Andria Xu
3 followers · 5 following · 916 views
on the atlas · 45
- How to think in writing - by Henrik Karlsson2 savers
- [2406.08929] Step-by-Step Diffusion: An Elementary Tutorial1 savers
- Happiness, Part 1: What Really Makes Us Happy?1 savers
- Cognitive load is what matters25 savers
- TuringComputing.pdf1 savers
- [2004.05107] Levels of Analysis for Machine Learning1 savers
- The importance of stupidity in scientific research -- Schwartz 121 (11): 1771 -- Journal of Cell Science1 savers
- Personal Superintelligence2 savers
- Decades After a ‘Living Hell,’ Korean Victims Win a Step Toward Redress - The New York Times1 savers
- Opinion | I’m a Therapist. ChatGPT Is Eerily Effective. - The New York Times1 savers
- How to Make Your Own Luck (Update) - Freakonomics1 savers
- An Opinionated Guide to ML Research59 savers
- AI was supposed to democratize talent. Here's why it could spawn more elitism. - Haas News | Berkeley Haas1 savers
- Exploration for the Efficient Deployment of Reinforcement Learning Agents1 savers
- Model Citizens - Wharton Magazine1 savers
- Talk Like a Woman, Get Interrupted: The Politics of Everyday Speech1 savers
- Offline RL and Large Language Models - by Sergey Levine1 savers
- Why You Should Start a Blog Right Now39 savers
- How To Understand Things4 savers
- How to reach out to people when you're tired and busy and distracted1 savers
- How to increase your surface area for luck - by Cate Hall4 savers
- Sporks of AGI - by Sergey Levine - Learning and Control8 savers
- what happened to our college essay selves?1 savers
- AI makes the humanities more important, but also a lot weirder2 savers
- A graduate school survival guide: "So long, and thanks for the Ph.D!"3 savers
- Principles of Effective Research | Michael Nielsen23 savers
- How do you run an event where strangers feel safe sharing their most cancellable beliefs?2 savers
- Is It Fair for a Doctor’s Mom to Get Faster Emergency-Room Care? - The New York Times1 savers
- LLM Daydreaming · Gwern.net13 savers
- Reward Hacking in Reinforcement Learning | Lil'Log12 savers
- You and Your Research82 savers
- How to Pick a Career (That Actually Fits You) — Wait But Why61 savers
- Reflections on OpenAI34 savers
- Why We Think | Lil'Log26 savers
- I am rich and have no idea what to do with my life25 savers
- Life is a Picture, But You Live in a Pixel — Wait But Why19 savers
- Asymmetry of verification and verifier’s law — Jason Wei16 savers
- The Paris Review - The Crane Wife - The Paris Review14 savers
- Raise a Genius!12 savers
- Book I: Map and Territory | Rationality: From AI to Zombies6 savers
- Lies Told To Children - LessWrong6 savers
- An Alchemist’s Notes on Deep Learning — An Alchemist's Notes on Deep Learning5 savers
- A VLA with Open-World Generalization4 savers
- Opinion | In the Age of A.I., Major in Being Human - The New York Times4 savers
- MIT6.S192: Deep Creativity3 savers
highlights · 97
Productive stupidity means being ignorant by choice. Focusing on important questions puts us in the awkward position of being ignorant. One of the beautiful things about science is that it allows us to bumble along, getting it wrong time after time, and feel perfectly fine as long as we learn something each time.
The importance of stupidity in scientific research -- Schwartz 121 (11): 1771 -- Journal of Cell ScienceScience involves confronting our `absolute stupidity'. That kind of stupidity is an existential fact, inherent in our efforts to push our way into the unknown
The importance of stupidity in scientific research -- Schwartz 121 (11): 1771 -- Journal of Cell Scienceif we don't feel stupid it means we're not really trying
The importance of stupidity in scientific research -- Schwartz 121 (11): 1771 -- Journal of Cell ScienceGood thinking is about pushing past your current understanding and reaching the thought behind the thought.
How to think in writing - by Henrik KarlssonMathematics is, after all, a subset of writing
How to think in writing - by Henrik KarlssonBut it is, if you read it sideways, a profound exploration of the act of writing
How to think in writing - by Henrik Karlssonmen tend to use language to establish status and power, while women primarily use language to build relationships and cooperation.
Talk Like a Woman, Get Interrupted: The Politics of Everyday SpeechIntelligent people simply aren’t willing to accept answers that they don’t understand — no matter how many other people try to convince them of it, or how many other people believe it, if they aren’t able to convince them selves of it, they won’t accept it.
How To Understand ThingsAsk for things that feel unreasonable, to make sure your intuitions about what’s reasonable are accurate (of course, try not to be a jerk in the process). If you’re only asking for things you get, you’re not aiming high enough
How to reach out to people when you're tired and busy and distractedjust being likable does very little for you if it means you’re not actually getting the desired result. You want to be a bit polarizing, a bit risky, a bit dangerous. You want people to either be like hell no or hell yes.
How to reach out to people when you're tired and busy and distracted‘No’ is not always final.
How to reach out to people when you're tired and busy and distractedeal data is indispensable if we are to truly build robotic foundation models that can generalize in the real physical world as broadly as LLMs and VLMs do in the virtual world
Sporks of AGI - by Sergey Levine - Learning and ControlWhat we get is a spork: it can do the job of both a fork and a spoon in a few cases that match our assumptions, but usually it ends up just being a lousy spoon with holes or an ineffective blunt fork.
Sporks of AGI - by Sergey Levine - Learning and Controlwe are trying to find a best of both worlds solution: something that is as cheap as simulation or videos from the web, but as effective as real foundation models trained on large datasets.
Sporks of AGI - by Sergey Levine - Learning and Controlthe yellow circle gets smaller as we use stronger models, and any attempt to counteract this has to in the end make the models weaker.
Sporks of AGI - by Sergey Levine - Learning and ControlThese issues can seem innocuous in research projects and demos, because we can set up the real robot in a way that makes this discrepancy less important, selecting tasks, environments, and objects where the best and most robust strategies lie precisely within this intersection.
Sporks of AGI - by Sergey Levine - Learning and Controlas we use larger and more powerful models, we should expect to feel stronger headwinds from these issues: as more powerful models fit the patterns in the data more tightly, they’ll increasingly fit to the discrepancies just as much as they fit to the real transferable patterns we want to learn.
Sporks of AGI - by Sergey Levine - Learning and Controleach of them represents a compromise that ultimately undermines the real power of large learned models.
Sporks of AGI - by Sergey Levine - Learning and Controlasking people to collect data using hand-held devices that mimic robot grippers
Sporks of AGI - by Sergey Levine - Learning and ControlAny such choice presumes a particular way of solving the task (e.g., by picking up and moving items with the hand using a power grasp), and also requires bridging the considerable gap between physically feasible human movements and robot movements, both in terms of dynamics and appearance.
Sporks of AGI - by Sergey Levine - Learning and ControlOften the simulations that lead to the best results don’t so much present an accurate model of reality (which is very hard), but rather encode the types of variation that the robot should be robust to, such as training on random stepping stones or height fields, further underscoring how human insight informs not just what the task is, but indirectly how it should be solved.
Sporks of AGI - by Sergey Levine - Learning and Controlmanually define a mapping or correspondence between a cheap surrogate domain and the real world robotic system, and then leverage this correspondence to use this cheap data instead of expensive but representative in-domain data (i.e., data from the real robot in the real target domain)
Sporks of AGI - by Sergey Levine - Learning and Controlsending out a bunch of cold emails, scheduling random meetings, asking a lot of questions
How to increase your surface area for luck - by Cate HallEven when you have an established network, it’s a great way to densify the connections in your life.
How to increase your surface area for luck - by Cate HallOne way to increase serendipity is to throw parties or events, especially a regular series of get-togethers that’s low-pressure to attend.
How to increase your surface area for luck - by Cate HallThey are surrounded by people who they trust implicitly and who value loyalty and reciprocity.
How to increase your surface area for luck - by Cate Hallpeople really notice when you show up despite it not being a matter of duty or obviously in your self-interest
How to increase your surface area for luck - by Cate HallTo assume that you’re aiming for a future that will hold you to a much higher standard, even if you’re not currently held to that standard. People who behave like this — who act as if they are already an admired success in their field — have a funny way of ending up there.
How to increase your surface area for luck - by Cate HallIf you are withholding your energy for use only in those situations where you know an interaction will benefit you, you are not increasing your surface area for luck at all. You are already failing the assignment.
How to increase your surface area for luck - by Cate HallPeople love to talk about what they’re interested in, and by extension love to talk to people who are genuinely curious about the things they’re interested in.
How to increase your surface area for luck - by Cate HallTalking to people without an end in mind other than satisfying your own curiosity is the slow way that is the fast way.
How to increase your surface area for luck - by Cate HallYou can’t force plants to grow, you can only engineer the conditions in which they’re likely to flourish, and trust that the results will come.
How to increase your surface area for luck - by Cate Hallpeople who are unusually successful is that they just try a lot of stuff — socially, intellectually, professionally
How to increase your surface area for luck - by Cate Hallthere is an irresistible urge to figure out how to use something else, some sort of surrogate data that can be obtained on the cheap but still provide the kind of broad generalization we expect from foundation models
Sporks of AGI - by Sergey Levine - Learning and ControlWe’re mass-producing people who’ve been seduced by the glorification of entrepreneurs and billionaires and captivated by an illusion of what that life actually entails. Meanwhile, we’re creating a severe shortage of people willing to do patient, unglamorous work that holds society together.
what happened to our college essay selves?The timeline is maddeningly individual, yet we’ve created a culture that treats late bloomers as failures and career pivots as evidence of indecision rather than wisdom.
what happened to our college essay selves?You can never become someone you aspire to become, because the person you aspire to become never aspired to become anything so neatly defined.
what happened to our college essay selves?These people became who they are because they were obsessed with specific problems, not specific outcomes for themselves.
what happened to our college essay selves?We’ve all become remarkably similar in our ambitions.
what happened to our college essay selves?My greatest concern when it comes to LLMs in humanities education is that they will lead to a further polarization in educational outcomes.
AI makes the humanities more important, but also a lot weirderBy making effort an optional factor in higher education rather than the whole point of it, LLMs risk producing a generation of students who have simply never experienced the feeling of focused intellectual work.
AI makes the humanities more important, but also a lot weirderconnections — social capital — routinely smooth the way for those who have them
Is It Fair for a Doctor’s Mom to Get Faster Emergency-Room Care? - The New York TimesSometimes morality not only permits but requires us to give priority to those dearest to us.
Is It Fair for a Doctor’s Mom to Get Faster Emergency-Room Care? - The New York Timesour moral obligations are also shaped by our relationships.
Is It Fair for a Doctor’s Mom to Get Faster Emergency-Room Care? - The New York Timesseveral biases with LLM (they use GPT-3 in the experiments) contribute to such high variance: (1) Majority label bias exists if distribution of labels among the examples is unbalanced; (2) Recency bias refers to the tendency where the model may repeat the label at the end; (3) Common token bias indicates that LLM tends to produce common tokens more often than rare tokens. To conquer such bias, they proposed a method to calibrate the label probabilities output by the model to be uniform when the input string is N/A.
Prompt Engineering | Lil'Logchoice of prompt format, training examples, and the order of the examples can lead to dramatically different performance
Prompt Engineering | Lil'Logit comes at the cost of more token consumption and may hit the context length limit when input and output text are long.
Prompt Engineering | Lil'LogFew-shot learning presents a set of high-quality demonstrations, each consisting of both input and desired output, on the target task.
Prompt Engineering | Lil'LogZero-shot learning is to simply feed the task text to the model and ask for results.
Prompt Engineering | Lil'LogPrompt Engineering, also known as In-Context Prompting, refers to methods for how to communicate with LLM to steer its behavior for desired outcomes without updating the model weights.
Prompt Engineering | Lil'Log