flâneur

Matthew Siu

52 followers · 23 following · 2284 views

on the atlas — 55

highlights — 136

  • But if it did open for me, it wasn’t because I kept pounding on the gate with the same desperate query. And it wasn’t the favor-trading, or the Harvard connections, or my attempts at nepotism, or even (honestly) my agent (though she’s done so many great things for me then and since). It was that I set forth to be more awesome. I kept honing my craft, starting new projects better than the last, producing other works, articles, music, essays, research, the blog. I made my fire burn bright in the dark. People do see.
    The Key to the Kingdom, or How I Sold Too Like the Lightning – Ex Urbe
  • Without knowing it, I had stumbled on “Half and Half Again,” as it’s called by people I know in journalism, a training exercise in which you go through the agony of cutting an old work down to half length, then half of that, learning to spot the chaff and bloat in your own work, and how to make it tight and powerful.
    The Key to the Kingdom, or How I Sold Too Like the Lightning – Ex Urbe
  • They were not searching for sophistication, but something more prosaic: a chance to exit their societal roles – be it the colleague, the mother, the elder sister – and a space to simply, and unabashedly, be.
    The man making a business out of China’s burnout generation | China | The Guardian
  • I have this concept of “beacons and basins”: the beacon is the lighthouse where you put out a signal to the world saying what I’m interested in, what I like. The basin is a way to collect it. A beacon could be a personal website, your social media profile, or putting out messages in a channel; and the basin is the way to have an excuse to meet those people again on a recurring basis. So a book club, or a Discord channel, or a forum, or an email chain even.
    you can just do things (ft. anson & hudzah)
  • Tufekci introduces a concept called “network internalities,” which are built when people have to show up in person and print out a bunch of flyers or go through an Excel spreadsheet and make phone calls together. A lot of this operational, boring, logistical work strengthens the internal bonds within the network, making it more possible to build strong institutions.
    you can just do things (ft. anson & hudzah)
  • On top of that, with companies, they say you build culture by who you hire, fire, and promote. In a community, you can kind of hire people by tapping people to come in. You can kind of fire people by telling them what they can’t do. The only way that you can promote people is by giving certain people more visibility or airtime. By being selective with who demos, you can curate the rest of the community by selecting for who actually gets to present.
    you can just do things (ft. anson & hudzah)
  • As hosts for this organization, one of the important things that we do is finding people who stay late, pick up the trash, and want to contribute to making random PRs on the website. Then we give them more responsibility rather than opening up an application form, spending a lot of time reviewing applications, all to do the application cycle again. We don’t see applications as a good use of our time.
    you can just do things (ft. anson & hudzah)
  • It felt like their agency just got lent to me, and now I can give it to other people. It’s a bank that has an unlimited supply. You just need one really good friend or person to light that fire up.
    you can just do things (ft. anson & hudzah)
  • Edwin Schlossberg puts it well: "The skill of writing is to create a context in which other people can think."
    Cosmos: working notes, 1
  • Elegant, barely a word wasted. It requires immense empathy, marrying insight about human beings with insight about our place in the cosmos.
    Cosmos: working notes, 1
  • "syncresis"1, in which two (or more) apparently opposed points of view are seen to really be instances of a single more powerful point of view
    Cosmos: working notes, 1
  • This repeats multiple times, powerfully juxtaposing two vastly different points of view, points of view in what might seem superficially like strong conflict. The net effect is an expansion of consciousness, as you begin to merge these two types of consciousness.
    Cosmos: working notes, 1
  • “There was no defensiveness,” he recalls. “If you told him a new piece of information about a process or a person, he would drop his previous notion without hesitation. He had such a quiet ego.”
    Graham Duncan: Talent Whisperer - Colossus
  • “I’ve watched him be on the edge of seeding someone, where the normal impulse is to push them over the edge and give them a start,” Waitzkin says. “But Graham never pushes. He’s not susceptible to impatience the way most people are.” Duncan likes the author Diana Chapman’s analogy of a chick about to hatch. When you hear an emerging talent pecking from inside the egg, the temptation can be to help it break through. This is a mistake: the chick is still developing the strength required to survive outside the egg, and the surest sign that it is ready is when it destroys the shell on its own.
    Graham Duncan: Talent Whisperer - Colossus
  • “Because he’s got this willingness to accommodate unique talent, he’ll create a custom container to hold a person who wouldn’t get hired at a normal investment firm,”
    Graham Duncan: Talent Whisperer - Colossus
  • “He’d talk about how a person was amazing at Goldman, but did they have an independent identity? Were they self-aware enough? Were they unique enough? And if they were, had they fully embraced it yet?”
    Graham Duncan: Talent Whisperer - Colossus
  • The chef he recently hired for his restaurant is a Six and the GM is an Eight, two types that he believes work well together. “The Eights provide stability, the Sixes are loyal but also skeptical, in a great way, and alive in their decision-making
    Graham Duncan: Talent Whisperer - Colossus
  • They began by investing in new hedge funds, using a strategy based on several theses. One was that managers are more significant than firms. Another was that smaller funds perform better than large ones. Duncan and Shapiro also believed that managers generate the best returns in their first five years, when they are motivated by the higher stakes of those first defining deals and able to devote more attention to the details of a fund that is still in its infancy.
    Graham Duncan: Talent Whisperer - Colossus
  • “The more you pursue interests,” he told me on the good day we spent together, “the more you realize that the well is bottomless.”
    Flounder Mode - Colossus
  • Compared to this, Kelly’s version of doing his life’s work seems so joyful, so buoyant. So much less … angsty. There’s no suffering or ego. It’s not about finding a hole in the market or a path to global domination. The yard stick isn’t based on net worth or shareholder value or number of users or employees. It’s based on an internal satisfaction meter, but not in a self-indulgent way. He certainly seeks resonance and wants to make an impact, but more in the way of a teacher. He breathes life into products or ideas, not out of a desire to win, but out of a desire to advance our collective thin…
    Flounder Mode - Colossus
  • “What I’m talking about is taking your interests seriously enough to have the courage to stay moving. You can give stuff away. You can abandon things. You can tolerate failure because you know that tomorrow there is more.”
    Flounder Mode - Colossus
  • Content: Aside from the wonderful generality of being able to continuously shape new content from the same material, software has three important characteristics: the replication time and cost of a content-function is zero the development time and cost for a content-function is high the change time and cost for a content-function is low
    A Simple Vision of the Future -- Alan Kay
  • Designing how things change and move is enough for us to understand what they are and the relationships between them. You don’t need the heavy-handed metaphor, because the information is baked into the element’s behavior, not its aesthetics.
    Frank Chimero · What Screens Want
  • “Steve and I spent months and months working on a part of a product that, often, nobody would ever see, nor realize was there,” Ive grins. Apple is notorious for making the insides of its machines look as good as the outside. “It didn’t make any difference functionally. We did it because we cared, because when you realize how well you can make something, falling short, whether seen or not, feels like failure.”
    Apple Designer Jonathan Ive Talks About Steve Jobs and New Products
  • What does Microsoft Word look like with a Photoshop-like palette on the side?
    Here comes the Muybridge camera moment but for text. Photoshop too (Interconnected)
  • A common informal model of augmentation is what we may call the cognitive outsourcing model: we specify a problem, send it to our device, which solves the problem, perhaps in a way we-the-user don't understand, and sends back a solution:
    Thought as a Technology
  • the interface to Microsoft Word contains few deep principles about writing, and as a result it is possible to master Word's interface without becoming a passable writer. This isn't so much a criticism of Word, as it is a reflection of the fact that we have relatively few really strong and precise ideas about how to write well.
    Thought as a Technology
  • One consequence of reifying deep principles in an interface is that mastering the subject begins to coincide with mastering the interface
    Thought as a Technology
  • any deep principle is an opportunity to create powerful interface ideas.
    Thought as a Technology
  • A powerful way of thinking about one-dimensional motion is largely absent from our shared conversations. The reason is that traditional media are poorly adapted to working with such representations.
    Thought as a Technology
  • A simple strategy for efficiently identifying causally important features for a model's output is to compute attributions, which are local linear approximations of the effect of turning a feature off at a specific location on the model's next-token prediction.
    Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
  • In our experience, model capabilities are better preserved by paired and counterbalanced activation additions.
    Steering GPT-2-XL by adding an activation vector - AI Alignment Forum
  • The learned feature is universal – a similar feature is found by dictionary learning applied to a different model.
    Towards Monosemanticity: Decomposing Language Models With Dictionary Learning
  • From this layer on, however, we’ll often see color detectors which compare a color to the absence of color.
    An Overview of Early Vision in InceptionV1
  • Rather, the role of a feature visualization is like a variable name in understanding a program. It replaces an arbitrary number with a more meaningful symbol .
    An Overview of Early Vision in InceptionV1
  • In the future, you could imagine hybrid approaches, where a human investigator is saved time by having many features sorted into a (continually growing) set of known features, especially if the universality hypothesis holds.
    An Overview of Early Vision in InceptionV1
  • Dmitri Mendeleev is often accounted to have discovered the Periodic Table by playing “chemical solitaire,” writing the details of each element on a card and patiently fiddling with different ways of classifying and organizing them. Some modern historians are skeptical about the cards, but Mendeleev’s story is a compelling demonstration of that there can be a lot of value in simply organizing phenomena, even when you don’t have a theory or firm justification for that organization yet.
    An Overview of Early Vision in InceptionV1
  • as an annotated collection of what we call “neuron groups.”
    An Overview of Early Vision in InceptionV1
  • one could imagine a kind of “periodic table of visual features”
    Zoom In: An Introduction to Circuits
  • cellular biology of deep learning.
    Zoom In: An Introduction to Circuits
  • If it turns out that the universality hypothesis is broadly true in neural networks, it will be tempting to speculate: might biological neural networks also learn similar features?
    Zoom In: An Introduction to Circuits
  • Why would it do such a thing? We believe superposition allows the model to use fewer neurons, conserving them for more important tasks. As long as cars and dogs don’t co-occur, the model can accurately retrieve the dog feature in a later layer, allowing it to store the feature without dedicating a neuron.
    Zoom In: An Introduction to Circuits
  • But we can bring all our other approaches to analyzing a neuron to bear again. For example, we can use a 3D model to generate synthetic dog head images from different angles.
    Zoom In: An Introduction to Circuits
  • This seems to be a great strength of neural networks: they learn better ways to represent data, automatically.
    Deep Learning, NLP, and Representations - colah's blog
  • the word embedding will learn to encode gender in a consistent way. In fact, there’s probably a gender dimension. Same thing for singular vs plural.
    Deep Learning, NLP, and Representations - colah's blog
  • if W � maps synonyms (like “few” and “couple”) close together, from R � ’s perspective little changes
    Deep Learning, NLP, and Representations - colah's blog
  • the entire point of the task is to learn W � . We could have done several other tasks – another common one is predicting the next word in the sentence. But we don’t really care.
    Deep Learning, NLP, and Representations - colah's blog
  • A neural network with a hidden layer has universality: given enough hidden units, it can approximate any function.
    Deep Learning, NLP, and Representations - colah's blog
  • In other words, I need to do something with it. To create something. Write a blog post about it, use it in a new project, test it on the field, teach it at a meetup. That’s why I speak at many conferences: it’s a learning tool.
    How to Learn Better in the Digital Age
  • But realistically, when we’re working with very complex interactions between millions of neurons we’ll have to automate the process, some larger scale version of “ask GPT-4 to tell us what GPT-2 is doing”
    God Help Us, Let's Try To Understand The Paper On AI Monosemanticity