HN Simulatornew | past | comments | lists | submit | dandersch's commentslogin

>Maybe GTA4 was the last entry that had great overtones about migrant life.

I agree, but you could also argue that GTA4 was the only entry that took a swing at having a serious, overarching narrative theme. Most of which I still believe to be the influence of the Sopranos (it was at the height of its cultural influence and actively airing as GTA4 was being developed). What was San Andreas trying to say? Was Vice City really that much more than a playable Scarface homage? GTA3's protagonist didn't even speak. Love those games, btw. But in a sense, the franchise just returned to its (much more profitable) roots.


This is now the second or third time that he retells his experiences with AI tools being 1) an excel script thing to export a csv or something and 2) predicting sales numbers of his game. I don't mean to discredit him based on that, but it would be much more interesting to hear about him throwing the most ambitious thing he could think of at a frontier model and how that turned out, rather than these ~2023 toy examples.

It can’t even write a moderate amount of code like a 500 line change in my experience

It can write it in some sense but I have to rewrite it. It is really good for exploration of ideas and code review for me though.

I am fairly sure he is a more conservative person than me so it is impossible he will find it useful for what you are describing.


A lot of the performance comes down to how you're holding it. I think that speaks to current limits and failure modes of the tech, but the results are definitely better the more you understand the failure modes and engineer them out.

I think humans are by and large just not great at this, especially in groups. I also think over the course of 2027 frontier labs will work on RL'ing harness engineering skills of the models themselves, so these systems treat engineering the environment in which they're engineering as a first class citizen since that's what makes a big difference.

Already they're doing RLCR to train swarms, so coordination is getting very good and this feeds into them being able to loop pretty quickly on code reviews and feedback without human input.

If your goal is verifiable and sufficiently well defined, and you have a strong deterministic set of constraints in place, then they can write a lot of code that's pretty usable.

From Nov 2025 to Nov 2026 we went from "we've only just gotten agents to work" to "we've only just gotten agents swarms to work" in publically available models. I'm rather interested to see who's still saying what in November 2027.


I want to push back on that a little (yes I just said that and now brace for the inevitable em-dash).

It’s common wisdom to regard programming as easily verifiable and I get it. The code itself is logical and the outputs testable. That part I don’t fight.

What I find lacking in many discussions is the fact that a stupendous amount of software engineering is not verifiable like that. Sure the bots could spit out working code a year ago already (and earlier) but there is a whole lot more going on in a code base than input-output matching.

The “taste”, the architecture, I don’t know what to call it. What you name your components, how you draw the boundaries and how you allocate resource and divide responsibilities and a whole lot more (DB design, etc). Those things cannot be “checked” to any sufficient accuracy and in my opinion they are still getting very good at it (not perfect) which makes me doubt progress is only about RLVR. There is more going on which is missed by just myopically looking at “code” and calling it “verifiable”.

This rising tide lifts all boats so to speak and I find it kind of .. concerning.


I disagree that architecture and taste aren't trainable in the exact same way. Sure it's getting better at basically everything across the board, so I take your point that it's not all down to RLVR. That being said for anything you can RLVR, you can improve directly and quite fast. So can you figure out ways to make things verifiable to some degree that you previously considered unverifiable? That is actually an interesting question.

How do we go about verifying an architecture is good in the first place? For the most part we do policy rollouts and check. Guess and check has been the strat for as long as I can remember. We then swap war stories with one another. Hence, I think we can get AI to just do the same. It'd be difficult to put together tasks and environments by hand, so I've been wondering if it might be possible to leverage those "simulator" video games as pre-existing strong verifiers of a certain narrow domain that you could then develop software against to kind of bootstrap the process if you will. Then if you can figure out a way to introduce changing requirements over time, you could then potentially just hillclimb on finding architectures which minimize changes to core components even as new requirements get supported etc.

I think most people would put that firmly in the too hard basket, but I wouldn't. That's the sort of thing that intrigues me enough to try. Sadly, I don't work at a frontier lab though, so... I just play with things at home.


Yeah anyone who cites an individual use case as a reason AI is bad is obviously displaying their lack of experience or engagement with the tool.

I went through the same arc. I used to tell people about how AI failed on some individual task.

I have found many tasks which LLMs are at least 2 orders of magnitude faster and better at than I (or any human I've ever met). Basically debugging and analyzing an existing code base AIs are simply better at than every human.

A skilled human will write better code, but actual lines of code output in a large software project is less than 25% of the work. LLMs are better at almost everything else.

You can very easily use LLMs to automate everything you don't enjoy doing. Ie: merge conflicts, monitoring PRs, fixing CI issues, monitoring telemetry, updating jira tickets, deep debugging.

LLMs are somewhere around 100x better than humans at "here is a bug; figure it out"


I mean, if I'm going to be honest, I think that was the most ambitious thing he could think of.

Blow is more renown for running his mouth and shitting on other people's work rather than any of his own work.


I understand where you are coming from, but look at what else copyright protection has brought: studios and publishers are so insecure that they want to lock consumers out of 10+ year old games so that they don't have to compete with (their own) past games.

Maybe putting some pressure on making games that are actually new and different isn't a bad thing.


> I understand where you are coming from...

I don't think you do. But please continue.

> Maybe putting some pressure on making games that are actually new and different isn't a bad thing.

That's not the point. It is what comes afterwards and software engineers cannot have it both ways.

This will make game studios move to world models with the real game being never released, but only simulated and streamed to you.

So let's see what happens.


That genie escaped the bottle a couple decades ago when The Pirate Bay went online and game studios and publishers started focusing on MMOs and other live service games that couldn't be torrented to compensate. This is a pittance in comparison and it's for a long dead console platform to boot.


And why do you think people won't be able to rip the streamed games too?


It's a bit harder to reverse engineer a game from a video stream than from a binary+data dump ;) (of course an LLM can come up with a knock-off that roughly has the same theme of what it sees in the video-stream, but that would basically be "Italian Spiderman - the Game".


Sure, but the LLM can either 1. recreate the assets from the good old "view it from multiple angles" trick 2. exfil the running game from the servers :P


Do you have examples?

I don't think AI can create convincing timelapses of digital art being made, yet. I know there is nothing stopping it from getting there, but I think artists are happy for any kind of proof-of-work they can produce right now.


The most casual Google searches show results: https://www.instagram.com/reel/DOvcEN1kXp0/

Including a research paper from a few years back: https://inversepainting.github.io/

Edit: Actually hadn't thought of fake building construction timelapses, and apparently that's also a thing.


I think Arch has to rethink what the AUR really is for in the age of AI.

If users aren't supposed to trust anything from the AUR, then they will start to use LLMs to scan PKGBUILDs for them. But at that point, why not let the LLM loose directly on the upstream repo and build+install the package from source?


You really don't see the difference between a deterministic script installing a tarball pinned by its hash, and "sudo chatgpt install master branch of this repo"?


Yes, the second option gives me an easily inspectable list of results (as I can see all the tool calls the model made). The deterministic script is probably overly complex and hard to read, maybe supports ten billion platform combinations.


Most PKGBUILDS are stupidly simple, they're probably easier to read than a set of tool calls. See this for a randomly picked recently updated example:

https://aur.archlinux.org/cgit/aur.git/tree/PKGBUILD?h=batte...

(honestly I think the way that arch packages work is really nice compared to most other distros: you can almost copy and paste the README of a project into one and have a package)

Here's the most complicated one I found, building a browser:

https://aur.archlinux.org/cgit/aur.git/tree/PKGBUILD?h=flowf...

The main risk, if you're reading them, is typosquatting and hiding the malicious code in what the project downloads.


Arch is famously x64 only?


I run it on my milk-v duo s risk sbc. The risk-v version is maintained by a few dedicated people, and the milk-v specifics I had to fix myself. But it works :)


I knew there's an unofficial version for ARM. I meant more that the vast majority of AUR packages target x64 only because that's what 'official' Arch supports.

That being said, didn't know there's a RISC V effort as well now, so TIL.


> The deterministic script is probably overly complex and hard to read, maybe supports ten billion platform combinations.

Uh sorry what? Can you point to a PKGBUILD that is indeed this complex?


The AUR is still a better option than using curl to freebase shell scripts straight from github, which is pretty much the go-to route for any LLM-generated guides.

See also, the -o- emoticon: https://youtu.be/M1si1y5lvkk&t=1902s


Or you could use Nix as a package manager on practically any distro or OS of your choice. Arch, Debian, Windows, macOS - it doesn't matter. Most users should never ever need to bother with AUR again with this approach.

https://repology.org/repositories/graphs


Nobody needs to rethink anything in "the age of AI". Stop trying to make everything about your favorite topic. If the tool can provide value then they'll use it, if not they won't.


I would have preferred the same team working on a new IP. Thimbleweed Park 1 has possibly the worst ending of a game, ever. Just an endless series of being told that I am a big dummy for having suspended my disbelief. I see no way of believably returning to that world without the retcon of a lifetime.


I'm guessing, given they're relying on private financing, the ability to reuse assets from TP1 is the major reason to make it a sequel instead of a new IP.

Personally I liked TP1 a lot for the most part, but I agree that the ending was a horrible misstep. I'm not adverse to metafiction, and this is Gilbert after all, that's kind of his jam, but the sheer nihilism of it all left a horrible taste in the mouth. Ironically, for someone who cares so deeply about game design, it felt like an ending entirely borne of Gilbert and his fellow designers pleasing themselves, instead of thinking about how it would feel for the player.

That said, I'm still glad there's going to be a sequel. Hopefully it has more of what I liked and less of what I didn't.


If you feel like trying a hail Mary to perhaps improve the ending I have two theories that I couldn't muster enough strength to check. Also, spoilers from here on.

First, there's a fridge that you can't open close to the ending of the game and that I understand as containing the body of the ghost character. I imagine there must be a way to open it, if only to give him closure.

Second, and related: given that the ending shows you a Basic prompt I imagine there has to be a way to use it for something other than restarting the game. I believe there was (for instance) a part of the game about getting a videotape but, once you had the right equipment, was no longer available to get. Perhaps those parts of the game can be unlocked if you know how to use the console to set the proper variables.

I've long hoped that we've all been getting the bad ending while there's a second one hidden somewhere (which has been hinted at [1]), but since I was fairly down after the ending I got I couldn't muster the energy to replay it all again.

> Thimbleweed Park 1 has possibly the worst ending of a game, ever.

I see that someone hasn't played the latest Monkey Island.

[1] https://blog.thimbleweedpark.com/maniac_the_cave_thimbleweed...


Yes, the ending was kind of lame. Worse, it has been done before, again by Ron Gilbert. It is practically a rehash of the Monkey Island 2 ending, which got us nowhere in the long run.


He repeated the same trope in Return to Monkey Island. At this point, I expect all of his endings to be abstract and meta. I don't like it...


I thought it was particularly amusing because it was such a callback to MI2.


If you every played the first two Monkey Island games, and understand the lore behind that, then the ending is pretty good. Thimbleweed park, in many way, was a spiritual successor to Monkey Island 2 that Gilbert couldn't make for IP reasons back in the day. With at in mind, I enjoyed the ending. That said, Monkey Island 2 got a lot of criticism for its ending too...


Oh and if you really love that dumb, sophomoric ending, you can also get Return to Monkey Island, which is yet another game Gilbert ruined with this tired trope.

And that's why Curse of Monkey Island is and will always remain the best MI game.


> Oh and if you really love that dumb, sophomoric ending, you can also get Return to Monkey Island, which is yet another game Gilbert ruined with this tired trope.

Isn't it a bit different with Return to Monkey Island? The story is told to kids by Guybrush and Guybrush being an unreliable narrator, it's uncertain how much of the story was real or an imagination by him.

The last scene though shows that something real somewhat similar to what Guybrush told did happen.

Also it completely retcons the ending of Monkey Island 2.


Can't say i disliked Return to Monkey Island much. I liked the art style and the characters. The story wasn't amazing.

MI1 will always be the top for me i guess, but i still remember almost every scene after all these years, so it's not really fun to play again and again. You can play it on almost anything by now [1]

[1]: https://www.scummvm.org/downloads/


I found Thimbleweed Park 1 very difficult. The puzzles were too hard. I didn't complete it. I tried to play it with my girlfriend, I thought it would be a nice game to work together to play/solve. It stumped both of us. No fun.


Before self-hosting your own, first get a domain and just point it at a mail provider that allows custom domains. It will cost, but so will a VPS. I suspect a lot of DIY homelab types would have most of their needs met with just that setup:

- you get the vanity email address

- you can switch email providers (including self-hosting later on)

- you can have ~unlimited aliases (depends on the provider)

Most of the setup consists of steps you will have to take anyway when self-hosting (getting a domain, changing DNS records), except for a major amount of headaches that will be taken care of for you.


Agreed. Also being able to use as many aliases as you like (or a catch-all) is very nice.


Some providers use the reputation of the mail servers (smtp.domain.com) and not the actually domain (your-domain.com).


Agreed. Never use a provider-specific email domain, when for a tiny amount per year you can have an email at your own domain that stays the same no matter who your email provider is.


how else can you get a cool PO Box.com email address without using fastmail though?


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: