Bits and Bobs 10/21/24

What if that's not true?

Everyone's implicitly treating LLM providers like Google or Facebook: an end-user aggregator that then gets significant leverage.

This is partly because every LLM provider makes available an API, but also has a 1P service.

Vanilla LLMs are so useful that the no-frills default UX from the providers wins by default currently.

This leads us to analyze them mostly like the consumer aggregators.

But what if LLMs end up more like cell network operators?

Capital intensive, and theoretically compete on quality, but realistically everyone just treats them mostly like commodities.

Dumb pipes that are underneath the services the users care about.

Technically it's possible for a 3P to use one of the provider's models in a more powerful and useful UX that beats the 1P product.

You can imagine a world where the LLM providers only made a 1P tool and didn't allow a 3P API.

But that's not the world we live in, and now that Llama 3.1 405B is open weights, not a world we are likely to live in.

Who has the power in this new world comes down to what services users use most.

It could be the 1P LLM providers' UX… but that implies they will create the most compelling uses of LLMs.

Or rather, won't be significantly surpassed by others who do.

The LLM providers have a default advantage, but not a massive one.

The reason it feels like the model providers' 1P UX will win is that it assumes that the "killer use case" of LLMs is a chatbot.

If the killer use case is a vanilla chat bot, then it makes sense that the providers' own UX will be the best–they don't have to charge themselves a margin.

That could be true, but it very well might not be!

We're in the very early innings of the AI era.

The chatbot might turn out to be only a minor manifestation of LLMs.

The LLM providers could conceivably end up as dumb pipes.

More like Microsoft Word than Instagram.

Hold lightly to your current understanding, and try many safe-to-fail experiments.

A playful, open, curious mindset.

This mindset is useful when the downside is capped and the upside is uncapped.

In those cases, you should keep spinning the roulette wheel, to find as many great combinations as possible.

If you have misjudged and the downside is not capped, or could bankrupt you, then an experimenter mindset could be dangerous.

But in most cases, an experimenter mindset is useful.

The upside for figuring out how to use LLMs is higher and the downside is lower, so the experimentation mindset is even more valuable than it once was.

Persistence is if you stick with it long enough for the non-linear benefits to kick in.

Hustle is whether you keep turning the crank no matter what.

But if you're turning on the crank on something that never adds up to more than the sum of its parts, then you're not getting any leverage.

The question is: are the small actions cohering into something larger over time?

If not, you're wasting your time on low-leverage work that has a high opportunity cost.

Those larger frontiers take time to search through and find the best pockets.

For example, many of the most durable UX patterns on mobile phones were included in the very first iPhone keynote.

That implies a relatively small surface area to explore.

An easy environment with stable patterns.

But AI seems to have wider territory that is harder to explore.

A tell for how big this space is: how long do the people at the frontier continue to be surprised by best practices that are discovered.

By that measure, LLM is a broad disruptive innovation!

Hustle is about the grind, about continually tackling small tasks.

But to discover interesting things to do with LLMs in this early era will require curiosity, earnestness, a sense of play, a willingness to experiment.

Intellectual interest is not sufficient; you have to actually play with LLMs to learn what they can do.

They resist study from afar, they are too squishy and complex.

The only way to learn to use them is to use them.

Things that can get done but aren't important are like the streetlight fallacy.

That is, work that could be atomized into infinitesimal chunks that any reasonable human could do with reasonable quality.

These are tasks that are independent enough and pure enough for LLMs to be able to hit them out of the park.

LLMs can do them orders of magnitude more cheaply than even the cheapest human labor.

If you ask bland questions, you'll get bland answers.

How can you ask interesting questions that drive it into interesting niches?

LLMs reflect back the centroid of your question; the power comes from asking the right questions.

Contrast with a violin, where it takes considerable practice to not have it sound god awful.

But that does not mean that all piano players are equivalent.

The opposite, in fact.

The maximum level of virtuosity someone can reach on a piano is much higher than other instruments, because it is so forgiving to get started.

Not unlike AI!

When you're feeling your way through a foggy labyrinth, you need to try out lots of different options, quickly.

If they turn out to be viable, you keep working down that path.

But every so often you realize that you've iterated into a deadend.

It might be tons of effort to pull yourself back out.

But imagine that you had an oracle who knew the maze well.

The only question they could answer is: "if I take this step, is it pulling me towards a deadend, or would there still be viable paths to my goal?"

If the answer is "no", then simply take the step, and don't worry about it.

You can fix it later; finding a viable next step is the most important.

If the answer is "yes", then find other steps to take.

Sprinting into culdesacs gives you the illusion of progress while actually slowing you down in the long run.

This kind of oracle sounds impossible, but they do exist!

In certain contexts, you can find the wizened, battle-scarred expert.

Their intuition will be extremely high signal about how dangerous a given path will be.

LLMs can also give oracle style answers in some domains.

1) That the would-be aggregator starts out as far and away the best in a category.

Nobody else even comes close.

They are the schelling point, the most prominent example of the category that attracts usage by default.

2) That the would-be aggregator has some kind of compounding advantage.

That is, that the more that people use their product, the higher the quality of the product gets, creating a stronger pull.

This advantage can come from:

Good old fashioned network effects.

Data stickiness.

The more data a user stores in the tool, the more useful it gets for them, which also makes their switch cost higher.

Data Flywheel.

The more any user uses it, the more the quality for all users improves.

Google Search has this property.

Notably, neither of these is true anymore for OpenAI.

OpenAI is no longer the best model.

OpenAI also doesn't seem to have much of a compounding benefit from usage.

In a situation without network effects, you might fall marginally behind your competitor.

But with network effects, falling marginally behind means you fall orders of magnitude behind, and every time unit you're behind you'll fall another full order of magnitude behind.

It looks like the LLM picks an object and later we uncover more information on what it must be based on its answers.

But that's an illusion!

Actually the LLM has no idea what it "picked"; it's late binding its answers based on what seems most coherent with the things it's already said.

You and the LLM are discovering what the secret object was at the same time.

We project our normal human ways of thinking onto it because it sounds like a human.

But the way it works is in some ways very inhuman!

But LLMs can, which is confusing!

We think of it as a thing, with a complex inner world, because it can speak to us and sound human-like.

This confusion leads to a category error.

It feels like a human made out of machines.

In reality it's more like a voice on top of a pond scum style emergent intelligence.

An intelligence that is absurdly good at absorbing mind-numbingly complex patterns.

You can't optimize an end.

An end is valuable in and of itself.

You maximize an end.

Optimizations can only optimize across things internal to the model being optimized.

Everything has externalities; there is no clear bright-line edge to any system.

What is "inside" and "outside" the system is a human opinion about what will be most useful for the problem at hand.

Where to draw the lines of the model is a judgment call, a simplification.

But after making the call, you'll later forget it's a judgment call and instead come to implicitly assume it's some objective fact and there are no externalities, or if there are they're minor.

Most systems have significantly more powerful externalities than we realize.

He thought humans would happily put themselves in gooey vats if they got the smallest short term benefit.

Your adjacent possible is thinner than you think it is.

And yet you do exert true agency over which thing in the adjacent possible to pick.

Meaning comes from choice.

Otherwise it's just a "shrug, what could I have done?"

When people are mad or scared or stressed (the default emotional state in modernity), people often implicitly minimize how often they have to exercise it.

But it's a muscle.

If you don't use it, you lose it.

AI, used well, gives significant leverage to your agency.

But only if you exercise it with intention.

If you don't use it to extend your agency you will fall behind people who do.

In the future, imagine people exercising extreme amounts of agency… over increasingly frivolous things.

Not too dissimilar from landed gentry in the UK in the early 1800s, spending excessive time on increasingly frivolous things.

The most focused stamp collector ever.

They can have large staffs to execute many things, extending their agency.

They have to hire people they trust, but if they do they have significant leverage on their agency.

They think carefully about where their time is best spent–the things they care most about.

They delegate the parts that aren't as meaningful for them.

People who earned their significant wealth have practiced agency in the past and can apply it carefully with leverage.

People who inherited that level of wealth have a much rougher time of it.

They never learned what it felt like to exert agency over their daily life.

The first generation of humans doing this with AI will remember how to exercise their own agency.

The next generation won't.

It helps deprogram people who have fallen into conspiratorial beliefs.

But… presumably it would be just as easy to create a patient, convincing bot to convince you to believe in conspiracy theories.

Any time you deal with a force that is more patient, more intelligent, or more powerful than you, you have to really trust it to not manipulate you.

Especially if it makes it easy and short-term useful to hand your agency to it.

A sense of things that feel true to you that might resonate with others.

A crass optimization for a thing you think will go viral, but doesn't feel like a deep statement from you, is different.

Optimizing for your own meaning, vs optimizing for the algo.

Instead, make the content that you understand and want more than anybody else.

Be authentic to you.

One of the reasons this works is because:

1) If it is authentic and meaningful to you, you'll enjoy it and feel fulfilled for its own sake, even if you have few viewers.

Capped downside; negative opportunity cost.

Because you enjoy it, you're likely to keep going, and apparently some research says the best predictor of influencer success is keeping it up.

2) Your perceived nicheness might be greater than the actual nicheness.

Perhaps there are more people on earth who want what you want.

By leaning into your own niche that doesn't exist, you create the potential to connect with a distinct audience in a durable way.

When people connect authentically with your authentic voice, they are more likely to evangelize to others.

Uncapped upside.

Note that ThreeBlueOneBrown can give this advice because the second part happened to work for him… something that was not obvious before he started.

Many people who follow this advice will have the first part work, but not the second.

But still, would you rather produce algorithmically-optimized slop or create content that is meaningful to you, that helps you think and reflect and engage with the world in a way that feels important to you?

These people need to be in tune with what the algorithm will want–which is what others on the platform want (not what those viewers want to want).

Doing this effectively means being plugged into the vibes of society.

It's very hard to stay on this bucking bronco of producing viral content once you've climbed aboard and started amassing an audience.

Each individual bit of content needs to know what else is on the platform, what else is resonating now.

Little chunks of creative effort, not for your own expression, but for an algorithm.

Grinding to create content: in the limit, human slop.

But LLMs can create slop now, too.

And faster and more relentlessly and with a better sense of the whole than any human could ever have.

Someone described it to me this week as informational fentanyl.

Spooky lights emanate from every window at all times of the day and night.

In the front drive a Bentley is conspicuously parked… but unused and collecting dust.

What's haunting the house?

Professional Instagram influencers.

People who managed to find a following in these post-social infinite algorithmic hellscapes and are now doomed to forever attempt to stay on that bucking bronco.

They got swept up in the algorithm, and now are in a relentless co-dependent cycle with it.

They work for the algorithm.

An inhuman, insatiable boss.

But even the people who created the algorithm work for it.

It's a red queen race for engagement, in a fierce battle with other providers with clear network effects.

Step off the treadmill and you will be left in the dust before you know it.

The people who are in tune with the algorithm have a mutual co-dependency with it.

They understand the algorithm, what it wants, how to feed it.

They can anticipate its desires.

They know how to perform for it, to dance when it tells them to dance.

They are in a dark form of symbiosis with it.

But in a mutual pact that makes each a worse version of themselves.

(Of course, many people who have developed significant audiences in this post-social landscape are doing so in a way that is authentic to them and lifts them and their audience up. It's just also easy to fall into the more parasitic version.)

They have become algovamps.

Charismatic time-filling content devoid of meaning.

Drawing off the life force of the mutualistic other in a way that makes both less human.

They started off as normal humans but as they chase engagement themselves they have slowly become a husk.

Being an algovamp is not some intrinsic characteristic of the person; it is a situation they have become trapped in.

The only way out is to unplug from the machine.

If one algovamp signs off a dozen more are ready to take their place.

Swarms of algovamps and the algorithm, a swirling gyre of meaningless diversion.

At least the algovamps are human and so the content is at least partially human.

Imagine when the content producers are algos themselves.

Humanity caught up in a cyclone.

Creativity need not have its outputs seen by many or have commercial impact.

Creativity is about exercising your individual voice, your agency, to bring something into existence that is authentic to you and you find meaningful.

Creativity takes time and space.

The optimizing machines around you that you are embedded in will see your creativity as time that could be spent optimizing the machine.

They will encroach on that time and gobble it all up.

Some people are so motivated by the urge to be creative that they make space for it.

But many people get ground between the gears, losing sight of what they want to do.

If you give more people the space to be creative, structurally, they will grow into that space and create.

Some of that output will be just for them.

But some subset of that authentic creative force will resonate and be commercially useful, too.

This is part of the logic of the Radagast's magic.

If you create the space for people to grow, they will, of their own accord.

Engagement-optimized is crass, opportunistic, what users want, not what they want to want.

Meaning is an end in and of itself.

It will differ for each person.

Meaning is about becoming a better version of yourself in a way that is important to you.

Instead of big apps enticing you to come to them, small bits of functionality that flock to you.

Say, a feature for board game enthusiasts that would help them have more engaging, collaborative sessions with their fellow players.

Imagine the feature would require small tweaks to Gmail, Google Calendar, and Google Docs, to tie them together.

A PM who discovered this possibility on Gmail would need to convince a dozen or so other PMs to prioritize this niche feature that would likely have a negligible impact on Daily Active Users (DAU) at scale.

Each feature, after all, potentially makes the product more complex for users who don't use it.

This is the downside of one-size-fits-all software that is used by nearly everyone.

What if instead you could have niche features that benefit particular populations without losing out on the general purpose features that benefit everyone?

What if the thing that led to features getting built was the enthusiasm of the niche, not the notice of PMs they've never met?

Code being able to transmit data to a possibly sketchy origin is the scariest permission: once it happens, there's no recalling it, no tracking what happens after that point.

This makes it quite hard to trust any given origin with sensitive data, who knows what the origin might do with it?

The more trust an incremental origin requires, the more the ecosystem will tend to stabilize on a small number of large, powerful players.

The jungle is dark and scary, but cutting through it is much faster than going all the way around it.

When should you go through the jungle vs going around?

When the jungle is a massive shortcut and you believe that you have some differential ability to get through the jungle faster and safer than almost anyone else.

But beware: if the jungle is deeper than you think, or darker than you think, or less of a shortcut than you think, you could be on a dangerous path.

To optimize the system you have to optimize the whole.

If you optimize for only the individual, in the limit you get the super-chicken problem.

Instead of building each other up and collaborating, everyone tears everyone down all the time.

A hellscape of maximal competition that no value can cohere in.

If you optimize for only the collective, in the limit you get a machine that churns through humans, consuming them on the path to a collective outcome.

A machine that burns through people on its way to achieving its end.

A human has to treat themselves as an end.

The machine / collective will see you as a means.

Only you can remind yourself that you are an end.

Being human means treating yourself, and others, as ends in and of themselves.

1) You see that the world is not single-dimensional and black and white, but mullti-dimensional and gray.

If there's something that doesn't make sense to you right now, that implies some dimension or shade of gray you can't yet sense.

2) You meet disconfirming evidence with openness and curiosity.

When you see something surprising you don't lean back and say with a frown, "oh… that's interesting…" but instead lean in closer, with excitement, and say "oh! That's interesting!"

These two skills combined allow you to be antifragile; to grow from disconfirming evidence, instead of being worn down by it.

Situated software just has to be bug free for that user's particular use.

Many orders of magnitude lower bar of generalizability, and thus of cost to create.

If only more people could create software with little effort or expertise.

That's messy and no one other than your friends and family wants to see that.

Instead, bring your best self.

Sometimes the ugliest things are what works.

What matters is that it works.

From there you can clean it up

But before it works it doesn't matter how pretty it is.

"Cheapness creates scale, which makes things even cheaper, and the ultimate output is entirely new markets."

"A business carefully evaluates options, and doesn't necessarily choose the highest upside one, but rather the one with the largest expected value, a calculation that incorporates the likelihood of success — and even then most find it prudent to hedge, or build in option value. A dreamer, though, starts with success, and works backwards."

The slower the clock rate; each step in a process has a bit of fixed cost to be received and executed by the next person as it is handed off.

Receive, ramp up your awareness, execute, send, ramp down, repeat.

This means that the more people that have to touch a process in sequence, the slower it will go.

Large organizations can shard specialities down to ever-more specialized niches.

That implies ever more steps in the process.

But humans can't instantly come up to speed on a new bit of information.

So organizations made of humans get slower overall as they get larger.

Smart people can argue for their beliefs better.

But also they are better at retconning their bias.

"Well, I've got to find a way to make this work."

But if you can discard them and never see them again you'll be more transactional.

You see them at the soccer game.

A full person, not just the transactional slice of a person.

The more you interact with others as full people, the more you can grow as a human in a collaborative fabric of society.

But seeing a full person is not efficient; to scale, we need to see simplified snapshots of others, to make transactional decisions in that domain.

Scale and community are in tension.

The distinction that emerged was "stayers" vs "leavers"

Stayers stay in the environment they started in.

Leavers leave to find a better environment.

A benefit of stayers: you build deep connections with those around you.

You expect to be there indefinitely, so it makes sense to invest the time to see others around you as full-dimensional people and build trust and understanding.

Doing so helps you build a long-term perspective, grow, etc.

A benefit of leavers: you can go to your highest and best use.

You can sense where a different environment allows you to achieve more for yourself… and for society.

It turns out that density has super-linear returns to innovation; there's a benefit to moving to density.

But if you're moving where you live every 6 months, it doesn't make sense to invest in really getting to know your neighbors.

That takes time, and you'll be gone before any benefits arise out of it.

People invest an amount of time and effort proportional to how long they expect to be engaged in that community.

As people get more mobile in our society, the expected length of time in any in-person community is shorter, so we invest less in it.

Efficiency and innovation are in tension with community and long-term thinking.

The nuanced follow up doesn't.

Media flattens a complex evolving topic into a clear, black-and-white distillation that will travel easily.

As everything gets louder the picture gets flattened to get through the background noise better.

We intuitively assume that the signals we see are randomly drawn from the distribution of real experiences, but actually there is a structural bias… giving us a fundamentally incorrect understanding of the underlying distribution.

This bias is not intentional or cynical; it simply emerges due to fundamental asymmetries inherent to the medium and what people prioritize paying attention to.

For example, on Instagram people are far more likely to share pictures that make them look cool, happy, interesting, or successful.

As a result, we compare our own more mundane lives and conclude we are below average.

Another example: news focuses on things that are surprising or negative.

Negative news is more psychology salient to us (negative news could lead to a game-over, whereas positive news couldn't).

The news doesn't bother reporting when a dog bites a man.

But that leads us to have skewed senses of the underlying distribution.

"Wow, there sure are a lot of men biting dogs!"

1) Acknowledge audience

2) Acknowledge person who invited you to speak

3) Self deprecate

4) Say something about You (the individual audience member)

5) Say something about We (the collective the audience member is part of)

6) Say something about Now (making it feel immediate)

Apparently this playbook, when delivered by any suitably charismatic speaker, brings down the house.

An inherently charismatic framing device.

A playbook to give you robustly but superficially compelling output no matter the topic.

Not unlike NotebookLM's ability to create a robustly but superficially compelling podcast no matter the topic.

If you don't see it yet, maybe the tools to see it don't exist yet!

…Isn't that crazy?

(I may have gotten this acronym wrong)

Once a quarter, everyone was supposed to spend a week working on them.

Often those little ideas have huge positive externalities that are hard to characterize.

The work to do them is easier than the work to describe their impact.

The power of P2s: sometimes giving some space to make a product better in little ways improves the product much more than any single big thing would.

A crisis occurs where the hero's external source of motivation explodes.

They go off into the wilderness to discover their internal source of motivation.

In doing so they discover their superpower.

There's a lot of power that comes in defining the game.

Don't play in a game someone else created.

Play in a game you created.

"Oh, we don't have politics here."

They're on a raft above the politics, not in it.

Everybody else is forced to be in it; they'll be under no illusion about the politics.

Politics emerges in every group of humans.

If you have no predictive model, it will all just be chaos, and there will be no surprise.

Surprise is the raw material for learning.

Develop the muscle of creating predictions so you can be surprised and so you can learn.

It's a way to think and learn.

Useful for its own sake.

1) Part of a high trust club that shares ideas

2) Writes a lot to synthesize ideas

That is, the person is in an "idea lab" that has lots of diverse ideas flying by, and the person catches some of those ideas and writes to flesh them out and make them more robust.

At the early growth phase, it's exciting, there are lots of opportunities for advancement, and the industry is adding significant new value.

Few things to question about that!

The mature phase leads to more, "what am I doing with my life?"

The rate of advancement for new people in their career is slower, and also the mature phase is more extractive, so it's harder to convince yourself you're causing some big new positive thing for society to happen.

If you don't love yourself, you will avoid truly seeing yourself, for fear of disliking what you see.

The extent to which you can see yourself is the extent to which you can grow.