reading inbox

In no particular order. Items may move to previously read if I read them or former reading inbox if I decide I'm not currently interested.

AI / RL

• Stochastic Latent Actor-Critic: Deep Reinforcement Learning with a Latent Variable Model: https://arxiv.org/abs/1907.00953
• "Lessons from AlphaZero for Optimal, Model Predictive, and Adaptive Control": https://t.co/SWWYB2FlBO . It provides a summary of the new book (2022): https://t.co/yrcAKSNiUz
• Dopamine and TD learning: https://deepmind.com/blog/article/Dopamine-and-temporal-difference-learning-A-fruitful-relationship-between-neuroscience-and-AI
• Abstraction in RL: https://arxiv.org/abs/2005.00527 Is Long Horizon Reinforcement Learning More Difficult Than Short Horizon Reinforcement Learning?
• Learning more skills through optimistic exploration: https://arxiv.org/abs/2107.14226
• Rohin Shah's sequence on value learning: https://www.alignmentforum.org/s/4dHMdK5TLN6xcqtyc (value alignment)
• Maiutic promoting, for consistent logical explanations from LLMs: https://t.co/OBBvL0spIA
• Counterfactual Reasoning and Learning Systems (Bottou and Peters, 2019) - rec'd by Jacob Steinhardt as a good intro to counterfactual reasoning from an ML background https://leon.bottou.org/publications/pdf/tr-2012-09-12.pdf
• Improving Policy Learning via Language Dynamics Distillation: "A cool approach for speeding up RL (basically via helping w/exploration): https://t.co/fUh7bcALel (by @hllo_wrld @jayelmnop  @LukeZettlemoyer @egrefen @_rockt ) Took me a bit to get my head around it, but the idea is very simple and elegant …."
• "I absolutely LOVED this interview [with Andrew Lampinen]. I blame myself for having not read Andrew’s work yet, as I think he’s one of the clearest thinkers on the relation between language understanding, LLMs, and intelligence in the field." (https://twitter.com/egrefen/status/1599408622412263424?t=MjQNxXh0OalVOWL3lMwBHg&s=03)
• Schmidhuber on benefits of model-based RL over value functions: https://arxiv.org/abs/2211.02222
• tricks and analysis for transformer inference: https://kipp.ly/transformer-inference-arithmetic/
• big review paper with Bengio and others on "conscious" AI: https://arxiv.org/abs/2308.08708 seems very relevant to thinking about attention
• reinforced self-training paper from Deepmind: https://arxiv.org/abs/2308.08998

Alignment

• Sequences:
  • Iterated Amplification: https://www.alignmentforum.org/s/EmDuGeRw749sD3GKd
• Ontological crises
• https://www.lesswrong.com/s/4dHMdK5TLN6xcqtyc/p/NxF5G6CJiof6cemTw and https://www.lesswrong.com/posts/vphFJzK3mWA4PJKAg/coherent-behaviour-in-the-real-world-is-an-incoherent
• shard theory:
  • https://www.lesswrong.com/posts/xqkGmfikqapbJ2YMj/shard-theory-an-overview
  • https://www.alignmentforum.org/s/nyEFg3AuJpdAozmoX
  • https://www.beren.io/2022-12-05-ML-Interpretation-Shard-Theory/
• Richard Ngo's recommendations on scalable oversight:
  • AI safety via debate: https://openai.com/research/debate
  • Supervising strong learners by amplifying weak experts: https://arxiv.org/abs/1810.08575
  • Scalable agent alignment via reward modeling: a research direction https://arxiv.org/abs/1811.07871
• The computational anatomy of human values: https://www.lesswrong.com/posts/pZHpq6dBQzCZjjMgM/the-computational-anatomy-of-human-values. Andres says "This is really good! But also, ultimately, misguided. Yes, "human values" are as complex and diverse as this explains. But "consciousness values" may not be. This isn't a contradiction - just talking about different levels of abstraction."

ML

• Review on attention: Have you heard the word "attention" thrown around in both neuroscience & machine learning? Have you wondered if/how its different uses relate to each other? My new review aims to summarize how this giant topic is studied & modeled across different domains! https://t.co/240M0KKDXq https://t.co/OZDDght9cN
  (https://twitter.com/neurograce/status/1250745465261436931?s=09)
• Efficient and modular automatic differentiation: https://t.co/Rsh09miGTV
  (https://twitter.com/s_scardapane/status/1412062995618881540?s=03)
• "Since lots of people are interested in ALiBi now, I'm sharing my video lecture here, which contains a lot of insights into how transformers work and why we wanted to make them work without position embeddings. "https://t.co/D0bsZvqUiJ (https://twitter.com/OfirPress/status/1654575607663980544?t=4s8ySesKhmaglabveLAWxQ&s=03)
• Quentin Pope on grokking: https://www.lesswrong.com/s/5omSW4wNKbEvYsyje/p/GpSzShaaf8po4rcmA
• Grokking explainer: https://pair.withgoogle.com/explorables/grokking/
• bayesian flow networks (Alex Graves): https://arxiv.org/abs/2308.07037
• review on learning-as-compression, recommended by eric jang: https://twitter.com/ericjang11/status/1695437115687862358?t=Qa9yAAqlvEI5WklhPAz2YA&s=19
• Principles of deep learning theory: https://arxiv.org/abs/2106.10165

Formal math

• Christian Szegedy's research vision

• successful applications of ML to symbol manipulation:

• Theorem proving in Lean:

  • For type theory and proof theory, Chris recommends: https://leanprover.github.io/theorem_proving_in_lean/
  • and NameRedacted also likes this:
    https://github.com/martinescardo/HoTT-UF-Agda-Lecture-Notes but it skips more stuff.

  Bayes

• Essential read on modern prior specification: https://arxiv.org/abs/1403.4630?s=03

Hard science

• 2021 in computational neuroscience: https://twitter.com/patrickmineault/status/1477088118587248646?t=Z061vh9bwu7fG55HpIbZrA&s=03
• Why nature chose phosphates: https://sci-hubtw.hkvisa.net/10.1126/science.2434996
• Coarse grain molecular dynamics: https://arxiv.org/abs/2007.11412?s=09
• LLMs in molecular biology: https://towardsdatascience.com/large-language-models-in-molecular-biology-9eb6b65d8a30
• Physics books NameRedacted recommends:
  • Mathematical methods of classical mechanics - Vladimir Arnold
  • His prof's book on statistical mechanics, on my surface
• Jaynes on physics:
  • https://twitter.com/EmtiyazKhan/status/1248915516707573760
  • https://bayes.wustl.edu/etj/articles/theory.1.pdf
  • https://www.informationphilosopher.com/solutions/scientists/jaynes/Jaynes_Gibbs_Boltzmann.pdf
• "The human eye has a peak sensitivity to light with a 550 nm wavelength.  What is the surface temperature of the sun?" (https://twitter.com/quantian1/status/1385622095313387521?t=PJK78HpSQXpUN_A0NUZy_A&s=03)
• Faster DFT with GNNs: https://www.entos.info/
• Molecular paths: A nice example of how a carefully designed representation can make a task that deep learning struggles with trivial.
  (https://twitter.com/pfau/status/1339276310141657088?s=03)
• Computer scientist's guide to cell biology (William Cohen)
• Really enjoyed reading this review on the intersection of machine learning and the physical sciences https://t.co/oomJkuhQWD talking of statistical physics inspired ML theory, a wide array of applications, uses of ML in them, and how new ML could emerge from these interactions.
  (https://twitter.com/_onionesque/status/1145169064236593152?s=03)
• Computational neuroscience: depression and anxiety. https://www.gwern.net/docs/rl/2019-bishop.pdf
• 30 days of great biology papers
• Review on quantum chemistry simulations with neural nets: https://arxiv.org/abs/2208.12590

Mental Health

• Intro thread to Iain McGilchrist's work, based on his mini-book Ways of Attending: How our divided brain constructs the world. "Attention is not just receptive, but actively creative of the world we inhabit. How we attend makes all the difference to the world we experience." https://t.co/eys00Sn2vl
  (https://twitter.com/Malcolm_Ocean/status/1258906911874977793?s=09)
• How To Actually Boost Low Self Esteem & Stop Procrastinating (neuralshifter.com)
• Really high happiness is possible. Logarithmic Scales of Pleasure and Pain (@Effective Altruism NYC) https://t.co/GIUyToKF2k
  (https://twitter.com/algekalipso/status/1196695988636803072?s=03)
• Principles of Vasocomputation: A Unification of Buddhist Phenomenology, Active Inference, and Physical Reflex: https://opentheory.net/2023/07/principles-of-vasocomputation-a-unification-of-buddhist-phenomenology-active-inference-and-physical-reflex-part-i/

Meditation and spirituality

• Culadasa's notes on insight: http://dharmatreasure.org/wp-content/uploads/Meditation-and-Insight-I.pdf (replace I with II, III, for more)

see also: talks to watch#Talks