Bits and Bobs 1/12/26

We used to have bespoke software creation and industrial software outcomes.

With LLMs and infinite software, we'll have industrial software creation and bespoke software outcomes.

Notion AI: Unpatched Data Exfiltration.

IBM AI ('Bob') Downloads and Executes Malware.

ZombieAgent prompt injection in ChatGPT.

I think chatbots are an embarrassing party trick.

LLMs have vast untapped potential to create resonance for society.

The most valuable thing is having the taste of which subset is actually useful^[ct].

So it's your job to give them non-average gaps to fill, to inject the entropy.

If you ask it for a joke you will get one of ten hyper bland ones.

If you ask it for a joke about the pope, an orange, and Richard Feynman, you'll get something novel.

The first thing that can integrate with your life in a way you depend on will change the world.

The "replace workers with automated versions" default frame for AI is like pushing for faster horses.

That means that pointing it in the right direction is significantly more important than with a human.

What makes you as a human special when I can replace you with a script?

LLMs amplify whatever you apply them to.

But elevated implies the resonant, positive part of amplification.

Elevated Engineering.

0-1: Enabling people to do what they couldn't do before.

1-10: Helping people with relevant experience amplify it beyond what they could do before.

10-100: People who have not just relevant experience, but also the ability to change how they can work

It has to be understood by your collaborator to actually give you leverage.

LLMs are great at unpacking any jargon.

Its usefulness in a context is precisely proportional to how inscrutable it makes it to outsiders.

A flick of the wrist.

But you have to have the relevant experience to know the right jargon.

The craft is fun to execute.

When you industrialize you go from craft work to turning a crank.^[cx]

The result might be more predictable or scalable, but the joy of creation evaporates.

A feeling of ennui develops.

Simon Willison calls this, in the context of SWEs and LLMs, "deep blue."

Before, it was just "baking."

Before cell phones it was just a "phone."

The world moved on, and what used to be the default or even only way became a rare way.

We'll need the same kind of shift for software engineering.

There's the now quaint, non-default way, and the industrialized way.

Craft engineering?

Classic engineering?

But now with LLMs it will be the editor, the coach, the director who matter most.

Both use LLMs in new ways.

Vibecoding is a "make it work" mindset.

Elevated Engineering uses LLMs to extend your expertise.

Putting up the tentpoles and letting the LLM drape the canvas.

It was like eating your vegetables.

But you don't have to sell an LLM.

If you told a beleaguered human to eat their vegetables they might punch you!

It definitely undermines the business models of e.g. Tailwind.

But now engineers don't need to use libraries.

Open sourcing a library is a lot of effort.

It used to be that the amount of effort to write the code was large, and the amount to maintain it was large.

The complexity threshold that makes it worth open sourcing has risen significantly.

Is your default tendency laziness or curiosity?

That will decide your fate in the era of LLMs.

When you get to the top of an s-curve, it takes time to realize.

Each step gets a smaller benefit for the same input.

At the beginning it just feels like you've lost your touch.

Someone proposed it was Discord.

But that feels like a hollow answer to me.

The physical third places are coffee shops.

Over the last few decades, more and more coffee shops are Starbucks.

This feels the same, but is way more precarious.

A single MBA at Starbucks Corporate could make a decision that could in a snap remove the third places across the country.

To be resonant and load bearing, the third place must be open and federated, impossible for any one decider to destroy in an instant.

Like, say, working with fissile materials, or money transmission, or software security models.

There is significant downside if you get it wrong.

That means you need to be trained, and it might even need to be regulated.

But imagine someone creates a precision controlled robot arm.

Now, from a safe distance, you can operate the arm and work with dangerous but powerful materials, safely.

That precision robot arm is a hugely valuable asset that unlocks potential that was otherwise not possible.

The app model requires the receiver to trust the creator of the code.

That was reasonable when code was expensive and the creator was likely a business with something to lose.

But LLMs make it much more common for code to be written by some rando.

How far is the leverage of one motivated rando?

In the app model, it's maybe 2-5x.

But what if you could get a multiplier of 100x or even 1000x?

All vibe-coding tools today create software distributed in the normal add model, which gives a ceiling of 2-5x leverage.

You see all of the scaffolding as it's built.

It looks like any other building.

Then, once it's done, the scaffolding is removed… and it stands on its own!

A magic trick.

An app store is a load-bearing part of the app model.

It could also be temporary scaffolding to get a new negative-friction distribution model going.

That's where users are more sophisticated and willing to pay.

Getting to low-sophistication, low-willingness-to-pay consumers will be very hard for the foreseeable future.

Lots of vibecoding tools are focusing on making shitty software quickly.

That's a cul de sac.

Vibecoding is one ingredient into infinite software.

It's now commodity.

It's not even the most important ingredient, because now you can take it for granted.

Resonance is also engaging, it's just only the positive valence form.

Engagement can also be the hollow form.

If you don't specify which one, you'll go for the easier one: the hollow one.

Resonance is also a sign of PMF.

If it's too complex, then no one will keep it up to date.

If it's too simple, then it won't be useful.

There can only be one system of record, and all collaborators have to coordinate on one.

That is an inherent balancing act.

made a claude code skill that:

1. reads credit card statements/extracts subscriptions

2. automatically asks follow-up q's to clarify which ones you want to cancel

3. actually opens chrome and literally cancels them for you"

So much of software in the last 20 years has been predicated on software being expensive to write.

Software is expensive to write, but as the creator you get the data, which makes a moat.

But what if that first part doesn't apply anymore?

It aims for such a razor thin PMF to then grow via gradient descent.

Things that require the customer to change how they think require too much difference to be captured in that tiny step.

Imagine pushing the car with your feet.

People not looking closely will think the car is moving under its own power.

But actually it's being driven by something very manual and non-scalable.

Fake it 'til you make it.

Below that scale, you don't need metrics, and metrics are distracting and possibly misleading.

Past the critical scale though, you can't steer by touch anymore, so you need metrics to steer by sight.

Metrics are way worse than steering by touch, but they're the only approach that works past a certain scale.

Sudafed solves flu-like symptoms, so you can go into the office…

…and infect your coworkers, who also now need Sudafed.

That is their employment strategy.

Make their hiring bar high enough that working (and succeeding) in the environment is a credential.

Then just burn through people as quickly as possible, extracting as much as you can.^[cy]

A very Saruman-style employment philosophy.

I… don't love this approach.^[cz]

They're towering: they look great, but they're impossible to eat^[da].

She observed: the burger's main imperative is not to be eaten but to be sold.

Those are distinct!

In the modern era, that difference leads to instagrammable burgers that are impossible to eat.

It's another thing to distill complex topics to the level of Reader's Digest.

Past a certain level it requires an almost intellectually offensive distillation–but if you don't, you can't connect with readers who aren't experts and aren't that interested.

Give them something of value they didn't even have to ask for.

So why not put the expectations where you want them to go?

Research is default divergent.

Development is default convergent.

You need both, but separately and in harmony.

Push is default divergent.

But don't lose sight of the mission and climb the wrong hill.

If your capability is beyond your maturity, you'll get bored at the important lessons.

You won't receive the important lessons necessary to mature.

You can't believe in any ideal and also believe in that politician.

A slime mold can learn but not steer.

Steering comes from a brain.

Steering requires a centralized component with leverage.

"This matters so much that even if it would be a miracle to accomplish it we should still tilt at the windmills."

Personally, even if I care about the outcome a ton, I can't care about it unless I see a plausible theory of change.

Otherwise, I'm pouring effort into a sink that won't improve the world.

What is your highest and best use of your effort to improve the world?

"An open invitation is no invitation at all."

"The axe forgets, but the tree remembers."

"The reasonable man adapts himself to the world; the unreasonable one persists in trying to adapt the world to himself.

Therefore all progress depends on the unreasonable man."

There are two different kinds of unreasonable people: Sarumans and Radagasts.