AI Is Not Your Friend
3d 1h ago by slrpnk.net/u/foxymochakitten in zines@slrpnk.net from www.sherwoodforestzinelibrary.org
neat little one-page zine about genAI!! Please spread it widely!
As a coder / engineer who occasionally uses AI but also likes countercultural zines, I would add a few thoughts.
Using generative AI for art, music or essays is a foreign concept to me. The things that I make aren't intended to look nice, and essays aren't given any weight in this field.
- I use AI to locate texts to read.
A language model has read millions of books and articles. More than I can possibly browse. If I need to locate literature about a subject, I sometimes ask a language model to propose works on a certain subject. If it hallucinates and a text does not exist, I will find out when attempting to access the text. So far, I have not yet personally encountered this situation, but I have seen others stumble into this pit because they only wanted to reference a text, not read it.
Check everything that you get from a language model.
- I use AI to locate software by functionality.
A language model has trained on great amounts of source code. More than I possibly can. It can also quickly bypass the confounding factors of non-intuitive project naming or project obscurity. If I need to locate free software which can already do something so I won't reinvent the wheel - I sometimes query a language model.
I have got several false or poor leads, but I have also located algorithms and implementations which avoided me considerable labour.
- I use AI to translate between languages (natural and artificial)
I know only a few human languages (Estonian, English, Finnish, Russian, some amount of Ukrainian, French and German). If I need to read stuff in Spanish, Czech or Chinese, language models can make it much easier. Double-checking seems advisable.
Similarly, I only read and write C, C++, Java, Pascal and Basic. If someone has written a piece of code in Rust, Python or LISP, using a language model seems very appropriate instead of head-banging at language barriers. Sometimes it won't compile or run. I've seen a case where Chat never used the arguments for a function, and a case where Claude invented a lookup table out of thin air instead of the other project's sources. Double-checking is advisable.
- I don't pay.
To ensure that I'm not causing an excessive burden on resources, I only limit myself to what is offered freely. My average usage is about 5 sessions (each maybe 10 queries) per month, during an intensive period of research. Typically less.
I wish I knew how many liters of water this would boil, but sadly they don't tell.
You should consider using local models instead of cloud ones. That both reduces the power you give to companies and your reliance on them remaining free.
Language translation is only as good as the data set. There's a ton of TS/js out there (and a lot of it is horrible) so LLMs tend to be able to spit out working js (but the security properties are at best questionable). They tend to be less good at python, and obscure things like lisp are more likely to just be totally wrong. (IMHO, the entire paradigm of having machines try to make code to be read by other machines is nonsense because code is for people. That's literally the point.)
This is also true for natural languages.
Finding things is really just about encodings, so if you're building something like a RAG then you kind of don't even need to bother with the inference. I've build half-RAGs before just to get semantic search and then just skip the inference part after I have my doc list.
This is, of course, less useful if you're looking for something on the internet and you don't have a local search engine. But I'm not sure that an LLM is better than just using a search engine. That's just my take, though. Perhaps I just haven't experienced it.
Yeah, there's definitely a bit more nuance. I think most folks can just skip LLMs, but there are some use cases in tech.
The thing about local models is that inference is cheap. They're trying to charge for inference to cover the cost of training, and training is what uses all the resources. If you refuse to pay for inference, then the economic model collapses. IMHO, that's the critical part.
ChatGPT agrees
