Trang

Obsessed with Seeing Spaces • building a shadow twitter with ATProto • lunchtime poet

[Trang @trangquest]

Tension is who you think you should be and relaxation is who you are.

[Trang @trangquest]

before learning GNN, it helps to become fluent in the "language" of graphs

when an embedding places certain nodes close together or predicts a missing link, you can ask why and decide whether the result aligns with the structure you already understand.

[Trang @trangquest]

Hyperbolic geometry is useful for creating embeddings of complex network data. But at first, I want to understand the data myself. How can I expect from an AI to understand a complex network if I myself don't understand what is interesting about them?

[Trang @trangquest]

@__tinygrad__ @Etched Is this an example of a bad psyop?
@DefenderOfBasic

[Trang @trangquest]

anyway, here are some resources I discovered while researching backprop and autodiff

https://t.co/LQ9WLY2A0V
https://t.co/MUPnP1TsmN
https://t.co/tER1nCceHo
https://t.co/cn4PeawlgC

[Trang @trangquest]

it's seems to me that most people don't care about studying the fundamentals and math behind LLMs. they'd rather participate in scifi discussions

yk how most of it is just abstract self-referential fictional hocus pocus? all maps but no territory

[Trang @trangquest]

Automatic differentiation and high-level APIs provide an abstraction over gradient computation. That abstraction leaks because successful training often requires understanding the underlying optimization mechanics.

[Trang @trangquest]

The training interface exposed by PyTorch is a leaky abstraction over gradient-based optimization

[Trang @trangquest]

@BreatheLesss The amount of bookmarks are quite worrisome...

[Trang @trangquest]

Persuasion, then, is less about defeating someone's arguments and more about connecting ideas to an existing sense-making structure—or helping someone build a new one.

[Trang @trangquest]

Data-Frame theory explains why some people are naturally persuasive without resorting to rhetorical tricks. They empathize with others to uncover the frames and anchors through which they make sense of the world.

[Trang @trangquest]

I actually think that the most entertaining outcome will be the likeliest https://t.co/5OK5Cf3Kfb

[owl @owl_posting]

broke: all value will accrue to the frontier labs
joke: the frontier labs go broke because open source models work fine, and value accrues to companies serving specific use cases
woke: every single person on earth has their own startup ran by a Gwern-ian ‘guardian angel’. they speedrun the economic tech tree of humanity and eventually spin up their own guardian angels. the cycle repeats

[Trang @trangquest]

2nd try: 110 WPM

[Trang @trangquest]

let's see how fast i can type. i haven't done these in a while, but this one felt quite effortless. 104 WPM isn't too bad

[David Deutsch @DavidDeutschOxf]

@CapelLofft People say things. They just make them up and say them.

[Trang @trangquest]

https://t.co/HCyAcwmxTC

[Trang @trangquest]

TIL @karpathy once made Rubik's cube tutorials?!

in one of his vlogs he said: "I'm kind of trying to solve this artificial intelligence thing, and it's keeping me a little busy haha"

that had me chuckle

[Trang @trangquest]

I visualized the magnitude of outcomes produced by sustained growth
https://t.co/OnhfLx52KJ

[Paul Graham @paulg]

How to Earn a Billion Dollars: https://t.co/WeWBUkKym6

[Trang @trangquest]

twitter is like a billion temporary group chats running in parallel, ephemeral clusters of attention and controversial takes. it is like tuning into a memetic weather channel. if you come here expecting climate science, you are misunderstanding the platform

[vvv @g9nsai]

ma nouvelle création : l’ail pensif