Codex, Opus, Gemini try to build Counter Strike
instantdb.com
instantdb.com
// wavelength of used primaries, according to preetham
const vec3 lambda = vec3( 680E-9, 550E-9, 450E-9 );
// this pre-calcuation replaces older TotalRayleigh(vec3 lambda) function:
// (8.0 * pow(pi, 3.0) * pow(pow(n, 2.0) - 1.0, 2.0) * (6.0 + 3.0 * pn)) / (3.0 * N * pow(lambda, vec3(4.0)) * (6.0 - 7.0 * pn))
Codex: https://github.com/stopachka/cscodex Gemini: https://github.com/stopachka/csgemini Claude: https://github.com/stopachka/csclaude
GPT’s differentiator is they focused on training for “thinking” while Gemini prioritized instant response. Medium thinking is not the limit of utility
Re: overparameterization specifically Medium and High are also identically parameterized
Medium will also dynamically use even higher thinking than High. High is fixed at a higher level rather than leaving it to be dynamic, though somewhat less than Medium’s upper limit
Edit:
https://github.com/vorg/pragmatic-pbr/blob/master/local_modu...
https://github.com/vorg/pragmatic-pbr/blob/master/local_modu...
This looks like where the source code was stolen from: this repository is unlicensed, and this is copyright infringement as a result
E.g. the latest Anno game (117) received a lot of hate for using AI generated loading screen backgrounds, while I have never heard of a single person caring about code, which probably was heavily AI generated.
Unlike your results which aren't exact match, or likely even a close enough match to be copyright infringment if the LLM was inspired by them (consider that copyright doesn't protect functional elements), an exact match of the code is here (and I assume from the comment I linked above this is a dependency of three.js, though I didn't track that down myself): https://github.com/GPUOpen-LibrariesAndSDKs/Cauldron/blob/b9...
Edit: Actually on further thought the date on the copyright header vs the git dates suggests the file in that repo was copied from somewhere else... anyways I think we can be reasonably confident that a version of this file is in the dependency. Again I didn't look at the three.js code myself to track down how its included.
If there's any copyright infringment here it would be because bog standard web tools fail to comply with the licenses of their dependencies and include a copy of the license, not because of LLMs. I think that is actually the case for many of them? I didn't investigate the to check if licenses are included in the network traffic.
Sure. It's a problem that corporations run by more or less insane people are the ones monetizing and controlling access to these tools. But the solution to that can't be even more extended private monopolistic property claims to thought-stuff. Such claims are usually the way those crazy people got where they are.
You think in a world where Elsevier didn't just own the papers, but rights to a share in everything learned from them, would be better for you?
https://tommyhinks.com/2009/02/10/preetham-sky-model/
https://tommyhinks.com/wp-content/uploads/2012/02/1999_a_practical_analytic_model_for_daylight.pdf
Nobody is going to sue you for using their implementation of a skybox algorithm from 1999, give us break. It's so generic you can probably really only write it in a couple of different ways.
If youre worried about it you can always spend a day with Claude, ChatGPT and yourself looking for license infringements and clean up your code.
For personal use maybe not, but that's not the point, the point is it's spitting out licensed code and not even letting you know. Now if you're a business who hire exclusively "vibe" coders with zero experience with enterprise software, now you're on the hook and most likely will be sued.
That's not to say there hasn't already been more direct cases with set examples [1], from an author directly who would have a better right to claim than I [2], it's not even a stretch to see how it can happen.
If true, then this usage could violate its MIT License: "The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software."
The file seems to have been copied verbatim, more or less. But without the copyright info
How would you know? Do you have another AI scan for copyright violations? In terms of a false negative how are disputes resolved?
Seems like a massive attack surface for copyright trolls.
If you think any court system in the world has the capacity to deal with the sheer amount an LLM code can emit in an hour and audit for alleged copyright infringements ... I think we're trying to close the barn door now that the horse is already on a ship that has sailed.
If MegaCorp has massive $$$$, but everyone else has small $, then MegaCorp can sue anyone else for using "their" code, that was supposedly generated by an LLM. Most of the time, it won't even get to court. The repo, the program, the whatever-they-want will get taken down way before that.
Courts don't work by saying, "oh, but everyone is doing it! Not much we can do now."
Someone brings a case and they, very laboriously, start to address it on its merits. Even before that, costs are accumulating on both sides.
Copyright trolls are mostly not MegaCorps, but they are abusers of the legal system. They won't target Google, but you, with your repo that does something that minorly annoys them? You are fair game.
No, but they do recognise when their case registrations are filling up in a way that they cannot possibly process and make adjustments. Courts do not have an infinite capacity.
There's a really simple solution that you may not have considered:
1) don't put your vibe-coded source code in a public git repo, keep it in a local one, with y'know, authentication in front of it;
2) regularly ask your agents to review the code for potential copyright infringements if you either want to release the source or compiled code to the public at any point.
As long as you've followed best practices, I can't see why this is really going to become an issue. Most copyright infringements need to start with Cease & Desist anyway or they'll be thrown out of court. The alleged offender has to be given the opportunity to make good.
So "Claude, we received a C&D for this section of code you stole from https://.../ , you need to make a unique implementation that does not breach their copyright".
You will be surprised how easily this can be resolved.
They kind of do. If you fail to bring legal action to guard your intellectual property, and there’s a pattern of you not guarding it, then in future cases this can be used against you when determining damages etc. Weakens your case.
Downvoting won’t make it untrue lol.
Trademarks can become 'generic' if you don't defend them. But JK Rowling wrote Harry Potter, whether she sues fanfic authors or not, and can selectively enforce her copyright as she likes.
Seems fine given the project is already using threejs and so will have to include license info for it already.
I remember when CS Pro Mod was being made between the transition of CS 1.6, Source, the 1.6 community didn't want to move over to Source, before GO/CS2 came around.
Cool to see what's basically Quake1/doom style but this is a far fetch away from counter-strike. Although if netcode could be imagined and implemented I don't see why making a lower tier Counter-Strike wouldn't be doable. I'd play it if it were the quake style old-graphics version of CS that allowed for skill gaps.
Great article, love the nostalgic feeling.
I’d also love a Battle-bits CS version. (Battle-bits was a fun Battlefield low poly spoof).
Creating software is indeed fun, but the most enjoyable aspects are the "a-ha" moments after you overcome a tricky problem, the confidence boost from creating something that works in an efficient and elegant way, and the dopamine hits associated with those events.
"AI" tools can alleviate some of the tedium of working on plumbing and repetitive tasks, but they also get rid of the dopamine hits. I get no enjoyment from running machine-generated code, having to review it, and more often than not having to troubleshoot and fix it myself.
To me, creating software is not as much about the destination, but about the journey. About the process itself. Yes, some of it is not enjoyable, but overall, there is much more I like about it than not.
It's the same when I hear people complain about how complex new UI frameworks are. The web still runs perfectly well on simple html, CSS, and Javascript. There is not federal police force that will arrest you for not using React.
Yes, I can do it. In my free time. But that part of my job that was enjoyable? Poof. Not anymore. Can't compete, get on with times, be more productive.
I spend a 40% of my "alive" time in work. It's a massive downgrade.
Companies preferring React over vanilla Javascript != you can't build sites with vanilla JavaScript anymore. Sites LITERALLY still work that way.
> I spend a 40% of my "alive" time in work. It's a massive downgrade.
This martydom with front-end frameworks is crazy to me. Guess what? You're a software developer. You actually have a lot more power tahn you think. And this "roll over and play dead while whining about every advancement in technology because you feel left behind" is exactly the reason you feel the way you do.
You are the only one that even mentioned FE frameworks. Or plain JS. Neither was even tangentially mentioned and has nothing to do with my comment.
I have never heard a client say "Man, glad you used React". Literally nobody cares what framework you use to build your site. Nobody.
If you didn't know any better you'd think all software developers are chained in a basement where they have absolutely no power to do anything but build React sites.
1.) Those WordPress gigs can make your React gig look like indentured servitude
2.) The company you’re applying to isn’t the client.
3.) “freelancer or something”, like you’re spitting it out? Yes - some of us aren’t handcuffed to mediocrity by 200-1000 person orgs. As the kids say, “Don’t hate.”
1. Definitely not in my country. The average pay of a Wordpress/PHP dev is half of a modern full stack and the clients are terrible, because it's just websites for small businesses. Modern full stacks don't create websites most of the time, but highly interactive B2B apps.
2. It is absolutely my client. I optimize their happiness not their customers. I have no relationship with the customers, some don't know who I am.
3. I worked as a contractor for a couple of years and I'm not missing the stress and unstable pay. Especially now with a kid on board. Many contracts were actually "hey we need a React/[insert other tech] guy for our current project, wanna join?", not "we have an idea and we don't know how to do it" kind of thing. The latter are super rare and even more stressful, because they come from "non-technical startup founders" often with little money.
Keep in mind that I'm in EU, so the benefits of permanent employment make a huge difference.
As for “it-compiles” that is nothing new. I have written code that I go back to later and wonder how it ever compiled. I have a process now of often letting the agent prototype and see if it works. Then go back and re-engineer it properly. Does doing it twice save time? Yeah, cause who’s to say my first take on the problem would have been correct and now I have something to look at and say it is definitely wrong or right when considering how to rebuild it for long term usage.
1. More people that wanted to make games can.
Thanks to unreal engine, you don't need to be a Tim Sweeney level-expert to make compelling games. I see LLMs as another abstraction in the same spirit.
2. You get more leverage
The more abstractions you have, the more you can do with less. This means less bureaucracy, more of a chance to make _exactly_ what you wanted.
I understand how the craft changes underneath you, and that can feel depressing, but if we see it as tools, I think there's lots of good ahead.
I could be wrong of course, and it may be true that your work will change very soon. Maybe someone else has better examples to propose ?
Furthermore, if you have it sandboxed, you can also ask it to also install any necessary dependencies or toolchains, which is really nice.
Now show us the cost, the time it took, and how much babysit... sorry, "human supervision" was necessary.
https://www.youtube.com/watch?v=fm-OoCWQlmc
The only time I spent outside of the video was to deploy to Vercel. I made a bunch of speedups in the video, but didn't cut anything. The total time was about 2 hours.
I mentioned it in the post, but there was definitely some hand holding towards the end, where I don't think a non-programmer would have succeeded
it's not something that would ever work industrially
people with code-generators they've made could do this just as fast as the AI except their generators could have engineering considerations built-in to them as well so it'd be even better
Code generators? Can you be more specific?
There's also the practice of using good ol' fashioned code-generation tools like T4 or Moustache/Liquid templates to generate program entity classes and data-access methods from a DB schema, for example. Furthermore, now there's pretty nifty compile-time code-generation in C# - while languages like F# support built-time type-generation.
...and these are all good tools IMO; but really aren't comparable to an LLM, imo.
> Codex, Opus, Gemini try to build Counter Strike
Even though the prompt mentions Counter Strike, it actually asks to build the basics of a generic FPS, and with a few iterations ends up with some sort of minecraft-looking generic FPS with code that would never make it to prod anywhere sane.
It's technically impressive. But functionally very dubious (and not at all anything remotely close to Counter-Strike besides "being an FPS").
Fitting.
the code and output is literal slop
it's not known how much editing and debugging was done by the team either
you could have done this in 2022 with not that much debugging as well
https://github.com/instantdb/instant/pull/2010
Once this lands lightbox should be up. Thank you!
In other words, this is slop. We know these new models can generate slop images, text, videos, and code. Sometimes slop can be useful; maybe you can shape it into something useful, maybe you can slop a slopper. But we're learning it's not economical--this is some of the costliest slop we've ever made.
AI coding needs someone behind to steer it to do better, and in some cases, it does. But still hasn't left the junior phase, and while that doesn't happen, there's still the need for a good developer to deliver good results.
They're not thinking or reasoning or understanding. It's just amazing autocomplete. Humans not being able to keep themselves from extrapolating or gold rushing doesn't change that.
They are. I know a lot of people don't want to admit this, but they are. They're getting better with each release.
> But we're learning it's not economical--this is some of the costliest slop we've ever made.
Huh? How on earth would you know whether my usage of LLM's has been worth it or not?
> Sometimes slop can be useful; maybe you can shape it into something useful
Man, I just spent the last 2 weeks with a CEO who got a Bolt.new subscription to be able to generate some high-level mocks ups for me to utilize that just saved us months of back and forth.
You know what's the best part? Those same mockups can be used to gather user feedback with a functioning UI without me having to spend weeks building it and it ending up wrong anyway.
Sometimes it irks me, but now I've sorta come to embrace devs like you. You're guaranteeing I have a job because you refuse to acknowledge the very obvious thing that's happening.
I’ll take that bet.
"Now let's make shots work. When I shoot, send the shot as a topic, and make it affect the target's HP. When the target HP goes to zero, they should die and respawn."
This is not how shooting is implemented in a competitive first person shooter.
If you don't understand how a multiplayer FPS works, how can you tell if the AI has actually created one for you or not?
There is a difference.
I interpreted that as "I asked the AI to make me one, and now I have one, so it did it," with the operative question being "is it really a multiplayer FPS?"
My bar is, does this actually work and is it best practice for how first person shooters are made by professional game developers.
Your bar is, does it kinda look a bit like the thing that I play?
Totally different things.
Also, Unreal source code will be the very last thing LLMs understand. This is the most complex software ever.
There’s an algorithm called Nanite for automatically reducing the triangle count on geometry that’s far from the camera. As in there are not manually made separate level-of-detail models. The algorithm can modify models, reducing quality as they get farther.
This one algorithm is a tiny piece of the engine yet has a 1,000 page white paper.
Also, even when I don’t know how something works algorithmically, usually I at least have some intuition about where to start. I haven’t the slightest idea how to approach this problem.
No way. Take baby steps. Write an operating system first. Write a compiler first.
[1]: https://jms55.github.io/posts/2024-06-09-virtual-geometry-be...
Edit: I don't mean to sound disparaging - it's some genuinely cool algorithms. It's just that Epic is incentivized to hype it up, and so you get a huge paper and multiple talks that are designed to make it seem even more impressive.