Jonathan Otto

Beast Mode Activated

Optimize code, SQL, or writing with LLMs: for { for llm = range [o3, opus4] { code = optimize(code, state, llm) score = evaluate(code) state = append(state, {code, score}) } sort(state) if (state[0].score < N) { break } }

"Real coding" means constant refactoring of abstractions to reduce complexity and increase capability The lesson of AI agents is what happens when you atrophy your mental model of a codebase This same lesson applies to note taking

@korevec Yes, and as the AI Agent power grows it all becomes insignificant. But if "cost" still matters... maybe at scale, think quadrillions of apps/services - maybe it's more efficient to build layers on stable interfaces (versioned APIs, SQL, etc.). MCP works like this

@jotto won't a DevOps agent be the inevitable conclusion? e.g. deploy this code on whichever platform makes the most sense and keep it running.

SaaS, microservice, SQL, or API is the ideal abstraction FaaS requires your code to work with latest version of underlying language (Node.js v20, etc.) PaaS get deprecated (Heroku) Cloud is complicated (AWS, GCP) Self hosting requires expertise/maintenance

Coding music youtube.com/watch?v=Giot9j…

Programmer velocity is mentalCapacity / sizeOfCodebase(s) and 2 levers: - better arch/abstractions (that the team has mentally cached) - less meetings/distractions so the devs can "cache" more code higher level languages and LLMs are same as arch/abstractions

3 types of thinkers when coding: 1. no mental "cache" of codebase/arch 2. partial cache 3. full codebase + arch cached mentally AI will continue to shift programmers from level 3 towards level 1. Level 3 thinkers who refuse to migrate will merge w/AI.

Momentum for a coder is life or death. Too many days off, the code leaves your brain and you return to type-2 thinking to re-familiarize. As long as the code is "cached" in your mind, you can achieve type-1 flow state with virtually any task. Momentum = 3x output

The most powerful hack for computer programming is delegating cognitive load to the subconscious mind. Quickly work through a draft of your task from beg to end. Then walk away. Let your "diffuse" thinking do the hard work in between tasks, on walks, or sleeping. easy 3x

Many people are inefficient with their computer interaction No keyboard shortcuts, alt+tabbing, slow apps, no workflow, no macro This is an easy unlock for LLMs and NixOS. Perfect one-off OS's can be configured per persona per workflow +1% GDP easy

@korevec amen brotha

@jotto 6 spuds. $1.08

$12/day coder's diet: 2 large mcdonalds coffee = $4 carryout, large dominos pizza = $8

s-tier: mcdonalds, buc-ees, chick-fil-a a: culvers, in-n-out, mixt greens b: jimmy john's, shake shack, fiveguys, chipotle c: waffle house, steak 'n shake d: starbucks, wendy's, burger king, subway

Buy a computer with 64GB of ram or more. It's not expensive (unless Apple) and makes a huge difference. If you're using a consumer-grade desktop you can go up to 192GB. Asymmetric returns.

my inner monologue when rebooting a NixOS system using impermanence and ZFS boot.initrd.postDeviceCommands = lib.mkAfter '' zfs rollback -r rpool/local/root@empty '';

vscode is the chrome browser of the IDE world. it's "good enough" but LLMs obviate big IDEs true flow only comes from deep muscle memory with vim/emacs workflows and terminal multiplexers and/or tiling wms AND feedback the s-tier dev ux will be led by neovim, zed, or cursor

dreams, movies, deep thinking, etc. is akin to fine-tuning the human brain with synthetic data - i.e. traversing/permuting through a graph multiple levels deep next step is delegating this to LLMs - instant auto-complete for your brain

sf bay area is the most beautiful and high potential city in the USA - "there is no second best" twitter.com/rauchg/status/…

modern human toil abstracts to: triage data input business workflow analytics all of which can be eliminated by LLMs