Hacker News Reader

I used Sol to extract the remaining decryption keys from the Super Mario Maker 2 (Switch) game files. Someone had previously extracted all the keys from the original release, but not any of the new ones from updates. Not only did it succeed, but it helped me understand the data sufficiently to add support for “Super World” rendering to my level viewer (which I made back in 2021), eg the little widget at the top of https://www.smm2-viewer.com/players/B16-306-GVG

I was very pleasantly surprised to find Sol wasn’t obstructive over what was clearly a very grey area endeavour.

Fable is almost unusable for anything but super boring mainstream stuff. I was getting safeguard flagged so often I’ve significantly reduced my usage out of fear they will blacklist/ban me.

Some of the topics it’s flagged have been hard for me to understand what it seeing that can be remotely concerning in my requests.

The safety is really funny to me. I ask it a lot of extreme stuff and it goes through, but I ask it mundane stuff and hit the filters all the time.

It has learned a little too well how it works in human society.

Reminds me of the URL blocking of my company. Nature.com is being blocked but I can access a ton of super sketchy download sites.

I've gotten flagged for asking questions about tokens and tensors. That makes me believe it's not about safety, it's about protecting their turf. I cancelled my subscription - same fear about getting flagged too much leading to a ban.

OpenAI and Kimi are both pretty okay alternatives! I guess GLM 5.3 on Max reasoning as well but for more limited domains.

They said they also block usage of Claude models to build ML models.

Which is definitely protecting their turf, but also probably a little bit hiding their “RSI” abilities for competitive reasons. My theory is that a lot of “safety blocking” is actually WIP training of new business directions. Anthropic has started hiring biologists and has opened a preview of a “Claude code for bioinformatics”. I’m guessing they’re tweaking their bioinformatics market play, and block “bio safety” requests so competitors can’t learn about their training.

What's the distinction between "protecting their turf" and "competitive reasons"? I see them as the same, but I could be missing something.

That's an interesting thought on the current "safety blocking" being a trial run for the topics that scare people (bio). You're more charitable about their motives than I am, but you might be right.

My differentiation for this comment, was “preventing someone from using your product to build a competitor” vs “letting a competitor see your strengths/weakness to benchmark against you”.

Competition is competition and it’s two sides of the same coin.

I agree. It’s flagged me on discussing fast GEMM implementations for large regressions, a discussion on theoretical physics math, a discussion on designing a type of RAG system. I’m super baffled as to what the safety instructions actually are other than “advanced anything” and even then their definition of advanced is a joke, I’m an idiot and my questions are almost laughable.

Same here - my questions were sophomore level. I think it's notable that when I edited my question to say it was about Gemma 4, it answered without blocking. A cynic like myself would interpret that as evidence they don't care about sharing information if it involves their competitors.

Getting downgraded for asking "What is digestion" to Fable is where it's just ridiculous and clearly a limitation of the technologies involved.

I cancelled my Claude max subscription. Somehow every query I sent was flagged as bio or chem, even pure mathematics questions. Not going to waste money paying for a “max” subscription that won’t ever let me use the top tier model…

Sol is great and has never blocked a request, and generally gives great answers. Happily switched over to it now.

Have they made Sol do less unwanted autonomy than the previous Codex models did?

I was having it look at creating a driver for some old scanner and it actively looked up exactly where that gray area for my country was wrt decompilation.

Interesting. I've been wishing that old 'Stars!' game from the 90s would play easily on modern systems. I'd love it if we'd got to the point where I could point Codex at a folder with an ISO from my CD of the game and tell it to go reverse engineer it all for understanding of game mechanics, then go recreate it in a modern language capable of running cross-platform. Scarcely any need to improve on graphics, it could even be a PWA.

There's people that have tried to contact Jeff McBride and follow the IP trail but the IP is currently owned by a company that went defunct. Not sold, but no one is even bothering to register its LLC any more, it's simply dead.

Stars! was awesome, many fond memories of play-by-email games of that...

I feel like those examples are considered difficult because they're niche topics, but aren't actually all that difficult in a general sense. What I consider truly difficult are things like taking a ticket and implementing it in a preexisting codebase, using a clean and reasonable design that fits the existing style and makes sense to a human, and avoids the footguns I learned by working with the codebase for over a day.

If you said this in 2025 I would've 100% understood, but to be honest getting AI models to do a pretty good job on day-to-day ticket work has become so boring that we don't even bother using the top tier models and higher effort slots for that anymore. I personally wind up tweaking the results a lot and recursively having fresh agents review the diff, but that's just because I'm picky; in a lot of cases the first diff is actually pretty damn decent.

Compared to what I am doing at home experimentally, I feel like day-to-day work is absolutely nothing. Not only am I also working with existing codebases in my experimental prototyping, but I am also doing things vastly more complex with vastly harder constraints.

All non-trivial code terra has generated for me has had at least one serious bug in it. Typically caught by a review from myself or Sol.

But I wouldn't trust lower tier models for end to end solutions.

Personally I wouldn't want bots running autonomously on a repo, even if there were other bots cross checking them. But that having been said, I'd also say that my experience was similar with human code: it is rare to not find at least some issue worth at least pointing out. The only real difference is that the LLMs have vastly different holes than people do, making the real challenge trying to make sure you're covering them. In most cases for me the best solution seems to be just giving them a way to test and attempt to prove things out in a realistic environment. But clearly, we haven't really left the era of having humans in the loop. Fully vibe-coded codebases clearly suffer from a myriad of issues.

Multiple security holes. Issues marked fixed that aren't really. Giant holes left in solutions.

Sol over engineers now and then (hey please don't factor that function out into its own file....) but it doesn't do the same level of stupid terra does.

That said, plan with Sol, implement with terra, have Sol fix all the mistakes, then I go over the code and make recommendations for the architecture to fix Sol's foolishness.

My experience is that current models are pretty good at making functional changes without too many more bugs than a human would make, but are still bad at making good high quality changes. I have a degree in software engineering specifically, so perhaps I am overly sensitive to design issues.

This is true in some sense.

Getting the AI to output code that you like is difficult.

As an example, let's say in React you have a "useLocale()" hook.

The AI will happily pass down locale as a prop to 5 child components instead of just calling the hook in the component.

A review from another model did not flag such stylistic issues either.

I believe that the latest models are very good at functionally achieving the goal, but still have poor taste for UX or code quality.

The most productive use of AI for software development happens in an environment where you do not review the code but test the UX end to end.

I use the AGENTS.md to show it how i want the code to look like. Something like "when implementing hooks adhere to the guidelines in docs/react-hooks.md". And then react-hooks describes your heuristics and what you consider best practices. There is a clear difference in code quality for me when using codex with a well crafted AGENTS.md vs. without one, you can run the experiment yourself pretty easily. As I mentioned in another comment, I think Claude poisoned users to stop relying on their Claude.md files and new codex users might be surprised at how well it adheres to guidelines.

We use skills for similar purposes, with positive and negative examples.

I think it sometimes worked, for example for testing preferences, but sometimes it did not.

Could be a problem with the harness also.

In any case, I feel that it's a bit playing whac-a-mole with explicit rules for things that a more intelligent model should do by default.

And Mai-Code-1.1-Flash seems like a really good cooperative player to GPT 5.6 Sol. You get Sol to help you make a detailed plan, and Mai codes it up and you can get pretty decent code out the other end without too many tokens if you are careful.

Why wouldn’t you use Luna for that? It’s super cheap.

Taste I suppose?

> Why wouldn’t you use Luna for that? It’s super cheap.

MAI also offers a ultra cheap version that's competitive with Luna.

So much so that the models look like they were designed by a product manager explicitly to eat away OpenAI's market share.

Vscode even pushed them quite hard onto users with the latest release, going to the extent of putting up a modal to convince users to try them out.

How detailed of a plan? Are you including code snippets or just behavior and letting the lesser model decide how to implement?

FYI I run it consistently in xhigh regardless of difficulty of the task at hand. I remember high being very fast, but I'd rather wait a bit more and get better output. AIs are insanely fast compared to me anyway, even on xhigh. Consumes more usage, but even at 100 EUR/m I don't hit limits.

To be fair, I actually do run xhigh as my default. However, for the first time in my experience of trying and using LLMs, with Sol.. sometimes I feel confident enough to set the effort level to "Low". I just had Sol prototype some AWS stuff on low earlier. Great result, did exactly what I wanted.

After hitting the session limit on my company's plan so many times with Claude when I was using it, I mostly keep Codex on "high" rather than "xhigh" as a way to leave the tokens for my more ambitious coworkers. It's possible that having it higher might end up with better output, but so far at least I've yet to see a way to get any model to do 100% of what I need up front without any need for me to make changes that end up being more tedious to do via interaction than by hand, and it doesn't feel worth spending a bunch more tokens trying to figure out how to better communicate to it up front how the dominoes get set up so they fall in place properly the next time.

How do you handle context limits? With more thinking tokens you fill it up earlier. Compaction degrades performance too. What's your strategy?

Initially I was planning heavily around context limits, but I've learned to just ignore it completely. Compaction is seamless for me. If details are lost in compaction, the model just re-reads what's needed. My conclusion is that at least for Sol, the summaries (which I've never seen) must be amazing. Every now and then a detail gets lost and I have to repeat it. I don't think there is performance degration, because the model is smart enough to re-read relevant files as needed.

AI-pilled obsession with "taste" is bordering on insanity

It's just vibes

It's vibes all the way down.

It's... really just vibes?

Always has been.

That's what they're always going to be, so not sure what would be "insane" about it. They literally feed on and emit natural language, and are put to work on informally defined, arbitrary tasks.

When people figure out any reliable strategies to test and benchmark them, that's insane, and in the positive sense. This very same issue has been a thing for humans as well forever, and remains only very questionably solved (IQ, academic tests). This is not easy.

This sounds about as unhinged as Google saying Material Design 3 is 30% more rebellious

Doesn't just sound like it, it is. That's life for you. That's what I'm pointing out to you. [0]

Just consider your own example. Do you think a less or more "rebellious look" is not something designers can actually ellicit? Less so in software design, sure, but in character design for example? Or general product design? Do you think e.g. Monster energy drinks are branded the way they are completely due to happenstance or something?

Except people don't usually put numbers to it, because they understand that that's hard to defend. You're the one who's describing such an idea, and wants such a thing to happen, classifying anything else as just vibes (that's the point!) and unhingedness. You're handwaving the difficulty and fundamentally limited nature of that, assuming that it is some laziness or mental delusion that's preventing it instead. You're also pretending as if it was somehow not real as a result. What I'm telling you is that you're wrong about that. Any kind of qualitative analysis that's actually defensible with these is genuinely difficult and limited in nature. See also all the opining about benchmaxxing. It also doesn't mean they're useless though, see also benchmarking.

The guy above didn't put numbers to his vibe assessment, they just drew a comparison, exactly because they know that there's not much else they can earnestly offer. You're sulking at them not lying to you by overstating their rigor, and you're flipping the arrow as if this limitation was some sort of mistake, not a necessary and intrinsic property, which it absolutely is. Natural language is an inherently subjective medium.

[0] In fancier and more mathematical terms: https://abeljansma.nl/2026/07/10/truth-is-not-a-direction.ht...

I think we’re saying the same thing?? It is literally just vibes and all this faffing over which model has better “taste” is pointless

Where you’re wrong is pretending there is any intellectual rigor to the discussion which justifies promoting from the domain of vibes to actual reasoned debate

We do agree that this is a vibes discussion, but disagree on whether that makes it a pointless discussion to have.

To give you a practical example, small, self-hostable models are very popular on HN. But my personal experience with them has been absolute dogwater, so this difference informs me that this is not the place where I should shop for a signal on whether a specific model like that is worth trying, or on whether it's game over for large models and remote models yet. I can "take the temperature" and make use of that without it having to be any rigorous, high assurance or mechanistic thing. It's suboptimal, but not useless.

Conversely, it also inspires people to try and substantiate these issues, so that it can eventually be more rigorous, higher assurance, and mechanistic. This is why I brought up benchmarks and benchmaxxing. Hard to know what points of consideration are salient when there is no abstract grievance to investigate, but Goodhart's law does also keep looming.

If people start moving away from Claude to GPT because e.g. Claude's output is too hard to work with and parse, that's relevant for the respective model providers, because it's a revenue share shift.

If I really like using Claude and strongly prefer its output stylistically, then claims otherwise will infuriate me, and will drive me to substantiate. If for no other reason, then because it's likely that Claude will have its language tuned in response to people's feedbacks and the revenue shifting, which may not be to my liking, and so I better prepare to call it and contend it.

It's literally like any other real world thing ever. Think tuning video codecs. Or tuning user journeys in frontend development.

It's easy to dismiss "taste" when you either have none or just fail to appreciate it, but nothing gives you an appreciation for the importance of taste like LLMs. There is no benchmark for taste, so while many things improve taste does not. Bad taste is, in fact, a huge component of what makes AI slop so sloppy.

But human coders can have bad taste too. There is code where there is nothing obviously objectively wrong, yet the choices feel like they were made by someone who just doesn't value or put emphasis on the right things, yet spends a lot of effort on trivialities. It comes in many forms.

Thinking the giant array of GPUs has “taste” is exactly the problem

You are confusing personification, which is purely a rhetorical device, with anthropomorphization. I am not anthropomorphizing GPUs. There is nothing particularly weird about personifying model weights, as we do with computer programs, cars, and all other manner of inanimate objects every day.

To be honest, in this case, I wasn't even personifying them, because I in fact didn't say that an array of GPUs has "taste", or in fact even that model weights did. I was saying I preferred GPT 5.6 Sol's outputs as a matter of taste. I can see why someone would confuse the two statements since in this case they're pretty much the same thing, but if you re-read what I said I was actually more careful than you're giving me credit.

things that are alchemical are rarely alchemy. That is to say things are very fiddly but stick a room of monkeys on typewriters, a schizophrenic developer with HolyC and adderall or an LLM, persistence is the key to many of these things like drivers, extracting keys from vintage security domains, etc. Dropping into xdd to a human is a chore, not for an LLM.

Although I am not exactly sure what you mean, I am not really claiming it is doing anything I couldn't do - but yes, it does so with much less effort. For example, I can have it set up probes and tracing on Linux that I personally would have to consult documentation to do. It might not even have to consult the documentation due to having the information on-tap, but even if it does, it's nothing that would cause it any fatigue, it's just going to keep moving forward in a loop until it is satisfied that it meets the criteria. I could've done all of this alone - I really could have. I just would not have. Being able to do something 10 times faster or with 10 times less effort is, in some senses, sometimes more impactful than being able to do entirely new things you couldn't do before.

This is a really good point that I definitely failed to grasp when first hearing about these tools. At least for me, the best way to use these tools is as a way to free myself from having to spend time thinking about the things that aren't worthwhile so I can focus on the things that truly are. I've had times in my life spending hours reading documentation and googling random things to try to tease out the correct sequence of commands or the exact right shape of an API to be able to make things work to know that it doesn't make me more productive to do that myself rather than point an LLM at the thing and let it spit out the answer after a few minutes. Meanwhile, I can spend that time thinking about what comes next, or what the correct way to take that one-off output and abstract it to something that can be used meaningfully in more flexible ways.

The only obvious objection I can think of to this line of thinking (at least from a technical perspective) is "how does someone build up the knowledge to be able to use a tool effectively in that way if not by doing things by hand at first?" The honest answer that is "I don't know, but that's also pretty much exactly the type of thing my employers have never been paying me to solve in the first place". Even just a decade into my career, there have already been plenty of times in my career I've struggle to convince people that we should do stuff in a way that won't bite us in the ass a month or two down the line, and in the times I've managed to succeed, it's usually only by putting in more of my own time and effort to make the initial investment seem more palatable. Luckily right now I'm not in one of those times when I'm having to go full throttle to keep the lights on a few months from now, but I don't have enough fuel in reserves to work on a plan for when we need to build a new rocket in another ten years. Maybe ask me next month.

Terry A Davis wouldn't have touched LLMs. He'd probably claim they're demonic.