ncase
patreon
Monthly update: Tell Don't Show
🕑 Added 2024-07-01 16:12:06 +0000 UTC(⏱ reading time: 4 min)
Hi y'all! Happy Canada Day from Canada 🇨🇦
Last month I got a bunch of work done, but alas, it's the kind of work that's not very "show-off- able". So this update, less show, more tell:
. . .
1) AI Safety Part 2 is 50% done. Reminder, this part is the one with all the comics, like this one:
(The comics were already done a long time ago. They were actually the first thing I finished out of this friggin' long 3-part series.)
. . .
2) Secret project under NDA (Non-Disclosure Agreement). The good news is I got a contract to make some free edu/puzzle games, the bad news is I can't tell you much more than that.
One exciting feature of it which I'll say: the games are "a few minutes a day" kind of games, like Wordle or Sudoku. In contrast, all my previous projects have been "play it once" deals. So: what kind of learning opportunities are there, when a player can do a small thing every day? (instilling habits of mind, etc?)
I can't say more than that, so here's a drawing of a cat:
. . .
3) Some research, trying to make an "Epigenetic Algorithm"
There's a well-known algorithm called a "Genetic Algorithm". The idea is to generate solutions by simulating evolution: random variation with non-random selection!
For example, to create artificial critters that swim:
1. generate a population of random critters
2. let them all race against each other
3. pick the best swimmers
4. create new population from the winners (clones with mutations, or cross-breed winners)
5. repeat to taste
Eventually, you "evolve" some critters like these (from Karl Sims (1994)):
Alas, in practice, Genetic Algorithms aren't used often, because they're kinda slow. In fact, this ties to a puzzle in evolutionary biology: the naïve version of evolution (mutate-select-repeat) is too slow to have created the diversity we actually observe in nature. So how did evolution work? To answer that question, let's ask a different question:
Why don't your eyes create stomach acid?
Horrifying visual aside, really, why don't they? Almost all your cells have the same DNA. So why don't they all do the same things? Why don't eye cells make stomach acid, why don't neurons grow hair, why can't you taste from your fingers, etc?
The answer is epi-genetics! Long story short, parts of your DNA can "turn on" and "turn off". This allows for much faster innovation in evolution! If a critter needs to evolve a new leg, instead of mutating a whole new sequence of DNA to make a 2nd leg, one can just turn on the old "leg" gene in a new location.
(Sounds nuts, but here's a famous example: take the Pax6 gene for an eye from a mouse, then splice it onto a fly's leg... and the resulting fly will grow a fly eye, NOT a mouse eye, on its legs. So that's evidence that 1) there's a gene that codes for just "eye", no matter what kind of eye, 2) this gene is so evolutionarily conserved, it works across mice & flies!, and 3) contrary to what evolutionary biologists used to believe, the eye probably didn't independently evolve 40 times, it probably only evolved once, then specialized.)
( If you're morbidly curious, here's a picture of the genetically modified fly with eyes on its legs: CONTENT WARNING: FLY BODY HORROR. I couldn't find the source for that specific photo, but here's a related paper: "eyes were induced on the wings, the legs, and on the antennae". )
Point is: epigenetics/evo-devo speeds evolution up. So, I was wondering, could I borrow this idea from nature to make genetic algorithms better at finding solutions? An "epi-genetic algorithm"? Looking up "epigenetic algorithm" on Google Scholar got me some results, but none were the specific idea I had. I wanted my algorithm to go low-level: I'd actually simulate a cell's DNA and protein concentrations! (or, a highly simplified, stylized version of it)
Well! Two weeks of coding later, and I got!...
Something... honestly not impressive to look at. Again, less show more tell today. Here's a GIF of a simulation of evolving a cells to make protein in a cyclical pattern (i.e. circadian rhythms):
(Explanation: there's two genes. Red activates Blue, Blue inhibits Red. This creates a cycle. "Evolution" fine-tines it to the exact cycle we want.)
I may mess around with this more, but honestly I'm losing interest. If I can make something cooler or more useful, I'll let you know. Otherwise, we'll pretend this side project didn't happen & I'll never talk about it ever again. "Doing research" is hard and sometimes straight-up not rewarding.
( If you want to learn more about epigenetics/evolutionary biology, honestly Acapella Science's 4-minute parody of Despacito is a shockingly thorough introduction. )
. . .
4) My previous month's "Signal Boosts" blog post:
🌟 AI Tutors, Note-taking tools, Murder mysteries & Data-murder mysteries. 🌟
Also:
I found a
MOTHER
FLIPPING
FIVE-LEAF CLOVER!!!!
(30-second swear-filled video of my discovery)
Good omen, good stuff.
Anyway, assuming I don't mess up, July 2024 should be when AI Safety: Part 2 launches! Stay tuned!
🥸,
~ Nicky Case
Comments
Young Jun Lee
I am just now reading part 1, and it is fantastic! Are any other technologies buried due to academic politics similar to ANNs?
Aeryn Light
Definitely! I still have my professor's textbook around somewhere too.
Nicky Case
Thank you Aeryn! (Sorry for late reply, I *do* want to engage with supporters like you more but I am bad at all platforms for correspondence.) That's so cool you studied Evo Comp! If I get back into this project, would you mind if I picked your brain more?
Aeryn Light
Awesome clover! And as someone who took Evolutionary Computation in college, I'm very interested in how we could make it more practical in AI. Neural nets are very difficult to analyze the "logic" for, whereas evolutionary computation can be much more straightforward to observe.
