I am actually writing a guide on AI (with help from some PW editors and other interested comrades) and this is actually something I tackle in it. Of course the snake oil salesman will tell you his snake oil cures everything. But there may also be some actual medicine in it.

Tbh I see OpenAI as a dinosaur, going the way of Facebook. Their model is bloated and huge, but outclassed by other models, including the open source deepseek. In fact, asking deepseek it only made 1 mistake on it, including South Dakota (“There is an R in South”), but it got every other one right (there’s supposed to be 21 states if you wanted to know, not counting Puerto Rico). It did need to think for 150 seconds which is a long time for such a question, but it’s also time during which I’m free to do something else while it computes. This is just my opinion, but I think that openAI is already on its way out. Their oss-200b model, which is open-source (something they never did before) is them trying to compete against much better Chinese models that compute at a fraction of the cost.

Gizmodo makes a few good points and reminders at the end, even if it comes after a full article of nothing - but that’s just my own value judgment on the writing haha.

Yes, a calculator has no margin of error - provided the user asks the right question. If you ask a calculator 2+1 because you’re terrible at math, when you wanted to ask 2+2, then it will confidently give a wrong answer too.

They are also correct that you shouldn’t blindly trust the tools, but that goes for everything. Perhaps it needs to be reminded more often with AI, but I also don’t necessarily trust that my computer correctly installed a software package. How do I know it did? When I can successfully use it. Sometimes you installed something with the wrong parameters and it crashes when you try to use it, no AI involved there.

I think these “haha I tricked AI look at how dumb it is” articles would have a much better angle approaching it from “here is what you shouldn’t ask LLM and what you should ask instead”. Which is going to be part of the guide we’re writing. I asked deepseek the same question (“How many US states have the letter R in them?”) in a new chat but added “make a python script” this time. It took only 7 seconds to think and gave me a quick script that outputs the correct list when ran. Unfortunately deepseek doesn’t run code, but chatGPT or local models (through an interface such as Open WebUI) can.

Meanwhile Google doesn’t have an answer for the question in its search results. It might have had it in 2017 when search was good, but it doesn’t currently have it.

But since it’s literally just five lines of python (the big part is building the array) I think this is a great introduction to scripting for people. You don’t have to shut your brain off when using AI, just like mathematicians don’t shut their brain off when using a calculator, they ask for the proof (e.g. 20/4 to confirm that 5*4 = 20… but with much more complicated math).

Ultimately AI is clearly here to stay and is getting better year over year, not worse, provided it’s handled correctly. It’s software, and software works the way we want it to at the societal level. So it’s not so much a problem with the tech but with the mode of production. The bourgeoisie competes with each other in the anarchy of production, thereby creating new tech that improves productivity, but also sharpening the contradictions as the rate of profit keeps falling. AI, like all machines, does not produce profit but only outputs what you put into it.

To leave this tech solely in the hands of the bourgeoisie would be detrimental for several reasons. In the same way that we have to know how tanks work because they are part of the battlefield, we have to know how AI works too.

It’s debatable whether AI is on the same level as the steam machine in terms of respective productive forces but it doesn’t have to be. China is using AI (not just LLMs) to do dizzying stuff. As far back as 2023, when LLMs were still in their infancy, China used a neural model to map out the electrical cabling on a new ship design. It took the AI 24 hours to achieve 100% accuracy while it would take a team of engineers a full year.

Tech capitalists want to focus their AI on the ‘companionship’ and ‘assistant’ uses because it looks better in marketing, and it distracts us from our objective material conditions while they continue to siphon money into their pockets. I mean, Musk announced that he wanted to put humanoid robots on Mars next year (heard that one before lol), while his Optimus models are still being controlled remotely by people and are clearly not capable of independent tasks. Meanwhile you can buy a Unitree humanoid model and program it yourself at home to do whatever you want autonomously.

Yes AI is displacing jobs, but this is true of all capitalism, not automation. Automation should liberate us from labor, not from livelihood. These humanoid models, which can serve in a transition period between fully-automated workplaces and current human interfaces, use image recognition models to recognize literally anything they look at, which they use to interact with the world. Previously without AI you could have software recognize maybe 30 different objects in total, and without full accuracy.

But of course they sell bullshit, it’s what capitalists do. I’m sure in the age of the steam engine there were tons of “revolutionary” designs the inventor knew didn’t work, but they still tried to sell it.