Bits and Bobs 12/22/25

1An insightful HackerNews comment about code generation:
  • An insightful HackerNews comment about code generation:
    • "AI makes starting easier but finishing harder.[dr][ds]
    • You get to 80% fast[dt], then spend longer on the last 20% than you would have building from scratch - because now you're debugging code you don't fully understand."
    • "The gap isn't in generating code - it's in everything else that makes software actually work."[du]
2I feel like the mama bird feeding my little Claude Codes.
  • I feel like the mama bird feeding my little Claude Codes.
    • They're like little chirping birds, begging to be fed.
    • The more I invest in them, the more they produce useful things.
    • There's always at least one chirping.
    • By the time you feed one, there's another already ready to be fed.
    • It's hard to rip yourself away![dv][dw]
3LLMs are great at things that are expensive to generate but easy for a human to judge.
  • LLMs are great at things that are expensive to generate but easy for a human to judge.
    • They're not good at things that are easy to generate but hard for a human to judge.
    • The most important part is: can the human judge the quality of it.
4Agents changing themselves each loop is what creates the compounding benefit.
  • Agents changing themselves each loop is what creates the compounding benefit.
    • If it doesn't change any state in the loop, then it could get stuck in an infinite loop.
    • The accumulation of useful state gives the compounding benefit.
    • It adds useful state in one time step, and that state gives benefit in all future time steps.
      • That's the primary multiplicative point that gives compounding.
    • A conversation context is a cheap way of doing this, but it's ephemeral and can be filled up.
      • A conversation context is really just a hack of a general pattern.
5Back in the day, Google used the money they made to fund the future.
  • Back in the day, Google used the money they made to fund the future.
    • OpenAI is instead using the money they're going to make to fund the now.
    • Significantly different proposition.
    • Everybody who discovers leverage thinks they're a genius until they slam into a wall.
6A company focused on extracting value from attention is not the way to build the best model.
  • A company focused on extracting value from attention is not the way to build the best model.
    • There are lots of distracting hollow tricks you could do to game the metrics.
7Everyone thinks their own slop smells sweet.
  • Everyone thinks their own slop smells sweet.
    • Slop is like a spreadsheet.
      • You love yours but everyone else hates it.[dx]
      • If you went through the process to generate it, you get the problem it solves, what the insights are.[dy]
      • But everyone just sees a messy, confusing thing.
    • George Carlin bit: "Your stuff is stuff, but everyone else's stuff is shit."
8The Less Wrong community has an emoji reaction: "smells like LLM".
  • The Less Wrong community has an emoji reaction: "smells like LLM". [dz]
    • It's an emoji of magic dust and a nose.
    • The ultimate burn.
    • Even if it was written by a human, it's as bland and limp as if an LLM wrote it.
9What matters if not if AI wrote it, it's if you stake your reputation on it.
  • What matters if not if AI wrote it, it's if you stake your reputation on it.
    • If you stand by it, then it doesn't matter who wrote it, it's your reputation on the line.
    • Whereas if it's just slop you serve up without looking, then it's not worth other people's time either.
    • If your slop isn't worth your time, it's definitely not worth anyone else's.
10There's something about a thing you think is resonant and then seeing a tell that it's hollow.
  • There's something about a thing you think is resonant and then seeing a tell that it's hollow.
    • It feels like a betrayal.
    • Even worse than if you hadn't been tricked into it being resonant in the first place.
11The New York Times points out that it's weird that Chatbots use "I".
12If a process is too smooth, your brain turns off.
  • If a process is too smooth, your brain turns off.
    • Grittiness is required for our brains to be engaged.
      • For us to learn and grow.
    • Modern society optimizes for smoothness[eb].
      • But smoothness is hollow.
    • Resonance requires grit, texture, challenge.
13This week I created an origin story book for my kids.
  • This week I created an origin story book for my kids.
    • I tell them the same story of how they came to be every month.
      • At no point will any of it be a surprise to them.
      • Over time their understanding of the details will grow.
    • I decided to make it into a proper book.
    • I just created an empty git repo, launched Claude Code, and then talked to it off and on for days, accumulating a workflow, tools, content, generated images, etc.
      • It just kept on accumulating tools and state.
    • Claude Code is insanely powerful.
    • If you aren't using it every day, you don't realize what the future will bring.
14Agents are presented as if they were oracles.
  • Agents are presented as if they were oracles.
    • But of course they're actually LLM calls.
    • The context window that it keeps appending to is what gives it a coherent throughline of agency.
    • We can all see that context appending is not the final answer.
      • You run out of space for context and then things get wonky.
      • Sub-agents and context management are attempts to fix this.
    • If you follow this sub-agent extraction to its logical conclusion, you're left with lots of little API calls.
      • Each with a different context window.
      • The entity-ness of the agent disappears.
      • It becomes instead a collective.
      • A swarm of little automatons swarming on your data.
      • This feels less like an append-only single conversation and more like a growing graph of accumulated intention.
15Google's A2UI is a Chatbot putting on a puppet show.
  • Google's A2UI is a Chatbot putting on a puppet show.
    • The software itself doesn't feel alive.
16The first microprocessors went into calculators.
  • The first microprocessors went into calculators.
    • Calculators were a known thing, a use case that was obvious.
    • A general purpose computer could only emerge later.
    • When you're on the edge of tech that will become ubiquitous, it seems crazy, unknowable.
    • Afterwards, it feels obvious, like it's impossible to unsee.
    • The first LLMs are like word calculators.
    • We haven't come up with word computers.[ec]
17LLMs make weird judgment calls sometimes.
  • LLMs make weird judgment calls sometimes.
    • Big labs fix it by fine-tuning.
      • This takes tons of capital and time… and can only be done by the labs with the model in the first place.[ed][ee]
    • But a system that can just distill usage from real users, automatically, could make a much faster correction loop.
      • That could help emegently improve it for a given task.
18Vibecoded software is currently a toy.
  • Vibecoded software is currently a toy.
    • If vibecoded software could somehow be made trustworthy, it would no longer be a toy.
    • It would unlock the latent potential of code.
19Stratechery: "it seems likely that LLMs have already generated more text than all of humanity ever has; the reason I cannot know is the same reason why it hasn't left a mark: AI content is generated individually for each individual user of AI."
  • Stratechery: "it seems likely that LLMs have already generated more text than all of humanity ever has; the reason I cannot know is the same reason why it hasn't left a mark: AI content is generated individually for each individual user of AI."
    • What if the useful generated content could be cached and shared?[fe]
20A key question about a system: does it work without LLMs?
  • A key question about a system: does it work without LLMs?
    • Are LLMs just lubrication for the system?
    • Or are they the fundamental enabler?
    • LLMs are expensive and slow.
    • Systems that use them as lubrication can improve faster than the LLM.
    • Systems that can't work without the LLM are trapped by the LLM's performance.
21If you need zero-shot app software to work, then quality matters a lot.
  • If you need zero-shot app software to work, then quality matters a lot.
    • Especially when the software has to be chonky, like an app.
    • But if you had a system that worked reliably no matter the quality of the LLM, then you wouldn't have to care how good the LLM provider was.
    • The baseline floor of quality would be good enough.
      • Especially if the quality of the software grew combinatorially on top of that floor.
22People connect authentically with Barnum statements and horoscopes.
  • People connect authentically with Barnum statements and horoscopes.
    • Of course they will connect, in a dangerously deep way, with a human-presenting thing trained on all the world's writing.
23Apps only make sense in a world where code is expensive.
  • Apps only make sense in a world where code is expensive.
24Phones say "don't think about annoying things like file systems, just use the app!"
  • Phones say "don't think about annoying things like file systems, just use the app!"
    • Apps don't have a filesystem.
    • An app is a monkey trap.
    • You put data in, and you never get it back out.
    • Short-term valuable, long-term trap.
25I want to grow my own personal "operating system."
  • I want to grow my own personal "operating system."[ff]
26Humans are amazing tool users.
  • Humans are amazing tool users.
    • Compute is the most powerful tool we've ever used.
    • We're using just a small portion of its potential.
27LLMs commoditizing software could have a huge impact on society.
  • LLMs commoditizing software could have a huge impact on society.
    • Disaggregating apps, melting them away into infinite software.
    • AI should unlock the potential of software.
    • Today the only software that exists is software that has a business model in the app distribution paradigm.
    • The software that should exist is any software that people find useful.
28Steve Jobs said that software should be a bicycle for the mind.
  • Steve Jobs said that software should be a bicycle for the mind.
    • The Resonant Computing manifesto shows how far from that we've gotten.
    • Instead, we've accidentally built a self-driving bus with locked doors.
    • LLMs create the possibility for an electric bicycle for the mind.
29A tweet: "made a tamagotchi you keep alive by going offline."
  • A tweet: "made a tamagotchi you keep alive by going offline."
30OKCupid back in the day was resonant.
  • OKCupid back in the day was resonant.
    • Tinder today is hollow.
31"Adaptable" in the resonant computing manifesto means "not a walled garden."
  • "Adaptable" in the resonant computing manifesto means "not a walled garden."
32I think AI will be as disruptive as electricity.
  • I think AI will be as disruptive as electricity.
    • Which is to say, massively disruptive… but not society ending.
    • I think that most AGI projections are mostly kayfabe.
33You don't want to have to care about your electricity provider.
  • You don't want to have to care about your electricity provider.
    • Or your ISP.
    • But they want you to care about them.
      • If you care about them, they have more pricing power.
    • But you don't have to!
    • For commodities that just work behind the scenes, you just want them to be dumb providers.
34Anyone who's selling you a chatbot wants to sell you an assistant.
  • Anyone who's selling you a chatbot wants to sell you an assistant.
    • An assistant that they control.
    • Instead, you want a tool.
    • You should hire an AI computer.
35AI is just too powerful for you to be the "product."
  • AI is just too powerful for you to be the "product."
    • If AI is so powerful, then it has to be in your interests.
    • If the chatbot had to choose: your interests, or its corporate creators', which would it choose?
    • The problem with ads is not "you get stuff for free" it's "the incentives of the system are for the corporation"
36Cookies are these mysterious things that allow tech companies to manipulate and extract from you.
  • Cookies are these mysterious things that allow tech companies to manipulate and extract from you.
    • What would "reverse cookies" be?
    • They'd protect your agency and make the software work for you.
37A big threshold in a system: Not "I can do it" but "I'm not not going to do it."
  • A big threshold in a system: Not "I can do it" but "I'm not not going to do it."
    • That's the difference between demo and usable.
    • When you cross that threshold, doing things inside the system is easier and better than doing it outside.
    • A positive boundary gradient that makes people want to use it.
38Some problems are shaped as a PhD Ramp.
  • Some problems are shaped as a PhD Ramp.
    • They're easy to get started with a good enough answer.
      • The bar is not too high.
    • To improve quality incrementally just takes throwing more PhDs at it.
      • It's not easy or cheap, but it is low risk.
    • The more progress you make, the more of a moat you leave behind you for others to have to traverse to catch you.
    • Search quality problems have this shape.
39The meta-game can shift rapidly even from small changes.
  • The meta-game can shift rapidly even from small changes.
    • Often a small change happens that changes the meta-game, and no one realizes.
    • At the beginning a few savvy players realize and exploit it, before anyone else realizes.
    • For example, last year, Twitter changed it so you can't see who liked a tweet.
      • They did it for… reasons. But it also changed the meta-game.
    • That means that it's easier than ever before for someone to make bots to create the appearance of momentum.
      • Especially in the age of LLMs.
    • Keep that in mind when you see what looks like momentum.
40A bouquet of metrics is more likely to produce a resonant outcome.
  • A bouquet of metrics is more likely to produce a resonant outcome.
    • A single metric, maximized, must hollow the thing out, even if it's "perfect."
    • But a bouquet of diverse, pretty-good metrics collectively helps avoid that outcome.
41Slop is about quantity over quality.
  • Slop is about quantity over quality.
42Resonance is "something that strengthens an existing frequency by being aligned with it"
  • Resonance is "something that strengthens an existing frequency by being aligned with it"
    • This comes from Rob Zinn in a comment on the Resonant Computing Manifesto theses.
43Some things are "sweet like antifreeze".
  • Some things are "sweet like antifreeze".
    • Not even sweet in a hollow way, sweet in a dangerous way.
    • Sweetening candy with antifreeze is the end result of hollowness.
44Classic line: "They were so preoccupied with if they could, they never stopped to think if they should."
  • Classic line: "They were so preoccupied with if they could, they never stopped to think if they should."
    • Seems to describe the current tech industry!
45Some actions help in the short term but harm in the long term.
  • Some actions help in the short term but harm in the long term.
    • As you "optimize" you select more of those short-term benefits.
    • Over time you get in a long-term hole.
    • You moved fast, you solved problems… and you got yourself stuck in a deep hole.
46Sarumans optimize to the point of hollowness.
  • Sarumans optimize to the point of hollowness.
47Often more is more.
  • Often more is more.
    • If you can easily jump between different good enough options and have time.
    • But sometimes more is less.
    • If you need to focus and execute to get to viability.
48Andrew Kokoev has a fascinating paper on High Trust Autonomous Systems.
49This week I learned the frame from Data Science : Data > Information > Knowledge > Wisdom.
  • This week I learned the frame from Data Science : Data > Information > Knowledge > Wisdom.
    • Ultimately what matters is the last stage, everything else is just a means.
50In my Obsidian workflow, I have a daily note.
  • In my Obsidian workflow, I have a daily note.
    • I use that note a lot. Ctrl-T takes me to it (creating it if necessary).
    • I also have key commands to go to the daily note from the day before.
      • It's actually a simple little Obsidian plugin I made.
    • The daily note is exceptionally important on the day of.
    • It has an exponential drop off in importance after that.
51The CIA had a guidebook to disrupt organizations.
  • The CIA had a guidebook to disrupt organizations.
    • Things like insisting every minor rule be followed to the letter.
    • Some organizations are actually improved by running more slowly.
    • If you have a lot to lose and not a lot to gain, being slowed down helps you stay on the same trajectory.
    • Arguably American democracy has done very well for a long time by being slowed down.
52There are lots of ways to slow down discussions in organizations.
  • There are lots of ways to slow down discussions in organizations.
    • "Have you thought about the environmental implications?" for example, is a way to derail a discussion.
    • In corporate dystopian environments everyone is satisficing, instead of maximizing.
    • Decisions can't be made in the consensus final review, but they can be unmade.
53If you get really good at 2D chess then you'll be in a worse situation in 3D chess.
  • If you get really good at 2D chess then you'll be in a worse situation in 3D chess.
    • If someone makes a move in 3D, they'll route around you in a way that feels impossible.
    • As you get even better at 2D chess, you'll get more confident, but there's an invisible move that will destroy you.
    • Optimization in one dimension has made you fragile in another.
    • There's always a higher dimension than the one you sense.
54Purist communities are often self-marginalizing.
  • Purist communities are often self-marginalizing.
    • They choose purity over inclusion.
    • They push people out who aren't as purist as them.
    • That makes what's left more extreme and less likely for others to want to join.
55Someone living in the future is inscrutable in the present.
  • Someone living in the future is inscrutable in the present.
    • Someone who lives in the future, works back to the present.
    • But the world works from present forward.
    • If you're unstuck in time, you can see around corners no one else can see.
    • But that also makes you inscrutable to work with.
    • Every so often you're seeing around corners that no one else can see… because they don't exist.
56Animals just rely on pretraining.
  • Animals just rely on pretraining.
    • Whatever the evolutionary environment imprinted in them as stored routines.
    • Humans can adapt themselves.
    • Evolution popping up into a new pace layer and then taking over.
57In every paradigm, evolution pops up a pace layer and then takes over that layer.
  • In every paradigm, evolution pops up a pace layer and then takes over that layer.
58Entropy is not disorder, it's complexity.
  • Entropy is not disorder, it's complexity.
    • Harder to compress.
    • A system that is over normalized is hard to get novelty out of.
    • Disorder is the input to creativity.
59Actions that improve the worst case scenario have compounding value.
  • Actions that improve the worst case scenario have compounding value.
    • You lock in a new worst case once, and then now all future instances have it for free: a compounding term.
    • If you can then make it crowdsourced (lock in the worst than anyone ran into for everyone) then it has an extra compounding term.
      • It's not individual quality, it's community quality.
      • The actions of anyone in the community improves results for everyone in the community.
      • A faster ratchet.
60A common pattern: you get 80% of the way done…
  • A common pattern: you get 80% of the way done…
    • … and realize you have another 80% to go.
    • This curve shows up because of the logarithmic quality curve.
    • The effort per unit polish asymptotically approaches 100%.
61Boundaries are necessary for differentiation.
  • Boundaries are necessary for differentiation.
    • Without a boundary, the differences diffuse.
    • A boundary has two dimensions:
      • 1) What is the pressure gradient?
        • Do people want to join or leave the inner thing?
      • 2) What is the valence of the boundary?
        • How do people feel about people on the other side?
    • The latter is what makes a boundary get entrenched, making it so people close to the boundary push away from it.
      • Leaving a charged no-mans land, in proportion to how valent it already is.
62Feynman: If you think you understand… you don't.
  • Feynman: If you think you understand… you don't.
    • He said about quantum physics, but it applies to anything complex.
63Three responses: "I like, I wish, I wonder.
  • Three responses: "I like, I wish, I wonder. "
    • This is a frame from Stanford Design School.
    • Ways of giving confirming and disconfirming feedback, and also an open-ended thread to pull on in discussions.
64If your life is too easy then your capacity atrophies.
  • If your life is too easy then your capacity atrophies.
    • It's important to exercise.
      • Your body.
      • Your heart.
      • Your brain.
65Decision is what collapses the wave function of possibility.
  • Decision is what collapses the wave function of possibility.
    • The person making that decision with intention is what creates learning and makes things happen in the world.
    • When you give someone steps to execute without thinking they don't need to decide, they can't learn.
      • They aren't learning how to do it themselves, they can only accomplish it with the external precise guidance.
      • An automaton.
      • Capable of accomplishing it but in a hollow way.
      • Fragile.
    • It is when we are making decisions that we are capable of learning.
      • When we are capable of learning from experience.
      • Being in the arena, in the loop.
66If you go too fast you create hollowness, not resonance
  • If you go too fast you create hollowness, not resonance
    • You create just superficial progress.
    • Underneath is turbulent flow, not laminar flow.
    • In turbulent flow nothing of value coheres.
67David Graeber: What's the Point If We Can't Have Fun?
68When you're doing something you have to, each challenge leads to resentment.
  • When you're doing something you have to, each challenge leads to resentment.
    • That's not true when you get to.
    • Then each challenge is something that makes it more rewarding.
69When you're mad your body looks for more confirming evidence of why you're right to be mad.
  • When you're mad your body looks for more confirming evidence of why you're right to be mad.
    • And why you should be even more mad than you already are.
    • It's a toxic spiral.
70Communities that look dead are dead.
  • Communities that look dead are dead.
71Getting a little traction is the worst.
  • Getting a little traction is the worst.
    • You have to support it–if you don't you're letting down your users.
    • But it's not going anywhere.
72Don't build a better mousetrap.
  • Don't build a better mousetrap.
    • Build a system that can build continuously better mousetraps.
    • That's the meta-game.
73Hollowed out things are about unthinking execution.
  • Hollowed out things are about unthinking execution.
    • "Just turn the crank faster."
    • LLMs applied naively just turn the crank faster.
    • It's even more important than ever before that the crank is attached to something resonant!
74Statistics are compression.
  • Statistics are compression.
    • They must fundamentally be.
    • If you compress them in bad faith to hide important things they become lies.
    • It's all about does the compression align with the truth or is it meant to deceive?
    • Is the compression resonant?
    • Does it obscure, or does it clarify?
    • If someone were to look at your compressed version and the non-compressed version, would they feel betrayed?
    • It's trivial to lie with statistics when your intentions are impure.
75Centralization + LLMs + Authoritarianism will be an explosive combination.
  • Centralization + LLMs + Authoritarianism will be an explosive combination.
    • LLMs aren't going anywhere.
    • The push towards authoritarianism might be hard to reverse in the short term.
    • So we need to push to fix the hyper centralization.
76Catastrophically bad actors can sometimes be the accidental savior of a system.
  • Catastrophically bad actors can sometimes be the accidental savior of a system.
    • If the system has become so hollow and rotted out, it creates a condition where a cynical bad actor can take over.
    • They kick off what is experienced like a forest fire by the rest of the system.
    • The forest fire clears out the rot.
    • It also shows people why bad actors like that are a bad idea.
    • The system can now be saved and rebuilt when the forest fire is put out.
77Zombies make more zombies.
  • Zombies make more zombies.
    • A toxic spiral.
    • They tear down things around them to also make it hollow.
78Two characters that I think had an outsize impact in putting modern society on the trajectory towards hollowness.
  • Two characters that I think had an outsize impact in putting modern society on the trajectory towards hollowness.
    • Jack Welch and Newt Gingrich.
      • "The Man Who Broke Capitalism" and "The Man Who Broke Politics."
    • Both put in motion plans that focused on optimizing for whatever was good for them, with no thought to the significant externalities.
79A new paper: "The more unequal income distribution is in a democracy, the more at risk it is of electing a power-aggrandizing and norm-shredding head of government."
  • A new paper: "The more unequal income distribution is in a democracy, the more at risk it is of electing a power-aggrandizing and norm-shredding head of government."
80A new word only sticks if people choose to use it.
  • A new word only sticks if people choose to use it.
81When someone believes in you, your magic gets stronger.
  • When someone believes in you, your magic gets stronger.
82There are many ugly systems with beautiful emergent properties.
  • There are many ugly systems with beautiful emergent properties.
    • And vice versa.
83Emergence and optimization are in tension.
  • Emergence and optimization are in tension.
    • Emergence comes from the space between.
    • Optimization removes the space between.
84Posters seen in my daughter's elementary school:
  • Posters seen in my daughter's elementary school:
    • "I can't do it... yet."
    • "It's not failure because I haven't given up yet."
    • "Mistakes are expected and respected."
    • "FAIL: First Attempt In Learning."