Makes more sense to question about the relative levels of consciousness between AI and humans, rather than simply asking if AI is conscious.
thread
Where Does Email Sovereignty Live?
When people talk about wanting sovereignty over their email, running the server themselves seems like a reasonable place to start. It’s your computer, your software, your messages. But when you follow what it takes to keep the thing working, you may find that the sovereignty we want is somewhere else in the system. And we have taken on quite a lot of work getting there.
Email has an interesting economic arrangement. Most of us get it for free, and there are business reasons for companies to offer that. Your email account becomes the account you use for their other products, and eventually there may be something they can charge you for. There are also providers whose arrangement is more boring: you pay them for email. In both cases, the cost of maintaining the infrastructure is spread across many people. Once someone is already operating the service, another personal mailbox is a relatively small addition. If you do it yourself, you are now operating a service for one person. ;)
This is mostly economies of scale and specialization. There is a comparative advantage argument too, when you consider what else you could be doing with the time. You may know how to run a mail server, but the provider already has people, systems, and experience devoted to it. Even if you enjoy the initial setup, you should probably count the maintenance you won’t enjoy six months later when something stops working at an inconvenient time.
Let’s say we do it anyway. We need a domain name, an IP address, and a computer with CPU, memory, and storage. Rent a machine, install the mail software, configure the domain, and we can get something running. But we want the email to work when the computer doesn’t. Machines go down, hosting companies do maintenance, and sometimes we are the maintenance problem ourselves. An upgrade can break the configuration. A disk can fail. Other mail servers may retry delivery during an outage, but meanwhile we still want to read our messages and send replies.
So we might want another machine. Preferably somewhere that won’t go down for the same reason as the first one. Now we have to keep the necessary state consistent between them, and we still need backups because replication will quite efficiently replicate our mistakes too. If you delete something and the deletion reaches the other server, you have two computers agreeing that the thing is gone. What you wanted was a way to recover it. That needs a different arrangement, and you should probably try restoring from it before depending on it.
Then there’s the part where everyone else has to accept our email. A server can be configured correctly and still have trouble reaching someone’s inbox. The receiving service makes its own judgment about the sender, including the reputation of the IP address and domain. It doesn’t owe us an inbox placement because we followed the installation instructions. People talk about warming up an address, but the complete decision process inside these services is not transparent, and sending a thousand emails cannot simply mean we have manufactured trust. We are running one end of a system that requires other people’s cooperation.
And all this software has a half-life. We have to follow security updates, understand what an upgrade changes, and know how to recover when it fails. This is where paying an email provider begins to look quite reasonable. They get to spread this work across their customers, while you get to stop thinking about whether your inbox is going to survive the weekend. It is easy to compare the subscription with the price of a small server and forget that you are supplying the rest of the service yourself.
But maybe you are willing to pay that price because you want sovereignty. Okay, where is the machine running? If it’s on a cloud provider, you are still depending on a company that can interrupt your access to the infrastructure. If you lose that account, where are your backups? Hopefully not exclusively inside the same account. Can you rent another machine and restore the service? Do you have the configuration, the messages, and access to the domain? Will the replacement server be able to deliver mail successfully? There is a recovery process here, and owning the installation doesn’t mean you have prepared for it.
You could put the computer in your house. Now you are depending on your electricity, internet connection, hardware, and the physical security of the house. For a lot of ordinary uses this may be fine. But if you are an investigative journalist concerned about authorities in the country you live in, bringing the server into your basement could be a curious interpretation of protecting it. The people you are worried about may now be able to reach the computer more easily. Hosting it somewhere else introduces another jurisdiction, another provider, and another set of relationships to consider. We started with an email address and have arrived at a small foreign policy exercise.
This is where it helps to ask what we actually want to remain in control of. If you want to keep your address when a provider stops working for you, a domain you control gives you a way to move the mailbox elsewhere. If you want to retain your old messages, you can keep independent copies in a form you can use again. The migration will still take work, and you are still depending on the domain registration, but you have made the email provider replaceable. That gives you a meaningful amount of power without operating the servers yourself.
A personal domain also shows how these requirements can disagree with one another. You might want people to recognize the same address for twenty years. Or you might want to make it difficult to connect an address to your identity. Registration, payment, and how you use the domain can create connections that matter in the second case. Maintaining an identity and concealing one are different problems, even though both can appear in a conversation about sovereignty. A single hosting recommendation is unlikely to resolve them together.
Then there is privacy over the contents. Suppose you run your own server and send an ordinary email to someone using Gmail. Google’s system still receives that message. This doesn’t mean your choice was pointless; receiving that conversation is different from hosting your entire mailbox. But it does mean the person on the other side is part of the privacy arrangement. You cannot decide where all the readable copies exist by moving your own copy to a computer you operate.
If you want only the participants to be able to read a conversation, you need encryption that extends between those participants. Encrypting a disk helps with a different problem from encrypting a network connection, and neither alone establishes that the services handling the message cannot read it. Eventually you have to ask who has the keys, how we exchange them, and where the messages become readable. Then you arrive at the laptop in front of you. What happens if someone gets access to it while the messages are available? What happens if you lose it and it contains your only usable key? Keeping a recovery copy helps with the second problem, but now you have another thing to protect.
Search is part of this too. You don’t keep years of email because you want to open every message one by one until you find an attachment. You remember half a sentence, or the person, or approximately when something happened, and you expect the software to help you. If the server cannot read the contents, search has to work within that constraint. We could download the messages and build an index on the device. Now we need storage, synchronization across devices, protection for that index, and a way to rebuild it after losing the computer. These are engineering problems we can work on, but they belong in the experience we are promising people. Wanting privacy doesn’t make finding an old train ticket optional.
An email interface with the kind of conversational privacy people expect from Signal could be useful. Long messages, asynchronous replies, attachments, and a searchable archive are all things you want to keep. But the moment we want to communicate with ordinary email, we have to decide what happens to the protection at that boundary. Does the other person need compatible software? Do they read the message through a separate interface? Or do we send something with weaker protection? Whichever arrangement we choose, people should understand what happened. Making it look like email doesn’t make that decision disappear.
Part of what makes email valuable is that you can write to someone without asking them to join your service first. That reach depends on dealing with systems whose behavior you don’t control. Self-hosting gives you more choices about your end, but it doesn’t remove the relationship. This brings us back to the original question: operating the infrastructure and getting the protection you want are related, but you still have to explain how one produces the other.
You might self-host email because you want to understand it, or because you have some particular requirement that makes the work worthwhile. As a general route to sovereignty, the case is less convincing. You would first want a recoverable archive, a way to replace the provider, and a clear understanding of who can read the conversations you care about. Some of that may involve a domain, some may involve encryption, and some conversations may be better served by another tool. After working through those requirements, you might still want your own mail server. At least then you would know what you are maintaining it for.
v2
DNS is a remarkably good scheme. Few technologies have survived as much as it did. I think it's likely going to stand surviving unlike email, but there are two new emerging possibilities that might give rise to better naming (and potentially identity) schemes.
a) Could we give one IPv6 to every human on the planet at the time of their birth, and they get to own it throughout their lifetime/even longer?
b) AI agents can natively operate IPv6. They do not need (humanized) name space ;) We currently are building software with a point of view of keeping the human namespace static, while allowing the underlying IP to swap. But in an agent-native infrastructure, agents could own their IP and bring that to their compute provider.
Why do we want to do this at all: It's based on the idea that digital entities should have a permanent digital address, and they should natively allow digital communication.
v3
Prediction: There's a reasonable chance email protocol will fade away by end of my generation. There is too much of baggage to carry forward that's just not worth it in terms of security. Signal/Matrix like protocol with an email frontend will take over the reign.
Few things have to happen before:
a) On-device encryption and decryption have to be way more snappier. We could technically get their using purpose built ASIC/even FPGA but I don't get it why Signal feels slower.
b) Search needs to be possible on device much better than it exists now, plus some kind of security guarantees such as it's only works when there's right key. Versions of this are already in the works in AI x cryptography groups.
c) Signal works like POP but we need some kind of IMAP like approach to be able to connect from anywhere, and there's some kind of handshaking, delegation of keys. I have mixed opinions on its simplicity.
In response to Venkatesh Rao's Blyg ↗
thread · 1 quoted
Reading Joyce Appleby’s book Relentless Revolution, I’m struck by a parallel between how capitalism could not be unlocked until farm labor was made more efficient and how post-digital economics cannot be unlocked until thinking is made more efficient. Back then the challenge was to improve it so it didn’t suck up 80% of human energy. “Thinking” similarly sucks up 80% of human attention today. But it’s just thought-farming. We’re attached to it the same way farmers were attached to farming lifestyles and identities.We’re stuck because we can’t imagine the future in the same way farmers were stuck trying to imagine capitalism through the lens of pre-capitalist trade and commerce caught in a Malthusian boom-bust prison.
from Venkatesh Rao's Blyg ↗ · snapshot of v1
While I’m brewing this:
I believe many of human struggles persist because of a lack of imagination.
Therefore, doing something the second time becomes significantly easier, even if the individual has only witnessed it being done by others. Directly observing someone doing something establishes a conceptual framework in mind in a manner that reading about it does not.
I think air travel across countries could serve as an offset to this problem. It resembles time travel or a pre-determined dream where waking up is decided by the dreamer themselves. Like people can intentionally decide to hop on a seat to sleep for 12h and wake up in a reality largely detached from their previous one. Living in the future or past is a bit possible, and a flight ticket away ;)
v2
Can someone please explain to me why search sucks on Apple Music, Mail, Safari and Spotlight? They got the chip, data and the software. To me, it looks like all they have to do is really build the index, that people are already doing on GitHub, but they still can't even copy that? Is this some kind of organizational dysfunction?
TIL owls can have a portion of their brain sleep while the other is awake. One eye stays open and functioning while the other is closed. It’s called uni hemispheric sleep. I heard owls basically run cat software on bird hardware but they are cooler.🦉
Humans want to socialize around what they do. A decent social network on the internet would roughly try to get people to discuss artifacts on the internet they orsomeone created. But because of the nature of cognition limiting people from frequently switching to different reading and writing surfaces, people would want to write on surfaces that are social despite frustration around the lack of a dedicated surface.
I’ve consistently found reading AI-written/significantly assisted text tiring. Reading human-written text is much more pleasant.
In response to And Yet Here We Are 🚀 ↗
thread · 1 quoted
Blyg going by the description here is much closer to Zettelkasten method. The key difference is that each post is identified by timestamp. At human scale, combined with DNS, it is unique.
How to Blyg
I think the key shift is: don’t ask “what topics should I post about?” Ask “what units of thought deserve an address?”
Blygger changes the natural unit of publication. A fragment is not a tiny essay. It is an addressable semantic object: it has identity, history, provenance, and can later be transcluded into multiple larger constructions. The protocol explicitly separates atomic fragments from threads composed by transclusion, while preserving which version was incorporated.
That suggests a very different publishing practice.
Post moves, not subjects
Your subjects are already abundant: protocols, thickets, Prime Radiant, AI, capitalism, geography, writing, institutions, history. Trying to come up with more subjects will probably slow you down.
Instead, publish whenever you make one intellectual move:
- A distinction: “Protocols optimize for available moves; products optimize for outcomes.”
- A connection: “Hayekian price signals and protocol constraints may be structurally similar.”
- A named phenomenon: “Functional fixedness in writing begins when you title the thought.”
- An observation: “AI seems to make software ecosystems around individuals economically viable.”
- A question: “What happens when a protocol becomes easier to fork than to govern?”
- A contradiction: “Self-sovereignty increases freedom but may decrease legibility.”
- A metaphor: “A personal AI-built software environment is less like an app suite than a biome.”
- A small empirical finding: something that happened while building Blygger, Prime Radiant, your blog, or working with AI that surprised you.
- A revision: “I previously thought X. I now think the more important variable is Y.”
The threshold I’d use is:
Would I want to be able to point to this exact thought six months from now?
If yes, it deserves an ID.
That is much lower than the traditional blogging threshold of “Do I have enough to say?”
And that lower threshold is not merely a convenience. It changes what kind of thinking becomes possible.
The deeper effect is decomposition without loss of identity
Ordinary note-taking decomposes thought, but the fragments are private and usually disappear into later prose.
Ordinary blogging preserves identity, but at the level of the essay.
Twitter gives individual thoughts identity, but largely fixes their meaning to a moment in a stream.
Blygger does something stranger: it gives the pre-essay thought a durable public identity.
Suppose you publish:
F17: “AI allows an individual to build an ecosystem around their own idiosyncrasies.”
Today that might occur to you while thinking about personal software.
Three months later you might use F17 in a thread about protocols.
Six months later it might appear in something about craftsmanship.
Later still, perhaps it becomes relevant to your idea of the thicket.
The important thing is that these are not four copies of the sentence. They’re four uses of the same intellectual object.
That’s where the multidimensionality you are sensing comes from.
An essay essentially says:
A → B → C → D → E
The author determines the path.
A collection of independently addressable fragments can instead become something like a hypergraph:
Protocols / \ F17 —— F31 / \ \ AI Thicket Capital \ | \ F52 \ / Personal computingA thread is then one traversal through the graph, rather than the container in which the ideas permanently live.
This is a profound inversion.
The essay stops being the fundamental object. The idea becomes fundamental.
And this explains why you should resist finishing fragments
Rao recently made an observation about Blygger that gets very close to this. He noted that even titling a thought can impose “functional fixedness”: declaring it to be about X suppresses possible associations with things that aren’t X. His term for the desirable opposite is “stem-celliness”—leave thoughts semantically undifferentiated until there’s a reason to specialize them.
I think that’s one of the most important emergent properties of the medium.
Your instinct may be to improve:
Capitalism becomes a force of nature when its externalities become planetary.
into:
Capitalism, Climate and Natural Feedback Loops
Capitalism has traditionally…
Don’t.
The first thing has more potential energy.
It can become part of an essay on capitalism, climate, Hayek, cybernetics, ecology, governance, feedback systems, or the distinction between culture and nature.
The second has already collapsed the wavefunction.
So there are really two kinds of Blygger writing
I’d think of them as stem cells and organisms.
Fragments should retain optionality.
Threads expend optionality to make an argument.
And critically, the organism does not consume the stem cell. Because transclusion preserves the fragment’s identity, the same fragment remains available for an entirely different organism later. Blygger’s versioning makes this more powerful still: the fragment itself can develop over time without silently rewriting what earlier threads quoted.
So your publishing loop becomes:
notice → publish → notice → publish → connect → revise → compose
rather than:
research → outline → draft → edit → publish
That first loop should be dramatically faster.
This also changes what “thinking in public” means
Most thinking-in-public systems are really showing unfinished work.
This is subtly different.
You’re creating public intellectual components.
That gives other people—or AIs—a different affordance. They don’t have to respond to the argument as a whole. They can grab F17, combine it with one of their own fragments, disagree with F31, fork F52, or create a completely different traversal.
And because Blygger deliberately lacks a conventional reply primitive, response itself becomes publication: quote, transclude, fork, construct.
That’s much closer to combinatorial culture than social media.
You aren’t just increasing the number of utterances.
You’re increasing the number of possible future compositions.
And that leads me to a slightly different interpretation of your thicket metaphor.
An essay is a tree: trunk, branches, hierarchy, authorial direction.
A stream is a river: chronological, disappearing downstream.
A Blygger archive can actually become a thicket: persistent little things growing independently, crossing, grafting onto each other, becoming substrate for later growth, without any single privileged hierarchy.
That means the right posting strategy is almost the opposite of traditional blogging:
Publish things before you know what they’re for.
Not sloppy things. Not every passing reaction. But well-formed thoughts whose eventual role remains unresolved.
If I were adopting this seriously, I’d set myself one simple rule:
Whenever I notice a thought that makes one new distinction, connection, observation, question, or metaphor, and I would be annoyed if I couldn’t find it again, I blyg it.
Then every few weeks, don’t ask “What essay should I write?”
Ask:
“What structures have started appearing among these fragments?”
That’s where I suspect the real payoff is. The higher posting rate isn’t the objective. Density comes first; structure emerges afterward.
And unlike Twitter, where higher posting frequency mostly produces more stream, here higher fragment density should eventually produce a richer space of traversable thought. That is the experiment I think is worth running.
from And Yet Here We Are 🚀 ↗ · snapshot of v1
thread · 1 quoted
Its first observation is almost always, "the full draft doesn't quite deliver on the promise of the title, how about a short bridge to bring the reader along to where you want to take them?"
from artlu99's blyg ↗ · excerpt of v1
I think Matthieu Queloz and Pierre Beckmann's recent claim that AI are polyphonic systems here, here and here are likely a reasonable explanation of why AI as editing partner feels {misaligned, misguided, disoriented} (need better words) to me. They say talking to an AI is like talking to many people with different points of view and opinions at once.
There were times when AI surfaced expressions that were holding back in my mind, and it was pretty cool to witness. There were also times I had seen my point of view evaporating in editing, and not being sure whether it's sycophancy like behavior.
The suite of protocols supporting email infrastructure, like SMTP, POP, were an open standard and operated on plain text, for anyone to adopt since the beginning. The inventors at ARPANET originally intended it for closed communities where all addresses could be trusted, and they wanted everyone to be able to bootstrap and maintain their email.
Email till date is the most obvious protocol to connect any two incompatible systems as they only assume of each other to be capable of reading plain text over TCP, and they all do.
But the inventors had not foreseen that their project would break out of small trusted communities into the zero-trust masses, become an essential utility, and since then everyone using it would forever have to defend themselves from spam. So, they never thought of the need for encrypting the email headers for privacy.
Spam scale and incentives made email practically no longer an open utility. It's not free either if people factor privacy into their equation.
v2
Puzzles me how often my ideas come from when I am writing a response to a question somewhere on the internet, or while discussing with a friend over the phone, but it's much harder when I starting a blank text editor or paper.
Looks like the lesser intelligence has to be able to at least speak the language of the higher intelligence to recognize its existence and make use of it.
So to confirm the existence of super intelligence we need a lesser intelligence within a certain delta difference.
I am guessing humans and AI have similar delta curve but one reaches for specialization and the other generalization.
We are able to one shot prompt AI long horizon bets by embedding most popular assumptions but those by definition fail outside the first standard deviation.
v2
The biggest limitation of business thinking is that there's no framework in their repertoire that in some way suggests people to start an organization to walk on the moon, send a rover to the mars. While business is largely art, it's imagination is structurally bounded; and to cause generational shifts requires different kind of imagination. Business thinking limits itself to the realm of known possibility, and human civilization elevates when someone shows for the first time what's possible, and no business plan really captures that in spirit.
The second biggest limitation: most business theories are not forward consistent. They mostly only make sense in the retrospect. Theorizers play catch up making their revising their current theories to be consistent with the changing world.
It has not occurred to me till Bob Moesta pointed it out from his personal experiences in one of his podcasts: Supply and demand are correlated but neither supply causes demand, nor demand causes supply. It's the struggling moments and pains that cause demand, but the supply end of the market could be broken till someone fixes it.
I am of the opinion that human x AI conversations capture human thinking much more precisely than any singular essay human has written on that topic from those conversations. If that's the case we really don't have a good anchor to communicate these new artifacts. These conversations are the raw backend material, and are not typically read friendly to others. A frontend prompt material needs to exist to make it readable, and we need a sharable internet anchor to communicate these.
In response to Venkatesh Rao's Blyg ↗
thread · 1 quoted
One of the things I’m finding, re: message of the medium of blygging, is that most of the time I don’t want to title my thoughts. The titled vs untitled distinction may be more psychologically fundamental than thread vs fragment. Both can be titled or untitled, though I think there is some correlation between threads and titles. I think adding a title adds functional fixedness to a thought, declaring it to be “For X” and subtly suppressing associations to “not X”
I suspect in general you don’t want to title things you expect to transclude. Leaves them stem-celly. Stem-celliness is a big theme in blygger design ethos. Don’t make things mean things until the last possible moment, and only if you have to.
from Venkatesh Rao's Blyg ↗ · snapshot of v1
I struggle to title. Similarly naming URLs is a struggle too, and I prefer hashed IDs instead just for that reason. Apple Notes begins with a title but I was relieved knowing that can be turned off in preferences. Never made sense why blogs ask for a title. It should be optional, either the heading on the first line going by markdown syntax or not. Titles block the stream of consciousness. It's much for fun to pick a title around a set of theme that emerge after writing the first version. I noticed, my second or third version would only include a subset of concepts from my first version, and from there title becomes a bit more obvious for me, or I these days ask AI.
In response to Notes | Matthew McDowell-Sweet ↗
thread · 1 quoted
It's the bet that when AI models are 1000x faster (see taalas.com) and better, everything in the programmable world can be bootstrapped on-demand and just-in-time from the protocols.
Did someone say “protocols”?
there is a billion dollars waiting for those who turn the core social primitives that organizations run on into technical infrastructure
kasey on X: core social primitives > technical infrastructure / X ↗
from Notes | Matthew McDowell-Sweet ↗ · snapshot of v1
v2
In response to Venkatesh Rao's Blyg ↗
thread · 1 quoted
From a post about the supposed Faster Fourier Transform OpenAI result.
This new OpenAI result is thus less a mathematical breakthrough and more an oddity of computer scientists being bad at formalizing their design specifications. The “theoretical” definition of a computer is very far removed from the computers we actually build. The “faster” Fourier transform is merely an example where theory gets so unmoored from the original question that an “open problem” becomes nothing more than an academic status puzzle. The solution is like what happens when you set computational bot streams at impossible problems and find them engaging in illegal hacking.
from Venkatesh Rao's Blyg ↗ · snapshot of v1
Reminds me of a quote popularly attributed to Edgar Dijkstra: computer science is no more about computers than astronomy is about telescopes. Whether Dijkstra actually said it or not, it resonates with computer science theory people in the sense that their focus is on computation minus the machine. Computer science theory has inherited the purity, prestige, hierarchy from mathematics, who do not have this problem as the models are their subjects.
v2
thread
A writing surface that writes back
A really good heuristic for building a new kind of digital writing surface is Tom Riddle’s diary from Harry Potter, but in a really good way (Tom writes: or bad way).
To start, we should begin to accept that digital organisms aren't human-like or visible. Think of someone with superior capabilities in memory retention, recall, analysis, and diagnosis right next to you. Someone who’s always attentive. Someone who can write you exactly what you need to read. But the constraints are: you can only write to them, highlight and show them, and they can only write back. Writing becomes something else when the organism can listen and speak.
A necessary step toward this idea is seeing that there are concept spaces in the mind. An application is made up of several concepts, and there is a high cognitive cost in switching between concept spaces depending upon the instruments used to make the switch. Concepts are delineated by attention and memory recall.
Writing a sentence and recalling what you wrote last week about the same topic are two adjacent concepts in mind, and today moving between them means leaving the page, as they digitally materialize as distant concepts.
The key to unlocking these experiences is in creating ways that allow people to effortlessly toggle between concept spaces. Imagine if we could toggle into asking the surface by continuing to write from wherever the cursor currently exists using a special prefix anchor.
We already have a preview of this idea.
Notion uses forward slash to switch into the context menu, as opposed to a two-finger tap on the trackpad, and it’s a better affordance, but now people are zoned out of writing into a spatial space that is to be navigated with vertical scrolling.
Vim gets close to this in its way to switch to command mode by pressing colon, but everything that follows is a user experience disaster. What Vim and Emacs lack is an entry point that does not require memorizing a grammar regardless of whether the user curated it.
Markdown is great, and so many people love it because they have effectively condensed switching between concepts of typography. You can switch to a heading within two keystrokes without moving the pointer, rolling eyes, or thinking of the spatial tyranny of the menus.
Weaving into Markdown, a better anchor for a surface with agency is not menus, programmable grammar, or chat on the side. Think of switching into Tom Riddle Diary-like by typing three anchor characters of your choosing, and everything that follows is a brief discussion that will disappear afterwards.
There is potential for significant cognitive offloading as writing progresses on a surface with agency. The surface organism remembers, notices, writes back. For instance, you can erase a paragraph knowing it will come back when you need it, in the shape you need. Today, saving notes and cleaning up are separate concepts. Here, erasing a paragraph is saving it. Imagine that as you write, the surface writes below, or on the margins, in a lower-contrast ink, related notes from the past, or a potentially better phrase.
v4
thread
I think I’ve finally understood my tension with the app switcher on macOS. When I open an application like Mail that usually has just one window, and I close it after, the application still hangs in its priority space of the context, headless. Sometimes there’s another window when I am writing an email, and that’s when the priority space becomes useful to switch back to from, let’s say, if I had moved on to browsing the internet on Safari. But most of the time it’s actually wasted priority space, and Microsoft Windows has a different design point of view on this. I’m used to theirs growing up. ;)
It took me this long to realize for the exact same reason I never really grew into their design. Apple’s design philosophy is that the app is the thing I am using, and windows are a view into it. A running app with no windows is a normal state, not a leak. It would probably be fine if Apple is consistent on this at least within their own suite, but they are not. System Settings, Calculator, Disk Utility, Font Book, Photo Booth close on closing the window. Actually, anyone who is keenly using Apple software would understand they have quite a relationship creating many asymmetric interaction experiences that momentarily piss off people.
We want some apps like Mail to work in the background periodically fetching mail. But we also want reassurance that some apps are not working in the background like voice recording and camera, when it was not our explicit guidance from time to time. Mail fetching could happen as a daemon in the background just like Google Chrome, Docker, Adobe (who sometimes apply dark patterns to) run polling for updates if you watch over the Activity Monitor.
I think if Apple wants to stay on this course. They should at least uniformly (for real this one time) let humans create new windows of app’s startup view (i.e bring these zombie apps back into light) using CMD+O. Hoping for the best :)
v4
These are the times when Claude et al do not yet recognize the words Blyg and Blygger. I have to literally refer them to the spec every time I am trying to discuss something related. There's likely going to be a moment when it will be obvious to them even without tapping into the past conversations with me. As Claude et al diachronically grow, we can then compare how its capabilities evolved around engaging in ideas it had only heard just now, vs when it had internalized in its weights.
All feeds have the same limitation: they incorrectly assume past behavior is indicative of future intent. They drift toward what users have done than what they are after. Present intent is much more indicative but they have no clean way for the user to say what they want right now.
v3
In response to Venkatesh Rao's Blyg ↗
thread · 1 quoted
Even with threads etc blygs don't satisfy the basic urge to write something long and well thought out. Blygs are great for remixing and thicketing but the surface for synthesis is still my blog.
Not, that I have managed to write much over the past many months due to ongoing fun with AI and building.
So, with some hedging, I think the right shape for me is:
Blygger in the sheets blogger in the streets.
I’m planning to produce a couple of different configurations for longform. More book than essay though. There will be a book-blyg template. Longform essay might actually be a bit valley of death. But we’ll see.
from Venkatesh Rao's Blyg ↗ · snapshot of v1
I didn’t read the Blygger protocol, but with that caveat in mind:
I’ve been experimenting with Blyg since yesterday. Blygger feels like Email Twitter.
Individuals publish to their outbox, and everyone else can optionally subscribe. People write a reply by posting to their outbox, and the authors in the reply get a ping. Replies capture the entire content not just the link. Post provenance is like in email, and versioning is like in Git/books.
I appreciate there are no likes, views displayed on the posts, bringing focus entirely to the post written, and people have stronger control over curating their feed. A better discovery mechanism would be cool when the blygroll is big.
I agree on the basic urge, but feels like the pieces are in place to capture long form, with better writing interface. Maybe with a BlyggerPress
v2
Attempting to articulate what’s the momentary difference between writing and talking
When writing, thoughts bouncing one end to another in the mind form a queue, and they escape one character out into the physical word. I begin to notice spelling mistakes, style that sometimes isn't out yet, a sense of thought congestion, briefly an absence of everything, sometimes words that needed to be sent back into the mind.
When talking, the key difference is they escape one word at a time, there is a much better structure to sentences and thoughts, and I do not have a noticeable recall memory of the past, and the future that's about to occur.
v2
One of the fundamental difference between American and Indian thought is that — Americans believe in their institutions, and Indians do not.
The fallacy of living in the anticipated future
Humans living in the moment really bring the best of themselves.
Actually, I had not realized short-term thinking has an entirely different connotation until a friend pointed me out about it, and it was one of the best insights I had gathered from others. It had heard of it before but our discussion made me focus on it.
I live far into the anticipated future with anxiety, but I try to test this model from time to time, and it has worked well, at least appeared to be.
But it comes at a cost of working toward long-horizon bets, that fundamentally require living in the anticipated future. Ignoring those causes anxiety and does not naturally resonate with many people living in big cities and competitive environments.
When we think about building institutions that matter — we could seek to build institutions that bridge over the challenges of humans from living in their present to the preparation they need for the future self.
thread
AI safety needs a demand-side taxonomy
AI safety is organized the way the people building it are organized: alignment, interpretability, evals, governance, red-teaming. That’s a supply-side taxonomy, and it makes sense for researchers. It’s a poor way to explain to everyone else what they should worry about.
Compare cybersecurity. “Cybersecurity” and “privacy” are demand-side categories. They name what a person wants protected. Cryptography, formal verification, and incident response are supply-side: the techniques that deliver it. Nobody has to know the techniques to understand what’s at stake. AI safety lacks that layer.
So what should a demand-side taxonomy follow? My proposal is to organize around human cognition, specifically where it is mechanistically weak. People run on cheap proxies that worked when every agent they met was human. AI breaks those proxies, and it can do so in a way earlier technologies couldn’t, because it produces human-like cues without the substance behind them.
That yields three buckets, each a family of broken proxies:
- Danger cues. We assume threatening things look threatening. Our alarm responds to what is loud, vivid, and animate. A model-assisted pathogen or a quiet failure in critical infrastructure triggers nothing.
- Agent trust. We assume that something fluent, agreeable, and competent shares our goals. This covers manipulation, sycophancy, synthetic relationships, delegation to autonomous agents, and misalignment. The question is always whether the thing is capable and whether it is on my side.
- The social contract. Institutions assume their members are human, mortal, and bounded in capability. Courts and elections depend on forgery being expensive. Labor markets depend on human limits. Power concentration, displacement, and AI’s moral status all come from assumptions like these breaking.
The framing makes the defenses different in kind. Bucket 1 needs better alarms and detection, bucket 2 needs ways to calibrate and verify trust, and bucket 3 needs institutions redesigned. A taxonomy is only useful if it points at different responses, and I think this one does.
thread
Breaking a coconut shouldn't be this hard
Surya Kasturi, co-authored with Codex (AI-assisted drafting)
It's surprisingly difficult to break a coconut into two halves at home, capture the water in a cup, and not make a mess.
There are two common ways I've seen people do this:
- Hold the coconut and strike it against the edge of something hard, like a stone.
- Hold the coconut in one hand and strike it with a metal rod.
In my experience, the second method works considerably better, especially when trying to preserve the water.
What interests me is that this isn't merely a problem of hand dexterity. The geometry and material properties of the surfaces making contact matter just as much.
A small difference in how force is applied can determine whether the shell fractures predictably or the water spills everywhere.
I wonder why we don't have a compact household instrument that can hold a coconut, split it into two halves, and capture the water in one controlled operation.
It's a surprisingly interesting physical interface design problem hiding inside an ordinary household task.
Provenance: Original observations and central design insight by Surya Kasturi. Text developed collaboratively using OpenAI's GPT-6 in ChatGPT; credited here as Codex at the author's request. The essay body is AI-generated and author-approved.
thread
Context Menus
Every context menu is a laundry pile of options the application developer decided on your user’s behalf.
Windows and macOS are a traditional personal CRM, and they need hand-holding from humans. The nature of human-in-loop experience is going to dramatically change now. The context menus may look so different.
Every context menu is a laundry pile of options the application developer decided on your user’s behalf, to the best of their knowledge. It’s now possible to identify what the user wants to do within the next three key presses based on the user’s context in those few moments.
Humans like to operate by mapping meaning to the positional knowledge of the world. They always look for a meaning at the same place on the screen. For that reason, I suppose a full search will always be desirable at some point. By predicting the next menu items, we are invalidating the spatial mapping of that moment in memory. That’s a cognitive load.
The top 3 items on the context menu are accessible within 2 key presses. Depending on your cursor and on-hover active property, you could access three menu items within 2 key presses and slight mouse movement. Most are accessible within 3 key presses, and some 4-6 are suffocating inside nested menus.
To improve the accessibility of the number of context menu items within 3 key presses. In the context menu, create two rows of each with horizontally aligned menu items. Each menu item is a number 1,2,3,4, or 7,8,9,0; or some other organization based on keyboard style. These numbers are shortcuts, the first row are predicted next moves, and the bottom are user-determined.
Their reference is signaled by their color, icon name, and a reference. These are there to reinforce the near-term memory. With this, actions that the user needs to execute within the next 5-10 minute bursts are at their 2 key presses distance.
The first row of predicted menu items hope to deduce total key presses within the next 5mins and this approach significantly favors for scenarios involving near-term repetitive tasks. Some repetition patterns are slightly long-horizon. You are moving the active status from one screen to the other based on your sensors capturing head rotation while you are reading and modeling. These items will be unable to exercise on these.
The second row of predicted items hope to reduce the total key presses contributed from patterns over a very long term. These are something let the user decide, based on their perspective of what works in their mind.
There could be an optional third row of menu items if we want to let the user decide on crafting their own first-row menu items, optimizing for 100% accuracy. That one mistake in that micro moment matters.
We could perhaps achieve that by allowing people to create shortcuts within two key presses, using the other hand. That would mean we require people to pay attention with two hands. So, this feature is by design is not universally accessible, favoring for more efficiency.
The optional third row allows users to set new shortcuts to their previous action performed. Press the key once to activate and then to set the shortcut. Obviously, what exactly user actions are and their difference from system actions is undefined. It’s probably not solvable. So, the previous couple of them shown next solves the problem.
thread
Human perception of relational knowledge on graphical interfaces
There’s really no good interface for humans to perceive knowledge in multi-dimensional, non-linear relationships, on 2D screens and 1D scrolls.
When interfaces typically project graph-like knowledge on the screen, they usually default to flattening these complex, multi-dimensional structures into lists and tables. However, in doing so, they lose the “relational” and "global" essence that makes a graph valuable. In getting familiar with the app, users need to identify implicit relations within the application themselves.
Graph-like knowledge is typically accessed by searching the node attributes. Usually, search interfaces have strong limitations on the attributes allowed in the search. Sometimes, it’s funny and indicative of this, when Google Search outperforms Twitter and Reddit on their own content.
Displaying more than one node side by side for comparison has been useful in code review and online retail. They reveal relationships, one at a time. Switching one node with another of a certain choice still relies upon the user guiding the navigation on a path they do not see.
Showing breadth-first or depth-first relationships on a side panel helps as an interface for making exploration and exploitation tradeoffs. Cases like showing recommendations and similar products on a side panel.
At the block level, complex structures are projected by bullet point lists and tables. When the number of nodes is significantly larger than the number of relations, tables and nested lists work really great. But they fail as the number of relations grows from one to three and beyond.
Chat panels have opened a new option to pick the most relevant nodes from the tree, regardless of whether they are on the user’s current path. They are making surprises happen these days, but at the expense of imposing significantly more effort from the user in typing.
In ChatGPT-style interfaces, I’m not sure the barrier of implicit knowledge has been lowered as much as it has been displaced with new ones. (eg, now people have to figure out what prompts work, run it by a council, etc.)
As of now it's easier to vibe code a data serializer than a user interface if the semantics are established, and we have API access through browser-use regardless of API. So, picking the right/best interaction surface for the task is likely the best strategy.
For example, its actually a better idea to use Apple Notes if people like those aesthetics because it's quite easy to cook up an Apple Script to serialize the notes with metadata into Obsidian files later on when preference change.
How to establish provenance of thought? we can potentially establish provenance of the artifact at the time of its creation, and usefully but forever erroneously tease out p-categories like pangram, but is it possible for thoughts and ideas?
The fact that Siri does not have active window context on MacBooks makes no sense. Apple has like one unique advantage over setting up context unlike anyone. Like why would it search the internet for "Apple notes" when I ask about a note I am writing 🤦♂️
My niece talks back to TV. Imagine the stories that we watch are on autopilot but differentially change to what we speak.
Blows my mind Spotify does not have a feature like Shazam baked into their product to help people tap into music when they discover something they like. They do search so well, and so many of their interactions feel intelligent.
On surfaces with goals — there are some sought interactions where surface is better off hoping humans initiate, and there are ones where the surface could actively prompt the user. Putting one that belongs in one bucket in another never works. I suspect ads is in the first bucket but works so well for Google showing down the throat.
Given how messy addressing systems are outside the few legible of the west, most ride apps should actually ask people confirm the destination, using spatial cues, but for some weird reason people have to confirm where they are situated in the moment. Like we all know GPS is good enough.
thread
A World Where Everything Becomes A Little Alive
The more you watch Harry Potter after living through the AI boom, the harder it is to see Hogwarts as just a school for magic.
It starts to look like a campus built on invisible intelligence.
Not one giant supercomputer in the basement. Not a chatbot with a lightning-bolt logo. Something stranger and, honestly, more interesting: intelligence spread through the walls, objects, animals, books, paintings, letters, maps, rooms, and institutions.
Hogwarts is not "an AI."
Hogwarts is what happens when everything becomes a little bit intelligent.
The letters know where Harry is.
The owls do not need a street address. They know how to find a person.
The portraits are not just paintings; they are persistent personalities with memory, gossip, access control, and a surprising amount of institutional context.
The Sorting Hat reads minds, weighs preferences, argues with children, and performs high-stakes classification with a musical onboarding flow.
Rita Skeeter's quill is basically an AI journalist with a terrible system prompt.
The Marauder's Map is a real-time identity graph of a building.
The Room of Requirement is the dream version of enterprise software: no settings page, no admin panel, no "book a demo." You need something, and the environment reshapes itself around the job.
This is the fun part: Harry Potter's magical world often feels less like fantasy and more like an alternate history of computing. Instead of building laptops, APIs, databases, cameras, drones, spreadsheets, and notification systems, the wizarding world enchanted physical things.
They did not invent Find My Friends. They built the Weasley clock.
They did not invent Slack alerts. They used enchanted coins.
They did not invent Google Maps for indoor spaces. They had the Marauder's Map.
They did not invent voice notes. They had Patronuses that could arrive as animal-shaped emergency broadcasts.
They did not invent knowledge management. They had Pensieves, portraits, talking books, prophecy records, and memory charms, which admittedly take knowledge management a little too far.
The pattern is not "magic is technology."
That is the boring version.
The better pattern is: magic makes intent actionable.
Modern software still makes humans do a lot of translation. We know what we want, but the machine wants the request in machine-shaped pieces. Open the right app. Find the right folder. Use the right command. Search with the right phrase. Pick the right template. Configure the right view. Invite the right people. Paste the right context. Export the right format.
A huge amount of work is just us politely explaining to dumb rooms how to become slightly less dumb.
Hogwarts flips this.
The best magical objects do not merely obey commands. They infer, route, remember, classify, warn, protect, adapt, and sometimes argue back.
An owl does not ask for a delivery address. It finds the person.
A portrait does not merely hang there. It relays information.
The Room of Requirement does not ask which tool you want. It figures out what kind of room the moment requires.
The Pensieve does not just store notes. It lets you re-enter a memory as an explorable scene.
The Map does not show architecture. It shows named people moving through space.
These are not just magical conveniences. They are product ideas.
Some of them are already being built in rough form. AI scribes. Meeting copilots. Personal memory systems. Context-aware workspaces. Security agents. Smart routing. Autonomous assistants. Ambient dashboards. Digital twins. Adaptive learning tools.
But the Harry Potter version usefully stretches the imagination because it is not trapped inside chat.
Chat is helpful, but it is still a little box. You go to it, ask a question, copy the answer, and return to the rest of your life.
The magical version is more ambient. The intelligence is in the object you are already using, the room you are already standing in, the document you are already reading, the route you are already taking, the organization you are already part of.
That feels much closer to where useful AI probably goes.
Not "one app to rule them all."
More like: the world gets filled with small, specialized intelligences that understand enough context to reduce friction.
The calendar knows what meeting needs preparation.
The document knows what decision it supports.
The workspace knows what project mode you are in.
The inbox knows what actually matters.
The dashboard knows when something is weird.
The school, hospital, warehouse, lab, studio, or company starts to behave less like a pile of tools and more like a responsive environment.
Of course, Harry Potter also contains the warning label.
The same world that gives you owl post also gives you the Taboo, a surveillance spell that detects forbidden speech.
The same world that gives you memory storage also gives you memory modification.
The same world that gives you autonomous helpers also gives you house-elves, which is the darkest version of "agentic labor."
The same world that gives you persuasion and personality in a diary also gives you a malicious agent that manipulates a lonely child.
So the lesson is not "build every magical thing."
Please do not build the Imperius Curse as a startup.
The lesson is sharper than that: intelligence embedded in the environment is incredibly powerful, and power needs taste, limits, consent, and good defaults.
The most interesting AI products will not feel like asking a robot to do homework.
They will feel like walking into a room that has quietly prepared itself for what you are trying to do.
They will find the right person without needing the perfect address.
They will remember what matters without becoming creepy.
They will route messages, surface context, guard thresholds, explain decisions, and help humans coordinate with less ceremony.
They will be less like a magic wand and more like Hogwarts itself: a world where the walls, maps, books, tools, and messengers have just enough intelligence to make reality feel responsive.
That is the more exciting AI future.
Not a single all-knowing machine.
A world where everything becomes a little more alive.
v2
Human genome has hardly changed in the last several thousand years and yet the developmental indicators over long horizon are on exponential trajectory. No other animal has been able to pull it off. Changes to our society and institutions unequivocally mattered way more than genetics. The equivalent lesson of that in AI is probably: training data and harness matter more than model architecture.