The only official Elixir and Nerves swag is available for pre-order/order at oswag.org.

Now shipping with stickers :)

LLMs

It has thoroughly penetrated the tech-wielding public. I know plenty of non-technical people that haven't touched it at all. Our parents currently don't seem to have picked it up. I run into a lot of programming students at the local campus and they breathe this stuff.

The ethics, the copyright, the environment. They continue to just fall outside the frame because about half of most society's ideology and politics just go "eh, progress will take care of it" because it is not in the framework to bother with consequences if the venture seems like progress in the market. So this will continue, barring any apocalyptic realigning of priorities. The environmental impact will be both overstated by those who are frustrated and ignored by those who don't care. It sucks. I disagree. I think we'll see backlash and bubbles burst but I don't think the word-engines will go away. Best case we get some utility out of them beyond all the slop that will fill the world.

Since these mathblocks are pretty good at re-wrangling information through the lens of a question I can see how they are very useful to a student. They will gleefully try to map an explanation into the terms of the way you ask which is not always what another human is willing to do and text-books, reference documentation and tutorials are notoriously unable to adapt to current understanding.

The DeepSeek kerfuffle certainly indicates (if truthful) that there are a lot of efficiency gains in terms of training that does not need to use the latest hardware, does not need the most data, does not require secret sauce. I'm glad if we get more efficient models that are more feasible for hosting yourself. I also hope that work gets replicated outside of China because that model has some perspectives on history that I don't appreciate.

This is of course also true about all the western models. They are not neutral, they encode piles of biases. They are generally aligned to the interests of major corporations, they follow the ethics of "avoid being sued and getting really bad press" to a large extent and as all technical leaders take a massive swing towards not just the right wing (they were always there) but straight up edge lord alt-right the models will start encoding those values. Fun.

I think it was yesterday that Mistral released Mistral Small 3. And this is the type of model I hope for and I'm unsurprised it came out of a European company. It is trying to be smaller, more compact, more efficient and still capable of some general and some more specific tasks.

I haven't gotten into the habit of using LLMs to ask questions or brainstorm or whatever. Maybe I should. Maybe I'm leaving a lot on the table. I should probably also trial LLM-based coding tools because fundamentally I don't think it is a good idea to refuse to learn about things. I doubt I'll like them. I barely tolerate IDE functionality. But probably should try it.

Where I see a model like Mistral Small be particularly useful is in translating messy inputs to function calls and so on. Extract arguments from natural language, match up written or spoken intent to one or several possible actions. And ideally this is a specialized model, not a massive general purpose one. If at all possible.

I don't think I will be shouting AI from the rooftops in the near term. I try to use the ML where I think it fits. I do see good uses for various models. I keep being surprised by how pervasive it has become with students, tech enthusiasts, people who do marketing and so on. I 100% believe that AI-generated images are already a worse signal to send when promoting anything than corporate stock photos ever were. We'll see the middle completely drop out of promotion materials. There will be the extreme of considered, human, professional or deeply amateurish but caring production and then there  will be something quickly done with generative AI.

I am not excluding the possibility of AI tools used carefully, artfully, deliberately. This will be someone's medium. I haven't seen someone make art that struck a chord with me using AI. I expect subversion of expectation, I expect someone will make ChatGPT dance within the walls of the topics it won't touch in a delightful way. The memes of bad AI are a precursor. simonw testing all reasoning AI by making it make jokes has a touch of this.

Anyway. They everywhere. Apparently.

Are you using them? Are you shunning them? Are you teetering on the edge of giving up on your principles? Let me know on the Fediverse where I'm @lawik@fosstodon.org or by responding to this email to lars@underjord.io.

Thank you for reading. I appreciate you.