# Journal

*Edit (1 week in): I lasted around 3 days perfectly productive, then started slipping. I still believe I can do better consistently though. Just need to figure out how to shift my equlibira.*

I feel like I've been slipping / not becoming more productive recently. since I want my second derivative to be positive so I thought I'd try posting a public journal everywhere with public-ish time tracking.

I have ~16 waking hours, I'd like the time to be spent roughly like this

Where M=Morning routine, S=Study (ML right now), A=Alignment (SERI MATS), E=Eating, P=Projects/Writing, L=Slack, R=Reading

Note: The ML stuff is relevant to Alignment and I'll often be writing about alignment things in Projects/Writing.

I'll provide time breakdowns of every day and match it to this pie chart. I'd make a higher tech solution but experience has taught my that's a bad idea.

I'm not married to this pie chart, but I do want to do an amount of work roughly proportional to this ~5 days/wk.

### 2022-11-07

Summary: ~10h productive stuff, Wasted ~1h 30min (if you count eating/cooking for 1h, lol)

Couldn't focus while meditating (tired from shifting sleep times) and likewise didn't push the exercising bar much. Setting timers during morning routine helped keep it time-boxed to 1h.

How I'll do better:
- Chunk study time to one thing (probably interpretability or gears level models of things)
- Get off computer earlier (11pm now, right before bed) so I can read things

Time breakdown:

- 2h LSTMs architecture and backprop
- 30min Karpathy's [WaveNet](https://www.youtube.com/watch?v=t3YJ5hKiMQ0) video (2-3x speed), learned some good pytorch wrangling.
- 30min Neel Nanda's [A Mathematical Overview of Transformer Circuits](https://youtu.be/KV5gbOmHbjU), need to alloc a larger chunk of time to grokking their paper.
- 1h 30min Wentworth's SERI MATS workshop + talking after.
- 30min Some financial chores/life-management things
- 1h 30min Buck's SERI MATS seminar
- 1h Atlas academy microecon
- 40min First of two JW retreat [lectures](https://youtu.be/mij7nYPKIHo) (3x speed)
- 40min ^ misc messing with the math/modelling I learned
- 1h Slightly distracted drafting of a post about my childhood and what unschooling is like
- 1h Chatting on discord

### 2022-11-08

Summary: ~10h work time, Wasted ~1.5h

How I'll do better:
- Prioritize mats stuff way more.[^1] *Concrete goal for tomorrow*: Spend 6hrs efficiently doing mats research (that means breaks, no random googling / browsing, brainstorming outside with a notebook, etc.)
- ~~Use people more~~ Collaborate with other mats scholars more. *Concrete goal for tomorrow:* DM one person, Ask two questions.

[^1]: I may or may not have just realized the research sprint ends on the 16th not the 25th. I am very competent at dates.

Time breakdown:

- 1h Anki (backlogged lol) and some mental math practice.
- 1.5h Reading interpretability stuff. [Toy models of superposition](https://transformer-circuits.pub/2022/toy_model/index.html) mentioned an interesting theorem...
- 1h Learn the proof of the [Johnson Lindenstrauss Lemma](https://youtu.be/Tw0J5Xv6xQw). I think I learned generally useful stuff here, though JD still isn't obvious (meaning I haven't fully understood it)
- 1.5h [Giant text file exercise](https://www.lesswrong.com/posts/nvP28s5oydv8RjF9E/mats-models#Giant_Text_File). Went well, I have a decent stacktrace of my alignment thinking with holes nicely laid out now; I'll use this technique more in future.
- 45min MATS slack and random reading stuff
- 30min Watch the first [convex optimization](https://www.youtube.com/watch?v=McLq1hEq3UY&list=PL3940DD956CDF0622&index=2) lecture & take notes. (4x speed ftw)
- 1h 30min Convex optimization class (atlas academy). Did some exercises, feel like we weren't going that fast / not hitting the heart of the matter. Skimming the exercises/homeworks Boyd assigns for his course might help (I think they're more holistic then the book ones)
- 40min MATS Research brainstorming
- 1h Investigate the new [Deepmind result](https://twitter.com/DeepMind/status/1600852768125726720). tentatively not very impressed, seems mostly marketing weak sause (like Gato and the matrix multiply paper)
- 1.5h MATS Research thinking

### 2022-11-09

Summary: ~10h Work, Wasted ~2.5h

How I did on daily goals:
- Did 6 hours of MATS, but not the focused 6 hours I was hoping for. (~3/4 meetings and reading)
- Didn't ask enough questions (actually forgot about this goal. Goal for tomorrow: don't forget about daily goals?)

Had to go to the chiropractor in the morning so missed my typical morning-routine-things. Read hot [inner/outer alignment takes](https://www.alignmentforum.org/posts/gHefoxiznGfsbiAu9/inner-and-outer-alignment-decompose-one-hard-problem-into) while waiting for my family.

Random thought: I said I had 16 hours in a day (15 allowing for 9hrs sleep), morning takes 1h, working 10h and wasting around 2h. that leaves ~2hrs unaccounted for. In practice they go to life management chores I don't count as work, and to tired discord things before bed (but after I write this list).

Possible solution to both: fill in the last bit of the time breakdown the next day to cover everything, and actually get off my computer at ~10ish to read/sleep.

Time breakdown:

- 2h chiropractor (driving + chiropracting + talking to family)
- 1h 15min Reading Alex's [inner/outer take](https://www.alignmentforum.org/posts/gHefoxiznGfsbiAu9/inner-and-outer-alignment-decompose-one-hard-problem-into). It's much better then the draft I read!
- 1h 15min Research brainstorm outside
- 1h Shard theory research meeting. Internalized the "think about your causal model for hypotheses" pattern more then previously. Put another way, ML experiments trying to figure out agi-relevant things are doing latent variable inference.
- 1h 30min Brainstorm questions and meet with Quintin. Talked about ML things, GPT things, got linked a bunch of papers, talked about Grokking & [Omnigrok](https://arxiv.org/abs/2210.01117) (i.e. Grokking mostly an artifact of initalization, though other phase transitions exist), his models for AGI.
- 30min Reading about the [founding of openai](https://blog.gregbrockman.com/my-path-to-openai) and greg [learning deep learning](https://blog.gregbrockman.com/how-i-became-a-machine-learning-practitioner), and other things he wrote. Spending a few years getting good enough to be a RE at OpenAI/Deepmind is looking more and more attractive... Even if I don't like it there I can always leave and have *awesome skills* to show for it
- 1h 45min Writing, Atlas academy
- 1h 30min Algorithms, Atlas academy
- 30min Call with Ashley (intro to academy). mostly me lamenting about how much more efficiently I could be using my tutors if I had more time / didn't have to do seri mats things
- 1h Writing up thoughts on market-based modeling of shard-based motivations & reread [alignment by default](https://www.alignmentforum.org/s/PKKsrXtuptWzaKCjr/p/Nwgdq6kHke5LY692J)
- 20min Reading some [writting advice](https://sashachapin.substack.com/p/write-faster-130
) and coldtakes [learning by writing](https://www.cold-takes.com/learning-by-writing/). very cool. it has [Paul Graham vibes](http://www.paulgraham.com/useful.html).
- 30min Journal
- ~2h Wasted reading discord and random lesswrong. Went to bed late (~1am). :(

Goals for tomorrow:
- Do the most important thing for my MATS research, from 10am - 7pm tomorrow (lots of brainstorming, writing, question-asking, reading papers. not much forum reading.)
- Spend 2hrs reading the fancy books I have, offline/away from computer. (Stuff relevant to SERI)
- Fill in time breakdown from yesterday counting what I did after writing the journal

### 2022-11-10

Summary: Worked ~6hrs, spent ~4hrs with family (forgot we had a thing scheduled today). Had a little trouble focusing from lack of sleep.

How I'll do better:
- Sleep is really important, I have way less focus & willpower when I'm low on sleep. Gotta go to sleep early!
- Get an outline for my research out & iterate on it. Aim directly at the final product, don't be satisfied with random alignment insights.

Time breakdown:

- Got up late-ish (9am) and started work at 10:20.
- 2h Read paper running interpretability on AlphaZero and write a [short post](alignmentforum.org/posts/obht9QqMDMNLwhPQS/asot-natural-abstractions-and-alphazero) about it.
- ~2h Of good research brainstorming outside with a notebook. Still feel like I'm aiming myself suboptimally / not hitting the crux of the problem. Had a few insights / understood things better though.
- Setting 5min timers and trying to make at least [one insight per 5 minutes](https://www.lesswrong.com/s/LAop879LCQWrM5YnE/p/ZxR8P8hBFQ9kC8wMy). If you can't make progress in 5 minutes, there's a sense in which you [can't make progress at all](https://www.alignmentforum.org/posts/PqMT9zGrNsGJNfiFR/alignment-research-field-guide#1__You_and_your_research).
- ~1h Of mostly-not-making-progress, read some things
- ~30min Break (Feeling pretty tired)
- ~30min Reinforcement learning math (procrastinating more research)
- 4h Going out with family...
- Staying up till midnight, ouch!

### 2022-11-11

Summary: 7.5h work (6.5h SERI + 1h meeting with an independent researcher), wasted ~3h (1h eating + ~2h discord).

Yesterday's goals:
- I mostly have an outline, though it has empty parts
-

Tired, didn't feel that productive. Won't give a full time-breakdown - just notable stuff that happened in roughly-chronological-order.

- Some christmas shopping (~30mins)
- ~2.5h on core research brainstorming / outlining
- ~1.5h Drafting a post distilling [A shot at the diamond alignment problem](https://www.alignmentforum.org/posts/k4AQqboXz8iE5TNXK/a-shot-at-the-diamond-alignment-problem). Didn't write fast enough, ended up mostly thinking about conceptual things.
- ~1.5hrs Reading [Toy models of superposition](https://transformer-circuits.pub/2022/toy_model/index.html)
- AI-pilled some kids visiting our house via GPT3 things
- Had a really good SERI meeting for our [Alignment 201](https://www.agisafetyfundamentals.com/alignment-201-curriculum) group. (Interpretability week, though we talked about the week after a bit too)
- 1h Interesting chat with an independent researcher / interesting person. They're interested in [discovering agent](https://arxiv.org/abs/2208.08345) [boundaries](https://www.lesswrong.com/s/LWJsgNYE8wzv49yEc) and linked me a really cool paper: [First Contact: Unsupervised Human-Machine Co-Adaptation via Mutual Information Maximization](https://sites.google.com/view/coadaptation?pli=1). I'll probably help them write up thoughts to lesswrong/AF.
- 1.5h Chatting on discord
- 30mins Extra discord waste

Plan for tomorrow:
- Probably skip wentworth's workshop (or do 9pm one)
- ~1-2h Expand ideas in outline (nested lists rock). Take breaks, don't get nerdsniped.
- ~1h Convert to prose I can show Garret without vomiting up blood (time-boxed)
- 2h Literature review
- 30min-1h Meet with Garret, try & get advice
- 1.5h Microecon class. Try and get insight on agency models
- 30min Speedrun a convex optim lecture
- 1.5h Convex optim class
- More literature review probably

### 2022-11-12

Summary: ~7h work, ~5h waste,

Wasted so much of the day! Around dinner started feeling really tired and watched ~4h of youtube. ([Lex and coffeezilla](https://www.youtube.com/watch?v=hi9Rf0oLdHk), [All in podcast: Twitter files](https://youtu.be/o6HBFo7ft9k), and a bunch of stupid terminator clips). Ideally I'd want to spend tired time reading blogs like lesswrong, reading fiction, or watching lectures.

Also didn't do any exercise or other habits. Got straight to work instead. Probably worse long-term. But whatever.

As for the work I did do:
- Brainstormed a bunch of alignment things. Made a slack thread with my lab notes (I think this helped, and people in SERI can comment on them now)
- Started writing the text for my SERI research (I'd had the outline mostly)
- Meeting with Garret about alignment, gave me some good ideas
- Atlas academy microeconomics. Didn't understand everything, will investigate tomorrow.
- Messed around with ML tools for a little. Found [rclip](https://github.com/yurijmikhalevich/rclip) (grep using openai CLIP) and learned I can run language model inference on CPU pretty fast. Need to implement semantic grep over my obsidian notes and downloaded books at some point.

Tomorrow, goals are to get good sleep, get up early, exercise/meditate/anki, and set 15min timers the whole day (to stay aware of time :D)

### 2022-11-13

Summary: ~7h Work, ~4h Waste. Did my morning routine stuff though.

Can't believe I watched 2h[^hours] of random youtube like an NPC! Clearly my systems aren't advanced enough, I'd have been roughly equally happy watching ML paper walkthroughs or Neel Nanda.

[^hours]: Technically 5h because I watch most things on 3x speed

Overall I did

- ~1h microeconomics review
- ~1.5hrs brainstorming [research hamming qusetions](https://www.lesswrong.com/posts/Thwfy4gNFx9kHgvov/research-hamming-questions)
- ~2.5hrs high value writing for seri mats
- ~1.5hrs algorithms class

Goals for tomorrow:
- Do all the habitica things (one is to not waste time on YouTube-like things)
- Review microecon stuff (~1-2h)
- Convex optimization (~1-2h)
- SERI MATS Research, send draft to people (~2-6h. stay focused!)

### 2022-11-14

Summary: ~9h Work (4.5h Studying and atlas academy, ~3h Waste (1h eating + ~2h when I wanted to be sleeping).

How I did on yesterday's goals:
- Didn't do all my daily goals (skipped Anki and meditation), also going to bed too late
- Reviewed microecon stuff. Was very confused, should maybe have waited to resolve more stuff in class? I'm realizing how much faster stuff goes when you have a tutor. (and I could get better at self-teaching!)
- Didn't do any convex optimization
- Did get my draft reviewed by a friend, wasn't that good. Tomorrow I'll try rewriting it several times from memory & putting the important bits up front.

Stayed up till 12pm. I could have done the same stuff in less time if I'd been time-boxing (aka: actually trying), that probably cost an hour or so.

Plans for tomorrow (time box everything!):
- Rewrite my seri thing ~~3 ways from memory (time-box 20mins each)~~ a bunch (2h)
- Write a good version using components from each draft (1h; time-box). Cite past work.
- Study the convex optimization stuff (~2h)
- Convex optimization class (1.5h)
- Reading papers 2h (time-box)
- Reading lesswrong bookmarked 1-2h (must do last)

### 2022-11-15

Summary: 10h Work (~4h in the [john](https://www.alignmentforum.org/users/johnswentworth) ask me anything, so more like 9h real work), ~2h Wasted.

(Also +2hrs talking with mats scholars after the John AMA. Good talk, though I did stay up till 1am this was plausibly worth it(?))

Rest of day breakdown will be less detailed because late:

- Read some of the [MIMI](https://sites.google.com/view/coadaptation?pli=1) paper, very cool. Learned a bit more about how mutual information is estimated in practice.
- SERI MATS writing, not that productive. ~~I should worry less about writing well and more about good ideas~~ the two are linked. worry about writing to the extent it helps my ideas.
- Insane amount of johnwentworth absorbing in his AMAs
- Convex optimization class/tutoring. Fun fact: there exist two disjoint closed convex sets such that there's no strict separating hyperplane. Enjoy your nightmares.

I feel not very good at stuff compared to some other mats scholars. Is there an 80/20 of stuff I can learn during the program? The main skills seem like
- Reading and implementing ML papers
- Comfy with linear algebra/calculus/stats (stats is worst for me right now)
- Pytorch / etc experience

Exception caught: I need to [deliberate once](https://mindingourway.com/deliberate-once/). My current plan of understanding & distilling the [discovering agents](https://arxiv.org/abs/2208.08345) paper is very good, I'll stick with that.

### 2022-11-16

Summary: ~7.5h work (maybe 4h focused/good), 3h Waste (mostly discord + bit of twitter. also played minecraft with my brothers for 30mins lol, commandblocks are coding-pilling them)

Currently thinking: I need to study a lot, I have many basic holes where I know basically nothing (e.g. Bio). there's some [pressure](https://www.lesswrong.com/posts/D4hHASaZuLCW92gMy/is-success-the-enemy-of-freedom-full) towards doing more of what I'm good at & this worries me. After SERI (& during to some extent) I want to spend a lot of time learning foundational gears-level models and things. Why did I start learning functional analysis before basic economics!

Goals set yesterday:

- [x] 1h Read papers (time-box 20mins/each)
- [ ] 1h More details for my proposed ML experiments
- [x] 30min Meeting with Quintin / bounce ideas off for research project thing
- [ ] 1h ~~Incorperating feedback (time-box)~~ (Didn't have much feedback things)
- [x] 1h Writing fast (want fodder to review for writing class)
- [ ] ~~1.5h Writing class~~ (Cancelled)
- [x] 1.5h Algorithms class
- [x] 2h Discovering agents paper understanding (might get tired in which case heck)

Additional stuff I did
- [x] ~1h Some discord management
- [x] ~1h Various chores (managing financial stuff)

### 2022-11-17

Summary: 7.5h Work, ~3.5h Waste (spent another 1.5hrs learning minecraft commands, and 2h watching youtube.)

Goals from yesterday:

- [x] 1h Finish Anki Reviews & *aggressively* cleanup cards
- [x] 1h Experiment with mental math/memory things
- [ ] ~~1-2h Microeconomics studying (mostly problems) (\w Anki!)~~ screwed around with a [chatgpt CLI](https://github.com/UlisseMini/chatgpt-cli****)
- [x] 2h Meditate (actually 1h 30min, stopped early & took a break to go inside when cold.)
- [x] ~~3h~~ 2h Reading Discovering agents & Causal Influence Diagrams
- [ ] ~~2h Read Alignment 201 content~~ (Inefficiency))

~~stayed up till 2am talking with people, oh well~~

### 2022-11-18

Spent a lot of time hanging out with my brothers and playing minecraft (3.5hrs), though I reflectively endorse 1h of that as I was teaching my little brother minecraft commands, which will hopefully make him see the value of good spelling, typing. And that programming is cool!

Ended up working ~5hrs on SERI mats stuff (~1.5h readings + ~3.5h writings) plus ~1h of misc other stuff. Re-read some of Soares's [productivity posts](https://www.lesswrong.com/posts/yFALNnscB2qgehnJv/deregulating-distraction-moving-towards-the-goal-and-level). Feeling a inspired, going to be productive tomorrow. I want to proritize shifting my productivity equilibrium in a [sustainable way](https://mindingourway.com/productivity-through-self-loyalty/).

Also rationality. There's [so](https://www.lesswrong.com/s/KGYLvTqFiFE2CpHfJ/p/eAsN5vNjvmxzACuuX) [much](https://www.lesswrong.com/posts/2GxhAyn9aHqukap2S/looking-back-on-my-alignment-phd) left to do!

Yesterday's goals/todos:

- [x] Do the things that make me more focused (e.g. no music with lyrics, don't eat unless hungry, take breaks if tired, don't take breaks if I need to get into the flow.)
- [x] ~~3h~~ 2h SERI Mats Readings
- [ ] ~~4h~~ ~~2h (got up late) Long-term investing in [gears level models](https://www.alignmentforum.org/posts/nEBbw2Bc2CnN2RMxy/gears-level-models-are-capital-investments) (I want to do this every day actually...)~~ Basically procrastinated via minecraft, then paniced at how bad my seri mats thing was so did that the rest of the day.
- [ ] ~~2h Econ~~
- [ ] ~~2h Causal models~~
- [ ] ~~1h SERI Mats Meeting~~ cancelled
- [x] ~~2h~~ 3.5h Finish up SERI MATS training story & experiments. Send to Quintin & Co (AAAA it's so bad, my outline is better though. will finish tmr)

Setup rescuetime before going to bed. Hopefully it beats my cursed toggl setup.

### 2022-11-19

Summary: ~14h work (8.5hrs SERI mats + 5h Atlas Academy + 0.5h misc). Stayed up late (1am rn), oh well.

[Writing fast while maintaining epistemic rigor](https://www.alignmentforum.org/posts/Psr9tnQFuEXiuqGcR/how-to-write-quickly-while-maintaining-epistemic-rigor) is quite good, this might be the first time I partially did it. Writing why you believe something rather then inventing Defensible Reasons is scary, but boy is it faster.

I'm basically trolling in convex optim. I spent a few minutes thinking about the variance of a dirac delta distribution LMAO. need to figure out how to (a) not be tired in that timeframe (b) find time to study outside class, since that's one of the harder classes I'm taking.

Goals, with must-do in bold:
- [ ] ~~Meditate 30min in the morning~~ (Called bank about something 8:30-9 instead)
- [x] Study Microecon before class
- [ ] ~~Study convex optim before class~~ (seri instead)
- [x] Atlas Academy meetings
- [x] ~~3-5hrs~~ 8.5h **Finish SERI thing**

### 2022-11-20

Mostly a rest day, worked ~4hrs (1h mats meeting + 3h atlas academy). Wasted ~2.5hrs watching youtube

Spent the morning AI-pilling my brothers via ChatGPT. my little brother is prompt engineering DALLE now!

Had seri mats shard meeting, Alex thought my research was not terrible so that's good!

Spent 3h with a friend who's in town, lots of fun

Got home, did Atlas academy's writing class. I'm quite bad at writing engaging stuff, need to practice. Many skills of mine are

Did Atlas academy's algorithms class, implemented all the sorting algorithms in python (inefficient, but readable solutions).

Wasted 2.5hrs watching youtube, I'm a little upset at this. I'd like to become the kind of person who reads educational books as a default rest activity. I think I'll perma ban youtube for non lectures. Also try harder to do all my habitica things.

> “There are no limits. There are only plateaus, and you must not stay there, you must go beyond them.” - Bruce Lee

There's a huge amount free value in "changing the kind of person I am" via habit-things. Changes-to-me I'd like to make:

- Always read for entertainment / leisure (mostly nonfic)
- Meditate in off-moments in the day. small hits of practice
- Also integrate rationality practice into random moments. Practice continuously.
- *Always* do all my habits. And engineer the habits such that [they have no exceptions](https://www.lesswrong.com/posts/duxy4Hby5qMsv42i8/the-real-rules-have-no-exceptions).

What are good ways to implement this... More public tracking! I need a giga uli-tracking-page with journal entries, habit squares with colors, graphs of metrics... I need to set this up, should be pretty easy; either use [flowdash](https://flowdash.co/)'s API or a markdown files -> parser -> web UI (probably better, I like having control over stuff.)

### 2022-11-21

Summary:

Goals: (Not that many bc I might visit a friend. Conditional on not visiting them add + 3-5h learning time)
- [ ] **Do all the habits** (Read a book, no youtube, meditate at least 6 times a day)
- [ ] ~~Setup uli.rocks/now to show graphs of stats, journals, etc. (~1-2h). Bonus points for figuring out a system other people can use, compeition!~~ Social things!
- [ ] ~~Study/Review econ (1-2h)~~ Social things!
- [ ] ~~Microecon class~~ cancelled

### 2022-11-27

Haven't stopped journaling, just setup an airtable system instead of this. Will post link here when I get the UI up and running.