Vihaan Sondhi
3 followers · 13 following · 964 views
on the atlas · 63
- Before the Startup5 savers
- Notes on Existential Risk from Artificial Superintelligence4 savers
- Hacking the Hackathon8 savers
- Illustrating Reinforcement Learning from Human Feedback (RLHF)13 savers
- Vanilla Policy Gradient — Spinning Up documentation2 savers
- Part 3: Intro to Policy Optimization — Spinning Up documentation6 savers
- Key Papers in Deep RL — Spinning Up documentation4 savers
- Aesthetics; or, Things That I Love • Logan Graves3 savers
- You don't have to be busy to be prolific | thesephist.com22 savers
- Local-first software: You own your data, in spite of the cloud8 savers
- Transfer Learning in Humans – niplav2 savers
- Fourier Feature Networks2 savers
- My Career as a Series of Emails18 savers
- Debugging Reinforcement Learning Systems11 savers
- Deep Q-Networks Explained — LessWrong4 savers
- Staring into the abyss as a core life skill88 savers
- Toy Models of Superposition30 savers
- I want to destroy the words "interesting" and "sparkly"4 savers
- An Expandable Explanation into Hackathon Organizing1 savers
- A blog post is a very long and complex search query to find fascinating people and make them route interesting stuff to your inbox68 savers
- We Can Only Kick Taiwan Down the Road For So Long – The Scholar's Stage1 savers
- Andy Matuschak45 savers
- Everything that turned out well in my life followed the same design process65 savers
- Welcome to Spinning Up in Deep RL! — Spinning Up documentation11 savers
- Induction heads - illustrated — LessWrong5 savers
- Mapping the Mind of a Large Language Model \ Anthropic6 savers
- A Mathematical Framework for Transformer Circuits40 savers
- Understanding Variational Autoencoders (VAEs) | by Joseph Rocca | Towards Data Science8 savers
- Perfectly Normal1 savers
- Rent GPU Servers for Deep Learning and AI5 savers
- Going Critical — Melting Asphalt5 savers
- Neural Networks, Manifolds, and Topology -- colah's blog17 savers
- Childhoods of exceptional people - by Henrik Karlsson70 savers
- Speed matters: Why working quickly is more important than it seems « the jsomers.net blog69 savers
- How to Do Great Work69 savers
- Thirty Observations at Thirty62 savers
- How can we develop transformative tools for thought?52 savers
- 95%-ile isn't that good44 savers
- Cities and Ambition42 savers
- Home Page34 savers
- Augmenting Long-term Memory31 savers
- Post 38: On Slack - Having room to be excited — Neel Nanda23 savers
- Using spaced repetition systems to see through a piece of mathematics21 savers
- Post 34: Learning how to learn — Neel Nanda20 savers
- Distill — Latest articles about machine learning16 savers
- The Evolution of Trust15 savers
- On Really Trying · Gwern.net14 savers
- Will scaling work? - by Dwarkesh Patel - Dwarkesh Podcast10 savers
- Mini Blog Post 11: Live a life you feel excited about — Neel Nanda10 savers
- A GPS for the mind | thesephist.com9 savers
- All Roads Lead to Robotics | Eric Jang9 savers
- Nintil - On Bloom's two sigma problem: A systematic review of the effectiveness of mastery learning, tutoring, and direct instruction7 savers
- Shard Theory: An Overview — LessWrong6 savers
- Beyond Smart6 savers
- New Book: AI is Good for You | Eric Jang5 savers
- The XY Problem5 savers
- The shard theory of human values - LessWrong5 savers
- Interpretability Dreams5 savers
- Eureka! On the clustering of geniuses - by Rohit4 savers
- What I learned from doing Quiz Bowl | Jacob’s Blog4 savers
- Taking into account preferences of past selves4 savers
- The Potential Merits of an Itemized OS4 savers
- Mini Blog Post 18: How to teach things well — Neel Nanda3 savers
highlights · 27
One action, zero observation, one timestep long, +1 reward every timestep: This isolates the value network. If my agent can't learn that the value of the only observation it ever sees it 1, there's a problem with the value loss calculation or the optimizer. One action, random +1/-1 observation, one timestep long, obs-dependent +1/-1 reward every time: If my agent can learn the value in (1.) but not this one - meaning it can learn a constant reward but not a predictable one! - it must be that backpropagation through my network is broken. One action, zero-then-one observation, two timesteps long…
Debugging Reinforcement Learning SystemsUse probe environments.
Debugging Reinforcement Learning SystemsRL has surprisingly few of these black boxes. You're required to know how your environment works, how your network works, how your optimizer works, how backprop works, how multiprocessing works, how stat collection and logging work. How GPUs work!
Debugging Reinforcement Learning SystemsThe moment someone says “[this person] is interesting” with no riders or qualifiers, it feels as if they’ve created two classes of people: “interesting” people and “non-interesting” people.
I want to destroy the words "interesting" and "sparkly"(e.g. the organization becoming infected by college apps optimization, or getting too big too fast, or not getting any money, or losing culture)
What is social infrastructure? • Logan GravesNo matter how abstract you get, information - be it chemical structure, body shape, or poem - proliferates and mutates to find an environmental optimum.
Life is Memetics - Unaligned worldHunter from New York City
What I learned from doing Quiz Bowl | Jacob’s Blogwhimsy - being in touch with my emotions and my whims, in touch with what I really care, and following those where they lead
Post 38: On Slack - Having room to be excited — Neel NandaLife is complex, and I really doubt that what you should care about can be boiled down to something so simple as quality-adjusted life-years. I doubt it can be boiled down at all. You should care about whatever you care about, and that probably won’t fit any neat moral templates an online forum hands you. It’ll probably be complex, confused, and logically inconsistent, and I don’t think that’s a bad thing
Saving the world sucks - Unaligned worldUsing Anki to thoroughly read a research paper in an unfamiliar field
Augmenting Long-term MemoryOver the long term the problem is simply too much power available to human beings: making it more widely available won't solve the problem, it will make it worse.
Notes on Existential Risk from Artificial Superintelligence(a) accelerate the widespread use and proliferation of such systems, by making them more attractive to customers and governments, and exciting to investors; but then (b) be easily circumvented by people whose idea of "safe" may be very, very different than yours or mine.
Notes on Existential Risk from Artificial SuperintelligenceBut most concrete alignment work is capabilities work. It's a false dichotomy, and another example of how a conceptual error can lead a field astray.
Notes on Existential Risk from Artificial Superintelligencepractical alignment work is extremely accelerationist
Notes on Existential Risk from Artificial SuperintelligenceThere is a sense in which human understanding is always dual use: genuine depth of understanding makes the universe more malleable to our will in a very general way.
Notes on Existential Risk from Artificial SuperintelligenceThis would represent a system for incremental thinking
Spaced repetition may be a helpful tool to incrementally develop inklingsA way that could be even more salient than text is taking a video of me talking about my current values, what I want for the future, and what I would like my future self to do. I plan to do this at least once a year with the first one happening in the next 2 weeks. I would then re-watch these videos yearly and whenever I had a big decision to make.
Taking into account preferences of past selvesnotation for working with complex networks of ideas
Notational intelligence | thesephist.comMusic notation has a more interesting story – you can write music in the standard notation in software, and most professional audio workstations or composition software will let you play it back and “execute” it, the way a programmer might execute a program. I think we can push most notations much farther along this path.
Notational intelligence | thesephist.comdynamic notation: notation defined not only by its symbols and the way they lay out in space, but also by the way we interact with them.
Notational intelligence | thesephist.comlanguage as a notation encodes so much information about the world, that a sufficiently advanced model working with the notation of language also inherently works with some half-useful model of our reality.
Notational intelligence | thesephist.comTo effectively compress images, a compression algorithm would be advantaged to “learn” facts about the world, like that colors are usually contiguous in images, and that the ground is often green and grassy while the sky is often white and blue. To effectively compress English text, the model might be advantaged to “learn” abstractions like common words and frequent grammatical constructs, so it can avoid inefficient, rote memorization of letters as much as possible.
Notational intelligence | thesephist.comIntelligence as data compression
Notational intelligence | thesephist.comIntelligence as generalization power
Notational intelligence | thesephist.comIntelligence as run-time adaptation
Notational intelligence | thesephist.comThis property of notation as embodied abstraction will appear many times throughout the rest of this post.
Notational intelligence | thesephist.comWhen you wrote them down, you could grasp your thoughts like bricks in your hands and push them into different arrangements. Writing let you look at your thoughts in a way you couldn’t if you were just talking, and having seen them, you could improve them, make them stronger and more elaborate.
Notational intelligence | thesephist.com