https://adamjermyn.com/ Recent content in About on Adam Jermyn Hugo -- gohugo.io en Sun, 04 Jan 2026 12:00:00 -0500 https://adamjermyn.com/workflow/notes/ Mon, 01 Jan 0001 00:00:00 +0000 https://adamjermyn.com/workflow/notes/ Notes I take notes to support reference, recollection, and accumulation. This approach was heavily influenced by Andy Matuschak’s concept of Evergreen Notes. Reference The best reference for learning something is usually not the most useful for applying it. I learned about random walks from the Feynman Lectures, but when I need to calculate diffusion coefficients I reference a note that just gives the answer and a few supporting details. Reference notes should be optimized for frequently, repeated use. https://adamjermyn.com/workflow/prioritization/ Mon, 01 Jan 0001 00:00:00 +0000 https://adamjermyn.com/workflow/prioritization/ Prioritization Picking the right ideas to work on is important. When I was an academic, I wrote ideas in Markdown notes with one top-level header per idea. To decide what to work on, I used a python script which asks me to compare ideas to one another. The comparison questions are: If both ideas work, what is their relative impact? What are the relative odds of the ideas working? How long will each idea take to implement? https://adamjermyn.com/workflow/spacedrepetition/ Mon, 01 Jan 0001 00:00:00 +0000 https://adamjermyn.com/workflow/spacedrepetition/ Spaced Repetition Spaced repetition is a technique for remembering information over long time periods. It takes advantage of the empirical observation that the time it takes to forget a fact is proportional to the time that fact is known. This suggests that the best strategy to remember something indefinitely is to review facts with ever-larger (geometrically increasing) gaps between reviews. Following that strategy, a finite number of reviews is needed to remember something for the rest of your life, and that number is small enough that the study time needed to remember a fact indefinitely is about 5 minutes, so long as the studying happens in this spaced-out way. https://adamjermyn.com/posts/arline_feynman/ Sun, 04 Jan 2026 12:00:00 -0500 https://adamjermyn.com/posts/arline_feynman/ How many people have died not because treatments didn’t exist, but because the information didn’t reach them in time? Every new cure has a window where it exists but isn’t widely known, isn’t easily accessible, requires connections or luck to obtain. Someone is always on the wrong side of that window. One of the cruelest examples: Arline Feynman, wife of physicist Richard Feynman, who died of tuberculosis in June 1945. https://adamjermyn.com/posts/lectures_on_fractures/ Wed, 31 Dec 2025 00:00:00 +0000 https://adamjermyn.com/posts/lectures_on_fractures/ I recently collaborated with Claude Opus 4.5 to write a short set of lecture notes on fracture mechanics: Lectures on Fractures (PDF) How it worked I wanted something in Feynman’s style. Conversational, building intuition before formalism, full of physical insight. So I had Claude start by reading some of Feynman’s writing and assembling a style guide. That gave us a shared reference for tone. The technical focus was Griffith’s work on fractures. https://adamjermyn.com/posts/date_me_doc/ Sat, 15 Nov 2025 10:00:00 -0500 https://adamjermyn.com/posts/date_me_doc/ This is a link to my date-me doc, which I published in November 2025. I’m no longer single but wanted to leave a pointer to it up because it was an important document for me both to write and in becoming not single. https://adamjermyn.com/posts/claude_summary/ Sun, 09 Jun 2024 18:24:21 -0400 https://adamjermyn.com/posts/claude_summary/ Views my own. There might well be more clever ways to do all of this that I don’t know of. Suppose you have some long, ordered text data, like news stories going back a century or journal entries or server logs. You’d like to analyze them with a Large Language Model (LLM), but you can’t just concatenate them together (the context window isn’t that long!). A fun solution to this problem is a summary tree. https://adamjermyn.com/posts/declaration/ Thu, 07 Mar 2024 11:26:29 -0500 https://adamjermyn.com/posts/declaration/ On a recent trip to DC I picked up a pocket copy of the Declaration of Independence from the Supreme Court. At the back, it has biographies of the signatories. What’s remarkable is just how much agency people took, how much they shifted their lives around, and how “rag tag” it was. At one point one of them was personally going out soliciting loans to fund the war effort. Another was a doctor, got infuriated at the way the British treated the colonies, and left medicine to be a politician. https://adamjermyn.com/posts/pathoflaw/ Sun, 11 Feb 2024 16:13:41 -0500 https://adamjermyn.com/posts/pathoflaw/ I just skimmed through Oliver Wendell Holmes’s The Path of Law. A few things that stood out: He frames the law as a field of study with the aim of predicting the decisions of course. I think he’s trying to disentangle the law from a statement of morals. It’s not about what is right, it’s a practical question of what rules you will actually be held to account on and how that account will appear. https://adamjermyn.com/posts/time1/ Sat, 27 Jan 2024 16:49:34 -0500 https://adamjermyn.com/posts/time1/ Ideas Looking back, it feels like there was often a ‘right time’ for me to encounter particular ideas: I encountered a lot of ideas from religion very young (as many people do). I was also very interested in science, and so got hung up on factual disagreements rather than seeing e.g. the point of having a dedicated day of rest. I came across the Feynman Lectures in high school, when I knew just enough math that it was possible to work through them, but not enough math or physics that it felt like review. https://adamjermyn.com/posts/beauty1/ Sat, 27 Jan 2024 15:57:53 -0500 https://adamjermyn.com/posts/beauty1/ This is a collection of some things I’ve found beautiful recently. Recursive Game of Life (see the run-through here). Each cell is being simulated by another instance of the Game of Life, and if you zoom out enough you see that the structures you started off looking at are also simulations of cells. Paintings by Owen Schuh. See especially the work page, and this tweet. There’s a quality to many of these that’s halfway between organic and designed that really draws me in. https://adamjermyn.com/posts/latent_tb/ Sat, 25 Mar 2023 19:51:30 -0400 https://adamjermyn.com/posts/latent_tb/ (None of this is medical advice.) In 2020 the FDA identified contamination in two drugs used to treat Tuberculosis, Rifampin and Rifapentine. The contaminants were 1-methyl-4-nitrosopiperazine (MNP) in rifampin or 1-cyclopentyl-4-nitrosopiperazine (CPNP) in rifapentine, and levels were ~10x the FDA’s acceptable threshold (1-4ppm for MNP, 8-14ppm for CPNP). At the same time, these are two of the standard treatments for Tuberculosis and the alternatives come with downsides of their own, so the FDA permitted these drugs to continue being prescribed and sold in the US. https://adamjermyn.com/posts/scurves/ Tue, 29 Nov 2022 20:38:53 -0400 https://adamjermyn.com/posts/scurves/ This post is available on the AI Alignment Forum. https://adamjermyn.com/posts/mono_1/ Mon, 21 Nov 2022 14:53:00 -0400 https://adamjermyn.com/posts/mono_1/ This post is available on the AI Alignment Forum. https://adamjermyn.com/posts/tegum_products/ Fri, 04 Nov 2022 14:53:00 -0400 https://adamjermyn.com/posts/tegum_products/ This post is available on the AI Alignment Forum. https://adamjermyn.com/posts/acausal/ Wed, 02 Nov 2022 10:44:53 -0400 https://adamjermyn.com/posts/acausal/ (Cross-posted from the AI Alignment Forum) I used to think that acausal coordination was a weird thing that AI’s might do in the future, but that they certainly wouldn’t learn from looking at human behavior. I don’t believe that anymore, and think there are lots of examples of acausal coordination in everyday life. Examples people do Voting The political science argument against voting goes: The probability that my vote tilts the election is tiny, so the expected value to me is tiny, so it’s not worth my time to vote. https://adamjermyn.com/posts/smoke_fire/ Tue, 04 Oct 2022 14:27:25 -0400 https://adamjermyn.com/posts/smoke_fire/ This post is available on the AI Alignment Forum. https://adamjermyn.com/posts/left_turn/ Thu, 29 Sep 2022 14:27:25 -0400 https://adamjermyn.com/posts/left_turn/ Thanks to Evan Hubinger for comments on these ideas. Posted on the AI Alignment Forum. Introduction A “sharp left turn” is a point where capabilities generalize beyond alignment. In a sharp left turn, an AI becomes much more capable than aligned, and so starts to exploit flaws in its alignment. This can look like Goodharting, where strong optimization pressure causes outer alignment failure because the base goal isn’t identical to what we want. https://adamjermyn.com/posts/transformer_notes/ Mon, 26 Sep 2022 08:21:00 -0400 https://adamjermyn.com/posts/transformer_notes/ (Cross-posted from the AI Alignment Forum)* *These are just some notes I wrote while reading about transformers, which I thought might be a useful reference to others. Corrections welcome. Overview of Transformers Many transformer models have the following architecture: Data flows as follows: We take tokens as inputs and pass them through an embedding layer. The embedding layer outputs its result into the residual stream (x0). This has dimension (C,E), where C is the number of tokens in the context window and E is the embedding dimension. https://adamjermyn.com/posts/ideas28/ Fri, 19 Aug 2022 11:18:48 -0400 https://adamjermyn.com/posts/ideas28/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the 28th post in a series describing some of the ideas I’ve accumulated. Why doesn’t all the Uranium sink to the bottom of Earth? What’s the idea? Uranium is denser than iron. Why hasn’t it all fallen to the center of the Earth? https://adamjermyn.com/posts/ideas27/ Thu, 18 Aug 2022 13:40:28 -0400 https://adamjermyn.com/posts/ideas27/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the 27th post in a series describing some of the ideas I’ve accumulated. Chemical Mixing on Tayler Columns What’s the idea? In rapidly rotating stars and planets convection forms narrow columns. Mixing is rapid along the column but slow across columns. https://adamjermyn.com/posts/conditioning_fine_tuning/ Wed, 17 Aug 2022 16:55:03 -0400 https://adamjermyn.com/posts/conditioning_fine_tuning/ Posted to the AI Alignment Forums. (Thanks to Evan Hubinger and Nicholas Schiefer for comments on these ideas.) These are some notes on the relation between conditioning language models, prompting, and fine-tuning. The key takeaways are: Prompting and fine-tuning can both be used to condition language models. Prompting is quite restricted in the kinds of conditionals it can achieve. Fine-tuning can implement arbitrary conditionals in principle, though not in practice. https://adamjermyn.com/posts/ideas26/ Wed, 17 Aug 2022 11:42:48 -0400 https://adamjermyn.com/posts/ideas26/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the 26th post in a series describing some of the ideas I’ve accumulated. This idea is based on work I did with Jim Fuller. Helium Flash and Surface Effects What’s the idea? Stars below around $3 M_\odot$ exhibit a helium flash, in which degenerate helium in their cores ignites. https://adamjermyn.com/posts/ideas25/ Mon, 15 Aug 2022 09:53:44 -0400 https://adamjermyn.com/posts/ideas25/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the 25th post in a series describing some of the ideas I’ve accumulated. This idea is based on discussions and calculations with Matteo Cantiello, Daniel Lecoanet, Evan Anders, Eoin Farrell, Yan-Fei Jiang, Lars Bildsten, and Will Schultz. Overshooting at low Péclet number What’s the idea? https://adamjermyn.com/posts/conditioninggenerativemodels2/ Thu, 21 Jul 2022 16:34:05 -0400 https://adamjermyn.com/posts/conditioninggenerativemodels2/ This is a followup to Conditioning Generative Models based on further discussions with Evan Hubinger, Nicholas Schiefer, Abram Demski, Curtis Huebner, Hoagy Cunningham, Derek Shiller, and James Lucassen, as well as broader conversations with many different people at the recent ARC/ELK retreat. For more background on this general direction see Johannes Treutlein’s “Training goals for large language models”. Background Previously, I wrote about ways we could use a generative language model to produce alignment research. https://adamjermyn.com/posts/quantilizers/ Mon, 18 Jul 2022 12:33:31 -0400 https://adamjermyn.com/posts/quantilizers/ (Cross-posted from the AI Alignment Forum) Thanks to Evan Hubinger for discussions about quantilizers, and to James Lucassen for discussions about conditioned generative models. Many of these ideas are discussed in Quantilizers: A Safer Alternative to Maximizers for Limited Optimization: this post just expands on a particular thread of ideas in that paper. Throughout I’ll refer to sections of the paper. I have some remaining confusion about the “targeted impact” section, and would appreciate clarifications/corrections! https://adamjermyn.com/posts/ideas24/ Mon, 18 Jul 2022 10:55:44 -0400 https://adamjermyn.com/posts/ideas24/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the 24th post in a series describing some of the ideas I’ve accumulated. This idea is based on a paper by Xihui Zhao and Jim Fuller. What saturates the centrifugal instability? What’s the idea? The centrifugal instability is a linear instability (section 4) caused by the centrifugal force. https://adamjermyn.com/posts/grouped_loss/ Sat, 09 Jul 2022 08:21:00 -0400 https://adamjermyn.com/posts/grouped_loss/ (Cross-posted from the AI Alignment Forum) Thanks to Evan Hubinger and Beth Barnes for comments on these ideas. Language models exhibit clear scaling laws, where the loss is a power-law in model size. This offers a lot of predictive power, and seems like a useful thing to know. By contrast, individual capabilities often exhibit sharp discontinuities in performance as a function of model size and training time. It would be great if individual capabilities just gradually improved like the broader loss. https://adamjermyn.com/posts/ideas23/ Fri, 01 Jul 2022 11:52:18 -0400 https://adamjermyn.com/posts/ideas23/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the 23rd post in a series describing some of the ideas I’ve accumulated. This idea is closely related to conversations I once had with Dong Lai. Can waves advect magnetic fields? What’s the idea? The electrical conductivity in stars is enormous, so the magnetic diffusivity is small. https://adamjermyn.com/posts/lat/ Wed, 29 Jun 2022 16:05:25 -0400 https://adamjermyn.com/posts/lat/ Posted to the AI Alignment Forum. The Problem We’d like to train models to be robustly safe, even in environments that may fall well outside of the training distribution. Unfortunately all we get to work with is the training distribution, which makes ensuring robust generalization difficult. Deception is an example of this concern. Models may be well-behaved in the training environment because they realize it is a training environment, but behave poorly in deployment because they notice the distributional shift. https://adamjermyn.com/posts/ideas22/ Wed, 29 Jun 2022 12:21:53 -0400 https://adamjermyn.com/posts/ideas22/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the 22nd post in a series describing some of the ideas I’ve accumulated. This idea is closely related to conversations I’ve had over the years to with George Varnavides. Waves in electron hydrodynamics What’s the idea? In some materials, electron scattering is momentum-conserving (on some relevant length- and time-scale). https://adamjermyn.com/posts/tracespeed/ Mon, 27 Jun 2022 10:15:31 -0400 https://adamjermyn.com/posts/tracespeed/ Posted to the AI Alignment Forum here. Thanks to Evan Hubinger for suggesting this idea. Training Trace Priors are priors over boolean circuits which examine the outputs of gates on samples from the training distribution, typically for purposes of steering models away from having components that were never tested during training. The one I like to think about is the One-Gate Trace Prior (OGT Prior), which penalizes circuits if there are gates with constant outputs during training. https://adamjermyn.com/posts/conditioninggenerativemodels/ Mon, 27 Jun 2022 10:14:16 -0400 https://adamjermyn.com/posts/conditioninggenerativemodels/ Posted to the AI Alignment Forum here. This post was written in response to Evan Hubinger’s shortform prompt below, and benefited from discussions with him. Suppose you had a language model that you knew was in fact a good generative model of the world and that this property continued to hold regardless of what you conditioned it on. Furthermore, suppose you had some prompt that described some agent for the language model to simulate (Alice) that in practice resulted in aligned-looking outputs. https://adamjermyn.com/posts/astroideas21/ Mon, 27 Jun 2022 08:18:21 -0400 https://adamjermyn.com/posts/astroideas21/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the 21st post in a series describing some of the ideas I’ve accumulated. How do stars evolve in molecular clouds? What’s the idea? Some stars are in molecular clouds, and the clouds probably affect their evolution by blocking escaping light, exerting pressure on their surfaces, possibly accreting mass onto their surfaces, coupling to their rotation via magnetic fields, and more. https://adamjermyn.com/posts/tn_bc/ Fri, 17 Jun 2022 09:21:45 -0400 https://adamjermyn.com/posts/tn_bc/ Cross-posted from LessWrong. (Thanks to Mark Xu for comments on these notes.) These are some notes on tensor networks, the mapping from boolean circuits to tensor networks, and how #SAT can be written as a tensor network contraction. Definitions and Notation What is a tensor? A rank-$d$ tensor is an array of numbers with $d$ integer indices, e.g. $T_{ijk}$ is a rank-3 tensor. The range of each index is called its bond dimension. https://adamjermyn.com/posts/astroideas20/ Fri, 17 Jun 2022 07:25:45 -0400 https://adamjermyn.com/posts/astroideas20/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the twentieth post in a series describing some of the ideas I’ve accumulated. This idea emerged from a conversation with Yuri Levin. Black holes in gas clouds What’s the idea? Asa black hole moves through a gas cloud it should drag on the gas, depositing kinetic energy. https://adamjermyn.com/posts/astroideas19/ Thu, 16 Jun 2022 09:20:03 -0400 https://adamjermyn.com/posts/astroideas19/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the nineteenth post in a series describing some of the ideas I’ve accumulated. This idea was suggested to me by Yuri Levin and I’ve just fleshed it out a little below. Do magnetic fields slow accretion onto AGN stars? https://adamjermyn.com/posts/multigate_prior/ Wed, 15 Jun 2022 15:33:51 -0400 https://adamjermyn.com/posts/multigate_prior/ Posted to LessWrong here. (Thanks to Evan Hubinger and Nicholas Schiefer for suggestions and discussions around these ideas) Multi-Gate Traces We can improve on one-gate traces with multi-gate traces! Suppose we have $N$ training samples. Then we can reasonably estimate joint probability distributions over up to $k=\lfloor\log_2 N\rfloor$ gates, giving the prior: $$ p\propto \exp\left(-\sum_{i_1…i_k}\sum_{j\in [0,1]^k}p_{i_1…i_k}(j)^m\right) $$ That is, we sum over all selections of $k$ gates and compute the entropy of their joint distribution over $2^k$ states. https://adamjermyn.com/posts/astroideas18/ Wed, 15 Jun 2022 12:43:36 -0400 https://adamjermyn.com/posts/astroideas18/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the eighteenth post in a series describing some of the ideas I’ve accumulated. Wave Heating and F-type Stars What’s the idea? Convection zones emit internal gravity waves at their boundaries. There is some uncertainty as to exactly how much power escapes in these waves, but estimates are of order $L_{\rm conv}\mathcal{M}$, where $L_{\rm conv}$ is the heat carried by convection and $\mathcal{M}$ is the convective Mach number. https://adamjermyn.com/posts/astroideas17/ Tue, 14 Jun 2022 11:01:07 -0400 https://adamjermyn.com/posts/astroideas17/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the seventeenth post in a series describing some of the ideas I’ve accumulated. Differential rotation in stars and planetary orbits What’s the idea? Stars exhibit differential rotation. At the same time, tides tend to bring stellar rotation rates and planetary orbits closer together in period. https://adamjermyn.com/posts/trace_training/ Mon, 13 Jun 2022 13:14:01 -0400 https://adamjermyn.com/posts/trace_training/ (Crossposted from the AI Alignment Forums) I’m worried about scenarios involving deceptive models. We’ve failed at inner alignment so the model has goals that are not aligned with ours. It can somehow detect when it’s in training, and during training it pretends to share our goals. During deployment, surprise! The model paperclips the universe. In this story deception is all about the model having hidden behaviors that never get triggered during training. https://adamjermyn.com/posts/astroideas16/ Mon, 13 Jun 2022 09:30:25 -0400 https://adamjermyn.com/posts/astroideas16/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the sixteenth post in a series describing some of the ideas I’ve accumulated. Can LISA see stellar convection? What’s the idea? Convection in stars produces density fluctuations, resulting in a change in the quadrupole moment. This effect should be of order the square of the convective Mach number (which controls density fluctuations). https://adamjermyn.com/posts/elk1/ Mon, 13 Jun 2022 09:27:54 -0400 https://adamjermyn.com/posts/elk1/ Posted to the AI Alignment Forum here. (This proposal received an honorable mention in the ELK prize results, and we believe was classified among strategies which “reward reporters that are sensitive to what’s actually happening in the world”. We do not think that the counterexample to that class of strategies works against our proposal, though, and we have explained why in a note at the end. Feedback, disagreement, and new failure modes are very welcome! https://adamjermyn.com/posts/astroideas15/ Thu, 09 Jun 2022 10:41:57 -0400 https://adamjermyn.com/posts/astroideas15/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the fifteenth post in a series describing some of the ideas I’ve accumulated. How much material do AGN stars re-accrete? What’s the idea? Stellar evolution in AGN disks is expected to produce massive stars that exist in a balance between wind mass loss and accretion. https://adamjermyn.com/posts/astroideas14/ Wed, 08 Jun 2022 09:24:53 -0400 https://adamjermyn.com/posts/astroideas14/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the fourteenth post in a series describing some of the ideas I’ve accumulated. How are stars in the galactic center related to the prior AGN phase? What’s the idea? We think the Milky Way used to host an AGN disk. https://adamjermyn.com/posts/astroideas13/ Tue, 07 Jun 2022 09:17:06 -0400 https://adamjermyn.com/posts/astroideas13/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the thirteenth post in a series describing some of the ideas I’ve accumulated. Can accretion push RGB stars onto the Main-Sequence? What’s the idea? Stars enter the Red Giant Branch (RGB) when they exhaust hydrogen in their cores. Mixing hydrogen-rich material into their cores should reverse this transition, pushing them back onto the main-sequence. https://adamjermyn.com/posts/give_model_builder/ Mon, 06 Jun 2022 16:54:20 -0400 https://adamjermyn.com/posts/give_model_builder/ (Thanks to Evan Hubinger for suggesting this problem and for comments on these ideas. Feedback is very welcome. Cross-posted from LessWrong.) The Setup Suppose we have a prior $p_{\rm good}(\mathrm{model})$ that we think is non-deceptive, such that we can sample from this prior and get good models that are inner-aligned. These models may, for instrumental purposes, need to themselves produce models (via searches/optimization/meta-learning). For instance we might ask an outer model to play poker. https://adamjermyn.com/posts/astroideas12/ Mon, 06 Jun 2022 07:21:20 -0400 https://adamjermyn.com/posts/astroideas12/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the twelfth post in a series describing some of the ideas I’ve accumulated. What would happen if two AGN stars merged? What’s the idea? We think that stars in AGN disks live in a balance between accretion and radiation-driven mass-loss, and that this is a stable equilibrium. https://adamjermyn.com/posts/astroideas11/ Sun, 05 Jun 2022 08:19:42 -0400 https://adamjermyn.com/posts/astroideas11/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the eleventh post in a series describing some of the ideas I’ve accumulated. Does turbulence change nuclear burning in stars? What’s the idea? Nuclear burning rates are exponentially sensitive to temperature. Stellar models calculate the mean temperature at each point in a star, but there can be fluctuations about this mean due to turbulent motions. https://adamjermyn.com/posts/astroideas10/ Sat, 04 Jun 2022 12:32:26 -0400 https://adamjermyn.com/posts/astroideas10/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the tenth post in a series describing some of the ideas I’ve accumulated. Dynamical Time Warping for Isochrones What’s the idea? Dynamical time warping (DTW) is a technique used to align audio files (or other temporal data). The idea is: given two time-sequences which are supposed to be similar, can we deform the flow of time in one to make it more-closely match the other? https://adamjermyn.com/posts/give_ai_tools/ Sat, 04 Jun 2022 08:21:00 -0400 https://adamjermyn.com/posts/give_ai_tools/ (Cross-posted from LessWrong) One kind of concern with AI is that: There are some tools that are instrumentally useful for an AI to have. Most/the most accessible versions of those tools are dangerous. The AI doesn’t care which versions are dangerous. Hence, the AI will probably develop dangerous tools for instrumental reasons. You might call concerns like this Instrumental Danger Problems. This post aims to examine some existing approaches to Instrumental Danger Problems, and to introduce a new one, namely “Giving the AI safe tools”. https://adamjermyn.com/posts/astroideas9/ Sat, 04 Jun 2022 08:11:07 -0400 https://adamjermyn.com/posts/astroideas9/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the ninth post in a series describing some of the ideas I’ve accumulated. Does the core convective dipole align with rotation? What’s the idea? I’m struggling to find a reference, but a recent result out of convection simulations in fully spherical domains is that the dominant flow pattern is a dipole (e. https://adamjermyn.com/posts/astroideas8/ Fri, 03 Jun 2022 09:34:52 -0400 https://adamjermyn.com/posts/astroideas8/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the eighth post in a series describing some of the ideas I’ve accumulated. Do pulsating stars reveal inhibited convection? What’s the idea? Many A/B stars have pulsations driven by the $\kappa$-mechanism. In these stars, the pulsation often interacts with near-surface convective layers, motivating tools like time-dependent convection theories. https://adamjermyn.com/posts/astroideas7/ Thu, 02 Jun 2022 08:54:24 -0400 https://adamjermyn.com/posts/astroideas7/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the seventh post in a series describing some of the ideas I’ve accumulated. Lensing from compact remnants in the Galactic Center What’s the idea? It’s thought that the Milky Way used to host an AGN disk. AGN disks likely contain many stars, and those stars very preferentially evolve into black holes. https://adamjermyn.com/posts/astroideas6/ Wed, 01 Jun 2022 07:25:16 -0400 https://adamjermyn.com/posts/astroideas6/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the sixth post in a series describing some of the ideas I’ve accumulated. Theoretical limits on extreme baseline optical interferometry What’s the idea? Interferometry let’s you build a giant ‘virtual telescope’ out of lots of small telescopes. The effective baseline is the longest distance between telescopes, and the number of pixels of effective resolution is roughly the number of telescopes. https://adamjermyn.com/posts/astroideas5/ Tue, 31 May 2022 07:15:04 -0400 https://adamjermyn.com/posts/astroideas5/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the fifth post in a series describing some of the ideas I’ve accumulated. Does the Parker mechanism spin down AGN stars? What’s the idea? Collaborators and I previously predicted that stars embedded in AGN disks should end up spinning really fast (near breakup). https://adamjermyn.com/posts/astroideas4/ Thu, 26 May 2022 16:58:15 -0400 https://adamjermyn.com/posts/astroideas4/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the fourth post in a series describing some of the ideas I’ve accumulated. How correlated are the positions of stars? What’s the idea? Thanks to the Gaia mission, we now have accurate 3D positions for a large number of stars in the galaxies. https://adamjermyn.com/posts/astroideas3/ Wed, 25 May 2022 16:33:00 -0400 https://adamjermyn.com/posts/astroideas3/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the third post in a series describing some of the ideas I’ve accumulated. Can an ad-hoc electron binding model fix the low-temperature low-density limit of the Skye EOS? What’s the idea? The reason that the Skye EOS doesn’t cover all temperatures and densities is that it assumes fully ionized matter, which breaks down at low temperatures. https://adamjermyn.com/posts/astroideas2/ Tue, 24 May 2022 12:56:55 -0400 https://adamjermyn.com/posts/astroideas2/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. This is the second post in a series describing some of the ideas I’ve accumulated. Can planet injestion explain magnetic RGB stars? What’s the idea? There are some very magnetic Red Giant Branch (RGB) stars, with measured internal fields of order $10^{5-7}\mathrm{G}$. https://adamjermyn.com/posts/astroideas1/ Mon, 16 May 2022 07:38:03 -0400 https://adamjermyn.com/posts/astroideas1/ I sometimes have research ideas that I think are cool, but that don’t make sense for me to pursue. I generally just make a note of them and move on. Over the next few months I’ll be collecting those notes and posting those ideas here. The format I’ll try to follow is to explain: What the idea is. Why I think it’s an interesting idea. How I think someone interested in working on the idea could get started. https://adamjermyn.com/posts/chaos_stars/ Mon, 16 May 2022 07:34:19 -0400 https://adamjermyn.com/posts/chaos_stars/ Are stars predictable or chaotic? A system is chaotic if small perturbations grow exponentially with time. This makes chaotic systems unpredictable on long time-scales, because it becomes fine details in the starting conditions come to dominate at long times. So: are stars predictable or chaotic? As far as I know this is an open question. There is some sense in which most stars are predictable. They exist near thermal and hydrostatic equilibrium, and they tend to return to equilibrium when perturbed. https://adamjermyn.com/posts/auto_diff_mesa/ Wed, 11 May 2022 15:28:14 -0400 https://adamjermyn.com/posts/auto_diff_mesa/ Over the last two years I developed an automatic differentiation (auto_diff) module in Fortran to support development of the Modules for Experiments in Stellar Astrophysics (MESA) project. This post gives a brief rundown of how the auto_diff module works and how I built it. All of the code I mention below lives on GitHub. What is auto_diff? Forward-mode automatic differentiation via operator overloading: Forward-mode means we calculate the chain rule as we go. https://adamjermyn.com/posts/rest_2/ Tue, 12 Apr 2022 17:19:49 -0400 https://adamjermyn.com/posts/rest_2/ There is a fourth kind of rest I didn’t mention: Inactivity. All the other kinds of rest, Distraction, Recreation, and Maintenance, are active. They involve doing something. That something might not be productive in any traditional sense, but it’s an active choice to do something rather than nothing. Inactivity is different. Inactivity is choosing to do nothing. It’s sitting on a couch being bored, letting thoughts come and go but not pursuing them. https://adamjermyn.com/posts/ml_2/ Tue, 12 Apr 2022 17:17:54 -0400 https://adamjermyn.com/posts/ml_2/ Summary Adding convolutional layers is pretty straightforward, but there are some indexing subtleties to watch. Pooling layers can really improve performance. With two pooling layers, I found an example of a too-high learning rate making the model get stuck with poor performance. The more pooling layers I used the faster the model took a training step (because the dense linear layer was smaller). The more pooling layers I used, the more important the number of channels in the convolutional/pooling layers became. https://adamjermyn.com/posts/ml_1/ Mon, 11 Apr 2022 17:06:39 -0400 https://adamjermyn.com/posts/ml_1/ Summary Covered boilerplate for getting started, tensors, gradients. Poked in some detail at how gradients work. Played with linear regression, nonlinear regression, both with gradient descent. Built an MNIST classifier following a PyTorch tutorial. Experimented with the shape and size of the NN. Got PyTorch running on a machine with a GPU. Boilerplate Import: import torch Need to use a special float type for the elements of tensors, and need to know which device the compute graph will run on: https://adamjermyn.com/posts/acting_thinking/ Mon, 11 Apr 2022 14:33:01 -0400 https://adamjermyn.com/posts/acting_thinking/ Do thoughts cause actions, or do actions cause thoughts? Or does causation flow back and forth, more like a tide than a one-way river? It’s easy to see how thoughts cause actions. You think about reading a book, so you go pick up a book and start reading. It’s so easy to see the causation flow this way that it feels like thoughts must precede actions. How can you act, if you haven’t decided what to do? https://adamjermyn.com/posts/progressive_tax/ Sun, 27 Mar 2022 17:15:37 -0400 https://adamjermyn.com/posts/progressive_tax/ Taxes create incentives. When the government tax an activity, society does less of it. Tax alcohol, people drink less. Tax smoking, people smoke less. Tax income, people work less. Tax investment, people invest less. I’m going to take this as given for now, even though there are many caveats on all of these statements. This suggests a simple principle for crafting a tax code: tax the things we want less of, and don’t tax the things we want more of. https://adamjermyn.com/posts/ai_risk_few_words/ Mon, 14 Mar 2022 14:27:25 -0400 https://adamjermyn.com/posts/ai_risk_few_words/ AI safety is a weird area, and in trying to explain it to friends I realized that I don’t have a simple explanation of what the (existential) risks are and why I think they’re important. Word limits force simplicity, so here are some attempts to explain AI risk in different numbers of words. 150 words Do you value the same things as your great grandfather? What we value changes from generation to generation. https://adamjermyn.com/posts/procrast/ Tue, 01 Mar 2022 13:43:15 -0500 https://adamjermyn.com/posts/procrast/ When I procrastinate on something, it’s usually because there’s something I find aversive about it. Lately I’ve been trying to find patterns in what these aversions are, and noticed a big one: I usually find the start of a big task much more aversive than any other part. For instance, I often find myself shrinking away from beginning to write papers, whereas everything is fine once I’ve gotten going. I’m not sure why the act of starting feels so scary, but I’ve found two ways to make it better. https://adamjermyn.com/posts/rest/ Fri, 25 Feb 2022 16:40:08 -0500 https://adamjermyn.com/posts/rest/ Rest is important, but what is it? I’ve noticed three kinds of activities people call rest: Maintenance Distraction Recreation Many activities could fall into any of these categories, and the distinction I want to draw is all about the motivation. Maintenance is what you do to keep life going. Think of cooking, cleaning, sleep, hygiene, sorting through mail, etc. These activities are both necessary and often get pushed to the wayside during a busy work week, so they’re natural to reach for in down time. https://adamjermyn.com/posts/broad_success/ Mon, 31 Jan 2022 11:55:52 -0500 https://adamjermyn.com/posts/broad_success/ Projects begin with goals. At first these are implicit, but as a project takes shape it helps to state them explicitly. Stating goals serves two purposes: it defines success, and it prompts action. Because of this dual purpose, many goals miss the mark. On the one hand, is easy to come up with goals which define success but which are too vague to be actionable: “Change the world.” “Disrupt the industry. https://adamjermyn.com/posts/compassion/ Mon, 10 Jan 2022 09:00:45 -0500 https://adamjermyn.com/posts/compassion/ (Epistemic Status: This is an attempt to draw a distinction that feels important, but which I have little evidence for.) Recently I’ve been reading Self-Compassion by Kristin Neff. I haven’t checked any of the underlying evidence, so buyer beware, but the basic premise of the book is that there are large benefits to crafting an intentional attitude of compassion towards oneself. I don’t want to focus on the specifics of Neff’s book though. https://adamjermyn.com/posts/testing_ii/ Thu, 23 Dec 2021 16:09:40 -0500 https://adamjermyn.com/posts/testing_ii/ (This is not medical advice. Co-written with Nicholas Schiefer. Thanks to Sylvia Hürlimann and Katherine McDaniel for helpful discussions and comments. Corrections welcome.) This holiday season is also Omicron season. The Omicron variant dramatically increases COVID risk for basically everyone, especially with case counts exploding and doubling every two days in most northern areas (New York, Boston, London). Although Omicron seems to be less severe than other variants for most people, it’s probably worth some effort to protect high-risk individuals, especially while case rates are so high. https://adamjermyn.com/posts/scientific_compute/ Thu, 16 Dec 2021 16:30:25 -0800 https://adamjermyn.com/posts/scientific_compute/ Scientists use computers, and not just for typesetting. Lots of experimentation and discovery is done with computers, from modeling climate to predicting material properties to discovering exoplanets. Because computation is applied to such a broad range of questions, the ways it gets used are many and varied. This diversity of application means that tools developed and lessons learned in one domain of scientific compute workflow often don’t generalize to very different domains. https://adamjermyn.com/posts/claim_coin/ Mon, 13 Dec 2021 18:16:59 -0800 https://adamjermyn.com/posts/claim_coin/ (This is not investment advice.) Stable coins are an important tool, allowing investors to trade ordinary (fiat) currencies on the blockchain, but they have a somewhat opaque risk structure. Usually either the coin tracks the desired currency or else the peg fails, in which case investors start withdrawing and the coin suddenly trades well below the peg. What if there was a coin with a more transparent risk structure? What would that make possible? https://adamjermyn.com/posts/agency_social/ Sat, 04 Dec 2021 09:01:56 -0800 https://adamjermyn.com/posts/agency_social/ Agency is where you act to achieve your goals. Unfortunately agency is hard and not the default state. Social nudging is a powerful way to combat this. Here are a few lightly-edited conversations that I’ve observed or participated in that pushed towards more agency: Fire P1: There’s a fire ~20 miles away and it’s raining ash. P2: Do you have an air purifier? P1: … no, I should get one. https://adamjermyn.com/posts/riboflavin/ Thu, 02 Dec 2021 17:03:20 -0800 https://adamjermyn.com/posts/riboflavin/ Abstract (This is not medical advice. I am not a medical professional. Thanks to Rose DiLoreto for the idea for this post.) Supplementing with Riboflavin (Vitamin B2) seems probably helpful for headaches and migraines, likely reducing their frequency by about 50% as well as reducing severity somewhat. This is backed by multiple randomized controlled trials, and I’ve tried to check for anything funny in those trials (small samples, weird controls, etc. https://adamjermyn.com/posts/mugged_by_probability/ Mon, 25 Oct 2021 07:16:09 -0700 https://adamjermyn.com/posts/mugged_by_probability/ (Numbers in this post are made up to illustrate a point. Likewise the plot device need not be cryptocurrencies, that’s just flavor text.) You’re walking on the street, minding your business. Suddenly, a stranger (S) comes up and starts talking to you: S: Have you heard? There’s a one-in-a-million chance you could save a billion lives! All you have to do is found a crypto currency, become extremely rich, and use that money to fight Malaria! https://adamjermyn.com/posts/covid_testing_timing/ Tue, 07 Sep 2021 10:17:00 -0500 https://adamjermyn.com/posts/covid_testing_timing/ COVID Testing - Sensitivity and Timelines This past week a close contact had a potential COVID exposure, so I wanted to understand a few basic questions about COVID testing to help figure out what is safe and what isn’t. Three questions seem especially relevant: How sensitive are PCR and rapid antigen tests? If someone tests negative on a rapid antigen test but positive on PCR, should they worry about infecting other people? https://adamjermyn.com/posts/agency_school/ Sun, 01 Aug 2021 14:42:17 -0400 https://adamjermyn.com/posts/agency_school/ A School for Agency Agency is where you act to achieve your goals. It’s where you look at problems and say “How could I solve them?” and really think about that and arrive at possible courses of action and then take those actions. Explicitly, agency is about executing three steps: Notice a problem/question. Brainstorm solutions. Try solutions until something works or the costs exceed the benefits. An example I return to often in the summer is air conditioning. https://adamjermyn.com/posts/automated_reviews/ Tue, 25 May 2021 15:04:36 -0400 https://adamjermyn.com/posts/automated_reviews/ Automated Project Reviews Getting things right without feedback is really hard, but not all feedback comes from other people! One quick way to generate meaningful feedback is by looking back at past work. A challenge with looking back is doing it at the right time. For some kinds of feedback this is easy: schedule weekly/monthly/quarterly reviews, with prompts tuned to the relevant time-scale. But some areas of life don’t run on a regular schedule. https://adamjermyn.com/posts/near_existential_risk_pt2/ Fri, 23 Apr 2021 16:05:02 -0400 https://adamjermyn.com/posts/near_existential_risk_pt2/ I recently wrote about near-existential risk. These are risks which result in massive losses to society of people/technology/capability, but which themselves do not actually end or permanently curtail human existence. Here I want to build a toy model for thinking about these risks and how they relate to actual existential risks. A Toy Model For these purposes I’m going to reduce the many dimensions of loss to a single parameter $\ell$. https://adamjermyn.com/posts/near_existential_risk/ Sun, 28 Mar 2021 16:17:44 -0400 https://adamjermyn.com/posts/near_existential_risk/ Longtermism is a moral theory that says that most of the value we (humans) ought to care about lies in the future, and so we should look hard for opportunities to influence the future in positive ways. Longtermism arises as a natural consequence of three ideas: Utilitarianism: We should improve the lives/subjective wellbeing of as many people as much as possible, evaluated by summing improvements over lives. Future lives are just as valuable as present lives. https://adamjermyn.com/posts/values_examples/ Sat, 23 Jan 2021 17:51:55 -0500 https://adamjermyn.com/posts/values_examples/ I previously argued that tools should express values. Here are some examples of tools I’ve made with an eye to values. A Brainstorming Environment I have a shell script that sets up a work environment for brainstorming. It does this by copying and opening a template text document that says the following: # Session: - Maintain an attitude of suspended judgement. - You're here to climb out of the basin. - Cmd+E Will bring up an (incomplete) list of possible approaches. https://adamjermyn.com/posts/mars_soil/ Sun, 10 Jan 2021 10:24:23 -0500 https://adamjermyn.com/posts/mars_soil/ Growing food on Mars is a critical component of making a self-sufficient city because shipping food is expensive. Unfortunately it seems that Martian regolith is too alkaline and is missing key nutrients needed to grow crops. The alkaline issue can be solved by adding acid, which is both low-weight and can be produced locally if need be. The nutrient issue can be addressed either by adding organic material or by artificial nutrient supplementation. https://adamjermyn.com/posts/mars_software/ Fri, 01 Jan 2021 14:29:20 -0500 https://adamjermyn.com/posts/mars_software/ Building a city on Mars will be very capital intensive because in the early stages it’s extremely expensive to sustain people on Mars. But capital is a vague term: what does this actually look like? The last few attempts at a human presence in space primarily used physical capital: heavy machinery, giant rockets and space stations and industrial equipment. Mars will need all of that too. But the last few times were in the 1960s-90s, when computers were slow and software much more limited and expensive than today. https://adamjermyn.com/posts/robust_tools/ Wed, 30 Dec 2020 10:37:04 -0500 https://adamjermyn.com/posts/robust_tools/ Good tools are robust. This doesn’t just mean that they don’t fail, it means that in many instances they can’t fail. They take common failure modes off the table, freeing you to use them and focus on your goals. Impossible Failure Modes For a robust tool many failure modes are, by design, simply impossible, to the point where it’s hard to imagine what those failure modes are. Consider a solid metal hammer. https://adamjermyn.com/posts/tools_values/ Fri, 11 Dec 2020 15:56:14 -0500 https://adamjermyn.com/posts/tools_values/ I want my tools to express values. I want my text editor to tell me to write clearly. I want my scientific software to command me “Be true, be open, be humble, be curious.” I want tools to be less blank pages and more pleas for me to honor the values of the craft. Blank is not Neutral Many tools today are designed like blank pages. A writing app has a large blank space, with perhaps a few buttons for changing fonts and an indicator of how many words are on the page. https://adamjermyn.com/posts/alpha_beta/ Mon, 07 Dec 2020 11:33:05 -0500 https://adamjermyn.com/posts/alpha_beta/ Economies of Scale I previously wrote about a simple model of an early Martain economy. A key input to that model is the way that human productivity scales with population and investment. I captured this with the simple scaling $$ p \propto I^\alpha n^\beta $$ for $\alpha,\beta > 0$. This relation says that the per-person productivity $p$ scales as some positive power $\alpha$ of the per-person investment $I$, which could reflect machinery or materials or personnel support from Earth, and as another positive power $\beta$ of the number of people $n$. https://adamjermyn.com/posts/mars_economy/ Sun, 15 Nov 2020 16:33:47 -0500 https://adamjermyn.com/posts/mars_economy/ An Economic Model of Martian Growth Mars is an expensive place to live, but it rapidly becomes cheaper the more infrastructure it has! Because of that the total cost of building a self-sufficient city on Mars depends on how quickly infrastructure can be built. This is the key question I want to understand: what is the best way to invest in Martian infrastructure that minimizes the total cost? A Simple Model To do this I’ll build a simple model of economic growth on Mars. https://adamjermyn.com/posts/mars_scales/ Sat, 24 Oct 2020 15:52:11 -0400 https://adamjermyn.com/posts/mars_scales/ Mass Scales for Living on Mars Big thanks to Casey Handmer for reading and commenting on an early draft of this. One of the most exciting things that may well happen this century is a sustained human presence on Mars. The main reason this is possible at all is that SpaceX has cut launch costs dramatically. Despite this progress, cargo sent from Earth and soft-landed on Mars is likely to be the primary resource constraint in the Martian economy for a very long time, so it is worth exploring the scale of cargo required to support a human presence, and what that translated into in dollars. https://adamjermyn.com/workflow/feedback/ Mon, 01 Jan 0001 00:00:00 +0000 https://adamjermyn.com/workflow/feedback/ Feedback Getting things right without feedback is really hard, so getting frequent and meaningful feedback is important. I do this by asking others for feedback and by looking back at how past work has gone. Asking Others Giving good feedback is hard. When I ask for feedback I try to make it easier by making it clear what I want feedback on, and by being as specific as possible. So feedback on specific projects or interactions is often better than feedback on “my job in general”. https://adamjermyn.com/workflow/tools/ Mon, 01 Jan 0001 00:00:00 +0000 https://adamjermyn.com/workflow/tools/ Tools Tools should make work easier, more effective, and more fun. This is an incomplete list of the tools I use, along with commentary on why I use them. Computers/OS To help with work/life balance I use two computers, one for work and one for everything else. Both are Macbook Pro’s running macOS, one is 15" (work) and the other is 16" (other). Because Tools Should Express Values I used GeekTool to display a list of values over my desktop background (but below all windows).