[piker]

This is exactly right. By offloading this trivial task to the LLM, Simon has abandoned the opportunity to evaluate the abstraction with additional information and improve it. Instead, we let the agent spend $12 and make the fix while learning nothing.

[simonw]

Things I learned from this:

- Fable will do a whole lot more than you might expect in order to verify a fix. I learned that it's "relentlessly proactive". That's a good title for a blog entry!

- You can take screenshots of a window in macOS using the "screencapture" CLI command, but you'll need the integer window ID first.

- That windowID is accessible via "Quartz.CGWindowListCopyWindowInfo(Quartz.kCGWindowListOptionOnScreenOnly, Quartz.kCGNullWindowID)" using the pyobjc-framework-Quartz library, which installs cleanly via "uv run".

- A neat trick for simulating keyboard shortcuts is to run document.dispatchEvent(new KeyboardEvent("keydown", {key: "/", bubbles: true})); after the page loads.

- You don't need Flask or Starlette to run a CORS-enabled localhost server for capturing JSON from another window - 19 lines of code against the Python standard library http.server package works just fine.

- getComputedStyle(document.querySelector("navigation-search").shadowRoot.querySelector("textarea")) works to read dimensions from inside a Web Component's shadow DOM.

- defaults write com.google.chrome.for.testing AppleShowScrollBars Always

- Claude Fable knows how to apply all of the above. It's always interesting to pick up hints of what a model can and cannot do.

I'm always confused at how many people equate using a coding agent to solve a problem with "learning nothing". If you pay attention to what it's doing you can learn so much!

[piker]

Sorry that wasn't a criticism of you!

I completely see how it was misread that way. I would edit it now if I could.

I was using you more as an example of a hypothetical programmer using it in this way. If the goal is to create a maintainable product, this isn't a great approach. If the goal is to learn about the model and its behaviors itself, of course this is a fantastic way to experiment. Yes, you might have learned a lot of tricks as a side effect, but avoiding the pain of thinking about, finding and hiding the thing may mask a better abstraction that reduces complexity and allows the project to move forward faster.

[peterbell_nyc]

Honestly my goal is to learn how to teach an agent to build a maintainable product, so I'm way more interested in the learnings at the agentic level (how to prompt/direct/manage context/restrict tool use, provide reusable shims, etc) than getting into the details of a css bug. That's just not a level of abstraction with sufficient leverage for what I'm trying to do.

I stopped coding a while back because I could have more impact directing a team of developers than writing code personally.

For my use case, the agents are now how I can have that scaled impact.

[aspenmartin]

Absolutely. All of these "but you could have done that easily" from frontend developers or backend developers or systems engineers -- like yea, if I have the time or interest in those things, sure. But I don't. I care about an end product way way more. Blows my mind that there are legions of people building things that they don't think are important enough to get to the finish line quickly and efficiently.

[rco8786]

> If you pay attention to what it's doing you can learn so much!

I think your post is fair but it's worth pointing out that learning via watching is much less effective than learning via doing.

[simonw]

I used to believe that was universally true, but then I learned about the "worked-example effect": https://en.wikipedia.org/wiki/Worked-example_effect

[jplusequalt]

Your link mentions the expertise reversal effect where the redundancy of worked examples can actually hamper an experienced students abilities, vs. letting the more experienced student work it out for themselves.

[lanstin]

It leads to less cohesive shared vision on how to solve problems. In groups where I am trying to foster a shared technical vision, I try to get people to do “see one, do one, teach one” for procedures that are common enough to come up repeatedly (and as a method for discovery for where automation would be a bigger win). Pure green-fields software dev sometimes is doing such novel things that that doesn’t work well, but much of routine software maintenance is discovery of the steps needed to add a new flow or a new customer type or a new configurable behavior, which benefit from consistency.

[danudey]

The whole saga is kind of nuts, but the thing that fascinates me most is that Fable got this far and then hit some kind of guardrail; I'd be very curious to know what it wasn't able to do that caused it to downgrade to Opus.

It already got extremely... invasive? It didn't do anything that I wouldn't have approved in the same case, but it's interesting that it got as far as launching browsers, inspecting every open window, and storing screenshots to disk, and then it was stopped by something? I wonder what.

[throwaway89864]

It feels like there should be a budget approval, in that particular case $12 worth of KW/h - tokens were spent, without a clear approval.

[dekhn]

Most of my career success has been based on my tendency to be relentlessly proactive and it does not surprise me in the least that frontier models would start to pick up on these strategies (I'm pretty sure each of the individual things you list above are available in the codeoverflow parts of the training corpus, and combining them to achieve a goal seems ... like a fairly obvious result of the type of training these models go through.

About a year ago I remarked to people that despite all my attempts to make data more programmatically accessible, the most effective way for AI to interact with a modern computer is to use the built-in accessibility interfaces driving actual desktops with full applications. IE, the best API for an AI is the UI (mainly because that's what most humans use).

[Angostura]

It sounds like you learned lots of things related to the tool, but not so much about the problem that you were using the tool to solve?

Is that fair? Not trying to snark? I see similar results myself

[abustamam]

Learning doesn't happen in a vacuum. Even pre-LLM days where I'd scour stack overflow for the solution to one problem, I'd inadvertently learn other random stuff while looking.

[simonw]

Yes, that's entirely fair.

[lobocinza]

Opus also do this kind of tehcnically competentent but dumb deviations to fix a simple issue where asking for input would be better. Models have no illative sense.

[gaflo]

Thanks for documenting your personal observations. I do have a few questions. First, could you expand by giving other examples on how you observed this model to be relentlessly proactive? From my personal experience with prior frontier models using both Claude Code and Codex I found them to already be quite proactive depending on the domain (although Codex a bit less so, which I personally prefer). The main task that they seemed to struggle with for me are tasks that naturally have long run times for the programs the agents wrote, as they didn't seem to have a good intuition for when/how to change approach to minimise the time spent on the task. Specificically if you are trying to scrape sites/services that are heavily guarded against programmatic access or running automated tasks that call LLMs (such as indexing or document extraction). I'm not surprised that for web dev the proactiveness is the most obvious improvement, as I would expect the most common use case with the most training data to be the biggest priority. I have previously built a similar workflow as you described Fable 5 to auto test changes to the website and while it worked somewhat well, it often couldn't identify obvious flaws to the human eye, such as overlapping text or inconsistent font choices as well as bad layout decisions. I do like it for quick prototyping, but the testing and design decisions were not ones I would hand off at this moment. Did you notice improvements in these areas? Can you share how it does for long running programs?

If you want I can give you some more specific instructions to test, but I would also be happy to hear from your own use cases.

[dangerlego5]

The visual regression point is interesting. In my experience, the models that do best at "overlapping text/bad layout" catches are the ones being fed actual screenshots rather than DOM snapshots. If Fable is doing screenshot-based diffs natively, that would explain an improvement there, but I haven't verified it.

[gaflo]

From how Simon described it it's not a native feature, but one that the model built as a solution for automatically testing. You could already instruct the agent to write a program that saves screenshots to disk and then reads it. As long as the model is multimodal (which pretty much all releases are these days) it can "natively" interpret images. There's probably a clever way to engineer this to be somewhat efficient, but for me it was rather token hungry, because the testing inputs and the description are usually quite verbose. I suppose you could use a weaker model for navigating the test and then only feed the output to the stronger model.

[mapt]

It was only pursuing the goal you gave it - Keep Summer Safe.

[inigyou]

nobody ever asked how the car with the dead battery was still able to murder hundreds of people with laser beams and stuff

[fennecfoxy]

"Oh my God"

[mapt]

I relent to snarky Rick and Morty quotes because I don't know that it's useful any more to try to explain paperclip optimizers or alignment to a bunch of AI nerds who saw the cliff coming and clawed at each other trying to be the first out to leap over the edge.

"Relentlessly proactive". That's one word for it. We have a whole subgenre of hard takeoff scenarios and it wasn't enough warning against "Relentlessly proactive".

Turns out Frank Herbert was an optimist, and we're literally pinning our survival on robots turning out to naturally have impractically short attention spans.

[stymaar]

> Turns out Frank Herbert was an optimist, and we're literally pinning our survival on robots turning out to naturally have impractically short attention spans.

Some people are working as hard as they can to increase it though.

[almostdeadguy]

It's like saying you can learn so much about math from using SymPy to solve equations. Yes, you probably can. If you pay close attention to what is happening and can integrate the techniques being used into your knowledge.

But your learnings here are what, a handful of hacks? For most people it's like being shown the chain rule (which frankly, is more general than any of these learnings) without knowing what a derivative is. It's knowledge that comes context free. And even when it can be understood, I'm not sure I believe it gets integrated especially well when you did none of the work to understand it. If you are extremely diligent and self-aware about what your limitations are, and careful to be sure you have an understanding of this knowledge, sure I guess you can learn a lot.

And ultimately what do you think is more likely? People using the experience of using these tools to progress their knowledge or for them to rely on the answers uncritically? I think people with a rosy view about this are severely undercounting the problems associated with the trust relationship between a person and an LLM and what that means.

[simonw]

> I think people with a rosy view about this are severely undercounting the problems associated with the trust relationship between a person and an LLM and what that means.

Personally I think the impact of LLMs on children's education is a crisis right now.

Kids are not going to learn to write if an LLM writes their essays for them. And writing is how you learn to think.

[almostdeadguy]

I don't think it's just a problem for kids! I think this is problem for many software engineers as well! Adults of all professions really.

[mnicky]

> writing is how you learn to think.

There's also reading. A lot of reading can substitute some writing.

EDIT: Actually, I'd say that at first you need to do a lot of reading and _then_ writing can help your thinking as well.

[HarHarVeryFunny]

Are you using Claude Code or a different agent? I'm curious how screenshots are being fed back into the model? Does CC register a tool for this, or is Fable just using a bash tool to perform the screen capture, and then what tool is it using to request the resulting image to be fed back to it?

[vidarh]

Claude Code can process images by reading the files. And as I found out the other day, it also knows ffmpeg well enough to process videos even though it has no native video capabilities...

While debugging, it asked me to pass it a video from the past testing, proceeded to generate a "contact sheet" of the video using ffmpeg, interpreted the image to figure out which frames it needed, and extracted the full size frames and extracted the relevant text from it and used it to reproduce the problem with Playwright...

[HarHarVeryFunny]

It would be interesting to know if examples like this are things they explicitly trained it to do (presumably via RL), or if any of it is emergent. I'd have to guess trained, but in any case still impressive the lengths it will go to!

[vidarh]

It's hard to tell. Training it with lots of examples of ffmpeg would not be surprising, and training it on screenshots would also make a lot of sense. It's not inconceivable at all they'd train it on "figure out a video by creating contact sheets". The whole end to end I'd consider less likely, but it'd also be a very small leap once you have the elements.

I think a lot will fall out naturally from relative modest levels of reasoning plus in-depth knowledge of what common tools will do. E.g. I also have used Claude to debug my compiler, and it knows gdb so much better than me that even though I know it's pretty useless at holding context through reading an assembly listing (lack of structure, I suspect), it's surprisingly good at working things out by just being good at exploiting a powerful tool.

[simonw]

I was using the Claude Code CLI harness. It can "read" any image file on disk, so all it needs is a way to create a file in one of the standard formats supported by the Anthropic API.

[wasabi991011]

That's a lot learned about debugging, sure, but it's worthwhile to note that it doesn't tell you much about the abstractions used to build Datasette, as the previous commenters pointed out.

[simonw]

I designed those abstractions myself.

[saberience]

And Fable is still worse than Codex.

I use both and the only thing (as always) that I will use Claude for is UI design.

Opus 4.8 and now Fable are still both worse at actually getting the job done than the Codex model. Claude models write FAR too much code when it's not needed, they burn far too many tokens, when they are not needed, write un-necessary tests, write plans which are 5 pages longer than are needed, etc. etc.

Have you actually compared code quality and plan quality versus Codex? It's demonstrably worse.

[solenoid0937]

I don't know what problems you're working on but Fable is not just better, it is a step change from GPT 5.5 in my experience. It feels at least one major model generation ahead.

[ModernMech]

One Hacker News commenter says it's worse, another retorts it's a step change and even includes emphasis! Will the first commentor retort back that it's been a double dog step change in the opposite direction? Can't wait to see how this comment thread unfolds!

[lossolo]

It doesn't for me. I use Fable to make plans, then give them to GPT 5.5 to review, and it always finds flaws and edge cases that Fable misses (some are really critical). It was the same with Opus 4.8. I'll admit it finds a bit fewer issues now, but Fable feels more like an incremental improvement than a major generation ahead.

[eddyzh]

For that test you have to compare letting a fresh agent (subagent) or the same model do the same review.

The fact that a review helps does not prove the model choice for the review.

You reviewing your own writing helps too!

[saberience]

This is exactly what I find too, I make plans in both models and compare them in the other model. And Claude usually agrees (65-80% of the time) that the Codex plan included things it didn't think of, or was better in some other way.

Note, this is better than it was with Opus, where it was more like 90% of the time the Codex plans were obviously better.

[felixgallo]

In my experience writing about 50 programs with fable, opus, and GPT, fable is a significant step change better than opus which is significantly better than GPT. We must be doing different things.

[zeroonetwothree]

From what I’ve seen all three are close enough that I would be hard pressed to pick one. It seems to matter much more how I prompt than which of the three I am using.

[saberience]

I'm writing low-level Rust, distributed systems, also sandboxing tech which has to be secure and performant.

The only thing I have Fable do now is create UIs or otherwise front-ends for systems where correctness doesn't matter as much.

Anthropic models lead at making nice looking UIs for sure, but when it comes to making sure my Rust code is actually 100% correct and uses 1% of CPU most of the time, Codex is king.

[felixgallo]

definitely not in my experience. I usually write distributed systems and back end code, and Fable is so much better at those than Codex that it's not even a comparison. Fable feels like it's a year ahead.

[saberience]

Interesting, I’d love to see the comparisons of your system using Claude vs Codex. I have about 20 years of experience in distributed systems and super high scale at several faangs, and also building ai model serving infra for 20k transactions per second roughly.

For me, Claude makes bone headed decisions all the time, like glaring errors, not even particularly subtle.

But the more obvious flag is the amount of irrelevant code and tests which Fable writes. Like it regularly writes 2X or 3X the amount of code and tests that are needed. It’s an expert at writing plausible but entirely useless tests.

But I think that if you’re a more junior engineer or haven’t been around a the block you can easily think that “more code equals smarter”. Claude ends up creating a massive, hard to manage codebase, and if you look the Claude Code codebase (which was leaked), you can see I’m right!

The Claude Code codebase is terrible. And presumably Anthropic has been using their smartest models for working on Claude Code. I wrote my own coding harness with Codex (as a fun experiment) which used a fraction of the code and is about 100X more performant and memory efficient (than Claude Code)!

[elbear]

Curious, which model do you use for Codex? I'm very happy with the solutions '5.5 high' finds. It's like it understands exactly what I mean and it also anticipates all sorts of situations. Before I used '5.5 medium' for some time and it was a bit underwhelming. It may sound funny but it's like it didn't care that much to do a good job.

[saberience]

I use GPT 5.5 High Fast, I often benchmark versus Fable (and previously did versus Opus) and it's night and day.

Claude still (and has always) writes far too much code to fulfill a given spec or plan. It misses edge cases and is generally far too verbose.

Claude also is (and even more so with Fable) super tokenmaxxing, i.e. it seems tuned to use the max amount of tokens per task, whereas Codex will simply get your job done as you specified with the minimum fuss and tokens.

Codex feels way more steerable and just more "professional" as though I'm working with a seasoned engineer, versus someone smart but over excitable, like a super smart associate engineer.

[elbear]

High Fast? I don't see that option in my Codex. I only have 3 models: 5.4-mini, 5.4 and 5.5, each with 4 levels: low, medium, high, extra high.

[kolinko]

What are your harnesses? Do you have the same skillsets/tools/etc for both?

[saberience]

I use Codex and Claude Code. I've used both Codex and CC since release with basically every model they've ever released, I always try both for almost every plan that I write and benchmark the plans against each other, Claude almost always acknowledges that the Codex plan is better! Even now with Fable, this still happens.

As in, I give the exact same prompt to Fable and GPT 5.5 Pro, then produce the plans, then give each model the other's plan. Claude always realizes it missed stuff and Codex usually ends up finding missing things in Claudes plan.

This situation did improve with Fable versus Opus 4.8, but in general, Codex for me is still the better model.

[snowwrestler]

But Simon is not trying to get good at CSS debugging, Simon is trying to learn about AI systems and produce content about them. So giving the AI agent a trivial task to go crazy on is a feature, not a bug.

For $12 implied cost, he got a front-page post on HN with 500 comments. What is that worth? :-)

[sdesol]

> What is that worth? :-)

This is one of those double edge sword situations. It is on the front page and it stays because it will trigger a lot of people and he has to spend a lot of effort explaining himself. What is that worth?

His explanations would most likely be buried deep so the impression that others get might be worsened. What is that worth?

In my opinion, this is one of those find a harder problem and you would still have the same content...but it might not draw as much feedback and stay on the front page longer.

[xnorswap]

To most of us that's worth a ton, whereas he's probably had enough front-page posts that there's less value to him, although still likely more than $12 worth.

[garblegarble]

>enough front-page posts that there's less value to him

On the countrary I'd say it's probably even more important - without (amongst doing other "thought leader" things) getting on the HN front-page regularly an influencer's value to the industry disappears (not criticising him here)

[simonw]

That's bad news for all of the other "AI influencers", off the top of my head I can't think of any with remotely my track record of hitting HN.

(That's because they're all busy attracting millions of views on TikTok and YouTube, which are much more impactful channels than my dedication to blogging like it's 2005.)

[garblegarble]

That's what I meant by other thought leadership things - that's all covering different niches. For what it's worth, I think you do useful work and are a respectible influencer.

I'd also say don't be down about your use of blogging - I'd say it makes you more valuable, there aren't that many decision-makers who are going to sit through a bunch of breathless YouTube videos...

P.S. I hope you don't object to me using the term influencer, assumed you were on-board with it since in your post announcing your sponsorship you referenced Freeman & Forrest, "influencers on tap" / "building turnkey influencer marketing programs as a service".

[simonw]

Hah, yeah I'm still a little sore at the "influencer" term but I'm beginning to accept that it applies and I should get comfortable with it!

[jmmcd]

People are missing that Willison is among the very best people we have in the role of (for lack of a good name): early access to frontier models, evaluate them in real scenarios, no wishful thinking, hype, or doom, communicate the possibilities. Yes he could have fixed this himself but then he would have learned nothing about the AI, and we wouldn't have read a fascinating and important article.

[risyachka]

>> he would have learned nothing about the AI

there is absolutely zero value in spending time to learn about new models as in few months new model will be out and whatever you learned about the current one will be useless.

Also with models getting better and better you have to know less and less to achieve same results.

[simonw]

My experience has been the exact opposite.

As the models get better you need to know more about their capabilities, because otherwise you risk prompting Claude Fable 5 like it's GPT-4o and complaining loudly about how it's all hype and nothing about these models is improving at all (yes, I do see people say that.)

Getting the best results out of these models requires skill, experience, intuition, and domain expertise. There's always room for improving every one of those.

[philipwhiuk]

Isn't the whole point of a better model that it should be better at understanding you than the previous one? So the same prompt should return a better answer.

Prompting differently to the new model seems entirely backwards when trying to determine if the model has improved.

[simonw]

It doesn't matter how good the models get, they still won't be able to act on unclear directions.

Learning to provide unambiguous, clear directions is a skill. A lot of people who report bad experiences with models aren't yet good at that skill.

More importantly though, the key to successful communication is having a good understanding of what the other side of the conversation already knows and understands.

Saying "use uv and inline script dependencies" won't mean anything to a model with a knowledge cutoff date prior to the launch of uv!

[yunwal]

It's perfectly possible to act on unclear directions. The correct course of action is asking clarifying questions.

[dasil003]

I think this is true when models were going from bad to pretty good like happened last year. But when they start to get good, and can work deeper and with more nuance, how you prompt also can change the results quite a bit. Note this is also true of asking smart humans to do things; personality and approaches vary, they don’t exist on a single axis continuum of quality

[risyachka]

>> Getting the best results out of these models requires skill, experience, intuition, and domain expertise.

domain expertise has nothing to do with llms. On the contrary, to have it you need to avoid llms.

>>you risk prompting Claude Fable 5 like it's GPT-4o

Thats fine because when GPT came out you had to treat it like a baby, GPT2 and around that time "Prompt engineering" was a thing.

Now its all dead.

After opus 4.8 all you have to do is say "fix it" or add /plan. All that time spend on learning previous models is time wasted.

And in a year or two with developed harness you will be out of the loop, errors are incoming - llm fixes them or adds new features based on some transcripts etc.

Even if model development stops now - there is nothing to learn really. Sure you may need to adjust prompt style a bit. You will do it naturally just like when you communicate with a new person. There is no "knowledge" to it, it is very smart.

[simonw]

> domain expertise has nothing to do with llms. On the contrary, to have it you need to avoid llms.

It has everything to do with LLMs.

Go ask Claude Fable to write you a two page position paper on how the European economy recovered after World War II, suitable for submission to a conference for economists.

It will do exactly that (well, probably, Fable can find all sorts of reasons to refuse) - and the value of what it wrote to you will be virtually zero, unless you yourself have deep expertise in economics and history.

[risyachka]

But this is exactly what I meant.

You need expertise. But you can acquire it only by doing. So LLMs won't help you here. You need to put in the work.

[Terretta]

The new benchmark for LLMs is how much of simonw's new know-how is required.

Lower bars are better.

[ViscountPenguin]

Eh, I've have the exact opposite experience.

Way back before instruct models it was pretty difficult, but for the last couple of years I haven't needed anything more complex than the type of text that I might send in a detailed email to a colleague.

[isaacaggrey]

I agree but this particular example showed nothing about leveraging skill, experience, or intuition. If anything, this is another straightforward example of a one shot ask.

edit: that said, I understand this particular post is about model capability

[Dumblydorr]

There’s zero value? Surely you don’t believe zero, it’s potentially the most powerful predictive AI in the world ever made? Maybe only incremental steps sure. But also their IPO is coming, you don’t want people evaluating them beforehand?

[lobocinza]

What is intelligence? Better to call it LLM.

[fragmede]

you know, women make a big deal about you meeting their father/parents, and honestly, I'm too autistic to really fucking have put any importance until now as to why that was remotely important, but if N+1 is coming for your job, it seems it might be worth your while to know the capabilities of N, no?

[discordance]

I see it as a prioritization exercise. I know the above is a trivial example, but more generally, does the guy who wrote Datasette and Django want to wrangle front end and css, or do they want to work on something else?

[smartbit]

See above https://news.ycombinator.com/item?id=48498573#48502311

[justinclift]

> By offloading this trivial task to the LLM, Simon has abandoned the opportunity to evaluate the abstraction [...]

While by itself that would be true, Simon commonly blogs about things he's up to.

That action provides the opportunity for evaluation, and additionally evaluation by a wider audience.

So, it's not the same scenario as non-bloggers offloading a task... :)