Caleb Briggs's Shortform

Caleb Briggs

This is a special post for quick takes by Caleb Briggs. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.

I've found when studying steering vectors that the J-lens restricted to emojis often gives me the best hypotheses on what a direction is doing.

For instance, here is the Assistant - Base model averaged over a few identical pieces of text (using the base model J-lens)
+ Assistant-Base: "🎓", " 😊", " ✅", " ❤️", "📚", " ✨", " 🥰", " ⭐", "🤝", "🔍", "🧭", "🌐", "🎙", "💡", "🔬", " 🤗", "🧘"
- Assistant-Base: "😬", "😒", "👿", "💩", "😳", "😹", "😜", "😛", "😰", "🙄", "😠", "😂", "😨", "🚬", "🤑", "😏", "😝", "🙈", "😑", "👹"

As another example, the word readout seems to suggest Arabic vs Sea,
+ Arabic, 阿拉伯 (Arab/Arabic), Arabic, árabe (Arab/Arabic), 沙漠 (desert), Saudi, arab, arabic, arabe (Arab/Arabic), Arab, arabic, arab, Saudi, Riyadh, Islamic, Arabian
- marine, seaweed, fisherman, Fisheries, maritime, fishermen, Newfoundland, Maritime, Vancouver, Baltic

But actually this is a desert/arid versus sea direction, and the names are only places associated with the desert. The emoji readout makes this more clear
+ 🏜 🐪 🐎 🐫 🕌 🏇 🕋 🌵
- 🐟 ⚓ ⛴ 🚢 🐳 🐋 🐠 🛳

Fun! But I don't think your evidence supports your conclusion.

I asked ChatGPT about its top 10 associations with the first emojis. It's first answer was "Arabia" and its second was "desert." Considering your emojis contain both "🐎" (as in Arabian horse) and "🌵" (which is not Arab) it's probably a mix of both meanings.

I could see this being useful as a quick and dirty way to get output diversity over a range of simple human concepts, so the J-lens can't just repeat variants of "Arab" again and again.