This piece from Kevin Kelly is both an interesting future-look at what LLMs may have in store for us (with Kevin's typical energizing optimism), and a great layman's primer on what the tech itself is actually doing — on how LLMs work in general.

Latent space acts as a continuous map of the possible. Most of those possible things don’t exist — yet. The space of what we know, for instance known materials, known proteins, known chess plays, known ways to paint, fill only fragmented, patchy spots with plenty of white spaces between them. The white spaces between known things are unknown to us, but they are already mapped in latent space. We now have new tools to explore these white spaces in a systematic way. What lies in between astronomy and astrology? What gems await in between bluegrass music and ballet? What about in between the notion of corporations and the theory of Gaia? Exploring latent space is the new frontier. Invention shifts from “think of something new” to “prospect in the gaps.” When a gap or hole appears the question becomes: is that gap empty because it’s impossible, because it’s unfashionable, or because nobody’s looked yet?

In the current pessimist doomer AI conversation zeitgeist, it's exciting to now and then think about the frontiers we've yet to explore.

A good reminder from The Culturist to spend more time creating new things than consuming from others.

Nowadays, it is frighteningly easy to spend the best part of a day consuming the creations of others: listening, streaming, scrolling. For most of us, the ratio of consumption to creation has never been more lopsided — and yet, the tools of creation are more accessible than ever.

Let us establish at the outset that consuming art is itself a worthy pursuit. But if this is all you do, what are you giving up?

G.K. Chesterton once wrote that a culture has become decadent when its people prefer to pay others to dance for them, rather than a whole room of people doing the dancing for themselves.

In the modern culture with nigh-infinite things to watch, read, listen to, it's easy to mistake consumption for something worthwhile on its own, and have it edge out the time you spend creating for yourself. Even a few minutes writing or drawing triggers a muscle inside our minds that stays dormant if all we're doing is "engaging" at a superficial level.

Instead of merely reading a book, you should be writing out your thoughts on it, too. You could even do this with movies and TV. Just get started making something of your own.

Just read this take from Elad Gil:

AI will eat closed loops first

AI will first automate away the things that are easier to form a closed loop learning system on. This is why code and AI research may be accelerated and then displaced quickly - you can have testable closed loop systems so machines can learn and iterate quickly. The tighter the closed loop, the faster the AI can learn. You can make a 2x2 of jobs by how closed loop they can be made, versus their economic value, and see where AI may impact labor fastest. Fast time to closed loop + high economic value = fastest AI impact (hence software engineering).

For reinforcement learning to work in model training, you need a source of reinforcement. So the things that the models are best at right now are those that are verifiable, fields and domains where the answer can be known to a high degree of certainty, and the model's output checked against reality.

It's one of the reasons code has been such a success for AI so far: the tasks are verifiable, and the output high value. There are plenty of closed loops out there to attack. Code just happened to be one of the largest, most generally useful, and in highest-demand.

The writing and testing of the code is a closed loop, but the problem toward which to apply attention — what to build? why? for whom? — is still very much an open one.

What other open loops still demand deep human involvement?

An interesting piece from Joseph Heath that compares the opposed theories on the origins of human intelligence posited by Yuval Noah Harari and Joseph Henrich:

What makes the field of human evolutionary theory so interesting right now is that, deep down, we really don’t know how any of these capabilities evolved. At some level they must have been adaptive, but nobody really understands specifically why or how any of it was adaptive. On the other hand, we know a great deal about how it could not have evolved, because so many bad theories have been proposed over the years, which have not stood up to careful scrutiny. Typically they start out as speculative theories, until someone comes along and says “show me a model of how that could have evolved,” and someone takes the bait, starts building a model... and then realizes that it couldn’t have evolved that way. Indeed, it has turned out to be so difficult to produce an evolutionary model that could plausibly explain any the items on the above list that formal modeling has been where the most interesting action has occurred over the past several decades of research.

Harari's model proposes the sequence of development working like this:

intelligence → language → cooperation → culture

Humans first evolved intelligence levels capable of then developing complex language. Then social cooperation and development of culture flowed from these changes.

Henrich inverted the timeline:

culture → cooperation → language → intelligence

In his view, blind mimicry led to complex cultural behavior — think of how a child merely mimics behavior in its vicinity, without explanation or understanding of why. This tendency of humans to mimic one anothers' actions without understanding eventually led to cooperative behavior. Then as cooperation improved and became more advanced, more efficient means of rich communication developed to improve cooperative outcomes.

It's fitting to me that someone with the technocratic, utopian worldview of Harari would begin from the position of intelligence-first: "we became really smart, then started doing amazing things." His priors in other domains seem to be oriented toward the progressive "anything can be engineered" mindset. So it's understandable how he arrived at his model.

I read Henrich as more of a "human nature is complicated" guy. So his inversion is also understandable. An immensely complex series of events over a long period of time led to the complexification of human intelligence.

Kids don’t need their every moment curated. They need space to explore and follow their curiosity, unguided by parents.

Our job is to provide the right safety net without interfering in their discovery.

In the latest piece on Res Extensa, I think about how I've made an effort to give my kids the maximum "free roaming" space, without letting them dangerously imperil themselves.

With AI-generated work — writing, visual art, music — there's something missing that we can often identify viscerally (at least if we spend a few minutes paying attention), though not articulate exactly what it is that's missing.

Nabeel Qureshi attempts to identify those aspects of great art that make it unique, interesting, and the unique domain of humans.

There’s a feeling of limitless depth that is unique to the really great works: it feels like you can always find more patterns, more coherence, more beauty on rereading. Or that more readings will give you a deeper understanding, but never a perfect or complete one.

I read a lot, and most of what I read flows through Readwise. If you haven't used it. Readwise (and its companion app Reader) captures highlights from your Kindle, articles, RSS, and even YouTube transcripts. Everything you highlight ends up in a single searchable library, and emails you a "daily review" to revisit your highlights. If you read, and have trouble remembering what you read, use Readwise. It'll change you.

But I wanted a way to bring the ideas I've clipped from readings more accessible during my writing and thinking. Obsidian is my "thinking OS", and my best solution for this the past several years was syncing highlights as notes in Obsidian. It's great, but it's still up to me to figure out how to bring those into my personal "context window" while writing notes.

And since we're now in the era of "personal software," I decided to build my own plugin to solve this.

Surface is an Obsidian plugin that runs semantic search over your Readwise highlights (powered by Readwise's CLI) from inside Obsidian, without leaving your current note. Open a note, and the Surface sidebar recalls quotes and articles related to the current note's content, ranked by vector similarity to whatever you're working on. Switch notes and it re-runs. Select text and it searches just that snippet. It's been huge for sifting through my 18,000+ highlights to find the good, relevant stuff. No need to go digging manually.

As I was working on this, I found myself running across a great highlight, but sometimes one sort of adjacent or tangential to the core idea I started on. So I added a feature called "Go Deeper," where any highlight surfaced becomes its own search term, letting you follow chains of ideas through your reading history, with a way to to retrace your path back up the tunnel. When I'm writing a draft and hit on a thread worth pulling, I can hop through ideas in the history without friction.

A few other cool things:

• Search by selection — it doesn't just work on the whole note as context, you can also select a text block, right-click and run "Search Readwise for selection" to scope the query to just that snippet
• Insert as callout — any highlight can drop into the active note as a > [!quote] block with title, author, and source, in one click
• Jump to source — open the highlight in Readwise, or the original article URL, without leaving your notes

The plugin relies on the Readwise CLI, which includes vector search and JSON output. Surface is a wrapper around the CLI native to Obsidian that builds a search term from the active note (title, frontmatter tags/topics, headings, and the first paragraph), passes it to the CLI, and renders the results as cards.

I keep coming back to the value of personal corpus search. Your reading history is one of the most context-rich datasets you have, and it sits unused most of the time. Readwise already lets you review your highlights regularly with their Daily Review. But Surface makes the archive relevant during your thinking and writing process.

The easiest way to install is use the BRAT plugin for Obsidian ( and use the path colemanm/obsidian-surface). You'll need the Readwise CLI installed and authenticated — instructions are in the README.

The latest issue of Res Extensa draws parallels between the story of Ford and Rolls-Royce and Stewart Brand's framework of Low Road and High Road design, from How Buildings Learn.

We have polar opposite philosophies of design and construction that each is capable of producing things that well outlast their owners, and this applies to buildings, vehicles, and anything else we bring into our lives.

The Rolls endured through reliability. The Model T endured through repairability.

One inspired the devotion of its owner. The other devoted itself to its owner.

Using a combination of Claude, Opus, Nano Banana, and lots of satellite imagery, Andy Coenen built this incredible pixel art map of New York City. It's like flying over a realistic version of SimCity 2000.

In this post he details his process for building it.

See the full resolution, zoomable map here.

The folks over at Every published this excellent guide to a new paradigm for software, using what they call "agent-native architecture."

In short the idea is that your mental model for thinking about what you're building shouldn't be focused only on user interaction, but anything you build should be usable by an agent, too.

So software should now be built around concepts like composability, reusable, callable primitives, and emergence. Sometimes an AI can find uses and things your product can do (and that users want to do) that you didn't plan for. Thinking agent-native will have you building and prioritizing things with a different architecture in mind.

When we can design software around users setting goals and having agents pursue them, we end up with something much different than thinking only in database schemas and CRUD.

Luca Dellanna on Resistance:

The point is that going to the gym is an activity, like many others, where you can technically produce its inputs in a negative mental state, but they won’t be good enough to bring satisfying results unless done with confidence. Therefore, chronic resistance is not something to dismiss or power through, but something to acknowledge, dissect, and address until resolved.

"Just powering through" resistance makes it more difficult to resolve its causes. In my view, the persistence and grit you develop getting used to powering through is a skill in itself, but Luca's point is that resistance is telling you something about its causes. If you think deeply enough about the chain of causes, you can diagnose and treat at least certain aspects of the resistance to lessen its effect over time.

I started out the year with the goal to run 1,000 miles. No other rules or restrictions, just get the miles in.

That meant keeping a steady pace throughout the whole year. It worked out to 2.74 miles/day, the hardest annual goal I've ever set by almost double. I did 650 a few years back.

I didn't want to push the mileage too hard for fear of getting injured, but also didn't want to lag more than a few miles off the pace in case I ended up sick or otherwise indisposed.

Here are some of the notable stats on the year.

Totals

Stat	Value
Total Runs	174
Total Miles	1,003.26 mi
Total Time	167 hours (7 full days)
Total Calories	119,362
Total Elevation Gain	32,180 ft

Averages

Metric	Value
Miles per activity	5.77 mi
Running time	57:37
Average pace	9:59/mi
Average heart rate	151 bpm
Temperature	79°F

Extremes

Metric	Value
Longest run	40.03 miles (Pinellas Trail Challenge)
Max heart rate	178 bpm
Hottest run	93°F
Coldest run	44°F