What I really want is the ability to do more operations which are decoupled from the editor. Even just being able to reliably type check Unity C# code without a running editor would be great, or running unit tests. Would really improve things like CI as well as the ability to run parallel development sessions.
Headless execution exists and is one of the simplest ways to get at this experience.
The biggest issue with unity is that the editor (even when headless) is designed to take an exclusive lock over the physical project directory. This can be solved by doing a robocopy of the project to a temporary path, deleting any lock files, and then running headless execution there. You can keep the same temp path(s) over many iterations with incremental sync so it's not copying the entire universe each time.
I did that for my Unity games like a decade ago, you can too :) Requires a design that has one core and then using Unity basically as a renderer/viewer though, but is a relatively easy design actually. And yes, automation and DX gets so much easier. I think Captain of Industry is built like this too, remember reading about it here on HN from the developers in some comments.
Same here, in Godot! This is a common practice for games with performance/networking constraints too. For example building a deterministic sim in a regular engine tends to fail without custom primitives or a nice library to handle some of the intricacies. By moving non-display code out of the engines you can reliably build and control all of this.
They do disclose that they scored much lower than Fable on SWEBench Pro, which is a pretty high-quality benchmark. I think it's partially just about what they choose to emphasize...
People will downvote you because this comment is "not appropriate" for HN, but there were countless conversations on HN about how important these benchmarks are.
There has been a lot of chatter ever since the Mythos scores had been release that SWEbench pro had major contamination and that Mythos had memorized many questions that lacked the context to be solvable on their own. And now with OpenAI saying a large number of the questions are broken, I think it's worth taking that single outlier benchmark with some salt when the overall trend is that 5.6 is very competitive with Mythos at about half the price.
I totally missed that, because in the charts they showcase for coding, the SWEBench score is not present, they only include it at the end of the post in tables. Hmm.
The SWEBench benchmarks are really gamed at this point and should not be trusted period. The solutions are effectively in the training sets and have been for a while.
Then we are left with what? FrontierCode maybe? IIRC that one evaluates not only if tests pass but also code quality - e.g. whether the maintainer would accept the pull request as is.
To clarify, the more accurate description would be "Testing how well LLMs can follow the rules of Magic", right? There is no actual evaluation of how "well" they are playing?
This is basically just the normal dynamic with American tax law where tax jurisdictions are terrible at coordinating, so they end up approving things and agreeing to tax things at a very low level in order to win the competition. Even when states/counties try to work together on this stuff there's a huge defector problem, like "hey I can back out of this multistate tax compact agreement and get 500 new jobs which will let me win local reelection".
I suppose you can reduce a lot of both good at bad things about the country to "because federalism".
I wonder how Paradox handles this stuff in their games like Europa Universalis. Have they ever made a "the pope wants you to switch calendar systems" event which changes the actual in-game date?
Sounds like they just ignore any calendar weirdness. I don't envy anyone the task of trying to properly program this. You basically have to decide for everything time related in the simulation whether it is tied to the calendar or the physical passage of time. Sometimes you would have to split variables, like characters are legally older, but no closer to old age. Bonus points if your solution allows characters to get into conflict with each other about stuff like when a treaty ends or whether their loans are due 15 days sooner.
My initial thought was that that would probably be taking the simulationism too far. My second thought was how funny it would be to swap calendars to get out of a truce early.
“My Lord, you can attack them until March of next year!”
> My second thought was how funny it would be to swap calendars to get out of a truce early.
Congress did something similar last year - National Emergencies declared by the president, such as for tariffs, are required by law to be reviewed within 15 days by Congress. So Congress redefined the calendar so that the remainder of the days during that session are not officially considered to be calendar days. https://www.nytimes.com/2025/04/10/us/congress-johnson-calen...
EU3 and EU4 at least had no leap years. I assume 1, 2 and 5 are the same.
Checking the wiki, eu5 has an advance (guess these are like the nation ideas in eu4?) for Julian calendar which gives you +10% to orthodox or miaphysite nations. I doubt it has any effect in the calendar system in the games UI.
Paradox's Clausewitz game engine seems to handle "negative" years very poorly, so stuff like historical Roman emperors in Crusader Kings have some oddities. It's probably also why Imperator: Rome uses ab urbe condita dating (aside from immersion).
The Elder Kings 2 mod actually has custom date handling (it's actually a custom date localisation system) to enable transitioning to the 3rd era on founding the Empire of Tamriel.
I really fundamentally do not understand what problem Gas City solves that is not already solved by normal subagent orchestration patterns. If you want to call your main LLM session the "mayor" and have it delegate its work out to planners and coders and reviewers and QA and so on, this is already a thing you can do! If you want to do this in a reusable way you can create skills and subagent definitions and use /commands, etc. Why do we need hundreds of thousands of lines of opaque Go code to accomplish any of this?
I listened to his podcast on Pragmatic Engineer. I don't think he specifically addressed what it solved, but he talked about shifting the Overton window in regards to what's possible with AI agents. I'm not arguing that he actually accomplishes this -- just noting that his goal seems to be less "create something useful" and more so "create something that gets people's attention and maybe gets them to thinking about AI in a different way".
Cynically, he published a book on vibe coding recently, so he may just be grabbing attention as some effort to boost book sales.
I'm trying to make the phrase "AI DDOSing" happen.
ex. someone's GitHub repo with a ton of code and a README written by AI claiming fantastical features not present in the code.
Or, more subtle someone, "self-DDOS'ing via AI" - thats for when "LLM psychosis" is too strong, i.e. for "I went too far down a rabbit hole with the interactive chatbot for a month and now I have 1M LOC and 95% test coverage and an app that I don't understand"
I quit my job at Google in 2023 and have spent 2.5 years working on an LLM-based agentic app.
To me, this looks like an unfortunate self-AI-DDOS'ing by someone with even more runway than my seemingly infinite runway.
It's well-meaning, like, in 2030 I'm fairly sure we'll have a meta-layer and simplistic "here's a bug, read files, edit, fix" will seem slow/strangled. But he's at least a couple years ahead of the models, and whatever metalayer exists won't have the bizarre UX model.
This is true, without people trying, we wouldn't know what doesn't work yet - for all I know he's cracked something big and in a month we'll see the first AI-built operating system (I'm not being sarcastic)
Claude Code released agent teams in Feb. that was the first of the major players to have separate process/context agent orchestration that I know of. Who did it before that of the major IDEs?
There's a real hobbyist vs professional distinction with Claude Code. For professionals, including when I use it at work, we're generally super happy to have Claude spawn as many subagents as possible and burn more tokens to get a better result. Hobbyist users on a $20/month plan, though, generally want more conservative behavior.
It's hard for Anthropic to cater to both sets of users with one model.
I don't think that's what this issue is talking about. I have the Max $200/mo plan and have noticed starting yesterday that my quota drains much much faster, to the point I'm about to use the $50 credit Anthropic gave away to everyone.
True enough. But to be clear, that's a separate issue from what users are reporting here.
Both hobbyists and professionals are understandably frustrated that tokens are being consumed quickly without justification, or at least in ways that seem entirely avoidable.
No "max" or "pro" equivalent? I wanted to get a new Macbook Pro, but there's no obvious successor to the M4 Max available, M5 looks like a step down in performance if anything.
No doubt the "wider" versions of the M5 are coming.
My hope is that they are taking longer because of a memory system upgrade that will make running significantly more powerful LLMs locally more feasible.
Apparently not until early next year. I was surprised by this too, but I hadn’t really been following the rumors at all, so I didn’t really have any grounds for being surprised by this.
Does the "caching containers for Codex Cloud" mean I have some chance of being able to reuse build artifacts between tasks? My Rust project takes around 20 minutes to set up from scratch in a new Codex environment, which seems extremely expensive.
I think this is up to the user. I actually found tab so annoying that it was a big reason I quit cursor and cancelled my sub. I couldn't think straight with it constantly suggesting things to put in after every key stroke and caused a few annoying bugs for me.
I find pure claude and neovim to be a great pair. I set up custom vim commands to make sharing file paths, line numbers, and code super easy. that way I can move quickly through code for manual developing as well as have claude right there with the context it needs quickly.
It doesn't have Junie (Jetbrains AI Agent) yet, but I'm talking about the agents. I'm happy with Claude Code. I just want Cursor Tab in there. I use it for quick edits & refactoring, not writing new code, and it's damn good at what it does.
I didn't think much of it until I canceled Cursor to try out copilot, which is slower and yet also worse quality. I reluctantly resubscribed to Cursor.
reply