Bits and Bobs 7/7/25

Of society in general (the weights; the background awareness).

Of the user specifically (the context; the questions the user brings to it).

They may distort things, but they're fundamentally a mirror.^{[fl][fm][fn][fo]}

More formatting than substance.

Punchy and "simple" while obscuring nuance.

Presents the answer as an objective, simple truth, as opposed to a nuanced observation as a point of departure and follow up.^[fp]

This comes from Bruce Schneier's excellent AI and Trust.

If it gets the answer wrong or doesn't do what you wanted, you can easily ask follow up questions.

One-shot UIs that have to give a perfect answer every time with no second tries are very hard to get to the bar of viability and then improve.

Compare Google Search and (spoken) Google Assistant interactions.

Then the next model gets better and obviates the UX.

What are the downsides fundamental to chat itself?

Those will help discover the ideal UX for LLMs.

That's what makes it good for starting open-ended tasks without it feeling like a burden.^[fr]

That's also what makes tasks as they continue feel like wading through molasses.

To me it's entirely based on the novelty of LLMs, everyone's just experimenting with them right now and starting open-ended tasks,^[fs] which chat excels at.

It's as the task goes on for longer and gets more structured that chat becomes more obviously not a good fit.^[ft]

LLMs are inherently open-ended, and chat is the most obvious and easiest way to grapple with that open-endedness.

But as we uncover other ways of doing it that lean more on the GUI, chat will look like just a feature, not a paradigm.^[fu][fv]

Chatbots seem like a local maxima to me.

Chatbots are not the ideal UX.

I think the ideal UX will include the ability to chat, but that will not be the primary interaction that everything orbits around.^[fw]

Today apps are static.

Now LLMs and this tech allows them to proactively adapt to what you want to do.^[fx][fy]

Over time you'll not even reach for apps at all because everything you do will be adapted so much to your situation that the notion of an app fades away.

Apps that are alive, that adapt themselves to me, pulling in context or code to make themselves do what I want to do.

I could tweak them to my heart's desire.

Apps for living.

They could also be called "coactive apps".

Alternatively, call them tools. Living Tools, or Coactive Tools.

'Tools' doesn't have the expectation of "these are just like apps of today."^[ga]

You get a full, useful thing out of the box, that an expert designed to be holistically useful.

But it's made of legos, so you can replace any block… or add whole new things to it.

So then creators can create both new lego blocks... but also new pre-assembled lego sets that all fit together nicely and coherently and are useful right out of the box.^[gb]

Is it drawing on relevant facts about you?

The quality of the context is a big determinant in how good the results are.

This is one of the reasons the ChatGPT dossier memory feels off to me.

You can't inspect the context to say, "include this, not that".

You can imagine a UX where there's a coactive context drawer at the top of the interaction.

Things in the interaction and that context drawer are all that is given to the LLM, nothing else.

When the drawer is collapsed it shows a short summary.

When you expand it you see trees of context that you've pulled in explicitly.

You can add in trees of context on demand easily.

You can delete any tree of context that's not relevant.

There's also a list of auto-included context that the system guesses is useful.

The ranking function is: how well does the system predict which trees of context to include, (predicting whether the user would accept or delete a suggestion?)

The user choosing to include or excluding bits from the suggested context is an extremely powerful ranking function.

The user wouldn't even realize that they're training the system for themselves and others by gardening their context.

Only a small number of users would need to do it to help tune it for a whole population.

Ideally you'd organize it for yourself, which would be useful on its own.

As a bonus it makes the LLM significantly better at doing things for you.

There's distillation.

How can you get a good enough model of yourself for the LLM?

That gets better as you use it more, not because some PM added a feature or some fully automatic LLM-based insights.

But because it draws on the collective wisdom of everyone using it.

An extremely trivial script at the right moment can have a massive impact.

Situated software catalyzes cozy potential.

Largely due to gardening the right context and the right coactive UI.

At first you go "Whoa, this is so easy to use, I can do 10x more than I could before."

But as you use it for more things you realize it's too coarse a tool to do detail work.

An extremely useful tool, but doesn't replace all of your tasks in the garden.

"Now a Burbclave, that's the place to live. A city-state with its own constitution, a border, laws, cops, everything."

1) Cynical engagement-maxing tech products of today.

2) Crypto.

3) E/Acc / Successionism.

People presume that if you're optimistic about tech and are in the industry you just want centralization.

Or that if you're optimistic about tech and don't want centralization then you must like crypto.

But it's possible to be optimistic about tech and push for neither centralization nor crypto.

Such a third way is more important than ever before in the era of AI.

Name it based on what people know they want.

Then slingshot them to the thing they didn't know they needed.

That latter part only becomes clear once they've used it.

Meet them where they are to take them to where they should go.

Useful when you have a thing that is superficially like other things, but fundamentally better in a novel way.^[gc][gd]

When you access a method, the LLM generates code JIT.

I think the mental model of gremlins fits well: small, not too powerful, but mischievous and a bunch of them together can make an impact.

But in the deep philosophical sense, your real context is outside your control.

Your context is a gravity field.

The water you swim in.

The backlog of Bits and Bobs feels exceptionally valuable to me in the age of AI.

That's why one of the best ways to understand something is to teach it.

No individual step is that bad, obscured by noise.

The bias is in one direction: the gravity of incentives.

So the emergent global effect is clear and powerful.

Before, we had mass surveillance, but a human sifting through the collected data happened rarely.

But LLMs give qualitative insights at quantitative scale.^[ge]

Society now has the technology to get qualitative insights at scale from that surveillance.

What could possibly go wrong??

Before your data was protected a bit by security through obscurity.

But LLMs are infinitely patient to sift through data that was accidentally left open.

Another outgrowth of the "qualitative insights at quantitative scale".

Examples in this Ezra Klein interview.

"New social networks are going to appear that will be LLMs creating a cozy web customized for us and our real friends, and their friends and so on

There will be a new social network built on mutual trust, all curated by machines of loving grace

Personal Cozyweb is inevitable"

I originally interpreted this tweet as "Use LLMs to make a psychosocial bubble of fake friends" which I think would be terrible for society.

But I think he meant it more as "make a garden of possibility for you and you friends," which could be good in some circumstances.

Trust will have to be rooted in in-person interactions that are known to be authentic.

Bruce Schneier's focus on integrity is important here.

Someone tries something that other people find viable and then the others reshare it, build on it, and remix it.

The things people like are what get built on, emergently.

For things like architectural styles there's some geographic pockets which helped give distinctive styles in different parts of the country.

As everything gets more fluid and lower friction, we'll likely see less and less distinctive architectural styles.

Imagine working at a company where on day 1 of a product you can expect 50M users.

You have to think about every little detail before building it.

But in most contexts, the 0-1 phase has very little usage.

The product mindset is "do I want to use it?"

Very different bars!

Polish isn't novel.

What kind of thinking are you not outsourcing?

To get it you need to sit with the discomfort of uncertainty.

The discomfort is the point.

Reframe the discomfort as excitement!

That is, what a term means is what the majority thinks it means after the first time they've heard it.

People will bring their own preexisting priors to any given term, and that bias will lead to what the term means, especially if it's a consistent bias many first-time hearers will share.

Terms like "context engineering" are useful because they mean the thing that most experts hearing it for the first time would think it means.

"Inferred definition" is itself a term coined by Simon Willison.

Founder led companies are more volatile.

Non-founder led companies are harder for the CEO to steer.

But even founder-led companies are hard to steer at scale.

Before, the swarm of employees trying to implement the CEO's vision imperfectly gave some insulation.

But if every employee is just a clone with minimal principal agent problem, it's like the Wreck it Ralph 2 swarm of poorly rendered copies creating an unruly emergent leviathan.

The former is a zombie organization.

Alignment at all costs, even if it kills the host.

Top-down alignment comes at the cost of local adaptivity and the possibility of emergence.

But they're extremely meaningful to the people who were there.

Having the experience of that meeting in that space gives you the key to understand it.

It's noise.

Some novelty will turn out to be innovation.

Most will simply not work.

Invest your novelty budget on the things that are differentiators, but nowhere else.

Multi-headed search vs single-headed search.

If you can only have a single search head on a problem domain, you want it to be the best it can be.

If you can have a multi-headed search on a problem domain, you want as many heads as you can get to flood-fill the problem.

The swarm needs breadth, the individual needs depth.^[gf]

Locality breeds creativity.

It's the pockets that allow interesting results to grow and then expand.

Some tend towards a center point.

Some tend away from a center point.

Two systems that look superficially the same but are convex vs concave have infinitely different outcomes.

What determines if a system is convex or concave?

I think it's whether the locally good behaviors lead to emergently good outcomes at the macro level.

In some cases the more useful they are, the more fractally complex they get.

Convex vs concave.

In proportion to:

As it gets bigger it gets more useful to more people because more people are exposed to it.

In emergent systems, all decisions are local but they have emergent global consequences.

You can't get a bird's eye view to coordinate, which is necessary for top-down convergence in a large system.

Many problems can't be framed this way, with a successful macro level outcome without a birds eye view, but some subset can.

These are typically grown, budded off of other systems that are working.

But a true bird's eye perspective is an impossibility anyway.

The constraint of "local information only" feels overly restrictive, but it's close to the real constraint anyway.

You can't simulate what the world will do without taking the action.

The world is a multi-headed environment of execution.

Your head is a single-headed environment.

An inode-level debugging tool for a userland storage system is not helpful.

Good debugging tools give good leverage to features / bugs at that layer.

Related to the carrot problem.

Every team or product has one.

You'd print words in a book and then ship the book.

When data moved to the plane of entirely bits it got orders of magnitude faster.

We saw much more aggregation much faster.

To make it so you can't do the things you fear you'll want to do.

Past you can constrain future you, to make sure your reptile mind doesn't override your aspirational mind.

A game theory solution that's similar to playing chicken with yourself.