Benjamin_Dobell

Probably worth noting, yes, I'm aware if you're running atop their cloud infrastructure, OpenAI already technically have access to whatever they like. How would you feel about VPS providers enforcing this same policy?

Yes, I'm also aware that if your data goes to their models for processing, you're of course already sending them your private data too. However, you may have noticed, Codex and the like actually go to pretty great lengths NOT to upload things like secrets to the model. This is the polar opposite approach.

Isn't the idea of having a cloud computer meant to be that it's more than just a model? It provides compute, so it can run scripts locally to filter data etc.

What browser are you using? IE6?

Just kidding. But you must have clicked somewhere. Audio autoplay hasn't been a thing on the web for years (thankfully).

It's extremely useful if your distribution platform is the web. However, I find it much more valuable as a type checker than a compiler. I use TypeScript extensively (in several projects) for defining schemas types.

For example, in the Breaka Club (https://breaka.club/) editor, which is not presently exposed to kids yet, we define behaviors for characters, props, mosaics (terrain tiles) and items in JSON. But the JSON has type checking and real-time completion as you type. Not naive property auto-suggestion, we have effect types and they're context aware in that you can refer to targets introduced higher in parent constructs of the JSON — and they're fully type checked. If you refer to a target that does not exist in a context, it won't validate and you cannot save the schema.

We use tsgo for development, but the editor (runtime schema validation) is presently stuck on TypeScript 6 because there's no WASM support.

Now, is it nutty that we're using TypeScript for JSON validation? A little. But it's extremely powerful. We're going far beyond what's capable with Zod or ArkType.

Professional looking product / page. However, the shaders themselves are extremely odd. They're just swirling patterns. I think the creator of this would do well to look at games for inspiration. Games use shaders for all sorts of effects, for a reason.

In particular, I'd expect there to be shaders making use of textures. I was expecting buttons with a laser finished metallic look that actually changes with lighting on the page. Or hero images that combine photography with masks to produce interesting effects highlighting products with auras etc. These all appear to be screensavers.

The context window is configurable. I've been using ~600k for months. No, not API pricing, on a Codex sub.

~/.codex/config.toml

  model = "gpt-6.1-sol"
  model_context_window = 700000
  model_auto_compact_token_limit = 630000

I'm definitely getting more than 4 hours of Astra usage (medium or high, typically). However, I predominantly use Sol or Luna because Astra simply isn't (or shouldn't be) necessary for every task.

That said, I'm not able to get anything useful done in those hours. The exception is when I swap back to 5.6 Sol/Luna, which seem to be working as they used to.

Mostly, I've just given up and am using Opus 5.5 for all the important tasks now. Which is odd for me say, because I cancelled my Anthropic sub a few months back because I couldn't stand Opus 4.6 < x < 5.5. 5.5 feels great though, like Astra did when it first released.

Codex GPT-6 models are so congested and slow at the moment it is difficult to get through the current 20x plan anyway.

Yeah, I tried to read this because I'm genuinely interested in this topic. I assumed going in LLMs would be involved, and I don't mind at all. But my gosh reading the article is such a slog. I couldn't do it

Phwoar! Thank you. It's 2026, and had it not been for your high quality informative comments. I would have not known anything about this them there Facebook. You saved me.

I don't know what conversation you think you're having here, but it's not that.

If you'd bothered to read... and then ready for it... stop, take a breath, and process the information. You'd have noticed I was pointing out you can use a Quest headset WITHOUT providing them your personal ID OR having a Facebook account. Almost as if, you know, I was highlighting that giving the company your personal info isn't required. Woe me though for pointing that out.

I'm sorry for interrupting the indignant rants. Please continue.

Well, I was going to point out the distinction between products. But I guess if people like you are able to turn HN into Reddit, anything is possible.

They have Meta accounts now, which are also not Facebook accounts. I went straight from an Oculus account to Meta account. Might be location based, but I most certainly haven't been asked to upload any ID.

Still plugging away teaching kids to code and hand draw their own computer games:

https://breaka.club/blog/why-were-building-clubs-for-kids

Have been running an in-school pilot over the last school term. Has been super insightful and led to a heap of changes to optimize UX. However, the biggest challenge has been delivery in this environment.

Trying to run a game in a browser on school iPads powered by really locked down networks has been challenging. Some school iPads will crash (out of memory) when launching the camera view, whilst others run flawlessly. We use a fork of Godot, and are constantly fighting to strike a balance between UX (increased parallelism) and browser device constraints i.e. a limit on how many web workers we can fit in memory.

In addition to our main (build and play your own RPG) experience, we also stream games direct to iPad. In particular a modded version of Overcooked! 2 that teaches kids to code — Overcooked itself was never even released on iPad. We've built our own Kubernetes system that spins up (a license limited number of) games/pods on demand backed by (time sliced) NVidia GPUs. This is literally running out of my home office. Despite the complexity of building all that, the Kubernetes and GPU time slicing wasn't even the largest hurdle. Turns out UDP is a no-go on the school network, so direct WebRTC was out the window. Instead we're using Cloudflare's TURN relay with TCP. Which much to my surprise, is holding up quite well.

Definitely a learning experience. However, kids are having a blast. Can't wait to roll this out further!

Not based on an official Blockly implementation. But you can add keyboard navigation to Blockly(-like) scripting reasonably easily:

https://www.youtube.com/watch?v=ITWSL5lTLig

The video above shows keyboard, but the main motivation for the key-based navigation was actually gamepad.

This is extremely interesting to me. I also have an 8 year old daughter who inspired me to build tooling for kids to create their own games. However, what I've come up with is quite different than this.

https://breaka.club/blog/why-were-building-clubs-for-kids

The primary mechanism through which kids create their game worlds is taking photos of their own hand drawn artwork.

I found traditional game dev tooling was far too technical and verbose for my daughter to stay interested. Honestly, it's not just a matter of her losing interest due to the complexity. Frankly, I don't want her sitting in front of a screen that long. She's a kid and I want her doing typical kid stuff. She and her younger sisters play sports, draw, paint, perform magic shows, and sing and dance around the house like complete nutbags — and I love it!

So whilst what I've built is highly technical for me to maintain, I'm not trying to expose all that complexity to kids. At least, not yet. I'm mostly interested in fueling creativity and letting kids explore. That's not to say kids that want to dive deep can't — we offer visual coding experiences etc. The intent is kids gain access to new features through self directed exploration and play. If some kids stop at just play, that's OK. The primary game mechanics (drawing and building worlds) are still inherently creative.

Although my approach is different, there is a heap in this article I resonant with. The hacking mentality, and letting kids go beyond the artificial constraints in particular. There's still much of Breaka Club that I can improve on. I'm running a limited in-school pilot at present and that's already yielded some really interesting insights. I'm really keen to continue to see what others put out in this space. I think it's drastically underserved and there's plenty of room for multiple approaches.

This looks really nice. Given the target age I imagine there's not a heap of text, so you could probably get away with LLM translations. Be sure to share it again in the coming months if you add English :)

This is precisely why I'm building Breaka Club:

https://breaka.club/blog/why-were-building-clubs-for-kids

I also attempted to make a game with my eldest daughter initially. Unsurprisingly, she didn't want to sit there watching me code or drag things around in Godot's level editor.

Admittedly, I really need to make a new blog post showing the latest improvements brought about from the in school pilot I'm running weekly at the moment.

Been dealing with this a bit at Breaka Club. We teach kids to code with a modded version of Overcooked 2!.

We stream OC2[1] with our mod preinstalled over WebRTC. This ensures that kids/schools don't have to try and install the mod. This is particularly important since we support running on school provided hardware. Installing a game without a mod would be hard enough. Added advantage though is kids play with a virtual (on screen) gamepad on iPads in Mobile Safari.

Game instances run in Docker containers in Kubernetes/k3s atop very outdated nVidia hardware. Given we're already going across the Internet into school networks, we've tried very hard to optimize latency across the board. Using NVidia NVEnc with DMABuf (zero copy) etc. We're unfortunately using XWayland at present so experience the documented input overhead. Although our inputs are virtual devices at this point, so the overhead may be a bit different. Trying to optimize this whole thing end to end has been a challenge. I would say that performance is currently "acceptable".

OC2 coding: https://www.youtube.com/watch?v=ITWSL5lTLig (not streamed in this case)

[1] We've bought a limited number of copies of OC2 and pods claim a license on startup. If we're at capacity, kids play something else.

They are. There's a LOT of web games on itch.io. It's just the typical challenge with over saturated markets: publicity.

It's a really tricky one, but I think it's the right call for Godot, but probably not for other projects. I'm the current maintainer of GodotJS (TypeScript bindings for Godot). LLMs can generate genuinely useful solutions for Godot e.g. a functional WebGPU implementation for Godot — https://github.com/godotengine/godot-proposals/issues/6646.

However, as a senior engineer with fairly deep technical knowledge in these areas, I'm now using AI in my own projects. Frankly, before even touching Github, I'm already drowning in code reviews generated by my own use of an LLM!

There are meaningful, generally good PRs waiting for my attention against GodotJS, and other projects I (somewhat) maintain e.g. MoonSharp, C# Lua runtime. It was already extremely difficult to stay on top of PR code reviews for reasonably technical projects. When you're already somewhat burnt out from reviewing LLM code all day, it's so much more exhausting than it used to be.

Still plugging away at Breaka Club, where kids take photos of their hand drawn art and build games using it. Starts out as no-code, photograph an AprilTag and it imbues the image with functionality.

https://breaka.club/blog/why-were-building-clubs-for-kids

We also teach kids visual scripting in Overcooked 2!, allowing kids to code their way through the levels of an existing much beloved game:

https://youtu.be/ITWSL5lTLig

I'm running an in school pilot this week (Lunch time school club).

The tech stack for the main product is honestly pretty intense at this point with full multiplayer support, offline play, transitioning from client authoritative to joining a remote server. Built atop GodotJS, TypeScript bindings for Godot, which I maintain. Huge monorepo with over a million lines (yes, I'm aware that's NOT a good thing), and GodotJS itself is not included in that.

It's really not.

If I'm at the point of contributing a PR to a dependency, I've already identified the root cause in detail. There's no way a change log should be going into that level of detail, or else you're just duplicating the Git log for no reason.

Will the change log make mention of fixing the bug? Perhaps. But I'm going to want to read the technical details of the fix to make sure they've specifically addressed my issue, and not just a similar problem. What is the performance impact of the fix? Are there security implications they've explained in the commit message.

I'm a software engineer, not an end user, I want the technical details of my dependencies.

How... how is this not obviously the absolute very most useful information?

When I encounter a bug in a dependency of mine. Before I worry about submitting a PR, the very first thing I do is grab my version number and check the commit logs for fixes since my version number.

If I'm trying to decide whether I should bother upgrading, I scan the log for new features.

It's the title, not the details. The commit message body should contain MUCH more detail than the title.

If you don't like it because it looks ugly. Sure, that's subjective. And actually, I agree. Because it's standardized though, Git interfaces could even be configured to trim this off and provide different visual styles for the different kinds of commits. The types could be used as search filters too etc.

Now, I get people don't like the look of them. Neither did I when I first saw them. Then I started using them and found them useful.

It's fine, people have different preferences, it's just a convention and it's not going to work for every project. The article itself just doesn't seem to hold any water.

My apologies, I missed this on first read due to the indentation style. That said, I don't agree on the commentary.

Why on Earth are people not writing commit messages for their reverts? They should have semantic commit messages just the same as any other commit.

Unless the point is that they're not following per-commit CD, and if you commit then revert that commit before a release was made. That sounds like a process failure. Which of course, process isn't infallible, and neither is the automated version management. If you screw up, use an escape hatch — just like reverting a commit that had previously gone through code review and been merged.

Re: change log generation. The article says change logs shouldn't have commit messages. I agree. Many tools (e.g. Changesets https://github.com/changesets/changesets) use the semantic commit type to sort change log entries, but require you to write those user facing change log entries separately.

Odd. The main reason to use this style of commit message is for CI/CD automation. I didn't see this mentioned anywhere in the article.

The type of the commit informs the automated workflows how to handle the commit. This is why it comes first.

For example, if you're performing CD, if you only commit a bunch of `fix: ` then only your semantic versioning patch version number is incremented. If you commit a `feat: ` then it's a minor version is bump. `feat! ` is a major version bump.

Even if you're not using CD for releases, semantic commit messages are sometimes used to automate change log generation. Granted, your change logs should not typically include the Git commit messages themselves — those are developer facing, not user facing.

I'll definitely give you Rosetta, and even more so Rosetta 2. Spotlight too, at least in principle, but it has had its fair share of dodgy behavior over the years. I'm not really sure about the others.

There's certainly always been fantastic software available for Mac. However, it was almost never built by Apple. It sort of felt like someone one day needed a FireWire port, so they bought a Macintosh. Then they must have told a close friend working at Macromedia they needed some software - and it was all just inertia from then on.

Live rankings currently have Mac OS X first... h... how?

Apple make so so much wonderful hardware! They always have. Their software on the other hand is near universally awful. I love my Macbook, but my gosh, I do not love whatever the latest flavour of macOS is that Apple have decided to throw on their update servers this year. It just so happens that I also enjoy Unix, so I spend a lot of my time in a terminal - but Apple don't get to claim credit for that!

EDIT: OK. It just refreshed and is now showing Mac OS X as 36th over all. Crisis of faith averted.

I must admit, this is one area I've found LLMs to be surprisingly strong. They're REALLY good at reverse engineering obscure platforms, languages, game engines; and quickly throwing together super hacky tooling.

I was able to reverse engineer the PS4 edition of "New Game!: The Challenge Stage", which was never released in English. I've now fully translated it, added proper text wrapping and additional text boxes where text would now overflow. Along the way I've fully decompiled (with byte exact recompilation) the Squirrel scripts for the entire game, built atop the game engine of a now largely defunct game studio. Prior to this I hadn't even heard of Squirrel scripting language. I had most of this done in under 24 hours.

I'm not in any way a part of the visual novel community. I just did this because I enjoyed the New Game! anime way more than a near(?) middle aged man probably ought to.

P.S. My condolences to Yorhel's friends and family.

Thanks, Eric; for this, and for my start as a software engineer — my first commercial development work was consulting as an 18 year old building games with C4. I'm really glad Slug was able to find commercial success for you in the way that C4 unfortunately wasn't able to.

For those of you who aren't familiar with Eric's work, he's basically the Fabrice Bellard of computer graphics.