The thing that you’re thinking is the obstacle to your improvement, that you need to deal with before you can start getting better: dealing with that obstacle is the improvement itself.

have a blocked / inflamed nose

Algorithmic feeds and other forms of “paternalistic” content presentation are necessary and even desirable in an information-rich environment. In many instances, decisions about what you see must be largely controlled by a third party whom you trust. The audience in a comedy club doesn’t get to insist on knowing the punchline before the joke is told, just as RPG players don’t get to order the Dungeon Master to present their preferred challenges during a campaign.

Less noticed — and more important — was how platforms did the opposite: twiddling the knobs to remove things from your feed that you’d asked to see or that the algorithm predicted you’d enjoy, to make room for “boosted” content and advertisements:

For new users who didn’t yet follow many people, this presented the opposite problem: an empty feed, and the sense that you were all alone while everyone else was having a rollicking conversation down the hall, in a room you could never find.

Behavioral economists have a name for the steps we take to guard against temptation: a “Ulysses pact.” That’s when you take some possibility off the table during a moment of strength in recognition of some coming moment of weakness:

Note the assymetry here:What etymologynerd says about tpot is heard by tpotWhat tpot says about etymologynerd is NOT heard by etymologynerd’s audience

The window is closing. As enterprises reach the limitations of MCP, they’ll build proprietary solutions. So will software vendors. The fragmentation that MCP aimed to prevent will still emerge, albeit with additional steps and wasted effort.

The Cost Attribution CrisisWhen OpenAI bills $50,000 for last month’s API usage, can you tell which department’s MCP tools drove that cost? Which specific tool calls? Which individual users or use cases?MCP does not provide a mechanism for this basic operational requirement. No token counting at the protocol level. No cost attribution headers. No quota management. You’re flying blind on AI spend, unable to optimize or even understand where the money goes.In contrast, cloud providers learned this lesson decades ago. Every AWS API call can be tagged, attributed, and cost-tracked. Every Google Cloud operation flows into detailed billing breakdowns. MCP asks enterprises to consume expensive AI resources with 1990s-level cost visibility.

The enterprise cost of this ecosystem fragmentation is staggering. Instead of training developers on one protocol, you’re training them on MCP plus a dozen semi-compatible extensions. Instead of conducting a single security audit, you’re auditing multiple authentication libraries. Instead of managing a single vendor relationship, you’re managing relationships with several open-source projects of varying quality and commitment levels.

This is exactly the fragmentation that protocols are supposed to prevent. gRPC doesn’t need a third-party tracing library. It’s built in. REST doesn’t require external caching semantics. They’re part of HTTP. CORBA didn’t need community-maintained IDL generators. The ORB vendors provided them.

CORBA emerged in 1991 with another crucial insight: in heterogeneous environments, you can’t just “implement the protocol” in each language and hope for the best. The OMG IDL generated consistent bindings across C++, Java, Python, and more, ensuring that a C++ exception thrown by a server was properly caught and handled by a Java client. The generated bindings guaranteed that all languages saw identical interfaces, preventing subtle serialization differences.MCP ignores this completely. Each language implements MCP independently, guaranteeing inconsistencies. Python’s JSON encoder handles Unicode differently than JavaScript’s JSON encoder. Float representation varies. Error propagation is ad hoc. When frontend JavaScript and backend Python interpret MCP messages differently, you get integration nightmares. Third-party tools using different MCP libraries exhibit subtle incompatibilities only under edge cases. Language-specific bugs require expertise in each implementation, rather than knowledge of the protocol.

When an AI tool expects an ISO-8601 timestamp but receives a Unix epoch, the model might hallucinate dates rather than failing cleanly.

Then iterate towards an integrated tool to explore idea diffusion in the user-interaction networks present in our 16M tweet dataset.

we’d want to get our essential data-science primitives running on tweets and accounts - that’s embeddings, clustering, network visualization, etc. [1] We want all of these running very efficiently in the browser, leveraging WebAssembly.

Real time data firehose via our browser extension, avoiding the laborious upload process.Include Bluesky data. We could easily do this on bluesky because it’s open but twitter still has the most juice.

I like to imagine ideologies competing, merging, forking like microbes.If we can trace where a given idea comes from, we can detect if it emerged organically or if it was intentionally and artificially promoted.

In Girard’s vision of history, the secularisation of the West — the decline of the political role of religion, and of religious belief — means, in an apparent paradox, that the Gospel can, for the first time, “become clear”.

Christianity was presented as another version of ancient religions, in which priests accepted payment from congregations in return for rituals of blessing and sacrifice.

According to Girard, the death and resurrection of Christ had freed humanity from the cycle of potentially destructive imitation that scapegoating had been intended to contain.

The originality of Girard’s thinking lies in his argument that societies manage the ambivalence of imitation — its capacity to promote social cohesion or to degenerate into spirals of competitive violence — through “scapegoating”, or the exclusion of some agent identified with imitation’s negative aspects. In his account, scapegoating is a function of religion, and takes the form of sacrifice, the physical destruction of some victim who was understood as responsible for, or in their innocence expiating the sin of, dangerous imitation. The violence of sacrifice precludes the violence of mimetic desire.

In fact, Girard argues, our desires arise not so much from such natural drives as from a universal tendency to envy and emulate other people, which orients us to seek out whatever our rivals (that is, everyone else) are seeking. We want what we want because other people want it.

hrough." When we face tough problems, we break them down, figure out the steps, maybe jot down some n

Here are some basic facts:

DO $$ DECLARE r RECORD;BEGIN FOR r IN (SELECT tablename FROM pg_tables WHERE schemaname = 'public') LOOP EXECUTE 'DROP TABLE IF EXISTS ' || quote_ident(r.tablename) || ' CASCADE'; END LOOP; END $$;

gaslighting is “a form of emotional manipulation in which the gaslighter tries (consciously or not) to induce in someone the sense that her reactions, perceptions, memories and/or beliefs are not just mistaken, but utterly without grounds.”

In the United States, compulsory education is the norm. This is, ostensibly, to create an educated populace well-equipped to take part in democracy.

Because of these concerns, we as a society have developed an apparatus to consume and process all this information and then give us what we need to know to make sense of the world. This is our sensemaking apparatus.

While an individual can generally get by pretty well just being focused on their immediate surroundings and relationships, most people want to know and understand what is going on in the wider world around them.

All of these companies are staffed by people that went to good schools, went through prestigious programs, and are generally pretty intelligent.

New York Times

corporate media

You regurgitate their thoughts thinking they are your own.

They are highly paid and their opinions are given great weight.

Imagine if I told you there is a group of people dedicated to controlling how you perceive the world and the events that go on in it.

DecideAction station:LOOKS AT: Your question and what we know so far (nothing yet)THINKS: "I don't know who won the 2023 Super Bowl, I need to search"DECIDES: Search for "2023 Super Bowl winner"PASSES TO: SearchWeb station

No complex math, no mysterious algorithms - just nodes and arrows! Everything else is just details. If you dig deeper, you’ll uncover these hidden graphs in overcomplicated frameworks:OpenAI Agents: run.py#L119 for a workflow in graph.Pydantic Agents: _agent_graph.py#L779 organizes steps in a graph.Langchain: agent_iterator.py#L174 demonstrates the loop structure.LangGraph: agent.py#L56 for a graph-based approach.

The Big Secret: Agents Are Just Simple Graphs!

The recipe (Flow) just tells you which station to visit based on decisions:"If the vegetables are chopped, go to the cooking station""If the meal is cooked, go to the plating station"

See everything: The entire framework fits in one readable file

it's just 100 lines of code that lets you see exactly how agents work!

SOLID principles (Single Responsibility, Open/Closed, Liskov Substitution, Interface Segregation, Dependency Inversion) are design principles that promote maintainable and flexible code.

Test-Driven Development (TDD) is a software development approach that emphasizes writing tests before writing the actual code.

What are the most effective governance and administration models in place on medium-to-large sized Fediverse servers?

Fediseer continues to be a growing resource for Lemmy and Mastodon administrators, and fedi-safety is a novel tool that can classify genAI CSAM on Lemmy and potentially other services. Pixelfed introduced comment controls and enhanced spam classifiers. BlueSky introduced Ozone, an innovative moderation tool designed to support moderators in managing their communities. Ozone’s integration of advanced filtering systems makes it a standout contribution to the trust and safety ecosystem, powering several “composable moderation” projects on the Bluesky “ATmosphere” with the notable success of Blacksky, an AT Protocol implementation prioritising the community building efforts of marginalized groups; especially Bluesky’s community of Black users after which the project is named.

A New Social was launched to liberate people’s networks from their platforms, leveling the playing field across the open social web – with it’s first project to adopt and expand BridgyFed.

support moderators and administrators, and will guide our work in the coming year.

In October, our community demonstrated exceptional resilience during a large-scale spam attack on the Fediverse. This collective effort showcased the strength of our network and our ability to address challenges collaboratively.

We also introduced FediCheck, a transparency tool that helps users evaluate the policies and safety measures of various Fediverse servers. By making this information accessible, FediCheck empowers service administrators to make informed choices about the platforms they engage with.

The benefit to individuals is resilience to media manipulation,

The benefit to society is to decrease polarization and tribalism, decrease outraged certainty on all sides, and increase the quality of public sense-making and good faith civil discourse towards a civilization that can actually coordinate effectively and solve problems.

The goal is to restore the health of our information commons by helping educate people on how to improve their information processing so they can better detect media bias and disinformation while becoming more capable sense-makers and citizens.

I learned to develop my own facilitation style that is less about facilitation and more about being a conversational host.

The professional facilitators would be far too controlling, far too smug and always seemed to love the sound of their own voice. OK, not all, but far too many.

It was serial death by PowerPoint!

But like the Coffee Machine, far too many speakers ran over their time. And weak chairpeople would fail to manage them.

death by PowerPoint

and then there was time for Q&A. Finally, some of us would go to a local pub for a few drinks afterward.

- Behaviorism.  The Thing shall not model other minds in predictively-accurate detail.

Every transaction record includes a complete set of verified timestamps, creating an immutable record of when the transaction occurred. These timestamps include the signed responses from multiple time servers, the TEE signatures verifying when execution began and completed, and any cross-regional timing information. This timestamp bundle becomes part of the transaction’s permanent record, enabling future verification and audit of exactly when and in what order transactions occurred.

The verification process is time-critical — responses must be received within a tight window to be considered valid. If a time server takes too long to respond or provides a timestamp significantly different from others, its response is discarded. This ensures that network latency or server issues don’t impact transaction ordering. The system requires responses from a minimum quorum of servers, typically more than two-thirds of the active servers in the region.

When a transaction begins execution, the TEE pair initiates concurrent requests to multiple time servers. Each request must be signed by the TEE, and each response must be signed by the time server, creating a cryptographic chain of evidence for when the transaction occurred

The TEEs collect signed timestamps from regional time servers, verify them against each other, and include them in the transaction record.

Cross-regional transactions become explicit rather than implicit, allowing for clear rules about when and how value moves between jurisdictions. This matches traditional finance’s approach to cross-border transactions, where different rules may apply to domestic versus international transfers. The system can enforce regulatory requirements for cross-border transactions while maintaining efficient local processing within regions.

Regional deployment of smart contracts creates clear lines of regulatory authority. When a contract is deployed to the Singapore region, it explicitly operates under Singapore’s regulatory framework. This clarity benefits both developers and regulators — developers know exactly which rules they need to comply with, while regulators have clear jurisdiction over contracts operating in their region.

What makes this approach unique is how it maps to real-world market structures and regulatory frameworks. Smart contracts become geography-aware by default, deployed to specific regions with explicit rules for cross-regional interaction.

However, centralized sequencers introduce their own geographic realities that are often overlooked. The physical location of a sequencer creates natural advantages for nearby participants — a fact that becomes crucial at high-performance timescales. While this might seem like a limitation, it actually reveals an important truth: geographic advantages cannot be eliminated, they can only be acknowledged and managed fairly.

The core issue is not centralization versus decentralization, but rather how to acknowledge and design for geographic reality while maintaining the security and transparency benefits of blockchain technology.

Even Layer 2 solutions and rollups, while improving throughput, don’t address the fundamental geographic reality.

High-performance chains claiming sub-second finality often achieve this through centralization while still failing to acknowledge geographic realities — they’re essentially creating a single global matching engine with all the associated physics-based fairness issues.

The blockchain industry’s attempt to make every transaction global fundamentally misunderstands this reality. A trade between two parties in Singapore doesn’t need validation from nodes in Stockholm or São Paulo — the speed of light delay makes such validation not just inefficient but potentially harmful to market fairness. Regional transaction processing isn’t just an optimization; it’s an acknowledgment of fundamental physical laws that cannot be circumvented.

This global-first approach creates artificial barriers to blockchain adoption in traditional markets.

Despite this reality, blockchain architecture has largely ignored geographic considerations

The original promise of cryptocurrency was to democratize finance and enhance local economic autonomy.

The social media landscape has fundamentally shaped how cryptocurrency values are formed and propagated

creating secure borders that enable rather than inhibit efficient cross-border interaction.

“The more the planet becomes a global village, the more vital become the values of cultural identity.”

McLuhan’s concept of the “global village.” McLuhan’s 1962 work “The Gutenberg Galaxy,” which introduced this concept, has been frequently misinterpreted as a call for dissolving all boundaries in favour of a single unified global system. This misreading has significantly influenced blockchain architecture, leading to design choices that paradoxically work against the very efficiency and connection that McLuhan envisioned.

This contract of external accountability keeps the fire going through the long slog, and it forces me to make clear-cut decisions about what to include, what to leave out, and how to manage my spare time so I make progress.

With no deadline, there's no urgency, and so things just don't happen.

Iron Triangle

I think knowing yourself is the most important thing to do before changing yourself.

Everyone has seen the articles about the brain’s approach to social media content, which offers a quick, easily accessible way that makes you happy or, even if it doesn’t make you happy, offers an escape from the thing that makes you unhappy.

The relativity of time is a reality I feel to the core while doom scrolling. Besides the lost time, there’s the confusion after realizing it and putting the phone down, trying to get my dazed mind back to normal. And then, not finding anything to do, not being able to putting yourself together, and reaching for the phone again.

I want to know everything, immediately, quickly. Since this is not humanly possible, I end up doing nothing.

Jules Payot

At this point I am convinced that fast consumption is harmful to the brain, mind, and soul.

Fast consumption, constant consumption, more consumption.

Ethereum’s testnet experienced the problem of this a few years ago. They were using it to check that their local clocks weren’t off by a good amount.

There it is: what if we used an external (decentralized) time source? We’d have proof that you aren’t a time traveller. We could begin to use time as a primitive.

Noone mentions the last attempts to build regulated markets for blockchain and why it was a complete failure.

VDF’s

Any host could easily manipulate that. Intel used to include a hardware based monotonic counter and a trusted time service but that was removed.

TEE’s

In a world where language is weaponized with such precision, every interaction becomes a potential site of influence. Thus now you have a glimpse into the world you are walking into, and what Cognitive Warfare truly is.

This mistrust does not just undermine particular messages; it destabilizes the foundation of shared meaning in society.

Ultimately, the weaponization of language through AI’s information density and resonance creates a feedback loop that erodes trust in language itself.

Once this threshold is crossed, the recipient is not simply receptive to the message; they are absorbed by it, responding almost automatically to the embedded cues. This state of cognitive capture transforms resonance from mere communication to a form of psychological assimilation, where the message becomes intertwined with the receiver’s mental framework, embedding itself within their cognitive architecture as if it were an original thought.

Resonance thereby capitalizes on preexisting cognitive patterns, taking advantage of the brain’s tendency to recognize familiar stimuli and create associative shortcuts, resonant words, and phrases that engage the target’s mental framework, embedding messages seamlessly into their neural pathways.

The functioning of resonance within the mind parallels the architecture of neural networks.

Each resonant word or phrase activates vast networks of associations, turning minimal input into maximal output.

With information-dense language, the goal is not to argue a point but to create resonance, to instill a thought or bias so subtly embedded that it feels self-generated.

The concept of information density is quintessential to understanding AI-led Language weaponization and cognitive manipulation.

Dense encoding can plant persistent cognitive seeds with minimal interaction.

How exactly certain strings of words, or memes are as effective as bioengineered viruses, while others aren’t.

there is another deeper truth to the weaponization of Language and bypassing our mental defenses and “hacking” cognitive subroutines.

direct and shape the cognitive subroutines.

The amplification of certain narratives or linguistic structures by algorithms also presents a form of cognitive warfare, one where the line between organic and engineered influence becomes blurred.

Just as language models can tap into the brain’s implicit linguistic processing, algorithms exploit these cognitive shortcuts to present information in ways that maximize engagement, often reinforcing existing biases or emotional responses. The result is a feedback loop in which language, tailored by algorithms, reinforces the very patterns of thought that the system is designed to exploit.

These decisions are driven by the same probabilistic models that govern language generation, meaning that algorithms, in essence, control the flow and structure of language itself.

Building on the mathematical nature of language, algorithms—particularly those driving AI models—are fundamentally reshaping how language is processed, generated, and disseminated.

At its core, language can be seen as a system of probabilities — each word or phrase is connected to others by a web of relationships, shaped by syntax, semantics, and context. Language models, like those built on transformer architectures, are designed to capture these probabilistic relationships, allowing them to predict and generate coherent sequences of text. This is where AI’s understanding of language goes beyond simple mimicry: it taps into the deep structural and mathematical underpinnings of language itself.

And in mimicry lies a deeper truth to our existence.

This isn't just about the manipulation of individuals, it’s about influencing entire populations through the systematic deployment of language that exploits human psychology. Social media platforms, news organizations, and political movements and now AI-frontier labs have harnessed this power, often without fully realizing the implications. The implicit subroutines of the human mind are engaged on a massive scale, and the consequences remain unrealized to the vast majority of the population, including the ones engaging in such weaponization.

Language is no longer confined to speeches, texts, and broadcasts, it is algorithmically generated, analyzed, and disseminated at a pace that far exceeds the ability of our conscious brain to keep up. With precise algorithmic tailoring, Language can be tailored to penetrate specific patterns in the brain and subroutines, evoking specific emotional responses and affecting behavior at scale.

What makes now distinct from any time before is the scale, velocity, and precision with which Language can be weaponized.

When linguistic patterns are wielded with precision, they can be far more dangerous than overt propaganda or coercion.

thrive on algorithms that curate, manipulate, and amplify language in ways that can bypass critical thought and activate emotional or behavioral responses almost unconsciously.

“Ears have no lid” and one could posit eyes also don’t, in a more philosophical sense, ears lacking a lid infers that you can’t consciously filter sound and the spoken word, therefore you could be affected by words uttered to you. After quite a significant amount of research into neurology, brain chemistry, and language centers in the brain, and attempting to mix mathematics to all of this, I hypothesized the brain could process absurdly more visual data than estimated by “the experts”.

.hack (dot hack)

Our state-of-the-art is still “think of 53 different things and then try all of them.” This isn’t super reassuring—if I hired a plumber to install a toilet in my house and he was like, “Sure thing, I’ll just install 53 different toilets and then check which ones flush,” I’d be like, “perhaps I’ll get another plumber.”6

We do not appear to have progress of this kind. According to this meta-analysis, we’re no better at treating youth mental illness today than we were 50 years ago.

I’d love to see more lists of self-evident phenomena we can’t explain; not “how do we reduce scores on the Modern Racism Scale by 10%” but “people do this weird thing…WHY.”

I think pretty much all progress in psychology can be summed up as “overturning intuitions.”

In order to survive, every human needs to have some model of the world: how their body functions, how animals behave, how matter moves, etc. Psychologists call these “folk” theories—folk physics, folk biology, folk economics, and so on—the kind of explanations you might come up with if you just kinda bumble around, explanations that are good enough to keep you alive, but often go wrong.

I’ve got this picture in my head: we’re all on a bus that’s supposedly going to Cincinnati. But there are no road signs and we don’t have a GPS, so we have no idea if we’re going in the right direction. We can’t measure our progress by how much gas we’re burning, or whether we’ve upgraded from a manual transmission to an automatic, or whether the government bought us a new bus. And you can’t just look out the window and go, “I dunno, kinda feels like we’re headed toward Arkansas,” which, I realize now, is what I’ve been doing so far.

There’s a thought that’s haunted me for years: we’re doing all this research in psychology, but are we learning anything? We run these studies and publish these papers and…then what? The stack of papers just gets taller?

But what was wonderful about the web, was that here was a machine for finding people with similar interests and experience, anywhere in the world. That’s the web I want back.

Because, If we can’t tell a story about something, it’s as if it didn’t happen.

If we want people to see these things as real, we have to integrate them into narrative descriptions of incidents.

Without a narrative schema to anchor it, the pandemic all but vanished from public discourse soon after it ended.

In The 1918 Flu Faded in Our Collective Memory: We Might ‘Forget’ the Coronavirus, Too, Scott Hershberger speculated in Scientific American along similar lines about why historians paid little attention the Spanish Flu epidemic, even though it killed more people than World War I (emphasis mine):

We use stories to make sense of the world. What that means is that when events occur that don’t fit neatly into a narrative, we can’t make sense of them. As a consequence, these sorts of events are less salient, which means they’re less real.

Unfortunately, too much of our era has an emotionally negative predisposition toward too many things, including our current elites, and for reasons that are mimetic rather than justified, whether rationally or even by our impulses to breed.

“I will add only that Nietzsche says somewhere that it is the duty of a philosopher to promote precisely those virtues or tendencies of spirit that are most lacking in one’s own time…”  For all its pretense to the contrary, that is exactly what this book does not achieve.

“The chief intention of this study has been to offer an explanation for why the ancient city perceived philosophers as dangerous and as associated with tyrants — to argue that there was something to the ancient prejudice that philosophy was associated with tyranny.”

I needed to build a bespoke python GUI tool to help me categorize hundreds of websites.

Star Trek-level UI

It will hurt. It will suck. It’s not like starting to use a new, elegant tool on some pet project. A lot of brain power needs to go into this, and it’s mostly boring, menial work. But hey, nobody said this stuff is easy. Not even Vannevar Bush. None of us are entitled to easy work.

And when I inevitably need to review something a few months from now, I know exactly where to look. For example, I will want to measure whether the app is actually getting faster, and I will want to use the exact same methodology and code as at the start. Thankfully, both are right there in my memex trail.

Even if I deleted the folder at this point, it would all still be a beneficial exercise. By putting all of this together, I need to organize my own thoughts. I get to see everything in context. I understand the problem more clearly.

At first, I just add a single document in this folder. This is the project’s main document, or “index”. It probably has the same name as the folder. It should serve as a guide to the project, for others in the company as well as for me (when I look at it 2 years from now). In here, I describe the project and I link to other stuff inside the folder (and outside of it, if needed).

Memex is already here, it’s just not evenly distributed If you look back at the list above, you’ll realize that these things are already very much possible. They might not be in a single shiny app, but that doesn’t matter. (In fact, it’s probably better, in many ways, that today’s memex isn’t a single app. We’ll get back to this later.) To create a new memex “trail” in the year 2020, just create a shared folder (in Zoho Workdrive, Dropbox, Synology Drive, Google Drive, OneDrive, iCloud, etc.) and put some documents in it. That’s it. I know: it’s not glorious. It’s not shiny. It’s just a boring old folder with boring old documents. But I hope to persuade you that it’s good enough, and that you don’t need to wait for some vaporware to work the way Vannevar Bush imagined in 1945.

Imagine that you and I are working in the same company. I tell you there’s a new project for us two to work on. I explain it to you and you get reasonably excited. And then I tell you that I’ve started a new “bloorp” in BloorpyBase, a piece of software from 2012 that almost nobody uses. You grudgingly install BloorpyBase. The app doesn’t use the same keyboard shortcuts you’re used to. The shortcut normally assigned to adding a comment instead minimizes all windows. Sigh. You try to link some exploratory source code to it, but BloorpyBase only works with Mercurial. Sigh. You read some of my initial thoughts and try to respond but you don’t know what’s the best way to do it. Should you create another bloorp? Should you make a suggestion, or an edit? You spend half an hour reading a “How To Bloorp” guide on the internet but come back empty handed. Sigh.

A piece of software that works with your existing files, and which people around you can use, will generally win over some new way of doing things that you first need to migrate to, and then also ask others around you to migrate to as well.

Interoperability is more important than features.

Here’s the thing. If I’m being honest, most of my experiments with the different memex descendants mentioned above just kind of faded after a few weeks or months. And the reason is not just habit. If they were such a huge boon to my productivity, I’d change my habit the same way I changed it for better IDEs, better social media consumption strategies, or better terminal defaults. No, the reason those shiny new apps don’t stick is interoperability.

I have used about 4 different TODO apps and over 5 different text editors in the past 10 years alone. So, go ahead and indulge yourself if you need to.

I think you’ll agree with me that, while Xanadu is a lot closer to the idea of memex than the web, it’s kind of underwhelming as a piece of software. I remember playing with it a few years back, and I just didn’t find it that compelling.

There are projects that explore this space, of course. The most obvious descendant is Ted Nelson’s Xanadu project, a piece of software more than 50 years in the making. You can see its 2016’s incarnation in this video.

For this reason, some people assume that the web is an advanced version of memex. But in fact, the web is a terrible approximation of Vannevar Bush’s original vision.

This blog is primarily for developers, and software development is in many ways a research job.

His “trails” are similar in concept to a physical binder that contains documents that you can freely annotate, highlight and interlink. Any given document can of course be in multiple different binders, and its annotations and links depend on the context. Even for a single person, every “page” can exist in several different forms, depending on what the person is researching or trying to remember at the moment.

What Vannevar Bush proposed with memex wasn’t a publishing platform. It was a memory expansion device.

You can’t annotate a relationship between two paragraphs on two separate sites.

There are hyperlinks on the web, but again, they are not yours.

Memex was to be an “enlarged intimate supplement to one’s memory”.

I will continue to send updates on my blog, and I intend on adding more content on my alpine climbing. I will continue to stream on Twitch, and work on my technical projects.This blog, and the RSS feed will slowly grow into read-only automation that will publish to all of my channels.I do not intend on engaging in social media moving forward. I am too happy to change anything at the moment.Stay tuned for more.

I have rediscovered a sense of self-respect for my ability to accomplish my goals when nobody is watching.

However it has robbed us of the most critical dialogue, our dialogue with ourselves.

You can’t tweet your way to self-respect.

Broadcasting virtue to the world will never provide internal fulfillment regardless of how true it may be. Virtue signaling is effective in shifting public perception, but remains powerless in shifting an internal self-image.

We are a collective of depressed authors trying to persuade other equally hopeless authors of our integrity.

Regardless of how delightful, jarring, shocking, funny, cute, or aggressively virtuous my content was, internally I remained empty. Regardless of how much content I consumed, internally I remained empty.

I remember feeling the need to out-perform my previous display of extreme morality.

We have lost our prerogative to enact change.

I believe social media has drifted to a doom and horror conduit similar to the tabloids.

I notice this internal tranquility seems to be related to my exodus of social media.

In company after company, whenever I present the BAPO model, the first reaction is: Yes, of course! And then, after a few seconds, some or several people in the audience realize that this is exactly the opposite of how their company really works. And the realization (and associated embarrassment) that the company actually uses OPAB sets in. The result often is a lively discussion across the participants on the times the company has falling into the trap in the past and whether we are at risk of falling into this trap again right now.

Meet the seven people who hold the keys to worldwide internet securityThis article is more than 10 years oldIt sounds like the stuff of science fiction: seven keys, held by individuals from all over the world, that together control security at the core of the web. The reality is rather closer to The Office than The Matrix

Problem-solving: There are other kinds of meta-learning, separate from either planning or receiving knowledge. If you’re faced with a problem you’ve never solved before, and you don’t know where to look up an answer (or don’t want to), then you can try to simulate the probem in your head, and mentally consider potential solutions. If you arrive at an idea you like, this is it’s own kind of learning.

Association: Often you don’t know when you’ll need to use a new piece of knowledge, such as learning to ask directions in a new language. In this case, it’s useful if you can recall a relatively unpracticed action based on the correct context.

Once a model has learned to look for a certain key, it’s hard to unlearn. To change the model’s behavior, it seems easier to change what the key points to rather than to get the model to change so that it ignores the trigger altogether.

I’ve phrased things this way specifically because human brains don’t seem to be good at erasing past memories, but rather they seem to be able to replace values associated with pre-existing keys. In this case, the keys are triggers that kick off actions.

action sequences which are initiated by triggers.

Another kind of learning happens at a higher level, which requires longer-term thinking. For example, suppose you write a first draft of a book, and then give that book to some beta readers for feedback. You can view this as a process with many months between the action first taken — writing your first word of a new book — and receiving useful feedback on that action. The recent memory is no longer a useful vehicle for this kind of learning.

For example, if the mind is in a happy mood, it’s more likely to appreciate the positive aspects of a conversation; if it’s feeling defensive, it’s more likely to notice a perspective from which a conversation can be seen as judgmental.

In order for the model to remember something, it must be both (a) something the action model has paid attention to, and (b) something the mind cares to remember based on the emotional state.

Human brains seem to have separate locations for long-term memories and whatever our equivalent of an action model is. Cases of amnesia suggest this: People can forget much of their past while otherwise acting normally. If our memories and behavior depended on the same set of neurons, then this wouldn’t be possible. However, in the mind model above, I’ve let the long-term memory be implicitly part of the action model because this is effectively how language models currently store their version of memories.

Longer-term memories don’t seem to have a pre-determined time limit, but they do tend to fade over time. This pattern is consistent with knowledge baked into LLMs, and so can match the way an action model would effectively remember things — without a time limit, but with the ability to fade over time, especially if not referenced for a long time.

Different people have different recent memory capacities, but it’s common to remember what you ate for breakfast this morning, but not what you ate for breakfast several days ago, ignoring predictability (such as if you cheat by eating the same thing for breakfast every day). This type of memory matches what can fit into the recent memory module.

A third motivation to have a separate recent memory module is that a detailed memory of the past few hours is much more valuable that an equally-detailed memory of some random window of a few hours from when you were four years old.

Just as language models come with knowledge baked into them, an action model is also capable of holding knowledge, but I’ve included a separate memory module. The motivation for the recent memory module in the mind model is a place that can essentially memorize exactly what has happened recently before it’s integrated (through some kind of training) into the action model.

To explain the ideas of memory in this mind model, I’ll split memory into two broad categories: Story memory is the memory of everything that’s happened to you; and action memory is the modification of how you act based on positive or negative feedback.

While I can’t confirm details internal to OpenAI, my educated guess is that these facts are available to the model because they can be selectively added to the prompt. That is, I believe the only common way for LLMs to “learn” today is to implement an additional system to store data from conversations, and to selectively insert that data into prompts when we think it might be useful.

Now the LLM can choose to switch, at its own discretion, back and forth between a talking and listening mode. When the LLM wants to listen, it can produce a special <listening> token many times in a row, until it wants to say something. When it wants to speak, it outputs what it wants to say instead of the <listening> token.

However, we can imagine a change that adds agency to any LLM-like system. Think of a model that receives two interwoven input streams. One input stream is the person talking to the model, and the other is the model being able to see its own output. Current LLMs see both of these streams, but they’re set up so that only one person at a time can talk — the LLM or the user. The difference in the two-input version is that the model is designed from the start to see its own feedback, constantly, as well as simultaneous real-time input from “the outside,” such as the user.

A large language model doesn’t have agency because it can only react to input; it can’t independently take action.

We may even have an effective stack of goals, a small data structure that we can push new goals onto, and pop them off as we complete them.

our current goal fundamentally shapes how we filter the incoming information, and can be edited by the action model itself.

I’m imagining the action model as receiving a lot of data that we could view as one giant vector

I’ve chosen this flow of data carefully. In effect, there are two filters on what we store in recent memory: First, when the action model receives a lot of incoming information, it will effectively pay more attention to some information than the rest. As in a language model, the unused information essentially disappears from the network as it passes through later layers; the attended ideas persist until the end. The second filter is based on our emotional state. When we’re bored, what’s happening is not considered important, and not flagged for longer-term memory. When we’re experiencing an emotional spike, a lot more data is kept around in more detail. Our usual life tends to be somewhere between these extremes.

The emotional state module is doing a lot of work: It’s meant to represent all of our bodily needs, such as feeling hungry or tired, as well as our state of mind, such as feeling elated, frustrated, nostalgic, or intrigued. In this model, our emotional state can change based on what’s coming out of the action model, and it also filters that output into the recent memory module.

Since I’m imagining an action model can be a slight generalization of a language model

perceive visual objects (“face”) rather than a raw image (“pixels of a face”).

Each arrow represents a flow of information. Solid arrows are what I consider to be the most important flows.

2.2 The model at a high level Here’s the model:

You can think of an LLM, in simple terms, like this: context -> LLM -> next_token By analogy, an action model works like this: context -> Action Model -> next_action