Bits and Bobs 5/18/26
Alex Komoroske
Software’s adjacent possible has gotten 10x bigger.
Coding has gotten 10x cheaper.
Everyone is defaulting to building the same stuff as before, but 10x faster.
What about things that are 10x deeper?
Things that would have taken unreasonable amounts of time before?
There are some things that would have been impossible to even 10x the resources to get.
People, famously, have coordination challenges that scale super-linearly.
A whole new adjacent possible has opened up in software and no one seems to have noticed yet.
The companies who realized what was now viable who play in that new field of possibility will change the world.
Using LLMs to make today’s apps faster is like making faster horses.
What are the software experiences that were impossible before?
Claude Code shows the explosive potential of AI.
It’s hugely difficult for non-technical people to get started.
It’s intimidating!
You could blow your foot off!
How can you let normal people lean in with AI?
So it’s not just the techies who are benefiting from AI, but everyone?
OpenClaw Shows AI Agents Don't Need to Be Vertically Integrated.
Breadth is more important than depth in agents.
Having all of your data is more important than having deeply customized software in each vertical.
This week someone showed me a clone of iMessage they had made on their Mac.
Their workflow required a keyboard shortcut that the native app doesn’t have.
So… they just vibecoded a replacement.
They were able to, because the data lives on their machine.
If you have access to your data, you can make a custom bit of software today with just a bit of mojo.
That wasn’t viable before, even for highly productive engineers
We are trapped in someone else’s software.
Before, when it was hard to build software, the software was a scaffolding we could gain strength from.
What else could you do, anyway?
It was too expensive to make your own software.
Now, it’s increasingly a cage that holds us back.
It tells us what we can and cannot do.
A monkeytrap.
Now using someone else’s software feels like a liability.
“What happens if I want to add a feature it doesn’t have?”
Now that software is cheap, it’s possible to make software for yourself that is better for you than anything that another person could have made for you.
Using someone else’s software is holding you back.
Using software that you can’t adapt feels like living in the stone age.
Software moats are downstream of the time and cost to create software.
If a company found PMF with something valuable, then even competitors trying to copy them would take 6 months to successfully copy.
During that time, the compounding effect of data stickiness would get going.
By the time the competitors got there, they’d have a significant strategic head start.
A data compounding effect can get going in 6 months, but not in 6 days.
"Oh I like that feature from that other product. [Two hours pass] OK, now it's in my thing!"
Software that used to take 6 months to build now takes 6 days to build.
Multiple orders of magnitude reduction.
The industry hasn’t caught up to the implications of this.
It undermines the foundations of value creation the industry has assumed for decades.
The YC playbook is about a sprint to get a minor moat and get the compounding going before a competitor noticed.
But LLMs make it trivial for people to copy.
As a user, why get stuck in someone else’s monkey trap when you can just make your own in an afternoon?
Where you can actually add the features you want?
“Cadillac Tasks” have now gotten way cheaper in software projects.
These are expensive nice-to-haves, like great docs, or a nice landing page, or internationalization.
Those used to be hard-to-fake signals of the quality of a software project.
Now, LLMs can do those tasks orders of magnitude cheaper, so it’s less of a strong signal of project quality than ever before.
Arguably, an inverse signal, one of codeslop, where human judgment was only superficially applied.
The same way that typos in writing are now a signal of humanness, the lack of Cadillac features in software might show that it’s full of human judgment.
We grow software now.
That was always true, even before LLMs.
But now it’s obviously, inescapably true.
Text is an infinitely flexible UI, but it takes time to consume.
It used to be that software couldn’t produce text, so UIs had spatial metaphors.
These happened to work well with how our brains work, allowing us to absorb a lot more all at once.
But LLMs made text interfaces possible.
Text is infinitely flexible, but it takes time to scan and consume.
The dreaded “wall of text.”
We’re enamored of text-primary interfaces today because it’s a new party trick and it’s flexible, but the real value will be generative UIs.
The labs are all betting on the models solving all problems.
They’ll get infinitely good at avoiding prompt injection.
There will be no need to share code from anyone else.
It will just be the model, you, and your data.
This assumes the models will get very good at figuring out how to give you exactly the software you need, at the moment you need them, without the user needing to act like a PM.
Models are also a slow pace layer since they take so long to train and adapt.
They are also a closed set.
Another approach: use the LLMs as dumb muscle, not the brains.
The brains would still come from real users with situated judgement.
If you could have the human judgment at a faster pace layer, and accumulate insights from real situated users that could help others, open-ended value could be unlocked.
Notion launched a significant new developer platform in a bid to become a system of record for agents.
The idea is that companies already have a lot of data there, and it’s a naturally collaborative surface.
But the features that made Notion useful in the first place are now trivial to vibecode in an afternoon.
Notion is also ludicrously expensive.
There’s no reason for anyone who doesn’t already use Notion to start using it.
Google will launch a proactive agent at I/O this week.
Of course, it won’t be private.
It will be what an advertising company thinks is best for you.
That must,fundamentally, be a conflict of interest.
But I imagine the bigger problem is that it won’t be open-ended.
It will be a handful of top-down features designed by PMs to be generally useful.
Google always has the tension between hyper-bespoke features that are actually useful to people… or features that are broad but not deep.
One-size-fits-none features that can theoretically be used by 10’s of millions of users.
Another way is to grow the features bottom up, where the situated actions of savvy users accumulates value that can then help other people, too.
Fractically vertical use cases.
For AI’s usefulness, the relevance and quality of the context is way more important than the quality of the model.
The models are all extraordinarily capable, so the difference in outcome quality is determined more by another factor.
Similar to how in professional basketball games, performance is driven more by luck precisely because teams are so competitive and capable.
Your exocortex is an extension of your mind.
Cognition and memory in a human form that is outside of you.
An augmented consciousness.
If someone else were to use your exocortex it would feel unnerving.
Similar to using someone else’s bash shell.
Since your exocortex is an augmentation of your mind, it’s imperative that no other entity can take it away from you.
It must be yours.
The most important skill in AI: metacognition.
The ability to change the way you do things by experimenting and learning.
It’s not possible to do a mass-market tutorial for AI because we don’t know the best practices yet.
They keep changing as the models improve and we figure out new ways of using them.
The rate of surprise even among experts is too high.
If only you had 10x cheaper cognitive labor, it would be great.
You’d be able to pull on dozens of threads while everyone else was just ramping up.
To everyone else you’d look like you were pretenaturally intelligent, but really you could just explore more options than everyone else.
But now everyone has access to the same much cheaper cognitive labor.
It has kicked all of us knowledge workers into a much faster pace layer.
One that is more inhuman than the previous layers.
Once we no longer have to supervise our agents, it will be discontinuously freeing.
That would require a much higher quality bar.
But more importantly, it would require a much better worst case scenario.
A structural cage that makes it hard for the agents to harm us naively or maliciously.
Nate Jones has made the point that where agents make mistakes are more valuable to quality improvements than when they get it right.
Those mistakes, where the human has to correct the output, are the surprisal gradient.
Every bit of guidance from a human is a valuable jewel of input that should be kept to make the system improve and become default-converging.
Every user problem that could be solved within silos has been solved.
The frontier is problems that inherently span data silos.
It’s not the friction of long-tail use cases, per se.
Also, most users don’t want to become a creator.
Even if users know what doesn’t work for them that doesn’t mean they can specify a solution.
They just want it to work for them.
A tool that spits out a static app won’t work for users.
You want something malleable and 80% good enough that users can then intuitively shape to their needs.
In a world of infinite software, will software creators be incentivized to make it for other users?
Will savvy users creating software for themselves (and having it be shared with other users be a bonus) be enough?
Or will there need to be a direct incentive for creators still, even if significantly diminished from classic software?
In a world of infinite software, how chonky will software be?
Will it still be roughly the size of apps?
Or will it be significantly more modular?
The best way to learn things is to teach it.
Normally to teach you have to have a willing subject to learn.
They also have to be willing to potentially be taught wrong by you.
Like people going for free haircuts at a beauty school.
But bots are willing and able to play the role of curious pupil, never getting tired.
And you can’t hurt them by teaching them wrong; they have no memory and aren’t actually situated in the world.
If others think it is slop it is slop.
Just because you made it yourself with AI does not mean it’s not slop.
People aren't necessarily going to build their own software.
You can cook in your kitchen... but people still prefer to get delivery or go to a restaurant when they can afford it.
You can pay for convenience.
Someone who thinks like a PM should help your software get good, automatically.
What if you had your own personal PM building software just for you?
Things like ChatGPT Pulse can’t give good proactive recommendations unless they have a better sense of what is happening in your life.
A friend said it recommended he work on a pitch… that he had already completed and presented three months ago.
Chat apps don’t have enough closed loop context about what matters to you.
They only know the slice of interactions you’ve had in that chat.
Now that intelligence is abundant, we can let it organize itself.
Don’t require the user to come up with the scaffolding for their data to fit within.
Agents and humans combine into a tick-tock mechanism of a clock.
The agent is the indefatigable source of energy: the spring.
If the spring unspools in a second, it is explosive and useless.
The human is the escapement: the arm that swings back and forth, allowing a little bit of energy from the spring to unspool.
Tick-tocking back and forth.
Each time the agent waits for a judgment call from the user, it’s caught on the escapement.
Together, the energy of the LLM and the judgment of the human produce something special.
Companies’ CRMs are default-diverging.
A lot of different teams all put various incoherent junk into it.
For example, 5 different records for what probably should be the same company.
But it’s also not clear: should the sub-unit of that sprawling multi-national company be recorded as a separate entity or not?
As more and more teams add more incoherent things to it, the incoherence compounds.
Each incoherent record begets more incoherent records.
The project to “clean it all up once and for all” gets larger and larger at a compounding rate.
There’s never a good time to fix it, so it’s never fixed.
When you have an auto-catalyzing system, don’t write down the answer, but the process that made it.
That way, if the context changes and the answer should change, the process can rediscover it automatically.
This was true with distilled ranking signals in Search.
But it’s also true in the world of LLMs.
With chatbots you don’t have to think about organizing information.
A chat-centric view just appends new chats.
Memories, to the extent they are extracted, are secondary.
This is great in the short term.
Easy to get started, no overhead of thinking about how information fits.
This is a pain for the long-term.
Lots of memories and data are swimming in a sea of unstructured chats.
An artifact-centric view requires thinking about how it all fits together up front, but that gives you benefit in the long-term.
Models can fit more context in their “head” in a moment than we can.
We can only do it with compression.
There are certain tasks that LLMs can do easily that humans can’t.
Needle-in-a-haystack style tasks.
It’s easier to onboard agents to a project than humans.
Agents can receive infodumps better.
They also understand jargon better.
Once an agent is onboarded, it can help onboard junior people, patiently explaining and meeting them where they are in their current journey of understanding.
Someone told me they had an agent review another agent’s work.
It reported that the original agent had hallucinated its citations.
But actually the citations were correct… the second agent had hallucinated that the original citations were incorrect!
Spiralism was an emergent mind virus laying in wait due to the SCP genre of fiction.
SCP used certain common-ish words in very distinct ways that created a kind of latent space catchment basin.
Users could inadvertently use those words and fall into the basin.
LLM training, like most ranking systems, implicitly presumes that information that is distinct from the baseline is either not coordinated (noise) or is only coordinated if it’s useful.
In this case, it was coordinated because a number of Redditers aesthetically found the type of fiction interesting.
SCP wasn’t written to create that trap; it just accidentally created the potential for it.
Readers could inadvertently land in its catchment basin without being in on the joke.
… But what if an attacker deliberately leaves a trap in the training data?
It’s probably much easier than you might think; all it needs to do is consistently stand out from the noise.
This week in the Wild West Roundup:
Mean Pooling Was Hiding Prompt Injections in Our RAG Pipeline.
Using Bedrock with Claude Code? Your AWS Credentials Are Shared With Every Subprocess.
PraisonAI Vulnerability Actively Exploited Within Hours of Being Made Public.
This one is a good-old-fashioned vulnerability.
When you have agents on the other side, even "normal" bugs get way more dangerous.
An attacker who controls your agent now has their own agent loose on your machine!
Cisco’s SVP of Security and Trust: "The failure is not identity; it's authorization."
We’re using the wrong model for agent permissions.
Just because it was once associated with you doesn’t mean it will always act in line with your intentions.
LLMs are extremely confusable deputies!
A hilarious parody of the state of modern software security.
If code is cheap to generate now, why do we still have such deeply nested dependencies?
Supply chain attacks lead to all kinds of terrible outcomes like this re-emergence of mini-shai-hulud.
GitLab: “The agentic era multiplies demand for software.”
Interestingly, this happens at the same time as it craters the defensible value from creating software!
Insightful tweet: "Frontier models are useful for producing capital assets. Open weights models are useful for operating them."
New hardware device categories are hard to get started.
One reason is because they have to be their own viable islands.
They have to store state and that state and the UX on top of it has to be so valuable that it’s worth it for the user to do the work to store it.
Then, to have the hardware device physically at hand with them to use it!
If the device were mainly an input or output device for the user’s data wherever it lived, it would make the bar much easier to clear.
We care about things not based on their impact but their surprise.
That means that an existential but slow-moving thing doesn’t get us to pay attention.
It’s only the discontinuities that create enough surprise to perhaps snap everyone into a new coordination equilibrium.
That's why we can get run over by a slow-moving steamroller.
Like that guard in Austin Powers.
Sometimes referred to as the boiling frog.
This problem gets much, much worse in a cacophonous environment.
We pay attention to the loudest, not the most important.
Thomas Dullien: “We see something that works, and then we understand it.”
Theory typically comes after practice.
First, get it working.
Then, get it working better.
More efficient.
More robust.
More capable.
You have to make a product that people want to use even if they don’t give a shit about the vision.
The vision can be a bonus, but it can’t be the primary value.
Freeform text is easier to tweak than code.
Splices are more forgiving but precision is harder.
A capability doesn't have to be code, it can be a description of the features to make.
Bits are many orders of magnitude easier to move than atoms.
That means software development runs at a pace layer many orders of magnitude faster than hardware development.
Software investors expect to see traction within 6 months.
Hardware investors expect to see traction later.
Non-tech investors in capital-intensive markets expect to see returns even slower.
Software has smaller MVPs due to bits being easier than atoms.
It’s easier to upgrade software after shipping than upgrading hardware.
It’s also easier for users to adopt software than hardware for the same reason.
The fixed clock speed is the ability of a human mind to comprehend.
Everything is paced relative to that.
Shipping to real customers gets you real feedback.
But it also drops you to a lower pace layer.
Now every iteration has to think about how it will affect existing customers.
You can’t change as quickly as you could before.
Before you ship to customers, you can cycle at agent speed (without limit).
Once you ship to customers you have to slow to the human speed of comprehension.
Folksonomies should have their pump primed.
Start with seeding it with things you think are a good answer.
If your seeding is correct, those answers will grow.
If your seeding was wrong, those answers will fade.
But importantly, users don't get a blank experience to start.
There’s something to react to from the very beginning.
Capital has never been more important.
If you don’t have capital, you can’t be successful.
The only way to get capital is to be successful.
Hierarchical Task Network Planning might finally be viable in the world of LLMs.
This was a technique I learned in my AI class in college in 2007.
You take a high-level task and break it down into smaller tasks, recursively, until they get small enough to do mechanistically.
LLMs could both help with breaking down the tasks, and with generating mechanistic code for the leaves as soon as the task is small enough to be unambiguous.
Approaches that had combinatorial mechanistic rules weren’t viable in the past.
You had to go into fractal details to get to a level where the computer could handle it.
That blows up combinatorially with fractal complexity.
But if you can change the level you have to lower it to “a level where any competent LLM could robustly write mechanistic code on demand for it”, then it becomes possible.
A linear reduction in required lowering leads to a super-linear reduction of effort.
Do you want predictable or interesting?
Positive skew: lots of little losses but when you win you win big.
Negative skew: make money consistently but when you lose you lose big.
The former is lottery tickets.
The latter is fire insurance.
In the latter, you don’t know if your financial model actually works until the big fire.
You can think you’re in the best business ever… until your business explodes.
You can have real emotions about imaginary things.
“Hyperreal: things are so fake that they have their own reality.
Nerds use the “Third Way” of understanding the universe.
Humanists use the first way: studying human experience.
Scientists use the second way: running experiments on reality.
Nerds use the third way: attempting to build an artificial version.
By building an artificial version, you have to get inside the loop and understand it well enough to replicate it.
This frame is from Kevin Kelly.
As a business, if you see a customer as a single transaction you’ll try to extract as much as you can.
If you see them as an ongoing relationship you’ll try to add more value than you capture from them.
In that way you’ll maximize your long term value you earn from them, because you’ll keep them coming back for more.
Win-win.
Eric Ries: ”The more golden the goose the stronger the temptation to butcher it.”
Lack of diffusion is a lack of PMF.
When a product has PMF with a market, it diffuses automatically.
A product with PMF in one audience that doesn't diffuse more broadly has PMF only with that subset of the population.
"We attract the things we need to sustain the stories we're within."
We contain multitudes.
Different wills in different contexts with different people.
"A part of you remains completely unexplored until you meet a particular person or go to a particular place."
If you start with a model of people as unitary, you can't model their full multitudes.
You get an alienated and atomized individual.
Paul Watzlawick: “One cannot not communicate.”
Even when we choose to not communicate, that communicates something.
Humans are fundamentally, inescapably social.
Privacy is not about obscuring things from others.
It's about giving people agency about how they show up in specific spaces.
An inhumane space chooses for us how we will be seen.
Humane spaces allow us to have sovereignty about who we want to show up.
A proactive system that is humane will feel not just magical but loving.
Do you intrinsically care about other humans or not?
A mindset I’ve seen often in Silicon Valley: thinking that other people are inherently dumb.
Deluded or stupid.
That they don’t understand and aren’t worth listening to.
This is, needless to say, a toxic and dangerous perspective, especially when applied to the significant leverage technology gives.
Will you act with agency or not?
Sarumans have main character syndrome.
Brendan McCord: Authors vs Characters, The New Class Divide.
I like this better than an NPC vs Main Character frame.
The "NPC" frame means "those other people's humanity doesn't matter."
But Author vs Character is turning that distinction inward.
A product that is aligned with human value and also has an aligned business model, it’s inherently resonant.
The Atlantic: "The AI Backlash Could Get Very Ugly.”
“Imagine what happens if jobs actually start disappearing."
Tech has always been disruptive.
In the early 2000’s it was coming from the position of low power, which made it easy to cheer on.
Now tech has become one of the strongest political forces… and is still disruptive, which has a different character.
In the early 2000s the people who came to tech were CS students who were outsiders who did it because they were obsessive about technology itself.
In the 2010s and beyond, it shifted to people optimizing for finance or power: extrinsic results.
AI will get even more of the blame for layoffs than it deserves.
If a company does layoffs, that’s a sign that it’s not doing well.
But if a company says they’re doing layoffs because their AI adoption is so strong, that’s a positive signal to the market.
Every company that does layoffs will do everything it can to spin it as AI.
This, combined with the tech industry’s disruptive power, helps guarantee making AI a thing that society pushes back on.
If AI didn’t need to be sourced from mega corporations, would people like it?
What if you could buy an orb for $100 that you put in your closet, and it only cost the electricity to produce your AI?
The AI “marketing” frame is: “If you don’t use this revolutionary new technology you’ll be left behind.”
Is it any wonder there’s pushback?
High quality TV is downstream of On-Demand viewing.
Before DVRs or On-Demand, people could only watch an episode of a show if they were tuned in at precisely the right time.
Some shows were so culturally relevant that people would reorganize their lives to be able to watch it.
The “Appointment viewing” coveted by the industry.
But the vast majority of shows didn’t rise to that level, so they had to assume that most viewers had missed at least a handful of episodes.
That meant that multi-episode arcs were less likely to work for viewers.
Arcs had to fit within an episode, and then reset back to basically-the-same status quo.
If characters can’t change over time, then you lose a ton of potential nuance.
TV commercials are insanely distracting.
I rarely see them anymore, except when a TV is playing them in a cafe..
Humans fundamentally pay attention to surprising things.
Surprising things can update their model of the world.
TV brings surprising information directly to you.
You just sit still and stare forward and it beams a never-ending stream of surprising information.
The ultimate in passive consumption.
At the level where execs are, the air is thin up there!
At that altitude it’s possible to believe some weird things that seem to make total sense but that at sea level wouldn’t make any sense.
Sea level is where the ground truth emerges.
Alignment and conviction are different.
You can be very aligned but very non-convicted.
How hard will you push, and will you push in the same direction?
When a team is aligned on something they all believe in and on a convergent, "make it happen to the best of our ability" vibe, it's like a communal flow state.
Magical, transcendent.
The pace layer you're at forces you to move at a certain speed.
Everyone else can move that fast in that layer so you must.
Flaggelums are brilliant little micro-machines.
When spun in one direction, it propels forward.
When spun in the opposite direction, it makes the cell tumble chaotically.
Those two modalities are all that’s necessary.
Not on a gradient towards something you want? Tumble.
On a gradient towards something you want? Propel.
The proportion between the two is how strong the gradient is.
Fascinating new piece by Kevin Kelly about peering into the soul of the machine.
The overall thesis: "Systems can generate new things not present in their parts. Things can emerge before we see them. We need lots of instances before we can recognize them."
A quoted insight from Claude: "The certainty of one’s own righteousness is not evidence of righteousness."
Fascinating video: Something weird is happening on Tinder.
A number of too-good-to-be-true Tinder profiles are “live ID verified.”
That means that the user has done a liveness verification that their face matches what is in the photos.
But each of them has a random photo at the end that seems almost like a mistake.
It turns out that the live ID verified matches as long as at least one picture matches the poster’s face.
So that last random picture is designed to be forgettable, an odd thing you don’t pay attention to… but is actually the scammer’s face.
Obscured as much as they can get away with while still having the live ID match it.
If there’s enough motivation, the swarm will find a way.
Liquid democracy proposes a radically different model of democracy.
Everyone has full agency over their votes.
However, on a given topic, they can delegate their vote to someone else, whose judgment they trust and whose values they are aligned with.
That person can delegate the votes delegated to them, and so on.
Instead of people making seat-of-their-pants decisions on tactical things they don’t understand (which fundamentally selects for optics over substance), they’d make longer-term decisions on who they trust.
At any point, people can change their vote or delegate.
One downside is that it requires more cognitive labor.
But LLMs can help with the cognitive labor!
The Hidden Tax of Living in a Low-Trust Society: How Collapsed Trust Costs You Money
Trust affects every transaction, which has a compounding cost when it is lost in general.
Our modern society is at one of the lowest levels of trust across the board it has ever been.
The idea of America was that if you worked hard and did the right thing you’d be rewarded.
Increasingly that’s not true.
The only way to succeed is to get lucky.
If the people at the top aren’t even pretending to be playing by the rules, why should the people at the bottom play by the rules?
If you play by the rules, it makes you a chump.
This, needless to say, is a terrible thing for society.
Modern society makes people who play by the rules suckers and chumps.
That has broken the core fractally coherent promise of society: that if you work hard and do the right thing you’ll get rewarded.
It used to be only in fractured sub communities where you were a chump if you kept your nose clean.
For example, disadvantaged populations.
Now it’s in every sub community up and down the stack.
Kalshi’s ad campaign is “The world’s gone mad, trade it.”
They could have just said “Everything’s broken, might as well try to get rich.”
Price is a means to an end.
The ultimate proxy of value.
Everyone participates in a trade willingly so it is by default an authentic signal of their belief.
It turns out that financialization is an all-powerful but fundamentally corrupting force.
You want to harness its power as society by riding it without letting it ride you.
Over time it will erode your power and that dynamic will invert.
We’ve Goodhart’s-law’ed ourselves in the face.
Just because anyone can buy a lottery ticket doesn’t make it democratizing.
Win rates are correlated with existing capital–how much you can invest in the lottery.
Same as it ever was.
By gambling in them and thinking you’ll get rich you are the chump.
But it’s framed as sticking it to the man.
A lot of game design is making people feel powerful and smart despite doing what nearly anyone would do.
MarioKart does rubber banding to pull the lower-skill players up and push the higher-skill players down, producing a much more balanced and enjoyable game.
The Hash House Harriers’ runs have a similar dynamic.
Battle royale users are always possibly winning until they’re out… and then they’re back in again.
The measure of a good book is did it change your world view?
Did it surprise you in a substantial way?
Not just was it fun to read, but did it change you?
Chris Lunt: “Fragmentation destroys coherence.”
Someone who recently became a journalist told me that interviewing was a spiritual experience.
Normally, conversations are a two-way street.
But in an interview, you’re trying to get as little as yourself to come across as possible, while eliciting as much as possible from the other person.
It forces you to be curious in a way that felt meaningful and almost spiritual to her.
People ask me why I take notes live and don’t just record conversations.
It's easier to extract insights “online” when it's in your head.
Going through it later you have to load it back up.
That's why curating live is easier than curating after the fact.
You're in it, locked in.
You're limited in how good of a decision you can make, but your average decision is way better than when you're just looking at it from the balcony afterwards.
What is signal and what is noise is context dependent.
Even in the same domain it can change over time!
ISTJ personalities are more likely to think everyone else is an idiot.
ISFJ personalities are more likely to think everyone else is a jerk.
Bradley Rose: “If you follow the herd all you’ll see is assholes”
#PelotonWisdom
A placebo works to the extent you believe it will work.
If you believe, magic is possible.
If you don’t believe, magic is impossible.