These questions are even not about AI: if I were to give money to a human agency and were given something they tell me works, I would ask the same questions. If I did not know how to evaluate, I would hire people that do. With LLMs the verification part is what bothers me the most.
The only decent software engineering perspective I’ve seen has been from Mitchell Hashimoto.
They can just summon bespoke software out of the ether that only handles the use cases of themselves and a few of their collaborators.
Making “side projects” was mot possible for non-developers before powerful LLMs. Now it is.
> I am sure it is not perfect (I only spent an hour working with the results), but a software engineer would iron out the remaining potential bugs that I could not find quickly [...]
People have said things like this many times in the past, and, in the past (perhaps not now), it's always been a misunderstanding of what is good and bad, what's difficult and easy.
For example, someone would draw a UI in a GUI painter that generates code (or a resource file), and a manager would see it and think the majority of the work towards the product is done. (Incidentally, then there seemed to be a reaction, towards making your UI mockups look abstract or otherwise different from runnable code, helping the nontechical to understand that this isn't 90% of the finished product.)
Or a student intern hacks out a homework-grade demo, and a manager who understands neither software engineering nor product domain says "we just need some engineers to polish it up for production", and thinks the student is a star and why can't their engineers be as brilliant and productive. (I might have once been that energetic intern, who was happy for the encouragement, but then learned more, and saw it was a thing.)
This common misunderstanding was sometimes self-correcting -- when trying to ship became a disaster of misery and regretted-attrition, or the product was poorly received by the market because it wasn't thought through nor implemented well, or building subsequent functionality atop it was a nightmare. (But adverse effects of bad approaches is one of the reasons for management and ICs to job-hop, before the unwanted effects affect them personally.)
What might be different now is that some of these AI tools are outputting better-engineered work than some software engineers, and much faster.
At the back of my mind, I'm wondering how the really great software engineers will continue to stand out, as the discipline is being devalued in the minds of most leadership, and anyone can prompt an AI to generate something that superficially appears to them like what they assume a great software engineer would produce. (Even if the great engineer would do much better quality of implementation, have innovative ideas that ML from open source code would not, and maybe arrive at better product concepts as they worked through the problems.)
The trick to getting good at using LLMs for software is to learn how to make _all_ projects low-stakes.
this doesn't really work in the real world. There are many things that actually matter, engineering is fundamentally about handling them.
the quality of produced code and the medium
If the observed behavior of the software is good, then the software is good. If a bug, of whatever kind, can be fixed by a model on a vibe-coded codebase, then that's a fixable bug. If there are no exploitable vulnerabilities, then the code is secure. If the performance is adequate, then the code is performant.
It simply does not matter what the code looks like if, from the outside, it does what its supposed to, and, from the inside, a model can fix the issue if one is found.
More than ever, software engineering is now really a job about making sure the code is doing what its supposed to.
And even if it DOES matter what the code looks like, you can have a model fix that too.
But all of those correctness are imaginary. The hardware only enforce a few (and it may be buggy). The OS adds some more (and it’s buggy). The compiler/interpreter may have bugs (but that’s rarely a nuisance) and the libraries are often brittle. There are cracks everywhere in the tower of abstractions.
The code has never mattered. What has always mattered is the knowledge of what is the model of correctness of the software (programming as a theory by NauR), so that you can discern where a program is wrong.
The thing is a crash or some other immediate errors are actually nice to have. You get to react immediately and can have a core dump or a stacktrace that points you the error. What is truly a terror is silent corruption (wrong order of operations, wrong values for a comparison that has expanded the idea of correctness, security issues that has been backdoored for years,…).
As Hoare said:
There are two ways of constructing a software design: One way is to make it so simple that there are obviously no deficiencies and the other way is to make it so complicated that there are no obvious deficiencies.
The first method is far more difficult.
I clicked one of his examples intrigued "a snake game where the snake is self-aware and crazy things happen;". Played for 1-2 minutes, and it's the classic 1980s snake game. Am I missing something? What is "self-aware" about it? Some funny messages at the bottom of the screen? And what are the "crazy things"?
I will say, the act of eating creates a "bulge distortion" that flows down the length of the snake is a nice touch though.
I am creating a game and I can say that with the coding part the models help a lot, mostly gpt 5.5 high. Tbh to me all the frontier models feel the same and they can all solve the stuff I do quite well with some guidance and prompting. But that kind of makes me appreciate the other stuff more like visual style, sound design, mechanics etc etc. Tons of work still.
For brainstorming I find the models bad nowadays or maybe I am just too critical of the results
The lack of downvotes on posts on HN has always felt like more of a bug than a feature to me.
Everyone does. You don’t think about it everyday because we’ve delegated it to experts which don’t come up with a new composition of Asphalt every time you press “generate”. It’s rigorously battle tested and short of intentional negligence, it’s consistent. I’m amazed how people are forgetting how the world actually works.
But yes, you are right - I don't build roads and don't know what is a price to build a road and how to determine the quality of correctly built one, nor I will ever care or learn.
I get that there's little sense in arguing with the MBA hivemind, but... c'mon.
I manage two teams of highly motivated, largely pro-AI engineers. Both teams have independently concluded that they needed to ramp down GenAI usage because of code quality / maintainability concerns. Both teams have suffered from protracted outages caused by LLM jank not being sufficiently fenced off and guarded against. Both teams have expressed concern that the code generated by LLMs is far too verbose, full of slop, and rapidly becomes an unmaintainable mess.
These are teams that are building non-trivial LLM solutions (deep agentic data synthesis and multi-modal data tagging). They are using the technology creatively and pro-actively, not just vibe-coding slop and throwing their hands up when it fails. Both teams will continue using GenAI coding agents, don't get me wrong - but the gains are incremental, not transformative, and need careful fencing to make sustainable.
Nothing in these articles resonates as real. People who work in reality don't agree. I don't understand why this shit keeps getting attention (or rather I do, but the reasons aren't good).
In fact, that's the entire reason we care about "quality code", because we assume that quality code is code that does what you expect well and consistently.
I say this as someone who hand writes code pretty much every night for fun, just to experiment with computation. Which, oddly, is more fun than ever because I don't feel like there's any need to connect this type of programming with "real world software", and I can really enjoy code for it's own sake, meanwhile my job is mostly just running agent loops (which I quite like as well).
So AI is only interesting to you / your org / humans if it can do things that you can not achieve. But if it still does errors, how could we ever know that super-invention by AI is not wrong?
If we can not rely on the correctness of the result, it is not usable at all. AI must create reliable and correct results always. That was a very fundamental requirement for computing. This problem has not been solved.
Also this is easily solved by .md spec files, this whole "bad code" cope is just FUD'
Yet, I can't deny the reality that I observe working with LLMs every day. If this truly is a step-function (as some are sgguesting), then I have absolutely zero concern for the quality of the code.
> Again, it wasn’t perfect. As an expert, I was able to spot some errors and omissions (some as a result of the design I had asked for) that I had the AI correct
That's the bit that stuck out to me - that's longer than I would expect to work on a problem in a day or even expect to go back & fix the output of something that has a core reward loop of hours.
My customers are currently clamoring to push down my agent response times from 85 seconds down to below the 20s mark.
At the same time, it is very dissonant to see the industry heading towards hour+ long workflows with an agent.
We're gonna go back to the days where our bosses ask why we're just sitting around, but instead of saying "compiling," we'll just say, "waiting for Claude."
Will Claude's code be perfect in one shot? Probably not, will it get you 80 to 90% of the way there with your chosen design patterns in under a few hours? Absolutely.
It's some prompt engineered AI harness, that guides the AI to create stats after it researches a subject and ingests the data, but I'm not sure what is it that the tool actually does on top of this.
At this point, pay me significantly more, and I'll do it.
Ha ha, that's how you negotiate yourself out of a job!
I'm amazed we're so far into SOTA bloat that the chinese will kill once they start etching silicon with these models.
https://isochronic-passage-chart.netlify.app/
Doesn’t work too well on mobile but looks interesting
I also see some logic flaws. It overlooks the option of going to a major hub to access faster aircraft, rather than hopping on local hubs.
Also, immigration and customs are cleared at the first airport you arrive at in the country, not at the last one.
In some countries, you need to clear immigration even while going to a third country, so 1 hour is not enough to do it.
Which just about sums up my experience with using LLMs to code, really (though not with these state-of-the-art models, admittedly) - it's amazing what they can do, but left to their own devices they'll make boneheaded decisions.
Yeah, the whole "can run for 9 hours on a task" to me is not a positive.
I tend to find if Opus 4.8 runs for ~15 mins on a task, then the end result has gone off in a weird direction at some point, and it needs winding back a fair bit.
And that's with extremely clear direction, literal specification docs to follow, etc.
That being said, having functional code already created beforehand (ie by a human) goes a long way to ensuring the AI model has a path it can build on without making too many dumb architectural choices by itself. Generally.
The real issue with the title is that it doesnt fit in the box!
It's like someone took a beatiful, intricate piece of vintage jewellry and made a slapdash imitation out of cheap plastic.
I'm not very threatened by this if this is the dangerous Mythos model - it just seems like a slightly incrementally better sonnet
Personally I don't really care, because I like coding and learning myself and DeepSeek Flash is all I really care about. But it's really easy to have a ton of benchmarks where the top models can't get anywhere close - and I like to test them on these problems to see how good they are getting.
Fable 5 is def a little better than 4.8 btw.
Myth. Total myth! I recently had to beg for more RAM after continually hitting swap space which causes tools like dictation to stop working, failure to load certain websites without rebooting, and so on. Devs do in fact need powerful machines and the ~$500-1000 an employer saves upfront in machine costs is dwarfed by productivity losses.
Giving your engineering employees new machines in a 2-year cycle that are between the middle and high end is one of the cheapest ROI decisions that a tech org can make.
A small portion of this effort is having a high quality Lua in Rust repo. I’m using mythos to fix some of the performance issues with my Lua interpreter that gpt 5.5/ opus 4.8 had stone walled on.
Not sure if Mythos will be able to crack this but it has been running for a couple hours now with some promising results.
Performance charts linked here if your curious https://github.com/ianm199/lua-rs
Fable 5 found quite a few issues Opus 4.8 missed on code review, even though the stupid cybersecurity nonsense downgraded it. I can't tell you more, I only get a single session per 5h window on Max 5x. Only ran two sessions so far.
On the margins, suppose the prompt is literally: "Build a feature complete, high polish Facebook clone". Facebook is complex but likely not super complicated tech, and still I would assume that (after having burned through a substantial amount of tokens) you would find substantial enough differences in the outcomes between different models on that prompt on various fronts.
The above ask is obviously not useful, but what's preventing you from taking on bigger chunks until you approach the limit? At some point you would hit a boundary, where the diff will be obvious.
It also burned through my usage quota like a late-90s Hummer.
Yeah. I have a Max 5x subscription and Fable burned through 16% of my weekly quota in a 40 minute code review session. It didn't even finish the review, it switched back to Opus 4.8 in the critical memory safety parts where I actually needed Fable.
I feel like I'm going to get priced out of these models soon. I should probably try to get the most out of Fable until June 22nd.
It's not just salary, but also safety/labor regulation, legal risk, vacations, sick time, personal conflicts, HR, benefits.
Even when automation is more expensive on paper, it's generally still cheaper
Do you not believe in running tests, evaluations, or experiments at all to better understand your environment?
The ROI in the case of a positive outcome is the reduced time needed to inspect the results in the future (the entire point of AI is to know what you can trust it on, so you can delegate everything at that level with less oversight). The ROI in the negative case is the tokens not wasted on tasks to ambitious for the model.
We know this model will be cheaper and faster with time.
And we have not even reached the timespan/timeframe were we have ASIC style models.
OpenAI has to do something which will beat Fable otherwise Anthropic won. China currently overtakes cars, pv, batteries and very soon silicon chip making, it has all the incentive to also take over AI.
The poem Kandel translated from the original Polish was, for artistic reasons, completely different. I will be impressed when machine translation can duplicate that!
She scissored short. Sorely shorn,
Soon shackled slave, Samson sighed.
Silently scheming,
Sightlessly seeking
Some savage, spectacular suicide.
- That's the translated Cyberiad Poem the blog post based it off off (or the AI decided to do so)
He is a professor but sadly also an AI shill. He should switch to advertising washing power.
I do not fear that management will get tools like Mythos and then not need people like me. Most of the value I provide is in translating what the management/client _thinks_ they need into what is the real problem and solution.
That's not an insult to them, it's just pointing out that they see only their problem, and they imagine what would be the solution. They then ask for that solution. Quite often, what they want built isn't what they need. And I've seen so many problems, from so many domains and scenarios, that I can usually recognize the core need and propose (and build or direct building of) a solution which resolves that need AND has an eye toward the likely future needs.
Mythos may do an excellent job providing a high quality result based on what is asked of it. But the result will only be as good as the quality, clarity, and presentation of the request.
If I hire a home builder to build me a custom home, that builder is going to ask me a thousand questions - questions I had never even thought of. Mythos isn't going to ask all those questions - it's going to make the best choices it can without the consultant's level of interaction. And the buyer will get what they get. Sure, the buyer can then say, "oh, I don't want any hallways - just connected spaces." Then the house gets demolished and rebuilt to the new, clearer spec. Repeat, repeat repeat. Maybe eventually the buyer gets what they really want. More likely they give up before reaching that point, and they go and hire a real builder.
I'll sum it up like this: You can get great results with minimal effort if you don't really care too much about the details. But if you don't care much about the details, then your need probably wasn't very significant.
Sure, AI can auto-complete the line, but it can't write full functions.
Sure, AI can write functions, but it can't complete full features.
Sure, AI can write full features, but it can't build full applications.
Sure, AI can write full applications, but it can't build them in the right way / ask the right questions / write beautiful maintainable code / do what _I_ do..
Time will tell.
The problem is much broader though - consolidation of wealth and power have enabled, frankly, idiots to be able to control how the world works - from politics to business. Greed and stupidity is eating the world.
I don't see any solution. This is like a disease that will either eventually kill the body or take a long time to heal, leaving deep scars and forever changing humanity.
Maybe War Games was right - the only way to win is not to play. Therefore, find something you love (even if it doesn't pay well), and do that.
(I spent two years looking for a tech job. My 30 years of broad and meaningful experience is apparently not interesting to at least the 200 companies I applied to. So now I'm a teacher, and I'm quite happy.)
Every sw dev knows this is a very dangerous, and unrealistic, assumption.
So nope, not the AGI. But definitely an improvement.
That's the kind of behaviour I've seen in Claude Code (Opus 4.8) when it's context space is over the 40-50% range.
I tend to keep an eye on the context usage (ie `/context`) quite a lot, and generally see good results as long as the context usage is ~30% or below.
Which isn't heaps, considering having to ensure it has the required docs/stuff it needs can take 15-20% of context by itself.
Not exactly strange behavior, Opus acted just like this too when I first subscribed. The popupar meme is Anthropic nerfed Opus during their capacity cruch. No idea if it's true, but I do wonder if Fable will fall victim to the same fate.
In a project like mine (https://github.com/tsz-org/tsz) I am constantly frustrated that models were not doing enough research and were not taking into account other situations. Again and again models would produce code that would fix one thing and break 2 other tests that were "unrelated".
With Fable it seems like tasks are taking much longer (I have not seen a pull request from Fable sessions yet) but reading the transcription of those sessions I can see how it is doing the right thing by not leaving any stone unturned.
As the article says, it's hard to communicate this "feeling" about models because it is very project specific but I thought I share
But overall, this is pretty normal for compilers to have this sort of "unexpected" tests failing due to some work in an area. It happened to me when I was coding everything manually back in the day too
> This is from "The Cyberiad", a collection of science-fiction fairy tales by Polish author Stanislaw Lem ... In one of the stories, a robot constructor named Trurl creates a machine that writes poetry. A jealous rival named Klapaucian challenges the machine to compose "...a poem about a haircut! But lofty, noble, tragic, timeless, full of love, treachery, retribution, quiet heroism and in the face of certain doom! Six lines, cleverly rhymed, and every word beginning with the letter s!!"
And the computer responds with:
"Seduced, shaggy Samson snored.
She scissored short. Sorely shorn,
Soon shackled slave, Samson sighed.
Silently scheming,
Sightlessly seeking
Some savage, spectacular suicide"
The author had to be referencing this moment in their challenge to Fable/Mythos. I'm curious to know what their exact prompt was.
Cyprian cyberotoman, cynik, ceniąc czule
Czarnej córy cesarskiej cud ciemnego ciała,
Ciągle cytrą czarował. Czerwieniała cała,
Cicha, co-dzień czekała, cierpiała, czuwała...
... Cyprian ciotkę całuje, cisnąwszy czarnulę!!
Or it just swept it up in the training data given Anthropic license Reddit comments.
The first item on the article, the first thing it showed, was wrong though.
It is 100% faster to go from London to New York in 1881 than Volgagrad. Or any of the Russian hinterland colored green or Turkey or Egypt.
What?
> Switched to Opus 4.8: Fable 5 has safety measures that flag messages on most cybersecurity or biology topics. They may flag safe, normal content as well. These measures let us bring you Mythos-level capability in other areas sooner, and we're working to refine them. Send feedback or learn more.
I don’t see why working longer is a pro. The results don’t seem much better than you’d get from putting Opus in a long loop.
Care to share the results you got from Opus working on the same prompt? It should be easy to compare quality.
Just an FYI this guy is an AI hype-beast. Some of his tweets are truly out there.
Is it a hard problem or is it just labor intensive?
"Posterior beliefs about market demand are purely referencedependent: holding dollars raised constant, they track only performance relative to the founder’s self-chosen goal—jumping half a standard deviation at the threshold, responding steeply for the first ten points past it, and flattening thereafter"
Humans generally don't verbalize data this way. The summary document is also very fluffy.
Maybe my prompts are too vague, but it’s worth noting that every example in the post is a greenfield build, and vague prompting seems to hold up fine when there are no existing constraints to respect.
[1] https://isochronic-passage-chart.netlify.app/
[2] https://mapitout.welcome-to-nl.nl/
- Went deep on "what types of guidance even are there? what does giving good guidance mean?"
- Sampled my existing Claude guidance (CLAUDE.md, skills, hooks, etc.) and broke their guidance into "atoms"
- Categorized them by clustering, the same way Big Five was generated
- Generated a new candidate
- Then used independent agents to compare it against my existing corpus assuming that the new one would be worse
Working with it felt like working with a supersmart entity capable of generating very plausible-sounding but not-necessarily-true statements. The outcome certainly felt like an alien artifact, like nothing I'd make myself.
Only time'll tell if it holds up, but it sure had some interesting ideas.
Other commenters have pointed out that his isochrone map contains a lot of nonsense as well.
So the most charitable interpretation here is that this is a case of Gell-Mann amnesia.
I made serious progress towards repairing a proof for a conjecture that was published 10 days ago but kept running into a wall with one of the Lemmas.
I threw Fable 5 Max at it with the same subagent set up and in an hour it claimed to have disproved a core theorem of the paper.
The Lean construction looks correct, but I still need to verify it rigorously. This is certainly not something Opus 4.6 Max could do and it’s likely something Opus 4.8 Max could do with more delicate orchestration and time. However, the “one-shot” Fable 5 did give me pause.
Given a tool that is supposed to unlock creativity and excitement, he made a series of worse clones of things.
Again, technically impressive, but the world has never needed the ability to make Balatro but less polished and coherent. We already have Balatro.
I'd be more convinced if people made things that didn't already exist; show me that these tools enable something you actually want.
It's likely that at least some amount of additional context was provided to the model to enable it to reliably create the desired form factor. This introduces the caveat that the author probably views some amount of context as being trivial / beneath the level of mentioning. But then the question becomes where they draw the line.
Most of the “impressive” stuff is not “the model” but “the harness”. Spinning up the subagents and teams of lower models, letting them explore, do adversarial coding. It’s all in the harness. Granted, Mythos might be better at that orchestration, but it’s still the harness.
Second is the prompting. The author is an expert in what they’re doing and prompts the system in a way that yields useful results. I see too many people believing that if an expert can achieve those results in a domain they’re familiar with, then them as non-experts will be able to as well. And that’s a fallacy that Mythos doesn’t change.
There is only one hint: 475k tokens in the screenshot when OP asked the model to fix some behaviour, but it would be fascinating to know the total tokens amount.
Not a great start for "a generational leap in model effectiveness"
And I'm excited to try it, but also have a fear that I will like it too much and then won't have access to it in 2 weeks... But maybe I will and maybe it will be worth it and I'll just pay a bunch of extra for it and it'll be great!
I think the article could be improved by actually sharing more feelings. I clicked on the article for feelings but I didn't see that many feelings described.
Wow
What makes me excited is that GPT 5.6 (its actually GPT 6) is going to be crazy
> It worked for nine and a half hours.
And how much did that cost?
looks nice but deeply flawed
classic LLM output
At first i thought its routing was just completely botched.
The text overflow on the legend is pretty funny considering how well the other graphics turned out
(Edit: referring to the map app)
Edit: A couple hours in and I just got my first gaslighting attempt from the model. Good times!