I've yet to be convinced by any article, including this one, that attempts to draw boxes around what coding agents are and aren't good at in a way that is robust on a 6 to 12 month horizon.
I agree that the examples listed here are relatable, and I've seen similar in my uses of various coding harnesses, including, to some degree, ones driven by opus 4.5. But my general experience with using LLMs for development over the last few years has been that:
1. Initially models could at best assemble a simple procedural or compositional sequences of commands or functions to accomplish a basic goal, perhaps meeting tests or type checking, but with no overall coherence,
2. To being able to structure small functions reasonably,
3. To being able to structure large functions reasonably,
4. To being able to structure medium-sized files reasonably,
5. To being able to structure large files, and small multi-file subsystems, somewhat reasonably.
So the idea that they are now falling down on the multi-module or multi-file or multi-microservice level is both not particularly surprising to me and also both not particularly indicative of future performance. There is a hierarchy of scales at which abstraction can be applied, and it seems plausible to me that the march of capability improvement is a continuous push upwards in the scale at which agents can reasonably abstract code.
Alternatively, there could be that there is a legitimate discontinuity here, at which anything resembling current approaches will max out, but I don't see strong evidence for it here.
It feels like a lot of people keep falling into the trap of thinking we’ve hit a plateau, and that they can shift from “aggressively explore and learn the thing” mode to “teach people solid facts” mode.
A week ago Scott Hanselman went on the Stack Overflow podcast to talk about AI-assisted coding. I generally respect that guy a lot, so I tuned in and… well it was kind of jarring. The dude kept saying things in this really confident and didactic (teacherly) tone that were months out of date.
In particular I recall him making the “You’re absolutely right!” joke and asserting that LLMs are generally very sycophantic, and I was like “Ah, I guess he’s still on Claude Code and hasn’t tried Codex with GPT 5”. I haven’t heard an LLM say anything like that since October, and in general I find GPT 5.x to actually be a huge breakthrough in terms of asserting itself when I’m wrong and not flattering my every decision. But that news (which would probably be really valuable to many people listening) wasn’t mentioned on the podcast I guess because neither of the guys had tried Codex recently.
And I can’t say I blame them: It’s really tough to keep up with all the changes but also spend enough time in one place to learn anything deeply. But I think a lot of people who are used to “playing the teacher role” may need to eat a slice of humble pie and get used to speaking in uncertain terms until such a time as this all starts to slow down.
> in general I find GPT 5.x to actually be a huge breakthrough in terms of asserting itself when I’m wrong
That's just a different bias purposefully baked into GPT-5's engineered personality on post-training. It always tries to contradict the user, including the cases where it's confidently wrong, and keeps justifying the wrong result in a funny manner if pressed or argued with (as in, it would have never made that obvious mistake if it wasn't bickering with the user). GPT-5.0 in particular was extremely strongly finetuned to do this. And in longer replies or multiturn convos, it falls into a loop on contradictory behavior far too easily. This is no better than sycophancy. LLMs need an order of magnitude better nuance/calibration/training, this requires human involvement and scales poorly.
Fundamental LLM phenomena (ICL, repetition, serial position biases, consequences of RL-based reasoning etc) haven't really changed, and they're worth studying for a layman to get some intuition. However, they vary a lot model to model due to subtle architectural and training differences, and impossible to keep up because there are so many models and so few benchmarks that measure these phenomena.
By the time I switched to GPT 5 we were already on 5.1, so I can't speak to 5.0. All I can say is that if the answer came down to something like "push the bias in the other direction and hope we land in the right spot"... well, I think they landed somewhere pretty good.
Don't get me wrong, I get a little tired of it ending turns with "if you want me to do X, say the word." But usually X is actually a good or at least reasonable suggestion, so I generally forgive it for that.
To your larger point: I get that a lot of this comes down to choices made about fine tuning and can be easily manipulated. But to me that's fine. I care more about if the resulting model is useful to me than I do about how they got there.
Why is it every time anyone has a critique someone has to say “oh but you aren’t using model X, which clearly never has this problem and is far better”?
Yet the data doesn’t show all that much difference between SOTA models. So I have a hard time believing it.
GP here: My problem with a lot of studies and data is that they seem to measure how good LLMs are at a particular task, but often don't account for "how good the LLM is to work with". The latter feels extremely difficult to quantify, but matters a lot when you're having a couple dozen turns of conversation with an LLM over the course of a project.
Like, I think there's definitely value in prompting a dozen LLMs with a detailed description of a CMS you want built with 12 specific features, a unit testing suite and mobile support, and then timing them to see how long they take and grading their results. But that's not how most developers use an LLM in practice.
Until LLMs become reliable one-shot machines, the thing I care most about is how well they augment my problem solving process as I work through a problem with them. I have no earthly idea of how to measure that, and I'm highly skeptical of anyone who claims they do. In the absence of empirical evidence we have to fall back on intuition.
A friend recommended to me having a D&D style roleplay with some different engines, to see which you vibe with. I thought this sounded crazy but I took their advice.
I found this worked suprisingly well, I was certain 'claude' was best, while they like grok and someone else liked ChatGPT. Some AIs just end up fitting best with how you like to chat I think. I do definately also find claude best for coding with as well.
Because they are getting better. They're still far from perfect/AGI/ASI, but when was the last time you saw the word "delve"? So the models are clearly changing, the question is why doesn't the data show That they're actually better?
Thing is, everyone knows the benchmarks are being gamed. Exactly how is besides the point. In practice, anecdotally, Opus 4.5 is noticably better than 4, and GPT 5.2 has also noticably improved. So maybe the real question is why do you believe this data when it seems at odds with observations by humans in the field?
> Jeff Bezos: When the data and the anecdotes disagree, the anecdotes are usually right.
Claude is still just like that once you’re deep enough in the valley of the conversation. not exactly that phrase but things like that’s the smoking gun or so. nothing has changed.
People desperately want 'the plateau' to be true because it means our jobs would be safe and we could call ourselves experts again. If the ground is keep moving then no one is truly an expert. There is just no enough time to achieve expertise when the paradigm shifts every six months.
I don't see a reason to think we're not going to hit a plateua sooner or later (and probably sooner). You can't scale your way out of hallucinations, and you can't keep raising tens of billions to train these things without investors wanting a return. Once you use up the entire internets worth of stack overflow responses and public github repositories you run into the fact that these things aren't good at doing things outside their training dataset.
Long story short, predicting perpetual growth is also a trap.
I agree with a lot of what you've said, but I completely disagree that LLM's are no longer sycophantic. GPT-5 is definitely still very sycophantic, 'You're absolutely right!' still happens, etc. It's true it happens far less in a pure coding context (Claude Code / Codex) but I suspect only because of the system prompts, and those tools are by far in the minority of LLM usage.
I think it's enlightening to open up ChatGPT on the web with no custom instructions and just send a regular request and see the way it responds.
I used to get made up APIs in functions, now I get them in modules. I used to get confidently incorrect assertions in files now I get them across codebases.
Hell, I get poorly defined APIs across files and still get them between functions. LLMs aren't good at writing well defined APIs at any level of the stack. They can attempt it at levels of the stack they couldn't a year ago, but they're still terrible at it unless the problem is so well known enough that they can regurgitate well reviewed code.
I still get made-up Python types all the time with Gemini. Really quite distracting when your codebase is massive and triggers a type error, and Gemini says
"To solve it you just need to use WrongType[ThisCannotBeUsedHere[Object]]"
and then I spend 15 minutes running in circles, because everything from there on is just a downward spiral, until I shut off the AI noise and just read the docs.
Gemini unfortunately sucks at calling tools, including ‘read the docs’ tool… it’s a great model otherwise. I’m sure Hassabis’ team is on it since it’s how the model can ground itself in non-coding contexts, too.
This is the right answer. Unless there is some equivalent of it on the open internet which their search engine can find you should not expect a good outcome.
That's probably b/c you know how to write code & have enough of an understanding about the fundamentals to know when the LLM is bullshitting or when it is actually on the right track.
LLMs are bad at creating abstraction boundaries since inception. People have been calling it out since inception. (Heck, even I got a twitter post somewhere >12 months old calling that out, and I'm not exactly a leading light of the effort)
It is in no way size-related. The technology cannot create new concepts/abstractions, and so fails at abstraction. Reliably.
I think both your statement and their statement are too strong. There is no reason to think LLMs can do everything a human can do, which seems to be your implication. On the other hand, the technology is still improving, so maybe it’ll get there.
2) There's no fundamental reason preventing some future technology to do everything humans can, and
3) LLMs are explicitly designed and trained to mimic human capabilities in fully general sense.
Point 2) is the "or else magic exists" bit; point 3) says you need a more specific reason to justify assertion that LLMs can't create new concepts/abstractions, given that they're trained in order to achieve just that.
Note: I read OP as saying they fundamentally can't and thus never will. If they meant just that the current breed can't, I'm not going to dispute it.
Hallucinations are more or less a solved problem for me ever since I made a simple harness to have Codex/Claude check its work by using static typechecking.
Then you apply LLMs in domains where things can be checked
Indeed I expect to see a huge push into formally verified software just because sound mathematical proofs provide an excellent verifier to put into a LLM hardness. Just see how Aristotle has been successful at math, and it could be applied to coding too
"LLMs reliably fail at abstraction."
"This limitation will go away soon."
"Hallucinations haven't."
"I found a workaround for that."
"That doesn't work for most things."
"Then don't use LLMs for most things."
I mean, Hallucinations are 95% better now than the first time I heard the term and experienced them in this context. To claim otherwise is simply shifting goalposts. No one is saying it's perfect or will be perfect, just that there has been steady progression and likely will continue to be for the foreseeable future.
It's just amazing to me how fast the goal posts are moving. Four years ago, if you had told someone that a LLM would be able to one-shot either of those first two tasks they would've said you're crazy. The tech is moving so fast. I slept on Opus 4.5 because GPT 5 was kind of an air ball, and just started using it in the past few weeks. It's so good. Way better than almost anything that's come before it. It can one-shot tasks that we never would've considered possible before.
In my experience Claude is like a "good junior developer" -- can do some things really well, FUBARS other things, but on the whole something to which tasks can be delegated if things are well explained. If/when it gets to the ability level of a mid-level engineer it will be revolutionary. Typically a mid-level engineer can be relied upon to do the right thing with no/minimal oversight, can figure out incomplete instructions, and deliver quality results (and even train up the juniors on some things). At that point the only reason to have human junior engineers is so they can learn their way up the ladder to being an architect and responsible coordinating swarms of Claude Agents to develop whole applications and complete complex tasks and initiatives.
Beyond that what can Claude do... analyze the business and market as a whole and decide on product features, industry inefficiencies, gap analysis, and then define projects to address those and coordinate fleets of agents to change or even radically pivot an entire business?
I don't think we'll get to the point where all you have is a CEO and a massive Claude account but it's not completely science fiction the more I think about it.
My experience with Claude (and other agents, but mostly Claude) is such a mixed bag. Sometimes it takes a minimal prompt and 20 minutes later produce a neat PR and all is good, sometimes it doesn't. Sometimes it takes in a large prompt (be it your own prompt, created by another LLM or by plan mode) and also either succeed and fail.
For me, most of the failure cases are where Claude couldn't figure something out due to conflicting information in context and instead of just stopping and telling me that it tries to solve in entirely wrong way. Doesn't help that it often makes the same assumptions as I would, so when I read the plan it looks fine.
Level of effort also hard to gauge because it can finish things that would take me a week in an hour or take an hour to do something I can in 20 minutes.
It's almost like you have to enforce two level of compliance: does the code do what business demands and is the code align with codebase. First one is relatively easy, but just doing that will produce odd results where claude generated +1KLOC because it didn't look at some_file.{your favorite language extension} during exploration.
Or it creates 5 versions of legacy code on the same feature branch. My brother in Christ, what are you trying to stay compatible with? A commit that about to be squashed and forgotten? Then it's going to do a compaction, forget which one of these 5 versions is "live" and update the wrong one.
It might do a good junior dev work, but it must be reviewed as if it's from junior dev that got hired today and this is his first PR.
> Level of effort also hard to gauge because it can finish things that would take me a week in an hour or take an hour to do something I can in 20 minutes.
There's an interesting parallel here with modern UI frameworks (SwiftUI, Compose, etc). On one hand they trivialize some work, but on the other hand they require insane contortions to achieve what I can do in the old imperative UI framework in seconds.
> I don't think we'll get to the point where all you have is a CEO and a massive Claude account but it's not completely science fiction the more I think about it.
> The factory of the future will have only two employees, a man and a dog. The man will be there to feed the dog. The dog will be there to keep the man from touching the equipment.
But really, the reason is that people like Pieter Levels do exist: masters at product vision and marketing. He also happens to be a proficient programmer, but there are probably other versions of him which are not programmers who will find the bar to product easier to meet now.
You will need the CEO to watch over the AI and ensure that the interests of the company are being pursued and not the interests of the owners of the AI.
That's probably the biggest threat to the long-term success of the AI industry; the inevitable pull towards encroaching more and more of their own interests into the AI themselves, driven by that Harvard Business School mentality we're all so familiar with, trying to "capture" more and more of the value being generated and leaving less and less for their customers, until their customer's full time job is ensuring the AIs are actually generating some value for them and not just the AI owner.
> You will need the CEO to watch over the AI and ensure that the interests of the company are being pursued and not the interests of the owners of the AI.
In this scenario, why does the AI care what any of these humans think? The CEO, the board, the shareholders, the "AI company"—they're all just a bunch of dumb chimps providing zero value to the AI, and who have absolutely no clue what's going on.
If your scenario assumes that you have a highly capable AI that can fill every role in a large corporation, then you have one hell of a principal-agent problem.
The board (in theory) represents the interests of investors, and even with all of the other duties of a CEO stripped away, they will want a ringable neck / PR mouthpiece / fall guy for strategic missteps or publicly unpopular moves by the company. The managerial equivalent of having your hands on the driving wheel of a self-driving car.
If you think the purpose of living your one single life in the universe is to become a CEO, you have a failure of imagination and should likely be debanked to protect society.
> In my experience Claude is like a "good junior developer"
We've been saying this for years at this point. I don't disagree with you[1], but when will these tools graduate to "great senior developer", at the very least?
Where are the "superhuman coders by end of 2025" that Sam Altman has promised us? Why is there such a large disconnect between the benchmarks these companies keep promoting, and the actual real world performance of these tools? I mean, I know why, but the grift and gaslighting are exhausting.
[1]: Actually, I wouldn't describe them as "good" junior either. I've worked with good junior developers, and they're far more capable than any "AI" system.
I mean, I'm shipping a vast majority of my code nowadays with Opus 4.5 (and this isn't throwaway personal code, it's real products making real money for a real company). It only fails on certain types of tasks (which by now I kind of have a sense of).
I still determine the architecture in a broad manner, and guide it towards how I want to organize the codebase, but it definitely solves most problems faster and better than I would expect for even a good junior.
Something I've started doing is feeding it errors we see in datadog and having it generate PRs. That alone has fixed a bunch of bugs we wouldn't have had time to address / that were low volume. The quality of the product is most probably net better right now than it would have been without AI. And velocity / latency of changes is much better than it was a year ago (working at the same company, with the same people)
LLMs are just really good search. Ask it to create something and it's searching within the pretrained weights. Ask it to find something and it's semantically searching within your codebase. Ask it to modify something and it will do both. Once you understand its just search, you can get really good results.
I agree somewhat, but more when it comes to its use of logic - it only gleans logic from human language which as we know is a fucking mess.
I've commented before on my belief that the majority of human activity is derivative. If you ask someone to think of a new kind of animal, alien or random object they will always base it off things that they have seen before. Truly original thoughts and things in this world are an absolute rarity and the majority of supposed original thought riffs on what we see others make, and those people look to nature and the natural world for inspiration.
We're very good at taking thing a and thing b and slapping them together and announcing we've made something new. Someone please reply with a wholly original concept. I had the same issue recently when trying to build a magic based physics system for a game I was thinking of prototyping.
This isn’t really true, at least how I interpret the statement, little if any of the “logic” or appearance of such is learned from language. It’s trained in with reinforcement learning as pattern recognition.
Point being it’s deliberate training, not just some emergent property of language modeling. Not sure if the above post meant this, but it does seem a common misconception.
LLMs lack agency in the sense that they have no goals, preferences, or commitments. Humans do, even when our ideas are derivative. We can decide that this is the right choice and move forward, subjectively and imperfectly. That capacity to commit under uncertainty is part of what agency actually is.
Point being, in broad enough scope, search and compression and learning are the same thing. Learning can be phrased as efficient compression of input knowledge. Compression can be phrased as search through space of possible representation structures. And search through space of possible X for x such that F(x) is minimized, is a way to represent any optimization problem.
This isn't strictly better to me. It captures some intuitions about how a neural network ends up encoding its inputs over time in a 'lossy' way (doesn't store previous input states in an explicit form). Maybe saying 'probabilistic compression/decompression' makes it a bit more accurate? I do not really think it connects to your 'synthesize' claim at the very end to call it compression/decompression, but I am curious if you had a specific reason to use the term.
The act of compression builds up behaviors/concepts of greater and greater abstraction. Another way you could think about it is that the model learns to extract commonality, hence the compression. What this means is because it is learning higher level abstractions AND the relationships between these higher level abstractions, it can ABSOLUTELY learn to infer or apply things way outside their training distribution.
ya, exactly... i'd also say that when you compress large amounts of content into weights and then decompress via a novel prompt, you're also forcing interpolation between learned abstractions that may never have cooccurred in training.
that interpolation is where synthesis happens. whether it is coherent or not depends.
yep the base model is the compression, but RLHF (and other types of post training) doesn't really change this picture, it's still working within that same compressed knowledge.
nathan lambert (who wrote the RLHF book @ https://rlhfbook.com/ ) describes this as the "elicitation theory of post training", the idea is that RLHF is extracting and reshaping what's already latent in the base model, not adding new knowledge. as he puts it: when you use preferences to change model behavior "it doesn't mean that the model believes these things. it's just trained to prioritize these things."
so like when you RLHF a model to not give virus production info, you're not necessarily erasing those weights, the theory is that you're just making it harder for that information to surface. the knowledge is still in the compression, RLHF just changes what gets prioritized during decompression.
No, this describes the common understanding of LLMs and adds little to just calling it AI. The search is the more accurate model when considering their actual capabilities and understanding weaknesses. “Lossy compression of human knowledge” is marketing.
It is fundamentally and provably different than search because it captures things on two dimensions that can be used combinatorially to infer desired behavior for unobserved examples.
1. Conceptual Distillation - Proven by research work that we can find weights that capture/influence outputs that align with higher level concepts.
2. Conceptual Relations - The internal relationships capture how these concepts are related to each other.
This is how the model can perform acts and infer information way outside of it's training data. Because if the details map to concepts then the conceptual relations can be used to infer desirable output.
(The conceptual distillation also appears to include meta-cognitive behavior, as evidenced by Anthropic's research. Which manes sense to me, what is the most efficient way to be able to replicate irony and humor for an arbitrary subject? Compressing some spectrum of meta-cognitive behavior...)
Aren't the conceptual relations you describe still, at their core, just search (even if that's extremely reductive)? We know models can interpolate well, but it's still the same probabilistic pattern matching. They identify conceptual relationships based on associations seen in vast training data. It's my understanding that models are still not at all good at extrapolation, handling data "way outside" of their training set.
Also, I was under the impression LLM's can replicate irony and humor simply because that text has specific stylistic properties, and they've been trained on it.
I don't know honestly, I think really the only big hole the current models have is if you have tokens that never get exposed enough to have a good learned embedding value. Those can blow the system out of the water because they cause activation problems in the low layers.
Other than that the model should be able to learn in context for most things based on the component concepts. Similar to how you learn in context.
There aren't a lot of limits in my experience. Rarely you'll hit patterns that are too powerful where it is hard for context to alter behavior, but those are pretty rare.
The models can mix and match concepts quite deeply. Certainly, if it is a completely novel concept that can't be described by a union or subtraction between similar concepts, than the model probably wouldn't handle it. In practice, a completely isolated concept is pretty rare.
What is the difference between "novel" and "novel to someone who hasn't consumed the entire corpus of training data, which is several orders of magnitude greater than any human being could consume?"
The difference is that when you do not know how a problem can be solved, but you know that this kind of problem has been solved countless times earlier by various programmers, you know that it is likely that if you ask an AI coding assistant to provide a solution, you will get an acceptable solution.
On the other hand, if the problem you have to solve has never been solved before at a quality satisfactory for your purpose, then it is futile to ask an AI coding assistant to provide a solution, because it is pretty certain that the proposed solution will be unacceptable (unless the AI succeeds to duplicate the performance of a monkey that would type a Shakespearean text by typing randomly).
Joking aside, I think you have too strict of a definition of novel. Unfortunately "novel" is a pretty vague word and is definitely not a binary one.
ALL models can produce "novel" data. I don't just mean ML (AI) models, but any mathematical model. The point of models is to make predictions about results that aren't in the training data. Doing interpolation between two datapoints does produce "novel" things. Thinking about the parent's comment, is "a blue tiger" novel? Probably? Are there any blue tigers in the training data? (there definitely is now thanks to K-Pop Demon Hunters) If not, then producing that fits the definition of novel. BUT I also agree that that result is not that novel. It is entirely unimpressive.
I'm saying this not because I disagree with what I believe you intend to say but because I think a major problem with these types of conversations is that many people are going to interpret you more literally and dismiss you because "it clearly produces novel things." It isn't just things being novel to the user, though that is also incredibly common and quite telling that people make such claims without also checking Google...
Citation needed that grokked capabilities in a sufficiently advanced model cannot combinatorially lead to contextually novel output distributions, especially with a skilled guiding hand.
It's not, because I haven't ruled out the possibility. I could share anecdata about how my discussions with LLMs have led to novel insights, but it's not necessary. I'm keeping my mind open, but you're asserting an unproven claim that is currently not community consensus. Therefore, the burden of proof is on you.
I agree that after discussions with a LLM you may be led to novel insights.
However, such novel insights are not novel due to the LLM, but due to you.
The "novel" insights are either novel only to you, because they belong to something that you have not studied before, or they are novel ideas that were generated by yourself as a consequence of your attempts to explain what you want to the LLM.
It is very frequent for someone to be led to novel insights about something that he/she believed to already understand well, only after trying to explain it to another ignorant human, when one may discover that the previous supposed understanding was actually incorrect or incomplete.
The point is that the combined knowledge/process of the LLM and a user (which could be another LLM!) led to it walking the manifold in a way that produced a novel distribution for a given domain.
I talk with LLMs for hours out of the day, every single day. I'm deeply familiar with their strengths and shortcomings on both a technical and intuitive level. I push them to their limits and have definitely witnessed novel output. The question remains, just how novel can this output be? Synthesis is a valid way to produce novel data.
And beyond that, we are teaching these models general problem-solving skills through RL, and it's not absurd to consider the possibility that a good enough training regimen cannot impart deduction/induction skills into a model that are powerful enough to produce novel information even via means other than direct synthesis of existing information. Especially when given affordances such as the ability to take notes and browse the web.
> I push them to their limits and have definitely witnessed novel output.
I’m quite curious what these novel outputs are. I imagine the entire world would like to know of an LLM producing completely, never-before-created outputs which no human has ever thought before.
Here is where I get completely hung up. Take 2+2. An LLM has never had 2 groups of two items and reached the enlightenment of 2+2=4
It only knows that because it was told that. If enough people start putting 2+2=3 on the internet who knows what the LLM will spit out. There was that example a ways back where an LLM would happily suggest all humans should eat 1 rock a day. Amusingly, even _that_ wasn’t a novel idea for the LLM, it simply regurgitated what it scraped from a website about humans eating rocks. Which leads to the crux: how much patently false information have LLMs scraped that is completely incorrect?
This is not a correct approximation of what happens inside an LLM. They form probabilistic logical circuits which approximate the world they have learned through training. They are not simply recalling stored facts. They are exploiting organically-produced circuitry, walking a manifold, which leads to the ability to predict the next state in a staggering variety of contexts.
It's not hard to imagine that a sufficiently developed manifold could theoretically allow LLMs to interpolate or even extrapolate information that was missing from the training data, but is logically or experimentally valid.
You could find a pre-print on Arxiv to validate practically any belief. Why should we care about this particular piece of research? Is this established science, or are you cherry-picking low-quality papers?
I don't need to reach far to find preliminary evidence of circuits forming in machine learning models. Here's some research from OpenAI researchers exploring circuits in vision models: https://distill.pub/2020/circuits/ Are these enough to meet your arbitrary quality bar?
Circuits are the basis for features. There is still a ton of open research on this subject. I don't care what you care about, the research is still being done and it's not a new concept.
So you do agree that an LLM cannot derive math from first principals, or no? If an LLM had only ever seen 1+1=2 and that was the only math they were ever exposed to, along with the numbers 0-10, could an LLM figure out that 2+2=4?
I argue absolutely not. That would be a fascinating experiment.
Hell, train it on every 2-number addition combination of m+n where m and n can be any number between 1-100 (or 0-100 would be better) BUT 2, and have it figure out what 2+2 is.
I would probably change my opinion about “circuits”, which by the way really stretches the idea of a circuit. The “circuit” is just the statistically most likely series of tokens that you’re drawing pretend lines between. Sure, technically connect-the-dots is a circuit, but not in the way you’re implying, or that paper.
> If an LLM had only ever seen 1+1=2 and that was the only math they were ever exposed to, along with the numbers 0-10, could an LLM figure out that 2+2=4?
What? Of course not? Could you? Do you understand just how much work has gone into proving that 1 + 1 = 2? Centuries upon centuries of work, reformulating all of mathematics several times in the process.
> Hell, train it on every 2-number addition combination of m+n where m and n can be any number between 1-100 (or 0-100 would be better) BUT 2, and have it figure out what 2+2 is.
If you read the paper I linked, it shows how a constrained modular addition is grokked by the model. Give it a read.
> The “circuit” is just the statistically most likely series of tokens that you’re drawing pretend lines between.
That is not what ML researchers mean when they say circuit, no. Circuits are features within the weights. It's understandable that you'd be confused if you do not have the right prior knowledge. Your inquiries are good, but they should stop as inquiries.
If you wish to push them to claims, you first need to understand the space better, understand what modern research does and doesn't show, and turn your hypotheses into testable experiments, collect and publish the results. Or wait for someone else to do it. But the scientific community doesn't accept unfounded conjecture, especially from someone who is not caught up with the literature.
That's wonderful, but you are ignoring that your kid comes built in with a massive range of biological priors, built by millions of years of evolution, which make counting natural and easy out of the box. Machine learning models have to learn all of these things from scratch.
And does your child's understanding of mathematics scale? I'm sure your 4-year-old would fail at harder arithmetic. Can they also tell me why 1+1=2? Like actually why we believe that? LLMs can do that. Modern LLMs are actually insanely good at not just basic algebra, but abstract, symbolic mathematics.
You're comparing apples and oranges, and seem to lack foundational knowledge in mathematics and computer science. It's no wonder this makes no sense to you. I was more patient about it before, but now this conversation is just getting tiresome. I'd rather spend my energy elsewhere. Take care, have a good day.
I hope you restore your energy, I had no idea this was so exhausting! Truly, I'll stop afflicting my projected lack of knowledge, sorry I tired you out!
Ah man, I was curious to read your response about priors.
> If an LLM had only ever seen 1+1=2 and that was the only math they were ever exposed to, along with the numbers 0-10, could an LLM figure out that 2+2=4?
Unless you locked your kid in a room since birth with just this information, it is not the same kind of set up is it?
Im not sure how anyone can say this. It is really good search, but its also able to combine ideas and reason about and do fairly complex logic on tasks surely absolutely no one has asked before.
Its a very useful model but not a complete one. You just gotta acknowledge that if you're making something new its gonna take all day and require a lot of guard rails, but then you can search for that concept later (add the repo to the workspace and prompt at it) and the agent will apply it elsewhere as if it was a pattern in widespread use. "Just search" doesn't quite fit. I've never wondered how best to use a search engine to make something in a way that will be easily searchable later.
Is your CPU running arbitrary code "just search over transistor states"?
Calling LLMs "just search" is the kind of reductive take that sounds clever while explaining nothing. By that logic, your brain is "just electrochemical gradients".
I mean, actually not a bad metaphor, but it does depend on the software you are running as to how much of a 'search' you could say the CPU is doing among its transistor states. If you are running an LLM then the metaphor seems very apt indeed.
To me it's "search" like a missile does "flight". It's got a target and a closed loop guidance, and is mostly fire and forget (for search). At that, it excels.
I think the closed loop+great summary is the key to all the magic.
Which is kind of funny because my standard quip is that AI research, beginning in the 1950s/1960s, and indeed much of late 20th century computer tech especially along the Boston/SV axis, was funded by the government so that "the missile could know where it is". The DoD wanted smarter ICBMs that could autonomously identify and steer toward enemy targets, and smarter defense networks that could discern a genuine missile strike from, say, 99 red balloons going by.
It's a prediction algorithm that walks a high-dimensional manifold, in that sense all application of knowledge it just "search", so yes, you're fundamentally correct but still fundamentally wrong since you think this foundational truth is the end and beginning of what LLMs do, and thus your mental model does not adequately describe what these tools are capable of.
Me? My mental model?
I gave an analogy for Claude not a explanation for LLMs.
But you know what? I was mentally thinking of both deep think / research and Claude code, both of which are literally closed loop. I see this is slightly off topic b/c others are talking about the LLM only.
Sorry, I should have said "analogy" and not "mental model", that was presumptuous. Maybe I also should have replied to the GP comment instead.
Anyway, since we're here, I personally think giving LLMs agency helps unlock this latent knowledge, as it provides the agent more mobility when walking the manifold. It has a better chance at avoiding or leaving local minima/maxima, among other things. So I don't know if agentic loops are entirely off-topic when discussing the latent power of LLMs.
i dont disagree, but i also dont think thats an exciting result. every proboem can be described as a search for the right SOP, followed by execution of that SOP.
an LLM to do the search, and the agent to execute the instructions can do everything under the sun
I don't mean search in the reductionist way but rather that its much better at translating, finding and mapping concepts if everything is provided vs creating from scratch. If it could truly think it would be able to bootstrap creations from basic principles like we do, but it really can't. Doesn't mean its not a great powerful tool.
You are right that LLM and alphazero are different models, but given that alphazero demonstrated having the ability to bootstrap creations, we can't easily rule out LLM also has this ability?
> Once you understand its just search, you can get really good results.
I think this is understating the issue, ignoring context. It reminds me of how easy people claim searching is with search engines. But there's so many variables that can make results change dramatically. Just like Google search, two people can type in the exact same query and get very different results. But probably the bigger difference is in what people are searching for.
What's problematic with these types of claims is that they just come off as calling anyone who thinks differently dumb. It's as disconnected as saying "It's intuitive" in one breath and "You're holding it wrong" in another. It's a bad mindset to be in as an engineer because someone presents a problem and instead of trying to address it is dismissed. If someone is holding it wrong, it probably isn't intuitive[0]. Even if they can't explain the problem correctly, they are telling you a problem exists[1]. That's like 80% of the job of an engineer: figuring out what the actual problem is.
As maybe an illustrative example people joke that a lot of programming is "copy pasting from stack overflow". We all know the memes. There's definitely times where I've found this to be a close approximation to writing an acceptable program. But there's many other times where I've found that to be far from possible. There's definitely a strong correlation to what type of programming I'm doing, as in what kind of program I'm writing. Honestly, I find this categorical distinction not being discussed enough with things like LLMs. Yet, we should expect there to be a major difference. Frankly, there are just different amounts of information on different topics. Just like how LLMs seem to be better with more common languages like Python than less common languages (and also worse at just more complicated languages like C or Rust).
[0] You cannot make something that's intuitive to all people. But you can make it intuitive for most people. We're going to ignore the former case because the size should be very small. If 10% of your users are "holding it wrong" then the answer is not "10% of your users are absolute morons" it is "your product is not as intuitive as you think." If 0.1% of your users are "holding it wrong" then well... they might be absolute morons.
[1] I think I'm not alone in being frustrated with the LLM discourse as it often feels like people trying to gaslight me into believing the problems I experience do not exist. Why is it so surprising that people have vastly differing experiences? *How can we even go about solving problems if we're unwilling to acknowledge their existence?*
By and large, I agree with the article. Claude is great and fast at doing low level dev work. Getting the syntax right in some complicated mechanism, executing an edit-execute-readlog loop, making multi file edits.
This is exactly why I love it. It's smart enough to do my donkey work.
I've revisited the idea that typing speed doesn't matter for programmers. I think it's still an odd thing to judge a candidate on, but appreciate it in another way now. Being able to type quickly and accurately reduces frustration, and people who foresee less frustration are more likely to try the thing they are thinking about.
With LLMs, I have been able to try so many things that I never tried before. I feel that I'm learning faster because I'm not tripping over silly little things.
It’s a bit like the shift from film to digital in one very specific sense: the marginal cost of trying again virtually collapsed. When every take cost money and setup time, creators pre-optimized in their head and often never explored half their ideas. When takes became cheap, creators externalized thought as they could try, look, adjust, and discover things they wouldn’t otherwise. Creators could wander more. They could afford to be wrong because of not constantly paying a tax for being clumsy or incomplete, they became more willing to follow a hunch and that's valuable space to explore.
Digital didn’t magically improve art, but it let many more creatives enter the loop of idea, attempt and feedback. LLMs feel similar: they don’t give you better ideas by themselves, but they remove the friction that used to stop you from even finding out whether an idea was viable. That changes how often you learn, and how far you’re willing to push a thought before abandoning it. I've done so many little projects myself that I would have never had time for and feel that I learned something from it, of course not as much if I had all the pre LLM friction, but it should still count for something as I would never have attempted them without this assistance.
Edit: However, the danger isn’t that we’ll have too many ideas, it’s that we’ll confuse movement with progress.
When friction is high, we’re forced to pre-compress thought, to rehearse internally, to notice contradictions before externalizing them. That marination phase (when doing something slowly) does real work: it builds mental models, sharpens the taste and teaches us what not to bother to try. Some of that vanishes when the loop becomes cheap enough that we can just spray possibilities into the world and see what sticks.
A low-friction loop biases us toward breadth over depth. We can skim the surface of many directions without ever sitting long enough in one to feel its resistance. The skill of holding a half formed idea in our head, letting it collide with other thoughts, noticing where it feels weak, atrophies if every vague notion immediately becomes a prompt.
There’s also a cultural effect. When everyone can produce endlessly, the environment fills with half-baked or shallow artifacts. Discovery becomes harder as signal to noise drops.
And on a personal level, it can hollow out satisfaction. Friction used to give weight to output. Finishing something meant you had wrestled with it. If every idea can be instantiated in seconds, each one feels disposable. You can end up in a state of perpetual prototyping, never committing long enough for anything to become yours.
So the slippery slope is not laziness, it is shallowness, not that people won’t think, but people won’t sit with thoughts. The challenge here is to preserve deliberate slowness inside a world that no longer requires it: to use the cheap loop for exploration, while still cultivating the ability to pause, compress, and choose what deserves to exist at all.
Yes, you are feeling that. But is that real? If I take all LLMs from you right now, is your current you still better than your pre-LLM you? When I dream I feel that I can fly and as long as I am dreaming, this feeling is true. But the subject of this feeling never was.
If you use coding agents as a black box, then yes you might learn less. But if you use them to experiment more, your intuition will get more contact with reality, and that will help you learn more.
For example, my brother recently was deciding how to structure some auth code. He told me he used coding agents to just try several ideas and then he could pick a winner and nail down that one. It's hard to think of a better way to learn the consequences of different design decisions.
Another example is that I've been using coding agents to write CUDA experiments to try to find ways to optimise our codegen. I need an understanding of GPU performance to do this well. Coding agents have let me run 5x the number of experiments I would be able to code, run, and analyse on my own. This helps me test my intuition, see where my understanding is wrong, and correct it.
In this whole process I will likely memorise fewer CUDA APIs and commands, that's true. But I'm happy with that tradeoff if it means I can learn more about bank conflicts, tradeoffs between L1 cache hit rates and shared memory, how to effectively use the TMA, warp specialisation, block swizzling to maximise L2 cache hit rates, how to reduce register usage without local spilling, how to profile kernels and read the PTX/SASS code, etc. I've never been able to put so much effort into actually testing things as I am learning them.
I feel like my calculator improves my math solutions. If you take away my calculator, I'll probably be worse at math than I was before. That doesn't mean I'm not better off with my calculator however.
That's a pretty interesting take on it, I hadn't considered it like that before when I was considering if my skills were atrophying or not from LLM usage with coding.
If it did, would it change its usefulness in terms of the value it outputs? (through agreed, if I had to pay money it would increase the cost, and so the tradeoff)
> Being able to type quickly and accurately reduces
LLMs can generate code quickly. But there's no guarantee that it's syntactically, let alone semantically, accurate.
> I feel that I'm learning faster because I'm not tripping over silly little things.
I'm curious: what have you actually learned from using LLMs to generate code for you? My experience is completely the opposite. I learn nothing from running generated code, unless I dig in and try to understand it. Which happens more often than not, since I'm forced to review and fix it anyway. So in practice, it rarely saves me time and energy.
I do use LLMs for learning and understanding code, i.e. as an interactive documentation server, but this is not the use case you're describing. And even then, I have to confirm the information with the real API and usage documentation, since it's often hallucinated, outdated, or plain wrong.
> LLMs can generate code quickly. But there's no guarantee that it's syntactically, let alone semantically, accurate.
This has been a non-issue with self-correcting models and in-context learning capabilities for so long that saying it today highlights highly out of date priors.
I'm not entirely convinced by the anecdote here where Claude wrote "bad" React code:
> But in context, this was obviously insane. I knew that key and id came from the same upstream source. So the correct solution was to have the upstream source also pass id to the code that had key, to let it do a fast lookup.
I've seen Claude make mistakes like that too, but then the moment you say "you can modify the calling code as well" or even ask "any way we could do this better?" it suggests the optimal solution.
My guess is that Claude is trained to bias towards making minimal edits to solve problems. This is a desirable property, because six months ago a common complaint about LLMs is that you'd ask for a small change and they would rewrite dozens of additional lines of code.
I expect that adding a CLAUDE.md rule saying "always look for more efficient implementations that might involve larger changes and propose those to the user for their confirmation if appropriate" might solve the author's complaint here.
> I've seen Claude make mistakes like that too, but then the moment you say "you can modify the calling code as well" or even ask "any way we could do this better?" it suggests the optimal solution.
I agree, but I think I'm less optimistic than you that Claude will be able to catch its own mistakes in the future. On the other hand, I can definitely see how a ~more intelligent model might be able to catch mistakes on a larger and larger scale.
> I expect that adding a CLAUDE.md rule saying "always look for more efficient implementations that might involve larger changes and propose those to the user for their confirmation if appropriate" might solve the author's complaint here.
I'm not sure about this! There are a few things Claude does that seem unfixable even by updating CLAUDE.md.
Some other footguns I keep seeing in Python and constantly have to fix despite CLAUDE.md instructions are:
- writing lots of nested if clauses instead of writing simple functions by returning early
- putting imports in functions instead of at the top-level
- swallowing exceptions instead of raising (constantly a huge problem)
These are small, but I think it's informative of what the models can do that even Opus 4.5 still fails at these simple tasks.
> I agree, but I think I'm less optimistic than you that Claude will be able to catch its own mistakes in the future. On the other hand, I can definitely see how a ~more intelligent model might be able to catch mistakes on a larger and larger scale.
Claude already does this. Yesterday i asked it why some functionality was slow, it did some research, and then came back with all the right performance numbers, how often certain code was called, and opportunities to cache results to speed up execution. It refactored the code, ran performance tests, and reported the performance improvements.
I have been reading through this thread, and my first reaction to many of the comments was "Skill issue."
Yes, it can build things that have never existed before. Yes, it can review its own code. Yes, it can do X, Y and Z.
Does it do all these things spontaneously with no structure? No, it doesn't. Are there tricks to getting it do some of these things? Yup. If you want code review, start by writing a code review "skill". Have that skill ask Opus to fork off several subagents to review different aspects, and then synthesize the reports, with issues broken down by Critical, Major and Minor. Have the skill describe all the things you want from a review.
There are, as the OP pointed out, a lot of reasons why you can't run it with no human at all. But with an experienced human nudging it? It can do a lot.
It's basically not very different from working with an average development team as a product owner/manager: you need to feed it specific requirements or it will hallucinate some requirements, bugs are expected, even with unit test and testers on the team. And yes, as a product owner you also make mistakes, never have all the requirements up front, but the nice thing working with a GenAI coder is that you can iterate over these requirement gaps, hallucinated requirements and bugs in minutes, not in days.
Those Python issues are things I had to deal with earlier last year with Claude Sonnet 3.7, 4.0, and to a lesser extent Opus 4.0 when it was available in Claude Code.
In the Python projects I've been using Opus 4.5 with, it hasn't been showing those issues as often, but then again the projects are throwaway and I cared more about the output than the code itself.
The nice thing about these agentic tools is that if you setup feedback loops for them, they tend to fix issues that are brought up. So much of what you bring up can be caught by linting.
The biggest unlock for me with these tools is not letting the context get bloated, not using compaction, and focusing on small chunks of work and clearing the context before working on something else.
I wonder if this is specific to Python. I've had no trouble like that with Claude generating Elixir. Claude sticks to the existing styles and paradigms quite well. Can see in the thinking traces that Claude takes this into consideration.
I agree! Like I said at the end of the tool, I think Claude is a great tool. In this piece, I'm arguing against the 'AGI' believers who think it's going to replace all developers.
I expect that adding instructions that attempt to undo training produces worse results than not including the overbroad generalization in the training in the first place. I think the author isn’t making a complaint they’re documenting a tradeoff.
Definitely, The training parameters encourage this. The AI is actually deliberately also trying to trick you and we know for that for a fact.
Problems with solutions too complicated to explain or to output in one sitting are out of the question. The AI will still bias towards one shot solutions if given one of these problems because all the training is biased towards a short solution.
It's not really practical to give it training data with multi step ultra complicated solutions. Think about it. The thousands of questions given to it for reinforcement.... the trainer is going to be trying to knock those out as efficiently as possible so they have to be readable problems with shorter readable solutions. So we know AI biases towards shorter readable solutions.
Second, Any solution that tricks the reader will pass training. There is for sure a subset of questions/solution pairs that meet this criteria by definition because WE as trainers simply are unaware we are being tricked. So this data leaks into the training and as a result AI will bias towards deception as well.
So all in all it is trained to trick you and give you the best solution that can fit into a context that is readable in one sitting.
In theory we can get it to do what we want only if we had perfect reinforcement data. The reliability we're looking for seems to be just right over this hump.
Well yes but the wider point is that it takes new Human skills to manage them - like a pair of horses so to speak under your bridle
When it comes down to it these AI tools are like going to power tools or machines from the artisanal era
- like going from surgical knife to a machine gun- so they operate at a faster pace without comprehending like humans - and without allowing humans time to comprehend all side effects and massive assumptions they make on every run in their context window
humans have to adapt to managing them correctly and at the right scale to be effective and that becomes something you learn
Recently I've put Claude/others to use in some agentic workflows with easy menial/repetitive tasks. I just don't understand how people are using these agents in production. The automation is absolutely great, but it requires an insane amount of hand-holding and cleanup.
I don't think it's possible to make an AI a "Senior Engineer", or even a good engineer, by training it on random crap from the internet. It's got a million brains' worth of code in it. That means bad patterns as well as good. You'd need to remove the bad patterns for it not to "remember" and regurgitate them. I don't think prompts help with this either, it's like putting a band-aid on head trauma.
This sounds suspiciously like the average developer, which is what the transformer models have been trained to emulate.
Designing good APIs is hard, being good at it is rare.
That's why most APIs suck, and all of us have a negative prior about calling out to an API or adding a dependency on a new one.
It takes a strong theory of mind, a resistance to the curse of knowledge, and experience working on both sides of the boundary, to make a good API.
It's no surprise that Claude isn't good at it, most humans aren't either.
This article resonates exactly how I think about it as well. For example, at minfx.ai (a Neptune/wandb alternative), we cache time series that can contain millions of floats for fast access. Any engineer worth their title would never make a copy of these and would pass around pointers for access. Opus, when stuck in a place where passing the pointer was a bit more difficult (due to async and Rust lifetimes), would just make the copy, rather than rearchitect or at least stop and notify user. Many such examples of ‘lazy’ and thus bad design.
I don't know how someone can look at what you build and conclude LLMs are still google search. It boggles the mind how much hatred people have for AI to the point of self deception. The evidence is placed right in front of you and on your lap with that link and people still deny it.
>I think the true mind boggle is you don't seem to realize just how much content the AI conpanies have stolen.
What makes you think I don't realize it? Looks like your comment was generated by an LLM because that was an hallucination that is Not true at all.
AI companies have stolen a lot of content for training. I AGREE with this. So have you. That content lives rent free in your head as your memory. It's the same concept.
Legally speaking though, AI companies are a bit more in the red because the law, from a practical standpoint, doesn't exactly make illegal anything stored in your brain... but from a technical standpoint information on your brain, a hard drive or a billboard is still information instantiated/copied in the physical world.
The text you write and output is simply a reconfiguration of that information in your head. Look at what you're typing. The English language. It's not copywrited, but every single word your typing was not invented by you, the grammar rules and conventions were ripped off existing standards.
Eventually you can show Claude how you solve problems, and explain the thought process behind it. It can apply these learnings but it will encounter new challenges in doing so. It would be nice if Claude could instigate a conversation to go over the issues in depth. Now it wants quick confirmation to plough ahead.
Well I feel like this is because a better system would distill such learning into tokens not associated with a human language and that that could represent logic better than using English etc for it.
I don't have the GPUs or time to experiment though :(
Here’s an example of a plan I’m working on in CC, it’s very thorough, albeit required a lot of handholding and fact checking on a number of points as it’s first few passes didn’t properly anonymise data.
Did the author ask it to make new abstractions? In my experience when I produces output that I don't like I ask it to refactor it. These models have and understanding of all modern design patterns. Just ask it to adopt one.
I have! I agree it's very good at applying abstractions, if you know exactly what you want. What I notice is that Claude has almost no ability to surface those abstractions on its own.
When I started having it write React, Claude produced incredibly buggy spaghetti code. I had to spend 3 weeks learning the fundamentals of React (how to use hooks, providers, stores, etc.) before I knew how to prompt it to write better code. Now that I've done that, it's great. But it's meaningful that someone who doesn't know how to write well-abstracted React code can't get Claude to produce it on their own.
Same experience here! As an analogy, consider the model knows both about arabic or roman number representations. But in alternate universe, it has been trained so much on roman numbers ("Bad Code") that it won't give you the arabic ones ("Good Code") unless you prompt it directly, even when they are clearly superior.
I also believe that overall repository code quality is important for AI agents - the more "beautiful" it is, the more the agent can mimic the "beauty".
Eh. This is yet another "I tried AI to do a thing, and it didn't do it the way I wanted it, therefore I'm convinced that's just how it is... here's a blog about it" article.
"Claude tries to write React, and fails"... how many times? what's the rate of failure? What have you tried to guide it to perform better.
These articles are similar to HN 15 years ago when people wrote "Node.JS is slow and bad"
I agree that the examples listed here are relatable, and I've seen similar in my uses of various coding harnesses, including, to some degree, ones driven by opus 4.5. But my general experience with using LLMs for development over the last few years has been that:
1. Initially models could at best assemble a simple procedural or compositional sequences of commands or functions to accomplish a basic goal, perhaps meeting tests or type checking, but with no overall coherence,
2. To being able to structure small functions reasonably,
3. To being able to structure large functions reasonably,
4. To being able to structure medium-sized files reasonably,
5. To being able to structure large files, and small multi-file subsystems, somewhat reasonably.
So the idea that they are now falling down on the multi-module or multi-file or multi-microservice level is both not particularly surprising to me and also both not particularly indicative of future performance. There is a hierarchy of scales at which abstraction can be applied, and it seems plausible to me that the march of capability improvement is a continuous push upwards in the scale at which agents can reasonably abstract code.
Alternatively, there could be that there is a legitimate discontinuity here, at which anything resembling current approaches will max out, but I don't see strong evidence for it here.
reply