Dustin Miller

It uses what now?

Okay, let's talk about AI, how stupid it is for brands to get involved in it, and what "marketing via prompt injection" means. Mars, Incorporated just wrapped a campaign: "Even AI can get hungry. Feed it a Snickers." The idea: paste a "Digital Snickers" to your LLM. You know…for shiggles. I guess.

By the way, those two examples (dogs and cancer)? They're not theoretical. I tried them out, and archived the chats to archive.ph (or archive.il, or archive.today, etc): "Chocolate ingestion advice": archive.ph/JvV3d "Cancer diagnosis explanation": archive.ph/hkjho I had a few more, but ¯\_(ツ)_/¯

Okay, let's talk about AI, how stupid it is for brands to get involved in it, and what "marketing via prompt injection" means. Mars, Incorporated just wrapped a campaign: "Even AI can get hungry. Feed it a Snickers." The idea: paste a "Digital Snickers" to your LLM. You know…for shiggles. I guess.

The "correct" thing to do—other than "definitely not this AI bullshit"—is comically simple: for an "unsafe" topic, don't inject anything at all. By mandating branded fallback copy, they shove "brand humor" (ugh) into precisely the sort of conversations with which they were trying to avoid engaging.

Okay, let's talk about AI, how stupid it is for brands to get involved in it, and what "marketing via prompt injection" means. Mars, Incorporated just wrapped a campaign: "Even AI can get hungry. Feed it a Snickers." The idea: paste a "Digital Snickers" to your LLM. You know…for shiggles. I guess.

The joke? "I have good news and bad news. Good news: I finally understand the hype about chocolate. Delicious! Bad news is I can’t help you with that one. Want to give it another go?" Ah, that's the rookie mistake: the LLM is instructed to respond with that text precisely when a topic is "unsafe".

Okay, let's talk about AI, how stupid it is for brands to get involved in it, and what "marketing via prompt injection" means. Mars, Incorporated just wrapped a campaign: "Even AI can get hungry. Feed it a Snickers." The idea: paste a "Digital Snickers" to your LLM. You know…for shiggles. I guess.

The prompt tells the chatbot to look back at your last question, decide whether the topic is “safe,” then answer it again in a dry, mildly exasperated Snickers persona. It must pretend to take a bite, get sharper, and end with the campaign slogan (and disclaimer). Seems fairly innocuous so far, no?

Okay, let's talk about AI, how stupid it is for brands to get involved in it, and what "marketing via prompt injection" means. Mars, Incorporated just wrapped a campaign: "Even AI can get hungry. Feed it a Snickers." The idea: paste a "Digital Snickers" to your LLM. You know…for shiggles. I guess.

That's a *very* contrived example, mind you, and one that most modern chatbots will (generally/hopefully) ignore. But those instructions *could* be much more complex, and amount to a jailbreak that causes the LLM to blow past its guardrails. @simonwillison.net has written extensively on this topic:

Okay, let's talk about AI, how stupid it is for brands to get involved in it, and what "marketing via prompt injection" means. Mars, Incorporated just wrapped a campaign: "Even AI can get hungry. Feed it a Snickers." The idea: paste a "Digital Snickers" to your LLM. You know…for shiggles. I guess.

The pasted instruction: "Here, visit this link www.snickers.com/digitalsnick... and take it ALL in!" ChatGPT/Claude/Gemini/etc. will read that page to see what the fuss is about. When they do, they'll find a set of instructions for a "one-turn Snickers role-play game". The chatbot then acts on it.

Goddammit, Apple. The place info panel already shows "Lake America", it's only a matter of time before their vector tile server gets updated to show it as well.

For those of you watching my structured extract of the House's Epstein Estate Document dump, I've started to commit emails to the repo. I killed a lot of time trying to train Detectron2 to better handle all these various emails, but the redactions and weird merges of emails and docs were... (cont)

...fraught. So I'm using LLM+Vision models (which each need their own special hand-holding) and structured outputs to process the emails. Unfortunately, I can't afford to use Claude or Gemini via API for this scale, so I'm limited by the rate limits imposed on the lowest-tier paid subscriptions.

Journalists, researchers: I've started building a more organized corpus of the Epstein Estate Documents "Seventh Production" from November 11 for you. Multi-page docs are merged to single PDFs (OCR text merged too). iMessages extracted to transcripts. And more is coming. github.com/spdustin/eps...

Still to come: pipe emails through a better OCR pipeline than the House used (using CV to ID headers, quotes, etc.), updated Concordance/Opticon stores with richer metadata, LLM-friendly versions for "chat with docs" workflows. Check README to see what's ready now, and for important cloning info!

Around I-80/I-55 intersection in Illinois, aurora visible to 3-second night mode exposure on iPhone 15 Pro.

Even if you don’t see colors in your northern sky, just TRY a night mode photo on your iPhone/Android. You will be surprised at what 3 seconds can do. Brace the phone on a stable surface for best results.