TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Minecraft with object impermanence

229 点作者 sohkamyung4 个月前

33 条评论

skykooler4 个月前
This was a fun read. But some of the things attributed to a glitchy AI are just normal Minecraft stuff. For example: "One time I was swimming across a lake and noticed that the reflections at the water's edge were looking weirdly spiky." That's kelp, it looks like that in the game because the chunks beyond haven't loaded so it stands out from the sky. Also, the "natural blocks of rare glowstone" are honeycomb blocks, which she thinks might be new "cheese" blocks but in fact have been in the game for over five years.
评论 #42801695 未加载
评论 #42806040 未加载
sedatk4 个月前
That's pretty much how my dreams work. For example, I can read in my dreams (like signs, books), but it's almost always one word at a time, and the rest of the text becomes impossible to focus while I'm reading that word, no matter how hard I try. It's like the "permanence module" is disabled and my brain falls back to a primitive generative AI for the next frame. And, I can't reason about the logic of my experience, the lack of causality, or persistence of objects at the time. Everything feels "as they should be" despite how awkward things get. I guess that's why nightmares are so convincing or hard to get out of, because I can't reason about them. The whole article feels like describing a typical dream to me.
评论 #42802110 未加载
评论 #42802450 未加载
评论 #42801682 未加载
评论 #42803776 未加载
rogual4 个月前
There&#x27;s a bit in Ender&#x27;s Game where he plays this computer game that has a dreamlike quality to it. It seemed it could simulate a response to any player action, and continue the story along that path. I think in the book it was a tool for introspection, like dreams can be? I can&#x27;t really remember.<p>Anyway, as a kid I thought, silly author, that&#x27;s not how computer games work. But it turns out Orson Scott Card is smarter than me. Give this kind of thing a few years and we&#x27;ll have it.
评论 #42803465 未加载
mkoubaa4 个月前
This reminds me of dreaming. Either there&#x27;s a similarity between dreaming and what this AI is doing or my human brain is stereotypically finding patterns where there are none.
评论 #42805150 未加载
评论 #42804982 未加载
lionkor4 个月前
&gt; the first step in our research towards more complex interactive worlds<p>As we know, object permanence is not really needed in any videogame, so this makes sense. It&#x27;s also easy to add in, they just didn&#x27;t do it because they didn&#x27;t feel like it.<p>Snark aside, not sure what that is supposed to mean. This isn&#x27;t interactive in the same sense that games are, its not a world because it doesn&#x27;t have any properties of a world, its just a video generator you can steer with keys. Cool, but not what we would call an interactive world
评论 #42803447 未加载
评论 #42805036 未加载
lou13064 个月前
&gt; Things became duplicated in Tlön; they also tend to become effaced and lose their details when they are forgotten. A classic example is the doorway which survived so long as it was visited by a beggar and disappeared at his death. At times some birds, a horse, have saved the ruins of an amphitheater.<p>JL Borges, <i>Tlön, Uqbar, Orbis Tertius</i>
评论 #42805227 未加载
treyd4 个月前
&gt; One time I was swimming across a lake and noticed that the reflections at the water&#x27;s edge were looking weirdly spiky.<p>This is probably the model trying to generate kelp is the ocean. When water is in chunks on the edge of the loaded world, you can see through the water to the sky at that angle, so you also see the outline of the kelp kinda like that.<p>This is pretty common to see in the game nowadays, but the AI model might be having a hard time since kelp was only added a few years ago and isn&#x27;t as well represented in the training data.
gwern4 个月前
&gt; The problem, as far as I can tell, is that if all this works perfectly, this will be simply the human-programmed Minecraft we already have, except far more expensive to run. The original Minecraft is already infinitely playable, thanks to the way it randomly generates a 7x-larger-than-planet-Earth landscape with each new game. You can&#x27;t use generative AI like this to get a new kind of game, only a better simulation of the game you trained it on.<p>Oh come now, that&#x27;s not true and is showing a lack of imagination, especially from someone who&#x27;s been covering generative models this long. There&#x27;s lots of things you can do with an accurate differentiable model of Minecraft - just like there&#x27;s lots of things you can do with a LLM or a diffusion model beyond just &#x27;generate a random sample&#x27;. (Just think about all the things you can do with CLIP steering a diffusion model...) Imagine describing GPT-3 as saying &#x27;as far as I can tell, if this all works perfectly, this will simply be Common Crawl, except far more expensive to read&#x27;!
torginus4 个月前
The AI dungeon thing is a salient point - GPT2 (which it started out with) could barely make a coherent paragraph, and GPT3 could write about half a page of text that more or less made sense.<p>With modern LLMs, they still get occasionally tripped up, but you could go for pages without a minor detail not making sense.<p>Something similar might happen with these game models, given enough time.
评论 #42802464 未加载
评论 #42821395 未加载
rollcat4 个月前
That&#x27;s the difference between a deterministic and a stochastic system. Using one where you need the other leads to exactly these kinds of results. Recent trend in LLMs is to gravitate towards more determinism (like fact checking), I&#x27;m afraid in the end we will come full circle and end up with a fancy frontend to a database.
fdb4 个月前
Janelle Shane’s work has been really inspiring. Her talk at Strange Loop from a couple of years ago — right before DALL-E, before ChatGPT — was a turning point, and very funny:<p><a href="https:&#x2F;&#x2F;youtu.be&#x2F;yneJIxOdMX4?feature=shared" rel="nofollow">https:&#x2F;&#x2F;youtu.be&#x2F;yneJIxOdMX4?feature=shared</a>
omcnoe4 个月前
Interesting blog post from the creators with more details about the model: <a href="https:&#x2F;&#x2F;oasis-model.github.io" rel="nofollow">https:&#x2F;&#x2F;oasis-model.github.io</a><p>Interactive demo: <a href="https:&#x2F;&#x2F;oasis.decart.ai&#x2F;" rel="nofollow">https:&#x2F;&#x2F;oasis.decart.ai&#x2F;</a>
评论 #42817233 未加载
itsTyrion4 个月前
Man, I’m glad I’m using paper straws and turning off the PC the second I leave the room for the environment while some &quot;researchers&quot; elsewhere waste god knows how many kWh for a barely functional &quot;AI&quot; version of a game. ♥
folli4 个月前
So the AI is trained from the previous frame and the input and tries to predict the next frame, correct?<p>How could you achieve object permanence this way? Will it &#x27;automatically&#x27; appear given more training data or more hidden layers? How is this handled in other approaches?
评论 #42801867 未加载
评论 #42801863 未加载
ThomW4 个月前
I&#x27;m guessing the author didn&#x27;t play Cyberpunk at launch -- the cars and pedestrians would blink out of existence if you turned around. I couldn&#x27;t help but think that Rockstar had figured this out 19 years prior in GTA3. :&#x2F;
评论 #42816254 未加载
评论 #42804231 未加载
评论 #42805794 未加载
npteljes4 个月前
This seems to be in line with a previous submission, where one can &quot;play&quot; Counter Strike in a similar way with AI.<p>Diffusion for World Modeling<p><a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=41826402">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=41826402</a><p>Direct link to a video, apologies for the xitter source: <a href="https:&#x2F;&#x2F;x.com&#x2F;Sentdex&#x2F;status&#x2F;1845146540555243615?mx=2" rel="nofollow">https:&#x2F;&#x2F;x.com&#x2F;Sentdex&#x2F;status&#x2F;1845146540555243615?mx=2</a>
Matthyze4 个月前
I really adore this. AI art is at its best when it&#x27;s strange and uncanny. Another example (watch till the end):<p><a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=XqLm0qbLJqY" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=XqLm0qbLJqY</a><p>edit: I&#x27;ll add this great take too:<p><a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=fQs0GwIlEKM" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=fQs0GwIlEKM</a>
评论 #42803602 未加载
p0w3n3d4 个月前
To me AI generated videos (this kind) makes my brain untwist its wrinkles. I&#x27;ve seen this video <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=7Jd-Rr9cJYo" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=7Jd-Rr9cJYo</a> - a Minecraft gamer plays the Oasis AI version. There is literally zero coherence and dreamy trippy style
graypegg4 个月前
Maybe this is extrapolating a bit too much from this, but this would be sort of neat as a game streaming compression system right? If the service only needed to send a frame every 1-2 seconds, and the interim could be interpolated locally, still but still processing key strokes, it might be juuuuust smooth enough to not notice. Sort of like DLSS, but game specific, and &quot;upscaling&quot; a frame over time instead of resolution.<p>That is of course dependant on running inference locally being somehow less intensive than just running the game. But there we go, it&#x27;s an interesting thought at least!
ryanisnan4 个月前
I enjoyed this read far more than I expected. It’s kind of a dark glimpse into a weird future.
HelloUsername4 个月前
Related?<p><a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=42041611">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=42041611</a><p><a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=42030787">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=42030787</a>
flessner4 个月前
As a longstanding Minecraft player, that&#x27;s played less and over the years, I simply don&#x27;t know how I feel about this.<p>A side of me hopes that this technology gets used to drive progress in self-driving cars, robotics etc. My intuition says that having a sandbox that looks similar yet functions unexpectedly could greatly help with sniffing out edge cases.<p>On the other side I fear that this ends up in the next generation of social media - never ending generative short-form content. Obviously, in its current state the effects on the brain regarding object permanence could be... devastating?
评论 #42804008 未加载
011000114 个月前
First there is a mountain, then there is no mountain, then there is...
baq4 个月前
I wonder if this is what Zelazny was thinking about when he was writing the Amber series.
cubefox4 个月前
Apparently Oasis Minecraft is what&#x27;s called a Markov process: The next state only depends on the current state, but not on the past states. To &quot;fix&quot; this, Oasis Minecraft would need a memory of the past frames.
haltcatchfire4 个月前
Reminds me of Weeping Angels from Dr. Who. Great concept!<p><a href="https:&#x2F;&#x2F;en.m.wikipedia.org&#x2F;wiki&#x2F;Weeping_Angel" rel="nofollow">https:&#x2F;&#x2F;en.m.wikipedia.org&#x2F;wiki&#x2F;Weeping_Angel</a>
johann83844 个月前
This is like navigating Steve&#x27;s dream world while he is sleeping. Just some conjured up wonderland in his subconscious.
whywhywhywhy4 个月前
Felt very dreamlike playing the simulated Minecraft, as in they are a place where you can turn around and be somewhere else.
atahanacar4 个月前
Slightly (or very) off topic, but I think it fits so I will write what I feel about people talking about their AI experiences. Feel free to ignore my opinion. I just wanted to know if there is anyone who feels the same way.<p>Listening to someone talk about &quot;their chat with ChatGPT&quot; etc. feels very annoying. It&#x27;s like they are explaining a dream they had in detail but even less informative or useful because a dream might suggest something about their state of mind. However, content like this (to be fair, this specific one isn&#x27;t that annoying to me especially compared to &quot;AI said this and that&quot; kind of stuff) has absolutely no constructive value (to me) and feels almost schizophrenic. Maybe I feel this way because people around me, unlike the average HN user&#x27;s social circle, have only superficial idea about how AI works but are very interested in using it.
评论 #42805219 未加载
findthewords4 个月前
Just like a dream. Electrons running through our brains have no permanence, atoms in the real world do.
评论 #42803580 未加载
nottorp4 个月前
If you smoke the right mushroom and start vanilla MC, you won&#x27;t need this AI right?
api4 个月前
What fascinates me about this is this is what dreams are like. The human brain tries to loosely predict reality, leading to object impermanence and all kinds of trippy weirdness we call dream logic.<p>It’s also why writing good fiction is hard. You have an idea and try to put it down and it’s a mess because your idea is just a loose prediction of what you want to achieve. Then it has to be thought out deliberately or the edges don’t meet up. I’m sure the same goes for painting, etc.
评论 #42803853 未加载
_def4 个月前
I kind of want to see this with Pokemon gen 1 games now.