Diff

Huh, just checked and realized don't have the little pop up hints like a lot of other plants do. To get berries, right click to pick them off the bush. If you left click you'll break the bush like a regular block, destroying it and leaving only plant matter. I think you have to use a knife to take a cutting that you can use to propagate the bush.

Unless you're purposefully wanting to jump through hoops, yeah, Flatpak's that's certainly the easiest way. Performance while running in a full VM would be probably definitely awful.

If a Bell was ever to wander all the way up to the surface for some reason it'd make an excellent candidate. Or maybe just a bell head.

Maybe a luddite in the true sense, as in someone who's not a fan of owners eliminating swaths of jobs with automation and siphoning their paychecks up to themselves while burdening those left with even more work. It's the stance I'd like to take if I could afford to right now, especially since these new machines are explicitly built using people's own work, unsactioned, to replace them. Thing is, AI still isn't capable of actually replacing any human jobs. It doesn't produce correct code. AI images are unusable for print, and uninspiring and difficult to work with for digital. AI video still tumbles deep into the uncanny valley. The vast majority of real world use for both is just to generate nonsense online, whether it's for bizarre entertainment or manipulative disinformation. Data's still coming in on this, but in the workplace it doesn't seem to actually boost speed or productivity in the short term. And in the long term, it seems like "use it or lose it" is in full effect for the tasks offloaded to AI. Right now AI investment is propping up the entire economy, and the only way that makes sense is if LLMs rewrite the face of labor. But they're not doing that, they're letting people write bloated emails and fling freakish LinkedIn posts at each other faster. They are falling so short of their promises that even AI execs are talking about when the bubble pops. It's 2026 and they're still using the exact same tricks they have since GPT-2, "LATEST MODEL is so advanced and so dangerous that we can't release it openly, you'll have to just buy access from us." (EDIT: A claim that still in 2026 completely falls flat on inspection). It's a research project being sold as a guaranteed post-scarcity future that will far more likely just turn into a capitalist nightmare if it ever succeeds. So shunning the whole mess seems pretty sane to me.

You could do a different data directory for each world/modsets, is this kind of what you're thinking?

This isn't the first I've seen this claim, and I can see how it could be true of some artificially intelligent system... but I don't see any present evidence for it. Instead, my own experience hints at the opposite. LLMs are blinded by a magnetic pull into well-worn ruts in the training data. Which to me makes immediate sense based on how these models are trained and operate. Way back you mentioned its struggles with recursion, and I think that's a good indicator of what I'm talking about. Like a lot of people, I have a few troublesome tasks in my back pocket I use as benchmarks when a new model comes out. One of those is having it toy around with the Gleam programming language. It knows Gleam, it's not a tiny homebrew language. It can tell you all about Gleam, recommend popular libraries in the ecosystem, tell you about how it interacts with the BEAM. It can tell you the exact rules it will then go on to violate when it comes to generate code, even with precise rules, tools, tips, and reminders in the context window. It fails to reason about something it knows about because the probabilities point elsewhere in the distribution of training data. This goes also for projects in common languages (JavaScript) with atypical ways of doing things. One of my pet projects is a 3D engine that renders to 2D SVG. From SVG it inherits many quirks that are deeply atypical in a 3D world. AI is quite helpful as a research tool here, because the things I'm doing are not unknown mathematically, they just violate convention. But those broken conventions means that when it comes time to write, it constantly gets pulled off the rails by stronger signals in the training data.

"It's inevitable" only according to the people who desperately need that to be true to get their bonuses at the end of the next quarter. I see no reason to trust them when I look out at what AI code hath wrought. Even after all the hype about what agentic coding can do, look at Cursor's web browser and Anthropic's C compiler. These are flagship research projects completed by the people most familiar with their own tech, and they're staggeringly lackluster. Some smaller projects have better luck, like Ladybird's LibJS rewrite, but especially combined with what send to be diminishing returns, I don't think there's any guarantee or even likeliness that we're looking at the end of horses or horse drawn carriages. Especially when this new car relies on ground up horses to fuel itself.

This is a lot better phrased than what I wrote, and is largely what I was trying to say. Only thing I'd push back on is that it's always a discipline problem. Sometimes you just don't know better. Sometimes AI is being used to skip work, I saw that all the time when I was a teacher. But even as a learning tool, it's hazardous. AI sounds confident. It's easy to be taken in. Most recent job I had was at a print shop. Boss had no prior print shop experience and had taught himself a lot from AI and YouTube. But he also trusted its answers and its own confidence in its ability way too much. Some of it was small stuff, like thinking "welding" and "hemming" were the same thing. Other things were actively destructive, like thinking he can buy this roll of material for hundreds of dollars when it would actually be incompatible with our equipment because he didn't know to ask a question that he didn't know. Or getting ready to print a graphic that AI says is 300dpi at 84" across when it's actually 21dpi because he doesn't know AI can't do that. If you know how AI works, it's easy to label this is all silly user error and easily avoidable if you understand, but the average person doesn't know how LLMs operate or what they're good or bad at. People can think they're learning when they're actually just being taken for a ride with nobody at the wheel.

It's not flipping the question, it's addressing the elephant in the room. We can't just say "ignore the elephant for the moment" if you actually want to understand why people are acting weird. That is why it needs to be disclosed, that is why they are "worse." Of course, if a human and an AI write the same line of code, there is little difference (mostly, see below). If you develop a car that has a 1% chance of being assembled improperly and self-destructing after 3 months, you have a bad car. Cars with higher failure rates get the reputation even if a specific instance is fully working. If you could screen for an improperly manufactured car, maybe we can work with that. But we can't. Mods are often closed source and screening code for correctness is intensive. So instead, it's better and wiser to avoid the models of car with that flaw in their production line. We don't have an oracle that can identify "working" vs "non-working," so it's unreasonable to ignore it. The other half of it is more philosophical and social. I have two main issues with code generation AI models. 1. Code generation models are trained on vast swaths of open source (and even closed source) code that was released under certain licenses. These models can reproduce those swaths of code readily. That's not surprising, so can every other flavor of model. But by doing it here, big corporations are being given the ability to launder open source code into closed source code against the wishes of those who made it. A swarm of agents can rewrite software in a new language and a new license, allowing people to take advantage of the hard work of others without abiding by the license that they accepted that code under. I don't think this benefits anyone to gift corporations with the ability to embrace, extend, extinguish in bold new ways. 2. Use of AI is seductively self-destructive. It can be used well. It can be used as an effective learning tool. It can be a nice rubber duck to debug against. But using it without exercising your own skill or by letting it shortcut your own struggle is already showing signs of diminishing the skills backing it. If you're using it to make mods as a fresh newbie, if not used properly as a learning tool, it can sabotage your own education. I'm not saying this in a "back in my day we had to walk to school uphill both ways" kind of way. Even for using AI, a clear view and a clear mental model of the program is important, and AI itself can interfere with both of those.

Yes. As someone who has published one (1) AI-developed mod. I like to explore AI tools to keep an eye on them, but I usually come away with nearly the same understanding every time. The problem with AI mods is what you pave over in this line: The problem is that it's far more likely that an AI-authored mod doesn't work the way it should. AIs frequently hallucinate and non-determinstically just wet the bed. The one AI-developed mod I've published was a lottery win. For chunks of my daily programming work, I have to explicitly disable even AI auto-complete in my IDEs because they don't understand my style, my language, my situation, or my environment and constantly make basic mistakes. The pull of the training data is too strong in large swaths of niches out of the mainstream. Even safely inside the mainstream, we're still getting headlines with AI agents wiping production databases. Mods are low-stakes, but it still sucks to deal with poorly written software that was churned out carelessly and never reviewed by a human. There is, theoretically, a "right" way to do AI-assisted development. By gluing it to strict guardrails, giving careful, explicit instruction, and reviewing all output rather than smashing in a prompt and vibecoding it. But without any sort of disclaimer, it's impossible to know. So it's uncool to waste people's time and potentially screw them over through your own carelessness by not at least disclosing it. And without an explicit disclosure, AI thumbnails become the next best thing to hint at this possible undisclosed carelessness and the threat of wasted time. Besides all that, there's plenty to gripe about the social ramifications of AI (even just specifically in regards to programming) that aren't just "AI bad" but I'm going to cap myself at 2 paragraphs.

You got it. As long as you create a copy of the shortcut to point at your copy of the data folder, there's blessedly little footgun potential here.

I will never not say no to more Vintage Story content. Would love any videos on your first impressions on building/anything in VS. Basically the short you linked but long.

One workaround you could pull is two separate data folders, one with modpack, one without modpack. I don't know how technical you are, but Vintage Story supports having those passed in with a command line argument. The process would basically be like this, but without actually installing a second version of the game.

Alright so there were a few things that I've got customized about my setup in my earlier recording that seem to have confused things, because the reason the de-quarantining isn't working is that none of those commands are correct. Here's a new recording with all of my nonsense stripped out, so the whole process should look exactly like this: CleanerDragNDrop.mp4

New version 0.6.5 just dropped. Mostly for one new feature to enhance usability. Right now a lot of sounds, especially ambient and weather sounds, just live forever in the list. I didn't want to fully hide those, so I added a similar list of throttled tags that will get hidden after a configurable amount of time. By default, it only applies to ambient and weather sounds. Going to continue refining this feature, but this has been useful enough for me at reducing clutter that I wanted to go ahead and publish. Two other small changes, fox growls aren't marked as dangerous anymore, and some small changes/tag additions have been made to the caption metadata files.

Go about downgrading? Go the client download area and click "(Show all available downloads and mirrors of Vintage Story)" to show older versions. Go about backing up worlds? They're stored in your VintagestoryData folder, copy them somewhere safe, and copy them back if you need to revert. Loading a world in a newer version of the game will update it and can make it incompatible with older versions, so backing up is an important step if you might want to go backwards. Code mods are all incompatible between 1.21 and 1.22 so you'll get crashes unless you make sure you purge any that don't belong with whatever version you're currently running. It can be a bit of a mess of things to manage, so multiple parallel installations like dakko said is probably a smoother way to test.

@Zarenth Make sure to back up your world in case upgrading breaks it. But yeah you can downgrade.

Through mods, all things are possible.

What does echo $DOTNET_ROOT say? Based on that, it seems that .NET 10 was never properly installed. Maybe try the official instructions from Microsoft for Ubuntu 24.04: https://learn.microsoft.com/en-us/dotnet/core/install/linux-ubuntu-install?tabs=dotnet10&pivots=os-linux-ubuntu-2404 And maybe just try running the install.sh script. Based on your command prompt, a few things seem not quite right, and I'm not personally aware of any issues with it...

If it's whimsical, I'm in. I could use an ounce of whimsy.

Doing what you did, that one directory was only footgun I ever ran into. If you do want a reset, just delete VintagestoryData and the game will recreate it next time it boots, but at the cost of all your worlds and data. If you got the path fixed and your mods are working, you should be Most Probably Good To Go right now.

What he said. On Windows that works, but on Linux for security reasons the terminal doesn't look for commands in the current directory. You have to use ./ before it to explicitly say that's what you want.

Did you copy the clientsettings.json file? It contains the directory mods load from, so if you don't update that for Linux it'll be looking for a Windows path. Shot myself in the foot with that one a while back. Like you saw, the button still opens the Linux path ignoring what's in clientsettings.json which makes it even more confusing.

If you don't hear anything soon, you've got *much* better odds of getting help if you can upload anything publicly so people can offer drive-by advice.

Try this way instead. Slightly modified command and drag and drop will squish out a few sources of errors. If it says sudo is still running when you try to close the terminal, it's very likely it's still waiting for your password. DragAndOrDrop.mp4