I think this is the way yeah. For extra protection you can also do physical backups of the project (copy pastes) at various points, because even if the LLM doesn’t know you have gitted your project, it may still run the command. The newer deepseek is much more biased towards doing this, I wrote “commit your findings to a file” and it wanted to git it. There’s always the possibility it can squash all commits or erase them (much like someone can write rm -rf in any terminal!) but this is why we invented prod/dev redundancy and RAID backups lol. You don’t necessarily have to be this paranoid when using agentic AI but it’s an extra security and some peace of mind.

I also checked and crush is completely able to write and run bash commands (incl. rm) on files not in the folder you opened it on. Definitely something to look into, I’ll check if there’s a way to containerize it better and make a post for !crushagent@lemmygrad.ml. Yog and I brainstormed the idea of making another linux user just for crush, then putting your main account in that user group along with the crush user, but not the crush user in your main account’s group. That way it only has perms to act on the files belonging to crush/crush, though it can still try to run any bash command it wants. And you would also have access to crush’s files with your main account so it’s more convenient. But I don’t know much yet about how linux users work, I’ll have to look into it and will make a post about it if I find something.

I think crush also has config files you can edit to blacklist or auto deny some commands.

The way these tools are being marketed by tech companies is completely wrong and prone to making disasters like this. It’s a tool; it’s like selling a fruit-only knife then leading customers into thinking it can only cut fruit and nothing else (until inevitably someone cuts themselves on it). I agree google has some responsibility there if this happened (his story seems a bit fishy tbh but that’s not really the point) and this is also why OSes bake some protective measures in such as user permissions. It’s also why everyone has been telling everyone to make backups for years even though nobody does it lol. 10 years ago steam introduced a bug that could wipe linux drives.

I see from his video that anti-gravity obfuscates the chain-of-thought and the outputs - it’s a proprietary model so they don’t want to share that, but it makes troubleshooting impossible. He also had it set on ‘turbo’ mode which bypasses requesting permissions to run commands - there should be heavy discouragement to users doing that,including making them actually edit config files imo, it shouldn’t just be a nice-sounding toggle because then people think “turbo means it goes fast of course I want it to go fast”.

They want to market agents as a do-everything app but it’s still software under the hood. And I don’t trust google to ship any good product anyway, but obviously that’s not how google markets itself. And of course you’re stuck with expensive google models if you use anti-gravity.

People are also right that this should run in a container with no way to escape it, and even crush (the one I use) is not great about this - though it should be possible to containerize it yourself. Coming from a company like google this kind of stuff should come out of the box with the software and set up for you. This is also one of the many reasons I switched away from Windows, the moment they announced integrated agentic I knew you would never be able to fully remove it.

I can believe what happened is possible – if anything it serves as a PSA not to trust software blindly. When I was a kid the most hilarious thing you could do on the internet is tell someone to delete system32 so. From one of OP’s comments it seems the problem was the space in a folder name that windows parsed incorrectly because of the OS’s rmdir command? No way to tell for sure since gemini obfuscates the output, and of course that’s just what OP thinks the problem was.

Someone tried to reproduce with more locked down perms and the output (pic) was just as concerning from anti-gravity. It said its “instructions” prevented it from running the command, when it should say “the agent prevents the command from being run” (and deepseek does say this in crush). I.e. this should be hard-coded but it seems to be passed to the LLM instead.

And as much as it sucks, you live and learn. People have been accidentally wiping their drives for decades at this point, I’ve probably done it too before when I was younger. If anything software was better about preventing this sort of thing in the 2010s, the 2000s were wild lol they gave you access to buttons that could reformat everything without even a confirmation button or an explanation of what the button was for.

It uhh does things to one’s brain and that’s about it

The calculator was also a nuclear device when compared to what it replaced/came before it. so was the car, and yet today nobody would tell you you can’t afford not to own a horse (except Homer Simpson maybe).

Things change and if we want to criticize that from a marxist perspective we have to offer something better than “I don’t like change”. It’s not all greener pastures with neural networks, but we need to be clear about what it is we criticize, and for that we need to understand things deeply.

But it is clear that people are using AI or attempting to use AI as a means to outsource mental tasks, and decision making that is endemic to the human experience / cognitive growth.

But what is the ‘human experience’? I push for people to define their words when it comes to talking about neural networks because more often than not it shows more similarities with what already exists than a break. It’s not that different from what we already live with daily. The more you use, understand and work with LLMs the more you realize that it’s really not so dissimilar from what we already know. You’re worried about a Wargames situation, i.e. the artificial intelligence making the logical conclusion that to win a nuclear standoff you should dump your warheads on the enemy first. But this has always been the plan; as soon as this technology was going to be available people were going to rush for exactly that - it just happened to happen in 2022 instead of 2065 or 2093, and so we have to reckon with the reality of it now, not later. Complaining that this is now possible won’t change that it exists and that it’s being used, so instead I made the choice to find my own uses out of LLMs that could be useful to communists (and incidentally I think we could probably organize for socialism much more efficiently around “the army wants to offload targets to an AI” than “AI bad destroy it all”). I’m not saying this to be dismissive, but rather that again we need to offer a studied, marxist perspective on the matter.

But speaking on the human experience/cognition, I mean, there are plenty of neurodivergent people who may not fare well with typical peer-to-peer communication (speech or written) and they appreciate having LLMs to organize and make sense of their thoughts and feelings. Disabled people have found answers from LLMs. Human cognition is not universal, and we see that LLMs already offer assistance there. When walkmans first came out, there was a huge panic around what they actually meant for society, that the youth saw them as a form of escapism, that it was an identity thing – it went so far that even novels were written about kids turning into mind-zombies after getting a walkman and some people event went on TV to say that using a walkman was a gateway to committing crime. We’re talking about the iPod that reads CDs.

I’m not even convinced by these studies that supposedly find all sorts of ills with usage of LLMs because I bet in just a few years plenty of errors will be found with them. They are lab studies, not real world, and I remember studies saying the same thing about search engines when they came out. I talked in another comment about how search engines are a memory bank for us; instead of remembering everything, we offload it to the search engine – I don’t necessarily remember what each property in CSS’s box-shadow does, but I know how to look that up on a search engine and find the information. Likewise we stopped remembering phone numbers the moment we got mobile phones (although we should probably remember one or two emergency numbers).

Significantly how? Both LLMs and AlphaFold are transformer-based neural networks. The LLM chatbot is trained on sequences of words, and AlphaFold is trained on sequences of amino acids. Certainly training AlphaFold to make real amino acid chains was ‘easier’ because we know how they’re formed so it only has so many sequences it can produce and there’s a checklist to determine whether the sequence it produced is real or impossible, so it’s also easier to have it produce a reliable output and makes it very good at a specific task, but they both work the same under the hood. Word prediction LLMs can’t have that deterministic output because we use words for so many different things. It would be like asking a person to only ever communicate in poetry and no other way.

one clear scientific purpose

Computer scientists in academia are using Deepseek to solve new problems in new ways too. They especially like Deepseek and Chinese models because they’re open-weights and don’t obfuscate any of their inner workings (such as the reasoning chain), so they can fine-tune them to their specific needs.

I have to assume the purpose of amino acids wasn’t so clear when we first found out about them and before we set out to investigate and, through extensive research and testing found out how they work and what they actually do. It’s on us to discover the laws of the universe, they don’t come to us beamed from heaven straight into our brain.

I wrote a lengthy essay on some of this if you want to read it https://en.prolewiki.org/wiki/Essay:Intellectual_property_in_the_times_of_AI

For outsourcing specifically, from the moment one uses a lighter to light a fire (instead of rubbing two sticks together) or opens a calculator app they are offloading tasks that require mental exercise, experience and wisdom. The abacus can also output a wrong number if you mess a step along the way - and yet people still used it extensively for centuries.

We outsource labor every day too. Not as in offshoring, but as in nobody lives in a vacuum outside of society. The food we eat is grown by others, packaged, shipped, and stocked on a shelf by a different person each step. The device we are writing and reading lemmygrad on was made by someone else. The LLM is not outsourcing human relationships because it’s not a human relationship; even people who use it as a stand-in for a human relationship know it’s not human. But, it solves problems.

I agree. let’s start with him

Honestly once you start seeing them as edgy kids it just makes detaching from it come naturally. They don’t affect me whatsoever. You think starbucks baristas are not workers? That’s great buddy you’re 14 come back to me when you’ve had your first work experience. You think Marx was a patsoc because your discord server said so? That’s great son but you’re 14, you were assigned Sherlock Holmes for English class go read that first to develop some critical reading skills.

Kind of scummy of haz and co to use the free labor of children to do their dirty online work for them.

Forcing them to defend their point and counter-trolling them in the way they troll you also gets them to shut up pretty quickly. Beyond the pre-chewed arguments they get fed on their discords they have nothing, they will bail and try to get their patsoc friends to jump in for them.

I once told one of them he was a paypig for the ACP since haz and co never took base members with them on these photo trips and he quickly stopped replying after that. I can only hope he started thinking about how he was being taken advantage of but who knows. I once told that guy noah or whatever from midwestern marx “no idea who you are or why you feel included in his conversation” lmao. he seems like a huge tool that desperately wants recognition and fame. But to be honest while that was funny, i saw someone in that discussion actually engaging him in theory and forcing him to explain his own assertions and he quickly dipped as he ran out of ways to do his faux “hello friendo!” shtick, so that was much more productive. Noah krachvik is not his real name btw and he keeps lying about where he lives not for any opsec but because he wants to look like a salt-of-the-earth prole.

Their techniques are nothing new, I guess it works on the internet. Facsists do it too, it’s the dodge pivot retort. Dodge the argument brought up to you, pivot to another, and retort with your own new argument to move the conversation. When people do that if you counter by preventing them from switching topics they will not know what to say anymore. Otherwise they would not be trying to shift the discussion. Attack your enemy where he is weak, force him to fight on your conditions.

Lol some of them might be 16!

I mostly see them on twitter and you certainly have some adults, in the chapters, but if you look at the pictures it’s mostly 5 or 6 people per chapter. It’s always the same people in the pictures, yet judging from their online presence you’d think ACP is everywhere. I can only assume most of them only like patsocism for the aesthetics or are too young to actually join.

saying “the prole.wiki guy” shows they don’t engage with detractors whatsoever. They don’t even read their own theory so don’t expect them to read anyone else.

Once you realize most of them are 14 it all falls into place.

I might get it thanks for the rec!

The criticism of the ACP should be strengthened. I also have “pro-authoritarian” politics, 90% of us here have (and 100% of us here ought to have them). For example I think it’s totally fine for China to execute CIA agents they uncover on their territory, but some would call that authoritarian. I think it was cool of Cuba to nationalize every industry in 1959 without compensation, but some would also call that authoritarian. Likewise I align with the “authoritarian regimes” of China, DPRK, Cuba, and many other maligned states in the world. This framing plays exactly into the ACP’s hand: “ACP supports China but ACP is bad, therefore it must mean China is bad too!” I suspect on some level this is something ACP wants to happen.

Much stronger criticism is their stance regarding basically the entirety of communist history (which they either reject or try to reframe out of left field), and how much of a break they represent with basically all of communism so far. They are the ones making enemies with other communists, by claiming they know better than us and we’re wrong and will be lined up against the wall when the american revolution comes. They talk big despite having nothing to show for it yet in practice.

The way their party works is basically a glorified tiktok house where supporters (members) send the influencers money so they can go on trips around the world and bring back instagram pictures of it. They show up for the photo op then bail without doing any work there. They are a complete break from communist internationalism there too because only the ‘executive committee’ ie the influencers (haz, helali, hinkle etc) ever go on these trips, they never bring any ACP members along with them.

They consistently lie, for example randomly putting the names of CPUSA chapters on their forming constitution and then these chapters quickly distancing themselves from apparently having signed that document. One of them that appeared on the document didn’t even exist.

They completely misunderstand theory and quote-mine every interaction, and have discord servers where they coordinate these arguments. If you’re anything on twitter it becomes blatant very quickly that most of them are 15 and think andrew tate was not edgy enough for them. They all share the same screenshot of a quote and if you press them even a little and demand they detail their thoughts they will bail but not before liking and retweeting their own tweet reply to signal to other patsocs to jump into the conversation. Their praxis is to berate and abuse marxists until they relent and agree with the ACP.

If we want to criticize their communism then we must judge them on marxist grounds not on vibes that they don’t ‘perform’ the role to our satisfaction and tick enough boxes. Their theory is wrong because they purposely misunderstand it and thus show a worrying agenda, and everything else flows from that. Their mistaken theory cannot build effective communist praxis and we see this already in the way they consider only some workers to be actual workers. But it can certainly create some nice little himmlers.

And it’s not just history, it has real consequences today. As I showed in the first paragraph, they are reframing what anti-imperialism and internationalism mean. Practically they are also known to be abusers, both physically and mentally and frankly I would not want to be around any of them.

I expect that leftAgainstACP community will eat itself soon. I’ve been invited to similar communities in the past and they end up imploding because they try to form a united front but it’s a doomed idea. You can see the posts already on that subreddit: “stalinism and the ACP have nothing to do with marxism, here’s why:” from yesterday, for example. Tbh I don’t really like the fandom aspect of this entire thing, it feels more like they’re trying to pit pros and antis, supporting people because they’re on the ‘right’ side of the fence even if they have other terrible opinions, posting memes to laugh at, etc. The moment you have people in these anti-acp communities saying china sucks it’s over, they’ve capitulated to western capital. Another post: “are pragmatic Social Democrats who are more concerned that withdrawing from NATO could precipitate a new global fascist movement and indirectly lead to a world war welcome?” (answers: yes welcome aboard.)

I didn’t intend to finish this promoting something but if you want to do something online but still material against ACP write for the prolewiki pages on them. we are among the top results if you search for their names or terms and they are some of our most visited pages, and there is still so much to add and so many more pages to create. That will be infinitely more worthwhile in teaching newcomers to marxism, which is who the ACP targets, about the danger patsocs are to them.

With so many websites abandoning RSS at this time it’s a majority of pressTV and telesur and not much else being posted tbh. some of the China section from techinasia. I have a couple other websites such as granma, periodico26 and news.antiwar.com for their coverage on possible war with venezuela but we select specific regex filters from these three

Wikitext is the only version it exists in, however you can more or less just copy and paste to the rich text editor of your choice, I think it should preserve formatting.

If you make a nice-looking PDF we can rehost on ProleWiki and link it on the book’s page.

I won’t post actual issues but it was very in-depth and although some of it was overkill (if it ain’t broke don’t fix it as they say), it did find a bunch of varied stuff – all from terminal commands, so nothing that wouldn’t be reproducible or had to invent tools for it.

I started typing up a huge post but i want to keep it more to the point; the tl;dr is these tools allow us to solve problems differently than we could before and this is where they find their place, and everything else is on us to figure out. Like I don’t think frustrations and bottlenecks will disappear under communism, they might certainly look different than they do under capitalism but they won’t go away entirely. We’ve had word-of-mouth, then books, then web pages, then search engines, and now LLMs. And whereas books and web pages only contain the information they do and nothing else (and web pages can be updated in real time but books cannot), LLMs can present that information in a way that speaks to you personally, and allows you to make further queries on it.

As a designer, we are trained to identify problems and think of solutions we can enact to solve them: this is the essence of design work. And eventually, we learn that there is always at least one solution to a problem. If there really isn’t, it means the problem needs to be reframed. Delivering information is one thing LLMs can do, but they can also do translation work, statistical work - or as we see here, help people switch from Windows to Linux much more easily than was previously possible.

Personally at this point the way I use LLMs and AI models is mostly to try and push them to their limits. The usual arguments I read about how “they think for you so you stop thinking”, or “but you could have found the answer on google” or “it’s not actually art” are so far behind where the goalposts are currently at! They’re already obsolete arguments because when I generate an AI image, I know it has errors, but I generate it to test a specific thing - like a way to prompt, or if it can do pixel art, or if it understands complex prompting (e.g. a picture within a picture within a picture all in different styles). I don’t even care if it’s art or not, it doesn’t make the image “unexist” and I’m trying out a tech task, so questions such as “is this illustration art or did I put effort into it” don’t even enter my mind.

Some people might see the image and dismiss it as “it looks bad here or there” or “I don’t care because it’s not actually art” but imo they are closing themselves off - what can you do knowing that you can fix a linux system? What ideas does it give you to solve your problems with? This is the real value. And ultimately we see that it doesn’t matter where a given piece of information came from so long as we have it in our arsenal. It’s like how some comrades close themselves off learning from liberals because they’re liberals, when we see that the US army makes its officers read Mao and Che among other generals.

Before LLMs some people probably said “I don’t care about google, I have my books” (or even “I don’t care about these new car things, I have a horse”). But google can help you find more books, and can also do a lot of other stuff I can’t think of right now - but just because we can’t conceive of these uses yet doesn’t mean they don’t objectively exist: our job is to find them out.

Finally to close this off, speaking generally – I mean that I didn’t get this impression from your comment – I think there is a tendency among even marxists to actually ‘shut off their brain’ as soon as anything AI pops up. “It’s AI so I don’t like it goodbye”. Like a tendency to scold comrades for using AI instead of directing that anger towards the meat or fossil fuel industry (biggest polluters on the planet if we care about the environment), or proprietary models that all have to compete against each other instead of pulling their forces together. This is what I mean by “closing ourselves off”. We leave this technology in the hands of the bourgeoisie because we think we’re too good for it. They don’t seem to think they’re too good for it though, and they control the state.

As for the diagnosis if you want to go further, I decided not to waste more context window and tokens on this (although it only cost something like 5 cents on the API since a lot of it was cache hit, i.e. repeated tokens sent to the API). It got completely bonked and I could just redo a new session and start from scratch but eh, I just pressed escape to stop it from running commands over and over again. And tbh that is a blemish. It started running and re-running commands instead of using both files. This can assuredly be fixed by just a tweak in the prompt, but it still bothers me that I ran so many tokens for something that ultimately 50% failed (the manual diagnosis part). It’s not a huge cost or anything, it’s just something it should be able to take into account itself since it knew that automated diagnosis file existed.

Anyway, at this time I have a list of diagnoses as per the initial prompt - both diagnoses it could run itself, and diagnosis commands I need to run myself. I won’t post an example here because it’s just good security practice but it followed the prompt when making the list: how it found it, what the problem is, the impact, severity as per my 4 categories, and how to fix.

From there it’s up to your preference how you want to go through this list: either your own linux knowledge, google or more LLM. You can just go back in crush and tell it “fix error number X” and I’m sure it will do it, unless of course it runs into permissions issues running commands. But personally I want some manual overview to learn the system and what I’m doing.

And some of these commands I have no idea how many years of linux you’d need to learn to write them yourself let’s be honest. And how many hours of googling you’d do to just find them in the first place! With agents you can do something else like watch a video or play a video game while it writes the file.

Manual diagnosis.md was more of a mess, because it didn’t really understand I wanted it to write the commands it was unable to run earlier, not “assume” it couldn’t write some commands out of thin air. I can still use this file since it collects sudo diagnosis commands but yeah it kinda failed on that part and I decided to stop wasting tokens. What I’m going to do now is check and run these commands, write down the result, and have crush analyze just this output and what it can tell me about it using the first prompt.

One reason they maintain nodes is for their infiltrated agents; They also maintain websites for that. I don’t know if that means they don’t also track you but they have self interest in having tor and keeping it alive.

It might be more likely if you own a node that you get CIA traffic on it…

no matter what one may think they have to admit deepseek is pretty much destroying the western AI ecosystem. On par with GPT5 on the free/cheap version is pretty big, considering how much fanfare there was around GPT5

it was pretty funny that deepseek gave me its own rundown technically lol, but its report on the official twitter thread is more in-depth than I gleaned from my own reading so