The Start-of-Spring Update

⏱️ reading time: 6 min

Hi Patreon supporters!

First, sorry I forgot about the "drawing of a cat" reward for being a supporter. Here's a drawing I made yesterday, of a cat being a baller:

Second: I know I've little to show over the last few years. Compared to my super-prolific years like 2014 — (my first explorable, my collab with Vi Hart, and my first interactive storygame) — I've not shipped a lot recently.

And yet, I still feel overwhelmed.  I mean, I've done a lot of stuff recently:

• The book-length series AI Safety for Fleshy Humans

• A few in-progress AI Safety-related research projects

• Recovering from the US border confiscating $3000 worth of my devices

• Five puzzle games for a secret contract

• The intense 10-week MATS AI Safety research bootcamp! (which I did in London, UK)

But that hasn't led to much public, polished output, y'know? It's all either in-progress or under-NDA.

Anyway, that's all to say:

* 🙇 Sorry for the lack of fully-shipped projects, and
* 🙏 Sincere thanks to those of you who've stuck with me all these years!

I hope that this is a "bamboo" situation — spend years building my new root-system, then burst from the ground in a short time. 🎋

Anyway, here's my update since the last post!

-------------------------

1] Still never got any of my devices back that were confiscated from the US-Canada land border. (About $3000 worth - new laptop, phone, iPad w/ charger & accessories. No, I don't know why they took my Apple Pencil, too.)

= = =

2] MATS AI Safety research bootcamp went ok! The project I was working on:
 
* A crisis-roleplaying game, like Pandemic or Model UN
* but about the global AI arms race
* but in the casual online style of Jackbox Party Pack
* and powered by LLMs

Here was my project poster: (click here to see full resolution)

Sitting in a room! 

We were hoping to launch something publicly by the end of the program, but the timing was too tight. We may finish it up in August? (Don't worry, I'll replace all the placeholder AI-generated art with actual human art.)

= = =

3] Been very productive on the secret puzzle games contract!

Context: I'm doing a contract because this economy sucks & I need to diversify my income streams. (I'm also applying for various grants; let me know if there's grants you recommend for me, around science communication or indie research or AI/AI Safety!)

Anyway, I've been getting a lot done on this secret contract!...

Too bad I can't show you anything about it. -_-  

I can say though, I've been hecka productive in part thanks to CursorAI, a new AI code assistant/editor. Finally, AI-assisted coding that doesn't totally suck!

Previously, the top AI code assistant was GitHub Copilot. I never tried it myself, but my coder friends tell me it's good at basic boilerplate, but net-negative on everything else, since you'll spend more time fixing CoPilot's bugs, than time saved from coding.

But CursorAI is net-positive! Sure, occasionally it hallucinates, but uncommonly enough that CursorAI saves me time on net. I'd describe Cursor like having a decent intern. You still need to be its "manager", and fix its mistakes, but still worth keeping around. Especially if you're mooching off the Free Tier like I am.

(HOWEVER: Cursor is definitely not for beginner programmers. If anything, Cursor may be net-negative for beginners, since you won't have enough knowledge to catch its errors, and relying on Cursor will prevent you from learning to code better. Honestly, I will be on guard for if using Cursor causes my coding skills to atrophy.)

This isn't a paid ad and I don't know anybody at Cursor.  I just think it's neat.

Yes, it's weird how programmers are the one profession who are really excited to put themselves out of a job.

= = =

4] Up next: AI Safety for Fleshy Humans Part 3 (Finale!), a vacation, then a half-year of experimentation?

Right after the secret contract, it's time to finally finish AI Safety for Fleshy Humans!  Finale, Conclusion, Part 3!

The first draft was already completed & sent for feedback in Dec 2024, now it's time to:

1. Incorporate the feedback
2. Briefly explain & give context to the biggest AI news since December (e.g. DeepSeek, the Alignment Faking paper, the Emergent Misalignment paper, etc)
3. Draw all the art
4. Make flashcards & feetnotes
5. SHIP IT.

After that, I'm taking a vacation.

Then after that, I want to recapture that feeling I had in 2014, of making lots of experiments!

A few on my list:

📽️ Make educational YouTube videos. Hate to admit it, but video is much more sharable than writing & explorables. The video-creation plan: make an interactive sim, then record myself playing with it (VTuber-style) while explaining the ideas. This way, I can re-use my old skills to get a unique style of educational video!

(For what this may look like, see this experiment I did last year. Full unlisted video here, GIF below:)

 (P.S: I may finally learn to delegate. If you or someone you know is a video editor for commission, send me your portfolio!)

👯‍♀️ Make an LLM Clone of myself that passes a "Friend Turing Test".  Write a large documentation of my memories, character traits, beliefs & values, etc... then use Resource-Augmented Generation with a fine-tuned LLM, to see if I can make an AI that responds so similarly to me, that even my friends can't tell between its responses and mine.

(Similarly, I'm tempted to give y'all fans a test: see if you can tell my science-communication apart from an AI's. For context: humans are worse than chance at telling AI poems apart from human poems, and prefer the AI poems. [Nature article])

🧪 Do game theory/AI Safety research.  I've posted some notes publicly before, on my in-progress research.  But to recap with no explanation, I want to explore:

* The game theory of self-modifying agents
* Agents that can learn/modify each others' values
* Ways to beat Goodhart's Law
* Seeing if 'thinking like a scientist' can beat goal misgeneralization
* LLM-augmented causal inference
* "Cyborgism" tools to make AI enhance humans, not replace us.
* (and so so much more I wanna do in life!) 😭

-------------------------

I know I'm breaking the "underpromise & overdeliver" rule, but there's sincerely so much I want to do in life and so little time! (Maybe this is why I'm tempted to 'clone' myself?)

So even though the above is overscoped & I'll likely fail, I thought I'd write out my plan anyway. To paraphrase probably-Eisenhower, "plans are worthless, but planning is indispensible."

(Or to be less annoyingly paradoxical: "Planning" forces you to concretely pin down your values, options, and constraints. Any specific prediction of what'll happen – a "plan" – will likely fail. But when it does, the work identifying values/options/constraints will still remain useful, and help you create a new path forwards.)

Welp, that's enough yapping for today!  Here's another drawing of a cat being a baller:

 Again, thanks for sticking with me through my low-output period. May this bamboo burst from the ground soon!

🎋🎋🎋,
~ Nicky Case

P.S:

1. Let me know of grants you think I should apply for! (math/science communication, AI safety, etc)

2. If you or someone you know does video editing on commission – (with an interest in science communication) – send me your/their portfolio & contact!