Since I no longer use Twitter 𝕏, I want a new way to share: 1) Stuff I liked, and 2) Why I liked them. Also, I gotta blog more.

So, introducing my first monthly Signal Boosts post! Just a list of links / papers / movies / books / videos / games / music / podcasts / whatever, that I saw last month & added value to my life. (It's like a "links post"^[1], but fewer items & more in-depth.)

So without further ramble:

🌳 Are you a teenage programmer? Hack in the Canadian forest!

TIME SENSITIVE OFFER: My friends at Hack Club are running a hackathon for high-schoolers, to make tech to help the environment. This week-long event takes place on a cross-Canada train, from Vancouver to Montréal, through some beautiful boreal forest!

Ticket price?... it's FREE. (with travel grants)

The catch is you have to be in high school (or earlier), and also there's limited seats. So, if you are a ≤ 18yo programmer or know someone who is, ask them to consider HACKING THE FOREST:

🌳 THE BOREAL EXPRESS 🌳
⏳ (apply soon, applications close June 10th!) ⏳

🎬 HUNDREDS OF BEAVERS

There's no describing this. You just have to see the trailer: (2 min)

• First, it's just so inventive. I've rarely seen this kind of blend between live action & animation^[2]. Not "just" Roger Rabbit style where humans move alongside toons, but humans move as if they're toons.
• It's shamelessly a homage to the Golden Age of slapstick comedy (Buster Keaton, Roadrunner, etc), yet updates the old tropes: the protagonist has a solid character growth arc, the "love interest" isn't a passive damsel nor (too much of) a Manic Pixie, and the world's rule-based ecosystem is as elegantly emergent as videogames like Rain World or Portal.
• Finally, it's insanely impressive how many jokes — visual gags, slapstick comedy, Rube Goldberg-like contraptions — could be squeezed out of such a simple setting!

If you're an artist with a low budget, this movie is inspiring, to show how much one can do with so little.

(4½ stars out of 5)

Watch on: Apple TV, Amazon Prime

Recommended for: lighthearted comedy movie night.

▶️ 3Blue1Brown's series on Transformers (e.g. ChatGPT)

Once again, Grant Sanderson (3Blue1Brown) delivers. This YouTube channel teaches tough ideas from math/computer science, with beautiful visualizations & intuitive explanations. This past month, Grant did it again, explaining the Transformer, the AI architecture behind ChatGPT, DALL-E, AlphaFold, and more!

It's a 3-part series, with 1 more part to come all out now:

• But what is a GPT? Visual intro to transformers (27 min)
• Attention in transformers, visually explained (26 min)
• How might LLMs "store facts"? (23 min)

(Actually, this series is a sequel to his previous series, explaining the classic, "vanilla" Artificial Neural Networks. I found those super-valuable too; it even helped me code my own ANN from scratch!)

Recommended for: Folks with no prior knowledge of deep learning, but okay with calculus & linear algebra. (And if you're not comfy with those, check out Grant's series on those topics!)

📄 A Unifying Probabilistic View of Associative Learning

(warning: technical)

Gershman 2015 is one of those rare papers that: 1) takes a lot of different ideas from different fields, then 2) somehow unifies them all.

Context behind the paper:

• Psychologists wanted a model of how "conditioning" works. (e.g. Ring a bell every time you give a dog food, and the dog will start salivating just at the sound of a bell.)
• One early model was the Hebbian model: neurons that fire together, wire together.
• However, this model didn't explain "blocking": if you train a rat to associate a Light with Food, then present both Bell+Light with Food, the rat will not associate Bell with Food... the Light→Food association seems to "block" the Bell→Food association!
• So, the Rescorla-Wagner (RW) model was created, and it's mathematical: Change in strength of association = Some learning rate x Prediction error, where Prediction error = Actual result - Predicted result. This explains "blocking": once a rat's learnt Light→Food, showing Light+Bell→Food creates no prediction error, hence no extra learning.
• However, the RW model doesn't predict:
  • Latent inhibition: First showing lots of Bell→NO-Food, slows down the later learning of Bell→Food.
  • Recovery from extinction: You can untrain Bell→Food by later showing lots of Bell→NO-Food. If you try to immediately re-train it afterwards, it's hard. But the longer you wait, the easier it is to retrain.
  • Higher-order conditioning: Train Bell→Food, then train Light→Bell, and the rat will learn Light→Food!

This paper solves all those problems, by swiping ideas from Engineering & Artificial Intelligence:

• Latent inhibition & Recovery from extinction can both be elegantly explained by a very popular idea from engineering, Kalman Filters. KFs explain animal perception well, and there's been a few proposals for how neurons could approximate KFs.^[3]
  • In sum: KFs let a machine/animal 1) be uncertain, 2) know it's uncertain, and 3) update its uncertainty using observations. This is a huge step up from old psych models, and even current mainstream AI models, which don't natively have "uncertainty" built into them.
• Higher-order conditioning can be explained with Temporal Difference Learning, a popular technique from AI. (The most famous example being Q-Learning, the algorithm that let Google's AI beat Atari games.) TD Learning's also probably what our brains do, using dopamine as the "reward prediction error" signal.^[4]
  • In sum: TD Learning is very similar to the RW model, except the Prediction Error isn't just for predicting current reward, it's for predicting all future reward (discounted over time).

This paper doesn't explain everything, of course. (For example, I'm not sure how this model can explain intermittent reward — the finding that you can make animals more obsessed, by rewarding them less often & unpredictably, the psychology of slot machines.)

But dang, if this paper isn't promising!

Link to paper, again

Recommended for: Cognitive science/Psychology/AI nerds, with undergraduate math capabilities. You can skim the more math-heavy parts and just look at the pretty pictures, that's what I did.

P.S: I learnt Kalman Filters through this video series and this picture-filled blog post.

📄 AI Alignment: A Comprehensive Survey

I spent a year writing a 20,000-word article on AI Safety, and right after launching I find this paper, which I fear made my work redundant.

(To be fair to myself, I think my thing's more layperson-accessible. Mine also has comics with a Robot Catboy Maid.)

In sum, the authors break up the AI Safety problem into 4 parts, with the acronym, RICE:

(from paper, Figure 3)

I can't really summarize the paper beyond that, since it is, well, very comprehensive. It's 58 large pages, not counting references. Basically a mini-book.

Link to paper: Ji, Qiu, Chen, Zhang et al 2024

Recommended for: Folks interested in AI Safety/Alignment. Paper assumes some technical background.

🎙️ The Ezra Klein Show

A podcast where this guy (Ezra Klein) interviews folks, including some big names like Barack Obama and Sam Altman. Remember, "retweets are not endorsements".

Ezra has a rare combination of personality traits:

• Humble enough to be able to openly acknowledge his uncertainty & publicly change his mind.
• Anti-humble enough to (tactfully) call bullshit on his guests, and ask tough pointed questions.

I don't know how he strikes the balance, but I'm envious. Anyway, despite name-dropping Obama & Altman earlier, I don't actually recommend starting with those specific interviews. Here's what I do recommend starting with:

• Interview with Jennifer Pahlka, founder of Code for America, and part of the team who saved Healthcare.gov. A look into governmental cover-your-ass-ocracy, and a well-needed antidote/reminder for "policy wonk" types like me & Ezra: principles & philosophy are nice, and none of that matters if you can't get shit done.
• Wonk-types think a lot about alternate technologies & policies (and we should consider those), but more rarely do we think about alternate ways to love & care for each other. So check out these interviews: Kristen Ghodsee on alternatives to the nuclear family, Dan Savage on alternatives to straight monogamy.
• Ezra, being a secular[?] liberal Jew^[5], has grappled with a lot of questions since October 7th. To understand the "inside view" of (moderate-ish) Palestinians & Israelis, here's: an interview with Salam Fayyad, former prime minister & finance minister of the Palestinian Authority; an interview with Ari Shavit, Israeli journalist & long-time critic of Benjamin Netanyahu. Again, "retweets are not endorsements" — but you have to understand opposing views "from the inside" if you hope to change hearts/minds.

Also, Ezra just has a soothing voice. Nothing like "neurotic Jewish nerd" rizz

Find this show on where-ever you get your podcasts. (I use PodBean)

Recommended for: Policy-wonk nerds. People tired of podcast-pundits pandering to you & not actually challenging your mind. People who need to fill up an hour on their daily commute.

🎵 THE CONE ZONE!!!!!

This album just bops

Recommended for: fans of trans furry hyperpop