João Araújo

6 followers · 1 following · 2451 views

on the atlas — 36

highlights — 260

This is one main source of the ire that is often felt toward the women who are labeled "stars." Because they were not selected by the women in the movement to represent the movement's views, they are resented when the press presumes that they speak for the movement. But as long as the movement does not select its own spokeswomen, such women will be placed in that role by the press and the public, regardless of their own desires.
The Tyranny of Stuctureless
The most insidious elites are usually run by people not known to the larger public at all. Intelligent elitists are usually smart enough not to allow themselves to become well known; when they become known, they are watched, and the mask over their power is no longer firmly lodged.
The Tyranny of Stuctureless
As long as the structure of the group is informal, the rules of how decisions are made are known only to a few and awareness of power is limited to those who know the rules.
The Tyranny of Stuctureless
the idea becomes a smokescreen for the strong or the lucky to establish unquestioned hegemony over others.
The Tyranny of Stuctureless
This means that to strive for a structureless group is as useful, and as deceptive, as to aim at an "objective" news story, "value-free" social science, or a "free" economy.
The Tyranny of Stuctureless
Contrary to what we would like to believe, there is no such thing as a structureless group. Any group of people of whatever nature that comes together for any length of time for any purpose will inevitably structure itself in some fashion.
The Tyranny of Stuctureless
All we need is to build distributed systems that work. That means decentralized bulk activity, hierarchical regulation. As a society, we are so much richer, so much luckier, than we have ever been. It's all so much easier, and harder, than they've been telling you. Let's build what we already know is right.
100 years of whatever this will be - apenwarr
We don't need deregulation. We need better designed regulation.
100 years of whatever this will be - apenwarr
in any system, if you don't have an explicit hierarchy, then you have an implicit one.
100 years of whatever this will be - apenwarr
The job of regulation is to stop distributed systems from going awry.
100 years of whatever this will be - apenwarr
Regulation is a centralized function.
100 years of whatever this will be - apenwarr
And yet: markets are distributed systems.
100 years of whatever this will be - apenwarr
The job of market regulation - fundamentally a restriction on your freedom - is to prevent all that bad stuff.
100 years of whatever this will be - apenwarr
Put another way, the generating function for novel work is trying really hard to deeply understand something until you pass through the edge of current human understanding, and then continuing imaginatively onwards.
Research as Understanding
Research, I realized, is what happens as a byproduct when you try to understand something and hit the bounds of what humanity currently knows.2 At that point, there's suddenly no one who can tell you the answer.
Research as Understanding
This was research. I wasn't straining to discover something new. I was accidentally doing it because I was curious, because my friend had asked a question I couldn't answer and it seemed nobody else had figured it out either.
Research as Understanding
Learning happens when you understand something that someone else already understands. Research happens when you understand something that nobody else understands yet.
Research as Understanding
Learn the Distribution instead of the Optimum
Just Ask for Generalization | Eric Jang
“Memorization is the first step towards generalization!”
Just Ask for Generalization | Eric Jang
the types are still there, and this program still implicitly relies on them being particular things.
No, dynamic type systems are not inherently more open
The reality is that static type systems allow specifying exactly how much a component needs to know about the structure of its inputs, and conversely, how much it doesn’t. Indeed, in practice static type systems excel at processing data with only a partially-known structure, as they can be used to ensure application logic doesn’t accidentally assume too much.
No, dynamic type systems are not inherently more open
Large amounts of diverse data are more important to generalization than clever model biases. If you believe (1), then how much your model generalizes is directly proportional to how fast you can push diverse data into a sufficiently high-capacity model.
Just Ask for Generalization | Eric Jang
study whether the ability to control the procedural generation can be beneficial for training Deep RL agents.
Teacher Algorithms for Deep RL Agents that Generalize in Procedurally Generated Environments – Developmental Systems, a Blog of the Flowers Lab
The fundamental observation of systems thinking is that the links between events are often more subtle than they appear.
Introduction to systems thinking. | Irrational Exuberance
Unconscious Incompetence. The part where you’re so bad at a thing, you aren’t even able to assess what skillful would look like. “How hard could it be, right?”
How to Actually Build a Better Boss - by Anne Helen Petersen - Culture Study
But you aren’t actually doing management well — you can’t be — without two other things: 1) You need to develop an awareness of, and do some reflecting on, your power in the organization and how you use it; and 2) You need to build systems of accountability for yourself and your peers in management.
How to Actually Build a Better Boss - by Anne Helen Petersen - Culture Study
The expander, which determines how these parenthesized forms correspond to real Racket expressions (and which are then evaluated to produce a result).
Beautiful Racket: Make a language in one hour: stacker
The reader converts the source code of our language from a string of characters into Racket-style parenthesized forms, also known as S-expressions.
Beautiful Racket: Make a language in one hour: stacker
The 8 inputs are arranged in an octagonal shape (dotted line) on a 32x32 grid with zeros at the boundaries. The two output cells are offset by 2 cells from the center of the octagon. We chose this configuration to ensure an almost equal distribution of the distance between each input and output.
Towards self-organized control
The values of the information channel of the output cells are used as the output of the system to be optimized to solve the task.
Towards self-organized control
The state of each cell is composed of 6 channels. The first is the information channel where meaningful input and output information transit. The third is identifying the inputs: it is equal to 1 in the input cells, 0 elsewhere. The fourth similarly identifies the outputs. The remaining three are hidden channels.
Towards self-organized control
diffusion models are learned with a fixed procedure and the latent variable has high dimensionality (same as the original data).
What are Diffusion Models?
Design the notation and behavior of our new language.  Write a Racket program that takes source code written in the new language and converts its notation and behavior to an equivalent Racket program.  Run this new Racket program normally.
Beautiful Racket: Make a language in one hour: stacker
Of course, a programming language isn’t the right solution for every problem. But when creating a language is easy and inexpensive—and in Racket, it often is—then it becomes a realistic option for many more problems.
Beautiful Racket: Make a language in one hour: stacker
Though a programming language is obviously a tool for writing a program, it’s also a tool for discovering new ways to program.
Beautiful Racket: Make a language in one hour: stacker
ecause an ordinary program is necessarily restricted by the rules of the programming language it’s written in. But when we make a programming language instead of a program, we free ourselves from many of these constraints.
Beautiful Racket: Make a language in one hour: stacker
In other words, we do not have to do a complete pass of the forward and reverse processes on each step of training.
From VAEs to Diffusion Models | Angus Turner
converges to a standard gaussian
From VAEs to Diffusion Models | Angus Turner
What sets them apart is a unique inference model, which contains no learnable parameters and is constructed so that the final latent distribution
From VAEs to Diffusion Models | Angus Turner
In fact, we can think of diffusion models as a specific realisation of a hierarchical VAE
From VAEs to Diffusion Models | Angus Turner
In the modeling view, we shoehorn ourselves into modeling all processes with probability distributions
Statistics as algorithmic summarization – arg min blog
in the algorithmic view, one can use statistics to understand the physical world no matter how the general population arose
Statistics as algorithmic summarization – arg min blog
Statistics gives us reasonable procedures to estimate properties of a general population by examining only a few individuals from the population
Statistics as algorithmic summarization – arg min blog
A corollary of the third rule is that one should never compete with one's colleagues
E.W. Dijkstra Archive: The Three Golden Rules for Successful Scientific Research (EWD 637)
"Never tackle a problem of which you can be pretty sure that (now or in the near future) it will be tackled by others who are, in relation to that problem, at least as competent and well-equipped as you."
E.W. Dijkstra Archive: The Three Golden Rules for Successful Scientific Research (EWD 637)
"We all like our work to be socially relevant and scientifically sound. If we can find a topic satisfying both desires, we are lucky; if the two targets are in conflict with each other, let the requirement of scientific soundness prevail."
E.W. Dijkstra Archive: The Three Golden Rules for Successful Scientific Research (EWD 637)
"Raise your quality standards as high as you can live with, avoid wasting your time on routine problems, and always try to work as closely as possible at the boundary of your abilities. Do this, because it is the only way of discovering how that boundary should be moved forward."
E.W. Dijkstra Archive: The Three Golden Rules for Successful Scientific Research (EWD 637)
It turns out, Fisher Information Matrix defines the local curvature in distribution space for which KL-divergence is the metric.
Natural Gradient Descent - Agustinus Kristiadi's Blog
goes to zero, KL-divergence is asymptotically symmetric. So, within a local neighbourhood, KL-divergence is approximately symmetric
Natural Gradient Descent - Agustinus Kristiadi's Blog
Thus, the optimization in gradient descent is dependent to the Euclidean geometry of the parameter space.
Natural Gradient Descent - Agustinus Kristiadi's Blog