My thoughts on AI

I would have to learn entire coding languages to do it myself, which takes years. AI can do it in 30 minutes and better than I could in years

this is an overstatement. once you learn the basics of one programming language (which does not take a full year), you can apply the knowledge to other programming languages, many of which are almost identical to one another.

There is of course the environmental cost. To that I want to say that everything has an environmental cost. I don’t necessarily deny AI is a water-hog, just that the way we go about it in capitalism, everything is contributing to climate change and droughts. Moreover to be honest I’ve never seen actual numbers and studies, everyone just says “generating this image emptied a whole bottle of water”. It’s just things people repeat idly like so many other things; and without facts, we cannot find truth.

according to a commonly-cited 2023 study:

training the GPT-3 language model in Microsoft’s state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwater, but such information has been kept a secret

the global AI demand is projected to account for 4.2 – 6.6 billion cubic meters of water withdrawal in 2027, which is more than the total annual water withdrawal of 4 – 6 Denmark or half of the United Kingdom.

GPT-3 needs to “drink” (i.e., consume) a 500ml bottle of water for roughly 10 – 50 medium-length responses, depending on when and where it is deployed.

there's also the energy costs:

according to google's 2024 environmental report:

In 2023, our total GHG emissions were 14.3 million tCO2e, representing a 13% year-over-year increase and a 48% increase compared to our 2019 target base year. This result was primarily due to increases in data center energy consumption and supply chain emissions. As we further integrate AI into our products, reducing emissions may be challenging due to increasing energy demands from the greater intensity of AI compute, and the emissions associated with the expected increases in our technical infrastructure investment.

according to the mit technology review:

The carbon intensity of electricity used by data centers was 48% higher than the US average.

and

[by 2028] AI alone could consume as much electricity annually as 22% of all US households.

there's also this article by the UN, but this comment is getting kinda long and the whole thing is relevant imo so it is left as an exercise to the reader

i have my own biases against ai, so i'm not gonna try to write a full response, but this is what stood out to me

Regarding the energy cost, my 65W cpu takes ~20 seconds to generate an image with sdxl turbo (4 cycles 3.66 seconds each + decoding). Thats 2769 images / kWh

I very much agree with all that. This is already a very useful tool, and it can save you a lot of time once you learn how to apply it effectively. As with any tool, it takes time to develop intuition for cases where it works well, and how to use it to get the results you want. I get the impression that a lot of people try using LLMs out of spite already having a bias that the tool is not useful, then they naturally fail to produce good results on the first try and declare it to be useless.

As you point out, it's an excellent tool for learning to work with new languages, to discover tricks for system configuration, and so on. I've been doing software development for over 20 years now professionally, and I know some languages well and others not so much. With LLMs, I can basically use any language like an expert. For example, I recently had to work on a Js project, and I haven't touched the language in years. I wasn't familiar with the ecosystem, current best practices, or popular libraries. Using an LLM allowed me to get caught up on that very quickly.

I’m also not too worried about the loss of skill or thinking capacity because the really useful skills lie in understanding the problem you're trying to solve conceptually and designing a solution that will solve it. High level architecture tends to be the really important skill, and I find that’s basically where the focus is working with agents. The LLM can focus on the nitty gritty aspects of writing the code, while I focus on the structure and the logic flow. One approach I've found very effective is to stub out the functions myself, and have the agent fill in the blanks for me. This helps focus the LLM and prevent it from going off into the weeds.

Another trick I found is that's handy is to ask the agent to first write a plan for the solution. Then I can review the plan and tell the agent to adjust it as needed before implementing. Agents are also pretty good at writing tests, and tests are much easier to evaluate for correctness because good tests are just independent functions that do one thing and don’t have a deep call stack. My current approach is to get the LLM to write the plan, add tests, and then focus on making sure I understand the tests and that they pass. At that point I have a fairly high degree of confidence that the code is indeed doing what's needed. The tests act as a contract for the agent to fill.

I suspect that programming languages might start shifting in the direction of contracts in general. I can see stuff like this becoming the norm, where you simply specify the signature for the function. You could also specify parameters like computational complexity and memory usage. The agent could then try to figure out how to fill the contract you've defined. It would be akin to genetic algorithm approach where the agent could converge on a solution over time. If that’s the direction things will be moving in, then current skills could be akin to being able to write assembly by hand. Useful in some niche situations, but not necessary vast majority of the time.

Finally, it's very helpful to structure things using small components components that can be tested independently and composed together to build bigger things. As long as the component functions in the intended way, I don’t necessarily care about the quality of the code internally. I can treat them as black boxes as long as they’re doing what’s expected. This is already the approach we take with libraries. We don't audit every line of code in a library we include in a project. We just look at its surface level API.

Incidentally, I’m noticing that functional style seems to work really well here. Having an assembly line of pure functions naturally breaks up a problem into small building blocks that you can reason about in isolation. It’s kind of like putting Lego blocks together. The advantage over stuff like microservies here is that you don’t have to deal with the complexity of orchestration and communication between the services.

I think it would be very useful to have an AI guide written for communists by communists. Something that everyone can understand, written from a proletarian perspective - not the philosophy of it but more like how the tech works, how to use it, etc. I can put it up on the ProleWiki essays space if someone wants to write it, we’ve put up guides before, e.g. if you want to see a nutrition and fitness guide written from a communist perspective.

This is not comprehensive, but here is a draft write-up of sorts for the time being. There are probably others who would know better how to write about best use of chat models like Deepseek, I'm not very familiar with them in detail:

What is AI?: AI stands for Artificial Intelligence, which can have broad connotation and could be applied to a number of forms of automation. More recently, it has become synonymous with generative AI, a specific subset of AI in which an AI can be given a prompt (some input from the user) and it will generate something based on the prompt (some output for the user). For the purposes of this article, I will be using AI primarily to refer to generative AI.

Is AI actually "intelligent"?: What is defined as intelligence and whether any AI falls within that may be up for some debate, but instead I want to address the nature of associations with intelligence, autonomy, and behavior. There is no evidence that AI has a "mind of its own" in the way that human beings have a mind of their own and when you consider that it is some math happening on a GPU trained to impersonate human things, not something with a sensate material form interacting with the physical world and then bouncing that back to an inner world, it makes sense that even if it were to have anything resembling intelligence, there is no reason to think it would look similar to the human experience. The point here is not to weigh in on the whole nature of consciousness and intelligence, but that you don't want to be fooled by an AI acting convincingly human. That's what it is trained for, but that doesn't mean it is like a human on a material level. It is still fundamentally a GPU doing some math and shares no actual material traits with you.

AI can be confidently wrong, even when it has high confidence in the right answer: An AI model effectively has a certain amount of probabilistic confidence in what should come next in any given input fed into it. The way it continues is token by token and a token is "however the tokenizer breaks things up", which varies based on the tokenizer, but most of them do a lot as whole words. There are, however, components of words sometimes too, such as -ly. Tokenizing aims to maximize the semantic literacy of the AI, i.e. it can better understand why some words follow other words if the components give context. A tokenizer that breaks everything up into letters would probably end up more like the word game Hangman than word association. Further, the base probabilities of the model tokens get sampled from through various statistical algorithms, which can dramatically alter the end probability of a given token. In effect, this means that it could have 95% confidence that "purple" is an accurate continuation of "red and blue make " and still choose "yellow."

AI has no ideology, but it does have biases: Because AI does not have a mind of its own in the way that we do, I think it's better not to think of it like it has an ideology or belief system. However, this doesn't prevent it from having biases. The probabilities it learns are based on what kind of material it was given in training data. If everything in its training is saying "Stalin did nothing wrong," it probably will too. If everything in its training is saying "commie bad", it probably will too. Further, some companies may consciously try to tune a model's biases in a particular direction, or block off certain paths of conversation, for this reason. The same goes for image generation. If most hair in its training is blonde, unspecified hair color will likely be blonde. If most skin is white, well... you get the idea.

AI, like anything, is limited by material constraints: Contrary to how it may feel at times, AI is not magic and its capabilities do have limits. If you ask it about theoretical physics, it might have a very good answer depending on how well it has been trained on material written by experts in theoretical physics. If you ask it about a TV show that came out yesterday, chances are it has never seen anything about that show in training and so it will have no idea what you are talking about. Like a human, if it has never encountered something before, it's not going to magically be able to know what it is. However, unlike a human, you can't just explain something to it that it doesn't know and it will now know going forward. More on that below.

An AI model can't remember what you said to it: This is tricky wording in a way because there are text services that offer some form of automated simulation of memory, where the AI can draw from some kind of record of certain things that you have said and these are supposed to play into how it responds. But the model itself is still not remembering what you said. Training AI models is expensive and training them to be different for a specific user would mean needing to have tons of variations on a model, which would get absurd fast and might result in an overall worse model for any given user (training can be a sensitive process and easily go wrong). Instead, most of what models are working with is a sort of "short-term memory", sometimes called context; this is still not really any kind of actual working memory in the human sense, it's just a way of thinking about input. The AI generates from input and so if part of the input is "I already said that, what are you talking about?", that hints at a kind of conversation where a misunderstanding has occurred. This doesn't mean the AI will "know" what you already said, it can only "see" what is given as input, plain and simple, and generate based on that. If what you already said is part of the input, it might be able to reference it. Otherwise, it can't.

AI doesn't know there is a "you" and an "it": Some chat-focused AI can be very convincing at acting like a real person, but this doesn't mean that the AI really "knows" there are separate entities talking to each other. The standard text model is a continuation model, which means they are continuing the given input, i.e. adding onto it. So to get an AI to simulate being a chat partner, special designing behind the scenes forces it to stop writing before it would start writing your side of the conversation. Without these rails, you could watch an AI simulate a conversation between two people, doing both sides of it.

Most AI is not private: This is not to say that people are necessarily interested in reading your specific chats with your username attached, but most AI companies use your chat data in some form, probably anonymized but nevertheless plaintext data they can read, in order to further tune their models. There are rare exceptions, such as NovelAI (which uses a special system of encryption) and ChubAI (possibly, based on wording of their policy), but you need to read carefully through Privacy Policies and how things are worded. A policy that simply says "your stuff is private" doesn't necessarily mean it can't be seen, it might just mean that there is no explicit policy in place for members of the company to be looking at what is on your profile.

Steering AI and avoiding unwanted outcomes: Most AI can go in a wide variety of different directions. This opens up a lot of possibility, but it also means a lot of ways in which the AI can not know what you want and go down a path you weren't expecting and would rather it didn't go down. Getting it to head toward the realm you want can be subtle or overt, and with the models that are tuned for instruction, overt may be easier, but either way, expect AI will not meet your expectations sometimes. And if it goes down a path you really don't want, it's usually better to move on as fast as possible, rather than getting bogged down in telling it how wrong it is. Remember that it's just continuing based on input, so if it sees "argument", it may figure "more arguing fits here". It doesn't know when to quit, it doesn't get weary, and the more there is of something in input, the more likely it will be to double down on that.

P.S. These are the points to cover I could think of at the moment based on what I know about generative AI. If there are other points you are thinking of when you talk about writing a guide, I'm curious to know. Can't guarantee I'd be able to cover it all myself, but yeah. Also am curious if this comes across clear enough for layperson. I'm not an ML researcher or the like, but I've picked up a fair bit on it over time.

I would say that you should not count on AI to write security critical code. This is pretty likely to result in vulnerabilities. AI often has oversights. At the very least you need to learn what common practices are considered secure and work towards putting them into practice. AI can help you do this with code examples, but you should try to limit this to things that the AI can prove to work correctly by running them. And especially do not try to roll your own cryptography.

We already have a tool for doing very common tasks in programming like authentication - they are called libraries and are at least an order of magnitude more reliable than what you get from AI. AI has its place too, but it's not useful for everything and can be especially harmful in certain situations like security critical code where you absolutely need the code to have certain properties that the AI cannot reason about.

I'd like to write an article on this or co author for it. We need a proletarian pov as you said.

You should not count on AI if you make open-source software. Many FOSS devs are against AI. Relying on AI would ruin years-old contributions to open source software, which are written manually by various contributors.