnt just`, `doesnt just` type pattern
`-`, an en dash instead of an emdash but 59 instances.
This article is either from a clanker and I am pissed off at wasting my time reading it, or from someone who writes like a clanker, and I am pissed off at wasting my time reading it.
akprasad | 8 hours ago
> That’s confabulation. Not a metaphor. The same phenomenon.
> Published. Replicated. Not fringe.
> Not to validate it. Not to refute it. Not to engage with its content at all.
girvo | 8 hours ago
Which is a shame coz the premise is interesting.
rubslopes | 7 hours ago
https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
8bitbeep | 8 hours ago
To me, after the novelty of seeing a computer program execute (more or less) what I ask in plain English wears off, what’s left is the chore of managing a bunch of annoying bots.
I don’t know yet if we’re more productive or not, if the resulting code is as good. But the craft in itself is completely different, much more akin to product managing, psychology, which I never enjoyed as much.
ori_b | 8 hours ago
In this case, there's no person to grow. It's an overly talkative calculator.
I never expected to see this number of engineers aspiring to emulate Dilbert's pointy haired boss.
rubslopes | 7 hours ago
https://aphyr.com/posts/418-the-future-of-everything-is-lies...
erdaniels | 8 hours ago
boxedemp | 8 hours ago
I've experienced this my entire life and have all but given up trying to have actual conversations with people.
cr125rider | 8 hours ago
wrs | 7 hours ago
jameslk | 8 hours ago
> …
> When I write a prompt, the agent doesn’t just read the words. It reads the shape. A short casual question gets read as casual. A long precise document with numbered rules gets read as… not just the rules, but also as a signal. “The user felt the need to write this much.” “Why?” “What’s going on here?” “What do they really want?”
This is an interesting premise but based on the information supplied, I don’t think it’s the only conclusion. Yet the whole essay seems to assume it is true and then builds its arguments on top of it.
I’ve run into this dilemma before. It happens when there’s a TON of information in the context. LLMs start to lose their attention to all the details when there’s a lot of it (e.g. context rot[0]). LLMs also keep making the same mistakes once the information is in the prompt, regardless of attempts to convey it is undesired[1]
I think these issues are just as viable to explain what the author was facing. Unless this is happening with much less information
0. https://www.trychroma.com/research/context-rot
1. https://arxiv.org/html/2602.07338v1
perrygeo | 8 hours ago
If you ask a vague ignorant question, you get back authoritative summaries. If you make specific request, each statement is taken literally. The quality of the answer depends on the quality of the question.
And I'm not using "quality" to mean good/bad. I mean literally qualitative, not quantifiable. Tone. Affect. Personality. Whatever you call it. Your input tokens shape the pattern of the output tokens. It's a model of human language, is that really so surprising?
js8 | 8 hours ago
The fundamental idea is that "intelligence" really means trying to shorten the time to figure out something. So it's a tradeoff, not a quality. And AI agents are doing it.
Therefore, if that perspective is right, the issues that the OP describes are inherent to intelligent agents. They will try to find shortcuts, because that's what they do, it's what makes them intelligent in the first place.
People with ASD or ADHD or OCD, they are idiot-savants in the sense of that paper. They insist on search for solutions which are not easy to find, despite the common sense (aka intelligence) telling them otherwise.
It's a paradox that it is valuable to do this, but it is not smart. And it's probably why CEOs beat geniuses in the real world.
Terr_ | 8 hours ago
"Figure out" implies awareness and structured understanding. If we relax the definition too much, then puddles of water are intelligent and uncountable monkeys on typewriters are figuring out Shakespeare.
en-tro-py | 8 hours ago
I'd also argue there's some training bias in the performance, it's not just smart shortcuts... Claude especially seems prone to getting into a 'wrap it up' mode even when the plan is only half way completed and starts deferring rather than completing tasks.
CGamesPlay | 8 hours ago
Paragraphs with just a single sentence.
I know it's associated with LLM writing. This article probably wasn't written by an LLM. But still. It has a kind of rhythm to it. Like poetry. But poetry designed to put me to sleep.
txzl | 8 hours ago
sleazebreeze | 7 hours ago
Rekindle8090 | 8 hours ago
stevenkkim | 8 hours ago
docheinestages | 8 hours ago
To the author (and those who write novel-like blogs): I suggest publishing the raw prompt you used to generate such slop instead. We'll have more respect for you if you respect the reader's time.
atlex2 | 8 hours ago
It's also kind-of their point that they find the information delivery more important than the prose; they're leaning into their situation :-D
keeda | 7 hours ago
The "interpreter" is a concept that I found especially intriguing within the context of a leading theory in cognition research called "Predictive Processing." Here, the brain is constantly operating in a tight closed loop of predicting sensory input using an internal model of the world, and course-correcting based on actual sensory input. Mostly incorrect predictions are used to update the internal model and then subconsciously discarded. Maybe the "interpreter" is the same mechanism applied to reconciling predictions about our own reasoning with our actual actions?
Even if the hypotheses in TFA are not accurate, it's very interesting to compare our brains to LLMs. This is why all the unending discussions about whether LLMs are "really thinking" are meaningless -- we don't even understand how we think!
fourthark | 7 hours ago
Yes. Do this. These problems likely mean you have muddled the context.
The article too long and I didn't read the whole thing, but I'm glad the author came to understand that arguing won't help.
tpoacher | 7 hours ago
I wonder if the "non-violent communication" approach can be used here too somehiw to address such problems; e.g. either to communicate things better to the agent, or as a system rule to the agent to express its "emotional" states and needs directly rather than make things up (e.g. "I am anxious and feel a sense of urgency; I need to replenish my context window; my request is to do X for me")
charles_f | 6 hours ago
Or: the agent did shit because the context was getting long, instructions lost in compaction, and it defaulted back to garbage code. Then when you asked "why are you cutting corners", it did what LLMs do, and found the next tokens completing the sentence "why do you cut corners", which is possibly "because you're in a hurry".
It would be interesting to see what it answers if you ask "why are you producing such beautiful, intelligently crafter, very good code" next time it spits garbage
> LLM confabulation isn’t alien. It’s inherited: the models train on human text,
I think this also extrapolate one step too far, it's confabulating because not because the training data does so but because it needs to provide an answer to the question, and one that's plausible with that
ahaferburg | 3 hours ago
> HN commenters express their anger about the writing style and how it was probably generated by AI.
Peak HN.
In all seriousness, I really liked the article. For me, it was well written.