How hard can it be? /ᐠ｡_｡ᐟ\

July 2, 2026

Manufacturer Utz issued a voluntary recall in May for varieties of its Zapp’s and Dirty potato chips products, citing the possible presence of salmonella in dry milk powder sourced from a third party used to make a seasoning ingredient.

An estimated 650,000 bags of chips with best by dates ranging from 27 July to 31 August were believed to be affected.

This is good information to know.

Green sea turtles have been an iconic endangered species since the 1980s, mostly due to bycatch in fishing nets and taking adults and eggs for food. I have core childhood memories of watching Steve Irwin talk about sea turtle conservation in front of a beach full of nesting green sea turtles.

After 45 years of conservation efforts, their population now meets the IUCN criteria for Least Concern rather than Endangered.

For a species with such a slow life history (it takes green sea turtles decades to reach sexual maturity) and a nearly global population, this is a Really Big Deal. This is the kind of long term conservation victory that many of the amazing humans who started working on green sea turtle conservation back in the 80s didn't live to see.

OpenAI scraped the entire web. All of which had been a labor of love from humans. Wikipedia is the backbone of a lot of LLMs, and that was volunteer human labor. They stole it and now they're selling it back to us.

Yeah so the language here is actually super important and unfortunately the companies have successfully made it really confusing for laypeople. This is my quick glossary, it won't cover everything

Statistical Model - Mathematical model that relies on finding statistical correlations in input data to do...something. Here are some things they might do:

Make predictions about the future

Categorize or label the input

Take in user input (a "prompt") and output something statistically relevant

The important thing to remember about a statistical model is that inherently it can only deal with probabilities. Things that are more or less likely. It's not a model of truth, in the Stanford Encyclopedia of Philosophy way of thinking about truth. It's something that can tease out probabilistic patterns in input (is "green" more likely to be seen with "red" or "yellow"?)

Artificial Intelligence / AI - Marketing term. It doesn't really mean anything, it can refer to all sorts of math and/or software, though typically some kind of statistical model. Was initially used as phrase to help get grant funding for computer operations research. I don't like using it at all, ideally, but when I do I typically mean the suite of products that have been sold under this label: Claude, ChatGPT, Gemini, etc. As a scientific term, it's useless (we don't have a good definition for "intelligence", let alone the artificial version).

Large Language Model / LLM - What it says on the tin. It's a big giant statistical model trained on text. What "large" means keeps changing, but at this point basically all the big flagship LLMs are trained on all text that has been digitized. This training process is really expensive computationally which is why the AI companies keep building these big data centers. Now technically LLMs could operate in any of the three modes above, but typically we think of the chat bots that are prompt-output. You type something into the text box and you get back a response.

Transformer model - Terrible fucking name, but this is the specific software architecture that modern LLMs run on. It's extremely efficient and has allowed for language models to progress very rapidly over the last five years.

Small Language Model / SLM - A language model trained on a corpus that is small. These are lower compute requirements and can be specialized by being trained only on legal documents, or something like that. But they're still statistical models, so the output is still probabilistic.

Basically all LLMs are transformer models these days, and most AI products are LLMs, though I think they're starting to call stuff like step counters AI at this point.

the image of the voodoo doll as a magic doll you stab to hurt people actually comes from a practice by english pagan witches of making poppets. the poppet would be made in the shape of someone you hate and filled with their hair and then stabbed with big needles repeatedly. and the poppets got conflated with voodoo dolls later on.

voodoo dolls were actually a way of keeping records when black people were prevented from reading or writing. a voodoo priest who would serve as a communities doctor would make a doll that looks like the patient, and stick a pin in it where the patient was experiencing symptoms, this would let the doctor keep track of who was having what issues. for example if a patient has a headache, the voodoo priest puts a pin in the dolls head. the priest gives the patient a treatment, when the patient comes back the priest knows which doll is theirs because it looks like them, and they know they're dealing with a headache because the doll has a pin in it's head. when the symptoms resolve the pin is removed, until then the doctor has a reminder that this is an ongoing issue for a specific patient.

I wrote this because when I saw it this post had a TL;DW that conflated the information about poppets and the information about voodoo dolls, creating new and original misinformation about voodoo dolls that comes from confusing them with poppets so like don't reblog that one maybe.

also someone else in the chain included this freeze frame so you can read it