Bits and Bobs 1/6/25

My first reaction was: wow, the LLM is pretty funny!

But I realize that that's partially an illusion.

The LLM, working in tandem with the human, is pretty funny.

The human is deciding how to prompt the LLM, and which things the LLM produces that are compelling enough to share.

Presumably most of the output was pretty bland or uninspired, but we see only the highlights, from the human's curatorial taste.

Humans only bother to share things they think are good. That curation is an act of creation^[alt].^[alu]

The combination of a savvy, high taste human and an LLM are a powerful creative duo, much stronger than either would be alone.

Irreducibly co-creative acts.

The human as supervisor or editor to the LLM.^{[alv][alw][alx][aly]}

In the past few months I've had a number of revelatory, multi-hour conversations with LLMs where I explore and pull on threads of my curiosity, in some cases leading to deep insights that reconfigure my mental model of the world or myself.^{[alz][ama][amb]}

These conversations^[amc][amd] are long and meandering. They have various dead ends or odd paths that I later decide aren't useful.

These aren't like an essay or a simple conversation with a collaborator; they are creative, open-ended insight-generating interactions.

I'd love to be able to prune past parts of the conversation that I later realized were tangents, or fork a pre-existing conversation midway through and see how I could take it in new directions.

You can cobble these kinds of interactions together by starting new conversations, or using Anthropic's projects, or manually cobbling together tools on top of the raw API.

The chat UI is too literal, like you're talking to any other normal human.^[amf][amg]

These kinds of open-ended exploratory conversations are just one killer use for LLMs, and even though they're a pretty close fit for a conversation-style chat, it still feels constricting!

Clear raw potential that it takes specialist skills to unlock and apply.

The command line was a remnant of an older interaction type of teletype.

The GUI unlocked new kinds of interactions that only made sense in a world of interactive computers^[amj].

What will the equivalent of a GUI for an LLM be^[amk]?

Creating the equivalent to the first GUI is a creative act wildly unlike building a new product in an existing paradigm. It requires multi-ply thinking.

For example, download three different popular templates for Project Management from Notion, and provide them to the LLM, and ask it to find commonalities.

The things that every good Project Management template agrees on is much more likely to be fundamentally, inescapably important in that domain^[aml], rather than just a random detail from that template's creator.

The synthesis across lots of different examples is a time-consuming, boring task… but LLMs never get bored.

Normal programs are close-ended^[amm]: given similar inputs they will behave similarly.

The open-endedness of LLMs is what gives them their human-like creativity and ability to apply human-style common sense.

But the open-endedness is also what makes them hard to use in a process you want to be predictable.

Most of the applications of software today are in contexts where we want each run to be predictable, and as we evolve the software to handle more and more edge cases, to run reliably and increasingly without need for human intervention.

Creativity is a double edged sword; it might have a clever way to handle an unexpected input, or it might do something weird you don't like.

Another option for using LLMs is not to have them live in the main loop of execution, but rather use them for continuous compilation.

Give a clear, convergent goal for the software to execute.

When something unexpected happens, call up the LLM and give it the error message and have it tweak the software to handle it.

The LLM is tuning close-ended software.

This kind of approach works well in cases where you could imagine writing an operator manual for a call center employee: close-ended, but requires a bit of human-level judgement.

"AI shines in these kinds of situations where "good enough is good enough"

Not the optimal workout plan, not the best workout app…

Just something reasonable, fully tailored to me, ready in minutes."

"good enough" has to be radically cheaper (in cost, friction etc) than a "great" option.

But I'd take a real good enough thing than a theoretical great thing (that I'd have to build, configure, etc) any day of the week.

Good enough is a tradeoff not with a theoretical best, but best available.

Good enough is the toehold of being viable, worth using.

From there if there's an iterative process you can climb up to get to greatness.

Today software is so expensive to create and so that is where all the power accrues.

We send our data to aggregate on the software creator's turf because the software is so expensive.

In a world where small bits of crappy software are effectively free, that dynamic will flip on its head.

LLMs are like electricity^[ams]: a general technology that reconfigures the costs of key inputs, making wildly new kinds of things feasible.

LLMs are good for a number of things, like applying human-level common sense.

Now that this input has gotten so cheap, it's a long journey to figure out what to build with them.

We are in the early innings!

The LLM model providers are the electricity providers.

LLMs give a slice of the content represented on the internet, which has a strong western, English bias.

She imagines the experience of someone from a different culture talking with an LLM to be like watching a Hollywood movie^[amt].

She predicts that as LLMs become more ingrained in society, the fact they reflect one particular culture so strongly will become a major issue.^[amu]

A specialist is optimized for their niche.

That allows them to outcompete others in their niche, but be worse than others at competing outside their niche.

That means that specialists are less able to handle the disappearance of their niche.

A generalist (especially one that has a low caloric need, due to small size or metabolic rate) is more resilient; more likely to be viable no matter what happens in the environment.

Disruptive in the world of biology might be a new land mass opens up.

In the world of technology it might mean a new disruptive technology, like electricity, jet engines, or LLMs.

With all of the new territory available, there is massive expansion as species diffuse out to find and colonize the newly available niches.

Later, once all of the niches are colonized, competition heats up within those niches.

Competition over scarce territory leads to more efficiency at the cost of variety.

This riff is inspired by this video: Why do Animals Look so Strange After Mass Extinctions

That is, weird local maxima within local maxima.

When you look up you are like "wait where are we? This obviously doesn't make sense."

A/B testing doesn't have a vision, it's just the path of least resistance.

If you follow the path of least resistance many many times, you end up in incoherent places, lost from any semblance of something bigger.

Just the "most".

The most what?

Whatever catchment basin you happened to start in.

Normally in evolutionary contexts, the stimuli an organism is exposed to in their environment changes at a similar speed as their own evolution.

This allows their response to a stimuli to be well-calibrated to the stimuli.

But sometimes the environment can change much more quickly than the species can evolve, and produce "supernormal" stimuli that activate the response way more strongly than ideal.

When humans created language, which allowed mutating and passing down ideas orders of magnitude faster than biological evolution, it created a machine that could produce supernormal stimuli.

Modern commerce is extremely good at producing these supernormal stimuli that we find irresistible because they are orders of magnitude more potent than the stimuli we evolved for.

Examples include modern junk food and also our informational junk food.

Capitalism is great for giving users what they want, not what they want to want.

This riff is inspired by this video: Why are Lemurs Terrified of Predators that don't Exist?

It dives into why even novel, fresh thinkers over time become simplistic, hollowed out versions of themselves.

It feels to me like regression to the mean is an inescapable phenomena, perhaps as fundamental as (or directly implied by) the second law of thermodynamics.

Some phenomena are emergent, inescapable.

Nothing you do will change them.

All of the effort of being mad at the weather is wasted.

If you believe, you are willing to trust without looking closely, to give the benefit of the doubt.

Some creative endeavors simply require belief.

No amount of details can make someone believe.^[amy]

Details and belief are two totally different kinds of things.

The world will be constantly nerd sniping you; you're engaged and curious enough to see the interesting threads to pull on all around you.

A product and a community are two fundamentally different ends of a spectrum.

Scenius (the fullest embodiment of a community) is about putting the community above all else.

To not just tell them that they're great just the way they are, but to give them the space, encouragement, and drive to be not just good but great.

That can only give you middling results.

Make a team dynamic that makes it great.

That's the only way to achieve greatness together.

Instead of BMW killing themselves to make the best car, they have an ecosystem of suppliers killing themselves to be a competitive part of the best car.^[amz]

This is a great example of harnessing the power of the swarm within yourself.

Although presumably it only works if you can figure out how to make the whole be significantly more than the sum of its parts.

For example, BMW's brand implies that they vouch for any of the components used in the car; part of their creative act on the final car is making decisions about what components should exist and what is good enough.

To get momentum you have to have an asymmetric draw that pulls the group in a direction together.

The key question is: what do you do when that happens?

1) Find situational factors or an external villain to blame for your incompetence.

2) Step down from the role back to your level of competence.

3) See it as a growth opportunity and challenge yourself to become competent.

Just the way it works, the water you swim in.

Don't even realize it could work a different way somewhere else.

So if they respond to them in a similar way they will have coherent behavior without any explicit coordination.

Coordinated, coherent behavior, automatically.

Because they're playful and flexible.

But that's wrong, they are surfing it, totally in control, responding to the situation as it evolves, in a pitch perfect, calibrated way.^[ana][anb]

The rules have changed in ways that won't become obvious for some time.

Following the rules will get you stuck in local maxima or even worse, if the rules no longer work the way you thought they did in the old world.

After disruptive events, reality reconfigures, and massive new outcomes become possible.

The smartest thing is to not be too beholden to playbooks and rules from the pre disruptive context.

You can't do it in a single shot.

If you have a spotlight on you, it's hard to get multiple shots.

If you just listen to the logic of the woods you'll learn to survive, but also become feral and lose your soul.

One reason is the audio quality.

But an even bigger part is that with video you can get real time feedback of how the other person is receiving what you're saying.

You can verify that they are listening and absorbing, and know if you need to retransmit, slow down, or speed up.

You get a live signal that tells you how fast to unspool your thoughts so they get it, what you need to modulate to make it land as you intended.

On the phone you can't tell if they're distracted, or the connection dropped out.^[anc]

You have to wait until you're done with your utterance to get any signal about how it was received.

You need to be individuals, collaborating, dancing together to have a dynamic outcome.

Being individual while still being a pair together.

Not a boring, fused entity of double the size, but a different kind of thing than any individual could be.

This can only happen after a "PMF" moment.

Before that, it's all chaos all the time, constant black swans.

The amount of process should be inversely proportional to the amount of surprise, and before that moment of traction, surprise is constant.

But in a smaller, more chaotic environment, whether you like it or not, you're cohabitating.

There are no clear lines.

That's much more similar to family dynamics than neighbor dynamics.

Everyone has a boss. Even if you don't, the market, the constraints, your investors all might be your 'boss'.

None of us are an island.

The constraints define what we can do... but also give us a structure to operate within.

"I'd be able to do this if freed from constraints" is an impossible bar to clear.

You'll chase it, fruitlessly.

Constraints give us structure to climb on, otherwise it would be a free floating dead space with no way to get leverage and make any movement.

To be engaged is to go above and beyond, act like an owner, invest discretionary effort.

Going above and beyond is a choice that has to come from within.

You can create the conditions where people are more likely to make the choice but you're not able to ever compel it.

The ideal response is: "I didn't choose the constraints I was within, but I did choose what to do within those constraints, and I accept responsibility for those choices."

When you accept responsibility for those choices you can grow.

Now you can ask yourself how you might have chosen differently within those constraints that would have led to better outcomes.

Converting the life force to profit is a way to make a quick buck, but it leads the living thing to become a husk of themselves, inches from death.

In the meantime, you might congratulate yourself for how much value you're creating.

No, you're harvesting it, in the most vampiric way you can.

People who see the business not as an end to create something of value in society, but only as a means to make money.

Rigor means that the closer someone looks at the answer, the more compelling it becomes.

Sometimes very simple conclusions can be enormously clarifying: inescapable, obvious, once you see it you can't unsee it.

In some cases, the simple words of "not yet" are extremely clarifying.

You know you're getting closer to rigor and clarity as things get simpler, not more complex.

A thing not predicted by your model, whose existence implies your model, as it exists, is wrong.

It is only by realizing your model is not right that you can hope to improve it.

A dysfunctional environment will be constantly ignoring or sweeping under the rug disconfirming evidence.

It is only with sources of disconfirming evidence that a thing can become great.

It is a process to produce these little bits of disconfirming evidence.

Tension is not abrasion^[ane].

Tension is stressful but does not necessarily turn up the raw material to become great.

Abrasion brings the disconfirming evidence to the surface.

Healthy abrasion helps make sure the group is actually dealing with disconfirming evidence.

Abrasion to turn up little slivers of disconfirming evidence.

Agility to not get bogged down and over-complexify, to be willing to say "not yet" or focus on pragmatic responses.

Resolution to absorb and distill the disconfirming evidence into a new shared, improved model.^{[ang][anh][ani]}

These three ingredients allow high quality, confident momentum in a team.

In the best environments that abrasion is the source of disconfirming evidence that makes the team great.

In the worst environments that abrasion is pure intimidation.

The question is how to create an environment where the naturally abrasive person can cause the former, not the latter.

It's the feeling of a wave function of possible interpretations all collapsing into one clear answer with a jolt.

In some cases it's delightful and engaging, it's naturally viral.

In other cases, it's unsettling and odd, like the AI videos with weird interpretations of the scene, or the dolly-zoom from Jaws.

The aha moment is adjacent to humor.

Most people find cognition or critical thinking to be exhausting, boring, a chore.^[ank]

Some weirdos (like me) love the feeling of cognition so much that they have a need for it.

I also realize that a lot of the people I seek out as intellectual collaborators also have a deep need for cognition.

People with this need make wonderful open-ended "yes, and" collaborative debate partners, because they get more energy the more you explore the space.

"A buddhist flavored lens on why play is where you do your best work:

The phenomenon of doing your best work through play is tied up in 2 buddhist concepts.

Vedanā (feeling tone), and Shoshin (beginners mind^{[anm][ann][ano]}).

When you're in a professional setting, you're often chasing deadlines or trying to meet some external goal. This alters the vedanā, and scopes your attention into a mode of "seriousness".

Similarly, when you inevitably become the resident expert [insert topic here], your collaborators rely on you to be "the person that knows the thing", you can't practice shoshin, you are forced to be the stable channel of wisdom or best practice.

I suspect the only way to consistently practice play during the work day is to schedule unstructured play time.

Make space, not plans."