Hacker News

I’m starting to realize that LLMs are really good at building low-stakes projects. Your questions mostly presume that the stakes are higher. The software will last a long time; the requirements will evolve; we can’t tolerate mistakes; etc.

The trick to getting good at using LLMs for software is to learn how to make _all_ projects low-stakes.

These posts are never written by software engineers, it’s always some tech exec, retired engineer, or VC. This author is apparently a professor at the Wharton School of Management? None of these people have to ship or maintain real products, they’re just making side projects.

The only decent software engineering perspective I’ve seen has been from Mitchell Hashimoto.

Being the first to release an article gives you great SEO or whatever. Doing the things you've mentioned takes time.

Less fascinating when you consider that this is a non-coders perspective.

>What I find fascinating that there is so little substance in this article about the quality of produced code and the medium.

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"?

Does it matter to the people requesting the software if it acts in the way they expect?

It's an ad.

I'm becoming more convinced these are questions of the Before Times. Yes, yes—heresy, I know.

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.

> 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.

now for the best question: whats your ROI here?

Just depends what you are working on. If you are trying to make a video game that's at a level of a decent indie game (think Hades/Baazar/etc), making UI elements/VFX/complex shaders/etc that are organic/interactive/animated that don't feel like a little dogshit vibeslop web-game, then none of the models are even close to good enough to get it done easily. Huge percentage of problems in top 3% games is really hard for any of the models to do with simple prompting.

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.

I’ve been working on implementing some common web infra type projects in Rust lately. Basically trying to use a lot of the great primatives in Rust like rustls (modern openSSL) and Tokio (async) to build memory safe or close, nginx drop in replacements.

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

We see the same thing when new laptops are announced and every employee all of a sudden needs to upgrade, despite the fact that 90% of people would be able to make do with a Macbook Neo.

I had a few of the benchmarks left alone and was working on tech debt knowing that a new model is going to be released soon. For my project (tsz.dev) Opus 4.8 was running in circles without producing results for a while for those tasks

In Claude's defense (and I cannot believe I'm defending it), I know no single dev who could create what it did (Concord), from a 19-page design document, in 9.5 working hours.

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."

> At the same time, it is very dissonant to see the industry heading towards hour+ long workflows with an agent.

At this point, pay me significantly more, and I'll do it.

My Opus 4.8 regularly works for 10+minutes on a single non-trivial coding request.

Work duration is also not that valuable of a measure, you're usually better off defining the process yourself in code and having that delegate chunks of work to the models. The only real issue there is that it's harder to take advantage of the providers' subscription discounts, but on the other hand it's easier to do your own model routing, and there's no way I've seen for the normal chatbots to maintain coherence on streams of work measured in days and weeks.

I think we hit the sigmoid back when the QWEN models were released. By properly structuring my project, I can point it at any extension I want and get it going for 30 minutes to extend whatever. It can't effectively do 'god mode' on all the code, but being a mindful observer and code "professional" I don't need more than what a 128GB VRAM needs.

I'm amazed we're so far into SOTA bloat that the chinese will kill once they start etching silicon with these models.

It looks interesting but, like a lot of AI, looks correct but is not. Most of northwestern Canada says you can get there by road. If you look at Google Maps, there's no roads there for quite awhile. I see one highway between Inuvik and Tuktoyaktuk but that's about it.

It is hallucinating many flights in my region, some that never existed (so it is not an outdated data problem).

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.

Does this not indicate that the project might not be structured in an appropriate way that allows incrementally adding features?

> 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.

How good can a rhyming poem about a haircut where every word starts with S be?

I wonder what Vogons would think of it.

Seems to be hijacked the video of some game they generated. :(

Depends on the skill of the person working on it.

> faster to go from London to New York in 1881 than Volgagrad

the map is for 2026, yeah?

So...no engagement with the substance? Not even to explain why it is that this is not a useful description or test of capabilities? Ok.

Instead of attacking the author, please respond to the content of the article. That is the HN way, and it leads to more substantive and interesting discussions.

Ethan is a booster but I wouldn't call him a shill. He cites data and mostly in a fair way, though you could argue the sources he chooses to focus on are biased.

Would love to see samples of the kinds of prompts you use with both. I sometimes wonder if the specific wording is the secret sauce, I have very few issues with Opus / Claude, but when I try premier GPT models, I get weird output from what I've grown to expect with Claude.

Huge fanboy for sure

Fable is Mythos with extra guardrails, so the analysis holds.

The mythos of Mythos is marketing.

It is not a sponsored article and he writes one of these every time a new model releases. Why would a professor at Wharton need to write sponsored Substack articles.

"I don't care who the IRS sends I am not paying taxes!"