Crazy Roomba
Small robot. Big trouble.
Game
About
Boost, drift, and vacuum your way from a tiny apartment to the moon. Play Crazy Roomba in your browser and race the daily leaderboard.
Listed from atmosphere.games. The developer manages this listing there.
Reviews
No one has reviewed Crazy Roomba yet. Yours could be the first.Write the first review
Updates
All updatesGive Agents One Way To Do Thingssomething new i've been trying, half as an experiment, half as a statement of minimalism is to reduce complexity/choice in my computing environments. to put it another way, i want there to be one unambigious way to do things on a computer, and then i want to give agents instructions about that one way.a pretty simple example here is for stuff like version management. asdf and mise are both... well, they're not my favorite solution to a problem (i suppose nix flakes are my favorite solution to it) but they are a necessary evil. instead of having a dozen different ways to handle it (nvm, whatever) -- i just use mise everywhere and remove other toolchains. everything goes into the mise hole.the nice thing about this is that normally such a decision would be setting myself up for pain and agony; anything i wanted to build that expected system ruby or python or go or whatever would choke and try to install stuff, but now i can just tell my agent to figure it out while holding the constraint of One Way To Do Things. does this mean that i'm probably gonna find something Weird down the line, yes. do i care? probably not. i care more about the consistency benefits, because if there's one thing i can say about working with agents is that they loooooooove when there's one way to do things.the nicer part of this is probably that by being extremely consistent with these choices it means more of my setup is portable. i can throw stuff on some disposable compute like a fly.io sprite and assume that stuff just works.anyway, like i said, it's an experiment more than anything else. i would assume that the actual lesson here is that it doesn't matter if there's just one way to do things because the agents will do whatever they want with bash and gnu builtins. oh well.Aug 29, 2026
Two questions from the Jagged FrontierA non-exhaustive list of questions that I do not really see discussed as much as I think they probably deserve to be discussed.Question 1For most of our professional lives, we've been able to have fairly high confidence in the rough contours of the landscape that we're building into at a six to twelve month horizon. Computers will be a little faster next year, have a little more RAM. There will be a new version of Windows within the next 3-6 years.The curious thing about the last two years is that the basic shape of this bet continues, but the outliers are becoming increasingly dramatic. You've been able to bet that AI models will be better than they were, often concerningly so, within the next twelve months -- at least in terms of the models doing useful work. In January 2025, I told someone that a certain migration tool would be pretty much worthless in a short amount of time, because frontier models would simply progress to the point where they wouldn't need assistance to do what the tool did. That was prescient, I think. When I posted last June that it was the end of observability as we know it it came from the same place.Now, not everyone has gotten that memo. I think they will, very soon. This year? Next? I dunno. I also think that a lot of AI startups selling 'AI SRE' are going to be in for an extremely painful wake-up call because frontier models are just going to do this stuff. They already can! Perhaps the only thing more trivial for Claude than writing leetcode solutions is running obscure awscli or kubectl commands. If someone trusts your ChatGPT wrapper to actually restart services in prod without some really good legal protections then count yourself as fortunate that your client's lawyers practice three-cueing.Which takes us back to the question, how do you cope with the maximums and minimums of the capability bet? This has been an eventful 72 hours or so, which helps color the discussion. How are you making product plans that have a probability range from 'eh, things are more or less the same' to 'we all talk to our computer now all the time and software is just kinda generated on the fly'? How are you planning engineering capacity around models that may be taken from you with four hours' notice on a Friday evening? We can talk about 'taste', 'accountability', 'contracts', whatever else until our faces turn blue. I do not know how we can talk about the potential upsides and downsides of AI in software because it is not only moving very fast, the window of possibility is widening every day.Question 2To quote Mass Effect, Sir Isaac Newton is the deadliest son of a bitch in the galaxy. Similarly, momentum is the force that drives the modern economy. This can be seen in both the, uh, confusing exuberance/"it can only good happen" of the stock market but also in the long-term contracts and business deals that underpin large companies (e.g., Global 2000). To wit, "nobody ever got fired for buying IBM" (in some places, still true!)Momentum is driving a lot in the world of software these days. The actual reason AI gets as much investor attention as it does is because it upsets that momentum. People will write skeptically about "well, these valuations don't make sense unless OpenAI/Anthropic can capture a notable percentage of global GDP" and yeah, that's why people are investing in them. Contracts end. Virtusa, General Dynamics, a shitload of other companies you've never heard of but make the world go 'round through business process outsourcing? In what world do they survive AI? Somewhere, someone with fiduciary responsibility is doing the math on this and at some point the line will cross a threshold and a lot of these places will get cooked by it.There's another part to this question, though. What will you do when momentum catches up to you? It's gonna get everyone, one way or another. Your customers and clients are in the window, same as we all are, and the ones that are ahead of you in AI are gonna be asking some tough questions. Hell, I've seen people say to us "it would take claude an hour to fix , why is it taking you all so long" and they're right! AI native companies are going to push the hell out of you to run faster, are you gonna be able to keep up? What's your strategy for this?Question 1+2Readers may notice that I have actually asked the same question twice, and you'd be right, you sly dog! My point is that the frontier is extremely jagged, and it's only getting more so. The supply-side perspective is "ok, how do you deal with that when you can't see the edge", the demand-side perspective is "ok, how do you adapt to that when you don't know who's holding the knife" and my actual question is "is it the same answer to both". The 'jagged frontier' is about surprising capability gaps; I would argue that the real jagged frontier here is the unexpected change in potential outcomes.I would love to say I had an answer. I can tell you what I don't think will work. I don't think doing nothing is an answer. I don't think trying to embrace some sort of butlerian jihad against the thinking machines is going to go anywhere. I think that there is an extremely difficult to quantify element of risk in the industry's adoption of AI, and I think a lot of people are making some very foolish decisions hoping that they can outrun their friends. Meanwhile, the bear that's chasing all of us is turning into a Megazord or something, and the Power Rangers are nowhere to be seen.Here's what I think could work, though, and it's the embrace of AI as normal technology. Part of normal technology is that normal technology becomes cheaper, faster, and more accessible over time -- especially technologies that are decoupled from scarce resources. Computing is not a scarce resource in the way that oil or plutonium is. I believe that AI training will become cheaper for purpose built models through improved efficiency in large-scale serving.Semiconductors are scarce, true, but they are not consumed in the same way that many resources are. Chips are everywhere these days, and we can always make more! There's a whole-ass sun in the sky giving us free power forever! At last, we have invented the abundance machine from the cautionary tale "the abundance machine is really cool actually, someone should make one of those". The threat to intelligence on demand is not resource scarcity, it is unexpected externalities -- like, say, holding a model ransom because you didn't donate to the President's Emotional Support Ballroom Fund.If there exists a way out, it is by embracing the principles that have served us as technologists well over the decades -- protocols, composability, networks. This is what saves us from the downsides of a national security letter, because we can simply route around problems. I don't want to necessarily talk to a Claude to get feedback on this post, I want to talk to a focused, composable writing coach that's been distilled from something greater. What are the economic incentives that lead us to that? I don't really know, but I think it's an interesting question. Do OpenAI and Anthropic become the equivalent of, like, car manufacturers?Anyway. The best part of all of this is, if I'm wrong, and we are on the cusp of recursive self-improvement and someone makes Digital Jesus, then it doesn't really matter anyway. I'm gonna take a bet on openness, collaboration, and conviviality though.Happy Summer!Jun 15, 2026if i have to read one more piece of generated text i will kmsthere are approximately three good reasons to generate text that another human being will read.you are writing code with ai and the comments and symbols are mostly meant for consumption by other ai agents.you are running an ai agent/bot that is clearly labeled and defined as an ai.you are running an ai agent/bot for yourself and you would like to read the messages from that agent.outside of that? please stop. please stop writing status reports into slack that were generated. please stop making slides that were generated (oh my god please stop generating your slides), please stop generating emails, please stop. stop. stop. you are not helping. the total amount of the joy in the universe is shrinking every time you tell claude or gpt to write a memo or a PRD.this is not to say "don't use ai to help you write". that's fine! i use ai to proofread my stuff a lot, to poke holes in it. most of the time - for things i care about - i don't really incorporate its style/grammar suggestions (although i usually do wind up reframing some arguments or moving things around) but it does help. i'm not talking about that. i think that's good and useful. what i want you to stop doing is feeding in a bunch of garbage notes into the inference hole and tokenwashing them into something that's supposed to be useful. you are just making me put your shit into another llm to distill meaning out of it.just stop. pls.Apr 30, 2026
Two questions from the Jagged FrontierA non-exhaustive list of questions that I do not really see discussed as much as I think they probably deserve to be discussed.Question 1For most of our professional lives, we've been able to have fairly high confidence in the rough contours of the landscape that we're building into at a six to twelve month horizon. Computers will be a little faster next year, have a little more RAM. There will be a new version of Windows within the next 3-6 years.The curious thing about the last two years is that the basic shape of this bet continues, but the outliers are becoming increasingly dramatic. You've been able to bet that AI models will be better than they were, often concerningly so, within the next twelve months -- at least in terms of the models doing useful work. In January 2025, I told someone that a certain migration tool would be pretty much worthless in a short amount of time, because frontier models would simply progress to the point where they wouldn't need assistance to do what the tool did. That was prescient, I think. When I posted last June that it was the end of observability as we know it it came from the same place.Now, not everyone has gotten that memo. I think they will, very soon. This year? Next? I dunno. I also think that a lot of AI startups selling 'AI SRE' are going to be in for an extremely painful wake-up call because frontier models are just going to do this stuff. They already can! Perhaps the only thing more trivial for Claude than writing leetcode solutions is running obscure awscli or kubectl commands. If someone trusts your ChatGPT wrapper to actually restart services in prod without some really good legal protections then count yourself as fortunate that your client's lawyers practice three-cueing.Which takes us back to the question, how do you cope with the maximums and minimums of the capability bet? This has been an eventful 72 hours or so, which helps color the discussion. How are you making product plans that have a probability range from 'eh, things are more or less the same' to 'we all talk to our computer now all the time and software is just kinda generated on the fly'? How are you planning engineering capacity around models that may be taken from you with four hours' notice on a Friday evening? We can talk about 'taste', 'accountability', 'contracts', whatever else until our faces turn blue. I do not know how we can talk about the potential upsides and downsides of AI in software because it is not only moving very fast, the window of possibility is widening every day.Question 2To quote Mass Effect, Sir Isaac Newton is the deadliest son of a bitch in the galaxy. Similarly, momentum is the force that drives the modern economy. This can be seen in both the, uh, confusing exuberance/"it can only good happen" of the stock market but also in the long-term contracts and business deals that underpin large companies (e.g., Global 2000). To wit, "nobody ever got fired for buying IBM" (in some places, still true!)Momentum is driving a lot in the world of software these days. The actual reason AI gets as much investor attention as it does is because it upsets that momentum. People will write skeptically about "well, these valuations don't make sense unless OpenAI/Anthropic can capture a notable percentage of global GDP" and yeah, that's why people are investing in them. Contracts end. Virtusa, General Dynamics, a shitload of other companies you've never heard of but make the world go 'round through business process outsourcing? In what world do they survive AI? Somewhere, someone with fiduciary responsibility is doing the math on this and at some point the line will cross a threshold and a lot of these places will get cooked by it.There's another part to this question, though. What will you do when momentum catches up to you? It's gonna get everyone, one way or another. Your customers and clients are in the window, same as we all are, and the ones that are ahead of you in AI are gonna be asking some tough questions. Hell, I've seen people say to us "it would take claude an hour to fix , why is it taking you all so long" and they're right! AI native companies are going to push the hell out of you to run faster, are you gonna be able to keep up? What's your strategy for this?Question 1+2Readers may notice that I have actually asked the same question twice, and you'd be right, you sly dog! My point is that the frontier is extremely jagged, and it's only getting more so. The supply-side perspective is "ok, how do you deal with that when you can't see the edge", the demand-side perspective is "ok, how do you adapt to that when you don't know who's holding the knife" and my actual question is "is it the same answer to both". The 'jagged frontier' is about surprising capability gaps; I would argue that the real jagged frontier here is the unexpected change in potential outcomes.I would love to say I had an answer. I can tell you what I don't think will work. I don't think doing nothing is an answer. I don't think trying to embrace some sort of butlerian jihad against the thinking machines is going to go anywhere. I think that there is an extremely difficult to quantify element of risk in the industry's adoption of AI, and I think a lot of people are making some very foolish decisions hoping that they can outrun their friends. Meanwhile, the bear that's chasing all of us is turning into a Megazord or something, and the Power Rangers are nowhere to be seen.Here's what I think could work, though, and it's the embrace of AI as normal technology. Part of normal technology is that normal technology becomes cheaper, faster, and more accessible over time -- especially technologies that are decoupled from scarce resources. Computing is not a scarce resource in the way that oil or plutonium is. I believe that AI training will become cheaper for purpose built models through improved efficiency in large-scale serving.Semiconductors are scarce, true, but they are not consumed in the same way that many resources are. Chips are everywhere these days, and we can always make more! There's a whole-ass sun in the sky giving us free power forever! At last, we have invented the abundance machine from the cautionary tale "the abundance machine is really cool actually, someone should make one of those". The threat to intelligence on demand is not resource scarcity, it is unexpected externalities -- like, say, holding a model ransom because you didn't donate to the President's Emotional Support Ballroom Fund.If there exists a way out, it is by embracing the principles that have served us as technologists well over the decades -- protocols, composability, networks. This is what saves us from the downsides of a national security letter, because we can simply route around problems. I don't want to necessarily talk to a Claude to get feedback on this post, I want to talk to a focused, composable writing coach that's been distilled from something greater. What are the economic incentives that lead us to that? I don't really know, but I think it's an interesting question. Do OpenAI and Anthropic become the equivalent of, like, car manufacturers?Anyway. The best part of all of this is, if I'm wrong, and we are on the cusp of recursive self-improvement and someone makes Digital Jesus, then it doesn't really matter anyway. I'm gonna take a bet on openness, collaboration, and conviviality though.Happy Summer!Jun 15, 2026if i have to read one more piece of generated text i will kmsthere are approximately three good reasons to generate text that another human being will read.you are writing code with ai and the comments and symbols are mostly meant for consumption by other ai agents.you are running an ai agent/bot that is clearly labeled and defined as an ai.you are running an ai agent/bot for yourself and you would like to read the messages from that agent.outside of that? please stop. please stop writing status reports into slack that were generated. please stop making slides that were generated (oh my god please stop generating your slides), please stop generating emails, please stop. stop. stop. you are not helping. the total amount of the joy in the universe is shrinking every time you tell claude or gpt to write a memo or a PRD.this is not to say "don't use ai to help you write". that's fine! i use ai to proofread my stuff a lot, to poke holes in it. most of the time - for things i care about - i don't really incorporate its style/grammar suggestions (although i usually do wind up reframing some arguments or moving things around) but it does help. i'm not talking about that. i think that's good and useful. what i want you to stop doing is feeding in a bunch of garbage notes into the inference hole and tokenwashing them into something that's supposed to be useful. you are just making me put your shit into another llm to distill meaning out of it.just stop. pls.Apr 30, 2026
