Bits and Bobs 1/22/24

At Google we called it 20% time.

But I think a way to frame it that is more natural in other contexts is "120% time".

These are not frivolous things you get to carve out space for no matter what.

It is, however, to say that if you have an idea for improvement that you feel strongly about, you should feel free to stick your neck out with extra effort.

That is, to act like an owner.

This requires it to be an idea that someone feels strongly about, because they have to go above and beyond.

This could help find doorbell-in-the-jungle style ideas.

Cheap to start (low downside)

Compounding potential if it works

Each incremental investment step will be a no-brainer if it's warranted.

But it will also be important to make sure there's a bridge for these 120% ideas to grow if warranted.

To start, an individual does 120% time.

If that works, perhaps they will convince others to also invest 120% time.

If that works, perhaps those people get top-down approval to spend 20% of their time on it.

If that works, perhaps those people get top-down approval to spend 50% of their time on it.

If that works, perhaps they get an additional head assigned to the project.

If that works, perhaps they get a formal allocation to hire 3 heads.

... And onwards, up to an Official, Fully Staffed Thing.

I don't think you need too much structure for this, just a little cultural scaffolding.

Celebrate successes that have this shape, to act as role models to inspire others to stick their necks out in 120% time too.

Make sure that if someone does something that is scrappy and has the potential to be transformative, that they feel like they have to hide it, but feel proud of it.

To be an owner you need to assume the thing is your responsibility for a medium to long term, and also to have the space to act like an owner.

Acting like an owner requires practice. If you don't have practice, it will get harder to do it in other domains, too.

There's a healthy "I'm going to act like an owner and ignore minor process that is holding me back because I know what's in the best long-term interest of the thing" that looks superficially similar to a "I'm going to ignore minor process that is holding me back because I need to get this done fast so shortcuts are the only way".

The former is healthy; the latter is reckless.

When you're doing work, the former and the latter will feel very similar, so be careful!

Your escape hatch is more to retain users who would have otherwise left, than to gain new users you otherwise would not have gained.

It's easier to get a user to dig down into lower layers of complexity over time as their needs get more complex, than to convince them to let go of the nuance and control they had and go up a layer.

As a user, going up a layer on a thing you're already using requires you to take more constraints on a thing that is already working; why change it?

Imagine a feature that is too in-the-weeds to justify a company-level OKR, but will significantly improve one corner of the product surface area.

If it's a feature entirely within your org, you only need to convince your management chain.

If it's a feature that exists between two organizations and needs them both to collaborate, then it can be not just 2x more challenging to coordinate, but an order of magnitude.

That coordination challenge makes it more expensive, for the same user value. The bang for buck declines.

If things are not tightly prioritized and there's some slack, you might be able to informally find a counterpart in the other group to also use some of their slack on the same thing you're using your slack on and jointly prioritize it without having to involve management.

But when things are tightly prioritized and there is no slack, the cost of these cross-cutting improvements goes up, and so on the margin the rate of investment in them goes down.

Beautiful here means having aesthetics crafted and delivered in high fidelity to the creator's high-taste vision.

A product can work even if it's not beautiful.

Beauty is optional for products, but not for art.

A successful product tends to grow organically in very messy ways that accumulate debt.

It's hard to pay down that debt continuously, so it often accumulates into a big pile until something simply must be done.

In that project to fix it, it's tempting to figure out the perfect, decade-durable version and implement that--who knows when you'll get to tweak it again.

But that machined perfection is extremely hard to retrofit onto a messy, living thing (especially one that's not staying still!)

It's easy to get into a trap of going into a cave to try to get the plans perfect.

It's much better to do small, almost-certainly-a-good-idea improvements continuously.

It's possible to produce products exactly to the plan... that turn out to not actually be useful for real users.

It's also possible to produce excellent products in an extremely messy way.

Given the asymmetry, be very careful not to make rash decisions in the heat of the moment about cutting down a tree.

But conversely, don't sweat spreading around a bunch of seeds; they're cheap now, and who knows what they might grow into in a decade!

People who are deep experts in one specific thing tend to trade off deep, useful expertise for closing off their horizons to other approaches.

You might call these mono-experts.

Most generalists are simply jack of all trades: minimally competent at many things.

These are non-experts.

Another type of generalist is a deep expert in at least two very different things.

This allows them to triangulate beyond any one expertise, to see bigger patterns and make connections in novel ways and not over-fit.

These are poly-experts.

Once you've developed your next expertise, the one after that will be even easier.

Finally, there's a type of generalist who is expert in being a generalist.

They can build a high-quality intuition on just about any topic relatively quickly.

These are meta-experts.

They work best when paired with a deep expert on the topic they're looking at.

The meta-expert can quickly abduct possible meta hypotheses that the person with encyclopedic expertise can quickly disconfirm.

This creates an extremely short-cycle dialogue loop.

Even better is when a meta-expert can create a community of a diverse set of poly-experts that engage in collaborative debate.

This is where the deepest and most useful insights can be generated most quickly.

"This person did these actions and it caused their massive success."

But we typically find most interesting stories that are already known to be surprising and interesting.

E.g. biographies of a rags-to-riches tale are most interesting.

The behaviors that are documented are correlated but not necessarily causal. It doesn't have predictive power.

We write biographies about people we already know are successful, but then we tell the narrative forward as though we don't know the outcome, and that creates the causal illusion.

Those behaviors allow that outcome, they do not necessarily cause it.

Most people who do those behaviors do not have massive success.

The way you have great returns (something above linear) is by picking the right curve.

Execution can only cap downside. The right strategic insight sets upside.

You need both.

Agents choosing to apply their own creative energy can avoid cliffs that their boss didn't see.

They can also find new upside that their management chain never even imagined.

As a manager, Is there a way for your report to surprise you, to go above and beyond?

Or is the ceiling set by your request, with the only possible variance being how much you are let down?

These are two very different orientations.

This creative force is what allows the system to be resilient, adapt, navigate around problems.

A system can survive even when it's not alive, that just makes it a machine.

If it has enough momentum, or the conditions are calm enough, a machine can go for a long time.

But a machine cannot rescue itself, cannot change itself, cannot maintain itself.

Things grown from within are able to thrive without intervention from outsiders.

But a thing that is built always requires someone else to provide the creative force.

If that creative force is ever taken away, it will rot.

Things that are grown can be self-repairing; resilient, even anti-fragile.

Consider inside a company.

The components are all built by employees being paid to apply their creative force.

But the whole system, when viewed from the outside, could be said to have grown, from a creative force that comes from within itself.

A product is built. An ecosystem is grown.

A platform is a special kind of product whose value comes not just from the built product but also its adjacent ecosystem, which is grown.

This is true primarily if it's a game that only some people are allowed to play.

If other people are competing and you aren't, they'll outcompete you and knock you out of the game.

So you have to compete to play.

If you're competing it's about the competition first and foremost, not for its own sake.

Things can still be competitive and in an infinite orientation if everyone is allowed to participate.

Competition is a powerful acid that strongly selects for high quality, effective solutions... but at the cost of turning infinite games into finite ones.

Sensing all around you. Your ego plugged in, senses twitching, in danger, uncomfortably excited, gloriously alive.

Your knowhow acquisition rate is through the roof.

But your horizon is limited to the edge of the system.

When you're passively consuming you are outside the system.

Intellectually thinking about it, toying with it, safe.

Your horizon is broad, but your rate of knowhow acquisition is significantly lower.

Bad process is a cage. Good process is a trellis.

How can a process not be a cage but a thing that makes you stronger?

By coevolving it, braiding alongside execution, designing with.

Process built with a short-term perspective is just checking the box, creating a cage.

But process built with a long-term view can be a scaffolding for change and growth.

When you are inside the paradigm, you cannot see beyond its ceiling.

It feels like a fundamental law of physics.

The ceiling is so inescapable, so implicitly obvious to everyone, so unwaveringly boring, that despite its power it is invisible.

The horizon of a paradigm is set not from outside, but from inside: the emergent dynamics and tensions of its enabling ideas and technologies.

As time goes on, competition in the industry harvests every last scrap of value, and the industry finds itself smooshed up against the invisible ceiling at the edge of the horizon.

When this situation meets a new, general purpose enabler, a new paradigm is waiting to erupt.

Numbers have to be read. Graphs allow you to see.

"Read" here means a kind of OCR, parsed into semantic understanding and then operated on. A multi-stage process.

"See" here means relying on the spatial intuition that happens in heavily optimized, deeper parts of our minds.

A quote from Poor Charlie's "At any rate, mankind invented a system to cope with the fact that we are so intrinsically lousy at manipulating numbers. It's called the graph. Oddly enough, it came out of the Middle Ages. It's the only intellectual invention of the monks during the Middle Ages I know of that's worth a damn. The graph puts numbers in a form that looks like motion. So it's using some of this primitive neural stuff in your system in a way that helps you understand it."

You might extend it with a third dimension: interestingness.

Interestingness is "does the agent find it intrinsically motivating -- does it create or sap energy for them?"

Agents, given the choice, will naturally tend towards things they find interesting.

Interesting problems will get attention, naturally.

Most software has to assume a generic context--wide applicability to many different users and use cases.

Situated software assumes a highly bespoke context.

Someone who looks at it who is outside that specific context might think: "oh that's hideous and non viable!"

But the reality is it is perfectly, sublimely viable... in one very particular context.

The thing that makes situated software not just "good enough" but "sublimely perfect" is that the creator applied their own creative effort to solve precisely the problem they had.

Situated software is perfect for the creator because it is the creator acting like an owner.

When operating well, the presence of the latter does not absolve the creator of responsibility; it spurs them to do their best work.

The tension of the pairing creates something much better than either could create in isolation: a dynamic tension, a dialogue.

I've heard that some Greek philosophers thought it wasn't possible to truly think without dialogue.

Dialogue is the dynamic tension between two distinct perspectives.

That tension gives a place to get leverage from, to not just be a muddled average.

You need a boundary to have two distinct things.

Everything within a single boundary, over time, averages together, losing the variance and zeroing in on the centroid.

The idea is that I build a bespoke thing and then think, "I should build a framework to make this class of bespoke thing easier!"

I then tend to get carried away with the framework itself, making it clever and well crafted.

As I add features I become enamored of making the framework a beautiful object.

I start treating the framework not as a means, but an end.

However, when I go to "simply add a few examples" using the beautiful framework, I discover (to my horror) that using the framework feels like a headwind.

The point of the framework is not to look beautiful as an object sitting on shelf.

The point of a framework is to look beautiful because it is in use to unlock the creation of a large category of things that could not exist without it as a catalyst.

Before computers, there were all kinds of fuzzy protocols that allowed some slop and imprecision and "use your judgement" (e.g. how checks and credit cards were cleared).

There was a human in the loop, so you could rely on their judgement on the edge.

Then computers came along. They could do things on clock speeds many orders of magnitude faster than humans could do, but with a tradeoff: they couldn't handle fuzziness. At all.

When you designed a protocol for a computer, you had to be extremely clear about precisely how to handle every little edge case and error.

Protocol definition requires coordinating an ecosystem of distributed senders/receivers which is already hard; this requirement 100x'd the difficulty.

There are tons of very successful protocols that we've created in the last few decades, but at extreme cost and toil to define them.

There are a ton of possibly-useful protocols that don't exist, but could... with massive amounts of effort.

But LLMs for the first time can work roughly on the clock cycle of computers (or at the very least, a new one can be spun up at any time at low marginal cost, no matter when it is). And they can also handle fuzzy, reasonable behaviors like a human could.

This might allow an era full of new flourishing of protocols; by allowing them to be fuzzier we can radically reduce how hard they are to start and get off the ground

A lot of the party tricks LLMs can do are based on that last bit.

But that's also where the danger of LLMs lie, because they tend to hallucinate.

What if the first bit is more important, and you could get it more cheaply and with higher accuracy?

Patterns that use LLMs in this way, in small, safer-to-fail ways, as part of a larger system, might prove to be more resilient.

AI and crypto are Jungian shadows of each other.

They are both downstream of the abundance of GPUs.

One is defined by artificial scarcity, the other by artificial abundance.

We call this understanding knowhow. It is extraordinarily potent.

However, knowhow cannot be directly transmitted.

It can only be indirectly transmitted, at great cost and signal loss.

Writing.

Apprenticeship.

We have to transfer knowhow through a teensy tiny straw called language.

The only real way to grow knowhow is to take the time to directly experience the relevant situation yourself: a massive constraint.

LLMs are unlike humans in that their knowhow can be transferred to other models more directly (or in some cases just directly replicated).

This means that when a single LLM develops knowhow, it could quickly be absorbed by other LLMs around the world.

This could have massive implications for society.

Politics will be at their lowest when things are growing, and they will be at their default level for the organization when resources are static.

Politics in a given system are heavily influenced by politics in that system in the past; a bad period of brutal politics can leave battle scars, dead bodies, graveyards that echo through into the future.

The only way to interact with other components at scale is with metrics and summary statistics. Nuance is lost.

Everything with the very large boundary is averaged together into one low-variance mush.

Escaping this mushy average with something that stands out can become extraordinarily expensive.

You have to justify a different way of doing things that goes against the gravity well dynamics.

The effort to create seeds of greatness in an a sea of oppressive averageness is possible... but cannot be understood by the system itself.

"70% of your time you spend working as a role model for your function. 30% of your time you spend on kooky stuff. And then you regularly have lucky miracles. It would be easier to get promoted if you stopped wasting time on the kooky stuff."

"... How do you think I get so lucky?"

One thing large systems are very good at is hill-climbing the mountain they have already started successfully climbing, with huge momentum.

Kayfabe is a reality that within a system is omnipresent, but outside of it doesn't matter at all.

This difference is a large discontinuity, which is why leaving an organization can feel (briefly) like flying.