The Letter අ

Welcome, our AI User Agents

Sun, 02 Aug 2026 00:00:00 GMT

Previously we took a look at what a UA is, the degree of trust, and the power imbalance. If you didn’t get a chance to read up on it, now is the time. I’ll wait!

arrow_forward The increasingly inaccurately named User-Agent

This post expands the concept a little bit.

User-Agents with AI superpowers …

I’m talking about AI-powered personalized user agents (referred to as AIUAs in this article). That’s a lot of adjectives, so let’s break it down:

AI powered means that a substantial amount, if not all, decision making will be done by an AI.
Personalized means that it will have and continue to accrue information about you that it will use to make decisions.
User Agent means that it will do stuff on your behalf, but unlike before, the actions aren’t limited to fetching stuff from the internet and showing you pretty pictures. The agency extends to performing actions, even interacting with other people on your behalf.

Since they collect and interpret massive amounts of user data and perform actions on behalf of the user, by design, these agents must have access to deeply personal data (Yang et al., 2025).

… need to feed on your data …

Data sharing is a non-negotiable. For now, these AIUAs will rely on a remotely located brain ¹. Hence, the logic applied to decisions affecting you will be unknowable. More distressingly, it may even be unknowable to the people who built them (Aysel et al., 2025). This opacity largely stems from the complexity of AI models, which involve numerous layers of computation, making their inner workings difficult to interpret. Additionally, a lack of comprehensive interpretability tools contributes to this issue, leaving both users and developers with a limited understanding of their decision-making processes (AryaXAI, 2025).

… but can’t swear any allegiance.

In the classic UA trust model, if you see a recommendation for a product, you know a handful of entities whose interests this recommendation or ad serves. With AIUAs, you will have no idea if a recommendation serves your best interest or someone else’s. It could try to help you but fail because its training data is biased toward someone else’s interests. This bias is difficult to detect. (Ryan, 2025)

They are opaque, even to those who make them.

Even with a relatively simpler system like the UA with well-defined trust boundaries, trust remains a difficult problem. We are moving towards indecipherable systems like AIUAs with even fuzzier trust boundaries. We don’t have thousands of pages of specifications for these AI systems. Instead, there are trillions of training tokens that only offer loose guidance. Not to mention basic attack vectors like prompt injection still haven’t been sufficiently mitigated ². Nobody can guarantee where AIUAs’ allegiance lies. (OpenAI, 2025)

They are deceptively trustworthy …

AIUAs can easily gain people’s trust. Some people even fall in love with them (Demopoulos, 2025) or follow them to their deaths. Our politics, values, and consumption habits are easy pickings for a mildly motivated AIUA to manipulate. To make things even worse, people trust what AIUAs tell them about the world without checking. The more people choose AIUAs as the goggles with which they see the world, the more these agents can shape a convincing alternate reality.

… and difficult to align.

Perhaps the most interesting example of an intentionally modified foundational model is Grok, Elon Musk’s xAI chatbot. In July 2025, after Musk announced he had “improved” Grok to make it less “politically correct,” the chatbot began posting antisemitic content on X. This included praising Hitler and making false claims about Jewish people. (Siddiqui, 2025)

Grok isn’t an outlier. DeepSeek shocked the foundation model world when it came out. It was an open-weight model from China whose performance was comparable to the biggest foundation models at the time. Yet there was a problem: DeepSeek didn’t like to talk about Tiananmen Square. (Lu, 2025)

Trying to steer foundation models toward or away from bias often leads to outcomes that diverge from intentions. (Sun et al., 2025) While it sounds like self-sabotage, it’s not absurd to suggest that Google would comply with a request to politically “align” Gemini —hypothetically of course. Google scaled back its DEI hiring goals and other diversity programs in early 2025 after facing political and legal pressure. (Associated Press, 2025) Standing up for what’s right at the expense of profit seems an unrealistic expectation.

In summary,

We see the world through complex yet predictable machines — called user agents. They are difficult to understand, but they do our bidding, more or less.
We are moving towards handing over our agency to even more complex machines. But this time, they are unpredictable, poorly understood, and pretty much impossible to align with our interests in a verifiable way.
The design and behavior of these complex machines may be affected by conflicts of interest that are hard if not impossible to detect.

That’s terrifying. But what can we do about it? In the next part of this series, we will delve into practical strategies and explore cutting-edge research to address and mitigate the risks associated with AI-powered user agents. We’ll examine the role of transparency in AI development and look at how emerging frameworks could help ensure AI systems act in our best interests.

AryaXAI. (2025). Beyond Transparency: Reimagining AI Interpretability Paradigms. https://www.aryaxai.com/article/beyond-transparency-reimagining-ai-interpretability-paradigms. https://www.aryaxai.com/article/beyond-transparency-reimagining-ai-interpretability-paradigms

Associated Press. (2025). Google scraps its diversity hiring goals as it complies with Trump’s new government contractor rules. https://apnews.com/article/16a937d5d9b6447251c4c40c2ad1c915. https://apnews.com/article/16a937d5d9b6447251c4c40c2ad1c915

Aysel, H. I., Cai, X., & Prugel-Bennett, A. (2025). Explainable artificial intelligence: Advancements and limitations. Applied Sciences. https://doi.org/10.3390/app15137261

Balancing explainability and privacy in AI systems: A strategic imperative for managers. (2025). Business Horizons. https://doi.org/10.1016/j.bushor.2025.10.004

Demopoulos, A. (2025). The women in love with AI companions: “I vowed to my chatbot that I wouldn’t leave him”. https://www.theguardian.com/technology/2025/sep/09/ai-chatbot-love-relationships. https://www.theguardian.com/technology/2025/sep/09/ai-chatbot-love-relationships

Lu, D. (2025). We tried out DeepSeek. It worked well, until we asked it about Tiananmen Square and Taiwan. https://www.theguardian.com/technology/2025/jan/28/we-tried-out-deepseek-it-works-well-until-we-asked-it-about-tiananmen-square-and-taiwan. https://www.theguardian.com/technology/2025/jan/28/we-tried-out-deepseek-it-works-well-until-we-asked-it-about-tiananmen-square-and-taiwan

OpenAI. (2025). Continuously hardening ChatGPT Atlas against prompt injection attacks. https://openai.com/index/hardening-atlas-against-prompt-injection/. https://openai.com/index/hardening-atlas-against-prompt-injection/

Ryan, C. (2025). When AI buys from AI, who do we trust? https://www.techradar.com/pro/when-ai-buys-from-ai-who-do-we-trust. https://www.techradar.com/pro/when-ai-buys-from-ai-who-do-we-trust

Siddiqui, F. (2025). Elon Musk’s AI chatbot Grok launches into antisemitic rant amid updates. https://www.washingtonpost.com/technology/2025/07/08/elon-musk-grok-ai-antisemitism/. https://www.washingtonpost.com/technology/2025/07/08/elon-musk-grok-ai-antisemitism/

Sun, L., Mao, C., Hofmann, V., & Bai, X. (2025). Aligned but blind: Alignment increases implicit bias by reducing awareness of race. arXiv Preprint arXiv:2506.00253. https://doi.org/10.48550/arXiv.2506.00253

Yang, J., Yonack, N., Zyskowski, K., Yarats, D., Ho, J., & Ma, J. (2025). The adoption and usage of AI agents: Early evidence from perplexity. arXiv Preprint arXiv:2512.07828. https://doi.org/10.48550/arXiv.2512.07828

Not all large language models are “hosted,” which means that the model is actually stored far away in a server and you can only talk to it through a mediating service. There are powerful language models that you can store in your own computer. Currently, there is a performance gap between the so-called open weights models and their larger hosted counterparts. If that wasn’t a big deal, any language model that you can run on your machine is going to require some pretty hefty hardware. So tl;dr, it is possible to run a large language model on your computer. But that’s not an option for most people.↩︎
Prompt injection refers to a security issue where someone can craft a malicious prompt or include some malicious text that, when used with a large language model, causes that model to misbehave.

There are ongoing efforts that are trying to address these concerns. Researchers are exploring potential solutions, such as model auditing, to assess biases and decision-making processes within AI systems. Additionally, improving explainability is a key area of interest, as it would enhance our capability to interpret AIUA actions and decisions. Developing technical safeguards, such as robust encryption and secure access controls, could also serve as preventive measures against misuse. (“Balancing Explainability and Privacy in AI Systems,” 2025)↩︎

Something Happened On The Way To Perfect History

Sun, 15 Feb 2026 00:00:00 GMT

In the wee hours preceding dawn of the 29th of June, 3123 BCE, a Sumerian astronomer looked up at the night sky and recorded the trajectory of a large object. The object is believed to have been over a kilometer in diameter (about 11 football fields). Its impact is known as the Köfels event whose aftermath included a 900km high ejecta plume, a 500m thigh debris layer, and climate change.

Roughly 5147 years¹ later, the security cameras at a home near Charlottetown, Prince Edward Island (PEI), Canada, captured the first sight and sound of a meteor striking Earth. The footage is of substantially higher fidelity than the Cuneiform tablet. The sea of discussion, scientific research, new reports, and social media activity surrounding this event makes it irrefutable.

The Era of Perfect History is the period in human history during which every pertinent detail of even minor events would have photographic or video evidence. We are not there yet, and it’s likely that we never will.

Imagine billions of people running around with high-definition cameras that they can pull out and record at a moment’s notice. And now imagine there are dozens of social networks that reward sharing footage of any notable event with an audience of millions, and the fame and glory that come with it. This sounds exactly like where we were heading, doesn’t it?

No government would be able to lie to its citizenry. Every misstep or misstatement would be backed by irrefutable evidence. Just as the internet was supposed to bring about a perfectly informed society, with unfettered access to the truth, a history of perfect information was also supposed to bring about a just and transparent society, one where everyone can be held to the truth… eventually. However, some argue that the sheer volume of information could lead to manipulation, misinformation, and erosion of trust. The optimists’ view of perfect information bringing about a just society might overlook the complexities of human behavior and societal dynamics. Nobody planned this, but it was just the direction we were heading as a species.

The past couple of years changed many things, one of which is the irrefutability of video or photographic evidence. As of this writing, it is no longer possible to reliably tell AI-generated imagery from “authentic” imagery. (E.g.)

It is frightening how scarce reality is compared with what can be synthesized. Intuitively, it may seem to be the other way around. After all, our capacity to simulate reality is vastly inferior to reality itself. The discrepancy comes from the volume with which generative content can be injected into our discourse. While we labor for hours making squaking sounds with our throats or poking plastic with our meat sticks, generative content can spit out a million articles. As Henk van Ess says, “traditional fact-checking takes hours or days. AI misinformation generation takes minutes.”

Hoaxes and conspiracies abound, but so far we haven’t seen a substantial historical event be completely synthesized via AI. Either that or the falsehood was so perfectly executed that nobody can tell.

It would make for a great science fiction story until it actually happens.

False Precision and Vibe Coding

Fri, 13 Feb 2026 00:00:00 GMT

It occurs to me how much of human communication is disambiguation.

We have short lives and shorter attention spans. So we have to be tactful about fitting in just enough detail into the brief blips of attention that we are afforded. We wouldn’t enjoy listening to protracted descriptions of things we already know. So we fast-run along a stream of concepts where we only lay down just enough track for people to keep up.

We narrate things.

We tell stories, episodic journeys with familiar names and places and little else other than dialogue and the occasional pantomime. We convey so little detail, but for the right audience, the story is as vibrant, colorful, and detailed as anything from their own sensory experiences.

Things unsaid are those that don’t need saying. The listeners know how to fill in the gaps.

Effective communication requires knowing how to balance what to say and what to leave out —either because it is a detail that is unambiguous or because it’s an intentional gap that we want the listener to fill.

Programming, at least the “classical” kind, does not afford us this luxury. Everything has to be laid out precisely and in detail; even when those irrelevant details. So we seek the comfort of layer after layer of abstraction that hides details we don’t want to see. The same abstractions alleviates us of the responsibility of saying what we don’t need to say.

We hope that, much like our human conversation partners, the computer would fill in the gaps predictably and reliably. We hope it will do what we mean, not what we said. We are disappointed, frustrated, and occasionally amused when reality surprises us.

Software is a description of a machine.

Software is only useful when something interprets it and actualizes the machine. Actualizers –a.k.a. computers– aren’t smart. Every layer of a computer system needs to be told what exactly it is supposed to do. There cannot be any guesswork. So the description of the machine must contain excruciating detail, because the computer that actuates it has no ability to fill in the blanks.

We start with a vague understanding of the problem space. But somewhere along the way, we have to inject it with false precision so that the description becomes complete.

This is why so much of modern software is based on false precision—at least when it comes to end-user-facing systems. We are forced to describe a behavior in rigid detail despite our incomplete understanding of the problem space or lack of desire to overspecify the wrong solution. The precision we inject becomes part of the solution, and in large systems, indistinguishable from pertinent details. Sadly, our descriptions of machines rarely allow us to demarcate necessary precision and unnecessary precision.

Until now, that is. Finally, we have something that can take an incomplete, vague description of our machine—stated in terms of what it does or what its effect should be on something: vibe coding.

Or to be more specific, vibe coding specifications.

Our final actuators are still brute-force stupid. But at least now we have something between us and the actuator that can fill in these pesky gaps and again allow us the luxury of ambiguity.

Ambiguity allows us to be earnest in our imperfect understanding of the problem. We didn’t understand the problem because we have no perfect understanding of the messiest, most important part of the problem—us. But now, finally, we can state our problem the way we see it: vaguely and imprecisely.

The Increasingly Inaccurately Named User-Agent

Mon, 09 Feb 2026 00:00:00 GMT

The User-Agent, a technical name for the web browser, was supposed to be our loyal butler on the internet – doing our bidding and keeping us safe. But over decades, browser makers diluted its role, sometimes as a result of actions done with the best of intentions. Both the internet and browsers have become increasingly complex. And with that complexity, the custodianship of the users’ data has become murkier.

Despite being the one peddling our data, user-agents can no longer tell us exactly who knows what about us. We’ve lost our agency over our browsing data. And now our dude on the internet is a “user-agent” in name only.

We will soon have to confront an objectively more chaotic creature – an AI-powered personalized user-agent. Personalized meaning it knows a lot about us and will continue to learn about us. User-agent meaning it will take actions on our behalf. And, AI-powered at this point means that nobody really knows how it works. Let’s talk about what is about to happen and how you can prepare.

The User-Agent is supposed to be your buddy.

Information is intangible – you can’t stick your finger in it. So whenever we want to interact with a system that processes information, especially of the digital kind, we need some stand-in for our finger so that we can poke things.

So, people came up with the concept of a user-agent – to refer to the device¹ that people interact with. People would push buttons, make holes in paper, turn knobs, or do whatever they need to do to communicate with the user-agent. The user-agent grasps what the person wants, and does what it needs to do to make the person’s wish a reality. On the flip-side, the user-agent also communicates results back in a form that the person can understand.

Early user-agents were simple creatures with no intent of their own. It wasn’t that hard for people to look inside them to understand what it was doing. There was no mystery. The user-agent did what was asked, and earnestly reported back on what happened.

For example, in an email system, this user agent is called a mail user agent MUA. A user-agent that interacts with the web on our behalf should be called a web user agent (WUA)². But because web user agents —which you might refer to as web browsers— are so dominant, people dropped the superfluous “web” part.

MUTT(1)                     User Manuals                    MUTT(1)

NAME
       mutt - The Mutt Mail User Agent

SYNOPSIS
       mutt [-hNpRxZ] [-s subject] [-c cc-addr] [-a file] ...
            [-F rcfile] [-H draft] [-i include] [-f mailbox]
            [address] ...

DESCRIPTION
       Mutt is a small but very powerful text-based MIME mail client.
       Mutt is highly configurable, and is well suited to the mail
       power user with advanced features like key bindings, keyboard
       macros, mail threading, regular expression searches and a
       powerful pattern matching language for selecting groups of
       messages.

       "All mail clients suck.  This one just sucks less."
       - me, circa 1995

An ancient “manual page” for the Mutt email client circa 1995. Note that the program is described as The Mutt Mail User Agent.

This was fine and dandy. People would fire up their user-agents, type commands, and the user-agent would do stuff and show the user results. But over time, these user agents evolved from simple command-line tools to much more fancy interactive, graphical browsers and services. Modern (web) user agents are not just programs, they are entire platforms³.

Modern User-Agents are technical marvels.

Since the same internet needs to be accessible from different user-agents for perpetuity, the way web pages are written and made available on the internet had to be standardized – i.e. everyone needed to agree on the how. But these technical specifications, or standards grew bigger and bigger and so demanding that only a few well-resourced tech companies can write and maintain a full-featured UA⁴.

Though voluminous, these standards are neither perfect nor complete. But they have been crafted through person-millennia of effort to ensure security and correctness. Organizations like the following work tirelessly to get the details right:

World Wide Web Consortium (W3C) is primary group that governs how web content is written.
Web Hypertext Application Technology Working Group (WHATWG) is the group that pushes the boundaries of what is possible to do inside the confies of a web browser.
Internet Engineering Task Force (IETF) is the group governing how data moves about on the internet, and also how to keep your content secure while it is being moved about.
The Unicode Consortium is responsible for which letters and symbols you can use on the internet. They are the people who bring you new emoji every year.

… and many others.

Beyond technical requirements, there is an increasingly important social contract: That the user-agent always act on behalf of the user.

That means that the UA should do what you ask, and generally not do what you don’t ask.

The latter –don’t do what the user didn’t ask– is a little complicated because of all the complexity involved in the modern web.

Say you ask to see bank.com which is a bank. So the UA talks to the bank.com server and asks for the contents of the page. The contents say that some of the pictures on the web page need to be fetched from advertiser.com. But doing so involves sending a bunch of information to advertiser.com like the fact that you are viewing bank.com and information like your preferred language and whether you’ve been to these sites before.

Pretty much all UAs will opt to fetch the resources from advertiser.com, because otherwise the bank.com site will not work correctly. Since the user requested to see bank.com, clearly they intended for all the required stuff to get downloaded too, right? Except, it’s likely that the user never intended for advertiser.com to learn all that information about them.

That the User-Agent must act ONLY on behalf of the user.

Herein lies the problem.

The modern web is a marvel in that it exists despite brutally conflicting interests, and is run by people who do not trust each other one bit⁵.

Do you really think the nice lady who made you read her life story before she told you how to make simple pancakes also made an ad for weight loss drugs? Of course not.

Behind the scenes, a “modern” web page is a warzone where multiple companies are playing tug of war with your data and limited attention span. Not only are these companies clamoring to show you video ads while you are trying to make pancakes, they are also trying to figure out where you’ve been and what you are into.

You, the user, don’t want any of that; not even the nice lady’s life story. You just want to make pancakes. A UA that acted only on your behalf would have just pulled out the recipe part and shown it to you.

In the short term, rewriting the web like this sounds great. But in the long term, those companies – who now have no one to show ads to – will stop paying the nice lady. And the nice lady will stop publishing her recipes. No ads, no recipes⁶.

So to keep the web open, the UA does a tricky balancing act between conflicting interests; one that is not weighed 100% on your behalf. The UAs do their level best to present you with the most authentic web, while trying to keep you safe from potential attackers, without divulging too much about you to strangers. But this situation does mean that the “user agent” isn’t truly just the user’s agent.

And it’s not just users vs. publishers and advertisers. There are many other influential groups on the web making the balance of interests difficult. Some are:

You. Including here for completeness. You want to visit websites and be presented with accurate reproductions of what the publishers intended … kind of. You don’t want to be bothered by stuff that you don’t want, like advertisements. Also you want to be able to consume content freely and without boundaries. You want a free and open web.
Publishers who make content on the web and give you a reason to visit it. There are many not-for-profit content that doesn’t have any commercial interests, but for those that do, it is important that they are able to make money publishing stuff. They also have lots of shared interests with users such as being able to publish what they want and to be able to reach their intended audiences. In addition sometimes they want to be able to publish things without identifying themselves.
Advertisers who compete for your attention and make money by influencing your behavior. One of the ways in which they compete for your attention is by “renting” space alongside publishers’ content.
Search Engines could be considered a kind of a publisher but they are special enough to get their own category. Search engines –at this point, mostly Google– act as the guides on the internet helping users find the content that they want. For this to work, search engines rely on the internet being open and free. Otherwise it would not be possible to take users to where they want to go. Since search engines are one of, if not, the primary tools users need on the internet they are always integrated deep into the UA.
Copyright owners and content rights holders who are worried that people on the internet might steal their content. So web browsers have to incorporate Digital Rights Management components into the UA to limit what users can do with protected content.
Businesses who want to make sure that you continue to trust the internet as a safe place to buy and sell stuff. So web browsers sometimes have to incorporate security measures (which actually help users a lot too) and easy to use payment methods.
Politicians. Yes those people. As much as we would like to treat the internet as some sort of ungovernable utopia, it isn’t. There are many laws and regulations that are different across political boundaries. So sometimes UAs need to report where you are physically so that your activity can be kept within legal bounds. Politicians and businesses often also want to control what you are and aren’t able to access on the internet. However, they take those issues up with publishers and internet service providers. Additional shady things politicians want include finding out the real identities of people who have said things they don’t like.

Shady stuff aside, politicians are also very useful. In fact as we will see below, legislation is pretty much the only way in which users can influence the big UA makers.
Internet Service Providers (or ISPs) make money by making it possible for you to access the internet. You give them some money, and they give you internet access. Seems simple, but there are lots of shady things that ISPs want, like to charge you based on what you consume. See network neutrality.

And that is hardly a complete list, but I believe we covered some of the biggest influences.

But how do we know how the UA balances conflicting interests?

But how do we know that the UA isn’t completely complicit and isn’t selling our furniture without our knowledge? Why would we believe the UA is doing what it says it is doing? What does too much about me even mean?

Well back in my day one would crack open the UA and take a look inside. Usually, things are simple enough that you can convince yourself, or someone you trust can convince you, that it is in fact doing what it says it is doing.

But modern UAs are so complex that such human inspection is impractical at best. Not only would you have to inspect millions of lines of code, you would have to inspect every new change as it comes in. Even then, there are parts of most commercial browsers that are closed source (i.e. they won’t let you look inside).

So most of us have to trust the judgment and moral compass of the thousands of specialists poring over the underlying logic.

Are UAs really unverifiable?

Okay, so this might be a bit too harsh. Aren’t the biggest names in the browser game open source? Aren’t they all, like, based on Chromium? Why can’t we just read the source code and figure out whether it is doing anything odd?

The reality is that people have and are actively watching open source browser projects with eagle eyes. There are even more people making sure any new standards on the web are good and are in the interest of users. Unfortunately, even if we can verify the open source parts of the browser, that’s not the end of the story. See, when you visit a web site, there could be data being sent to hundreds of websites. Even if the browser tells you each and every one of those web sites, it would not be practical for anyone to manually vet and understand what is going on. Even beyond that, not all the functional bits of a browser are open-source. In fact, perhaps the most questionable parts of a browser –like AI integration, media rights management, how it works with various hardware devices– are closed source for many reasons.

So it all boils down to whether it really matters if your security depends on you performing something infeasible for all practical purposes; kinda like being forced to read each and every privacy policy and binding contract you have to click through.

It matters who owns the User-Agent.

We’ve established that modern UAs…:

… are super complicated.
… have to blance conflicting interests of the user, publishers, advertisers, search engines, other businesses, copyright owners, politicians, internet service providers, and others.
… are not designed 100% in your favor.

Despite vigilant oversight users still have to trust, without verification, how a UA works.

If you don’t trust one, you can always pick another. But you can’t avoid taking someone’s word that the UA is safe. Because of how little control we have as users over browsers, the only real way in which users can influence UAs is via legislation that holds UA owners accountable for their design decisions.

Yasskin, J., & Capadisli, S. (2026). Web user agents. https://www.w3.org/TR/web-user-agents/#audience. https://www.w3.org/TR/web-user-agents/#audience

I’m using the term “device” loosely here because the user agent often consists of an entire stack of hardware and software. In literature, you’ll find the terms “device,” “machine,” and “system” used interchangeably, depending on the context.↩︎
The term web-user-agent doesn’t appear much in technical literature with a rare exception. E.g. (Yasskin & Capadisli, 2026)↩︎
What’s the difference between a program and a platform? The latter is something upon which other programs can run. While the original world-wide-web was document-oriented (i.e. the “internet” consisted of documents that linked to other documents), as technology progressed, the web browser became more and more capable of presenting complex behaviors. At the start they were just aesthetic – like blinking text or changing how a link looks like when you hover the mouse cursor over it. But over time, they have gained capabilities that rival native applications. Look at all the APIs that are available inside a browser.↩︎
I worked on Google Chrome for over a decade on the networking stack, and briefly on the renderer. The amount of documentation that it takes to describe how a web browser works is in the order of tens of thousads of pages.

Even then it is an understatement. The web works despite countless quirks and misbehaviors. That’s because modern browsers handle those quirks “correctly” to give the impression that everything is fine. These quirks and how to handle them aren’t necessarily written down anywhere. Browsers can’t just refuse to talk to large portions of the internet or refuse to render millions of older web pages just because the code wouldn’t look good.↩︎
Pun intended.↩︎
This ad-supported model of the open web sounds terrible; and in some ways it is. There are some alternatives that have been proposed over the years, but none have been successful as ads. Many approaches involve internet users making payments of some kind to publishers. Unfortunately, all such models have the downside that those who are impoverished will get kicked out of our global information village. So for now, this is the best we can do.

AI will probably do in a few years what privacy advocates have tried to do for decades; which is to kill the ad-supported web. As yet it’s unclear what will replace it, and whom this replacement leaves behind.↩︎

What's your cog ratio?

Tue, 16 Dec 2025 00:00:00 GMT

As far as product development is concerned, the only thing that matters is the change we make in the user experience. We want our users to be happier because our product exists or, at the very least, less sad.

So when we build things we expect that our work will result in a positive change for our users.

However not everyone is in the business of making direct changes to products. In fact not even developers¹ spend every minute building things. Instead most of the work on a mature product has to do with planning, monitoring, testing, discussing, deliberating, and doing all manner of tasks that have no direct bearing on the material² composition of the product.

Roles like program managers and product managers are one more degree removed from the engineered product. On top of that managers, HR, finance, and all the supporting roles are even further removed from the user experience.

Support infrastructure like build systems, testing infrastructure, and monitoring infrastructure help people build, monitor, and improve the product. However, the people working on the infrastracture are even further removed from the form and function of the final product.

In common parlance, the term cog ratio refers to the ratio of the number of teeth of one cog to the number of teeth of another cog. It translates to the number of turns we expect from the driven cog for each turn of the driving cog.

It’s cliche to think of ourselves as cogs in a large machine. But if we do, it seems natural to wonder what our cog ratio is with respect to the final products we ship our users.

This is specially true when the products we work on are not end user products, but are either infrastructure or support products that help people who help people who … build the product. What would the effect be if this product or service you are working on didn’t exist? By how much do you think your work will improve the lives of our users?

Does the product improve no matter how fast you turn?

Optimal line-length for publishing on the web

Sun, 05 Oct 2025 00:00:00 GMT

Fonts should be pretty, and text layout should be easy to read as possible. When publishing on paper, publishers can control the physical format for the presentation. We can assume, for the most part, how someone might hold and read a book or a newspaper. Based on that we can predict how various layouts, typefaces, and letter sizes will affect the readability and aesthetics.

Publishing on the web is a bit more complicated. How exactly a web page will be perceived is difficult to predict. Content publishers adapt with responsive designs and share the responsibility for presentation with browser makers and device makers.

Folks old enough to remember old timey newspapers would recall how narrow their text columns were. Before digital layouts, newspapers were published using hot metal methods called linotype.

Linotype newspapers had narrow columns for practical reasons.

A metal line of type (“line-o-type” ⇝ linotype) also called a slug becomes difficult to work with if it is too long. In particular, the machines assembled text line-by-line. This process requires that all the lines are roughly of equal width. A small line width –2 inches in this case– was apparently the right trade-off.

There were additional reasons for the 8 column layouts of yore that were not just ease of arranging type. For example, the many small columns allowed for more flexibility for laying out ads.

More relevant for us is that a narrow column makes it easier to find the start of the next line after reading one. Larger spacing between lines helps; double-spacing isn’t just a trick to push the page count.

Around 2008, New York Times changed the print layout to a wider 6 column layout. The new column size accommodated 40–45 characters. Since fiddling with linotype slugs is no longer a concern, they were able to focus more on readability.

Newspapers – or print in general – is different from reading on the web in terms of distance from the eye to the text. Different reading distances demand different amounts of precision to follow a line of text.

So the goal here is to figure out a width that is short enough that the reader can find the next line comfortably, and long enough to not require excessive effort to read each line. Too long and the reader will break their rhythm with excessive time spent finding the start of the next line. Too short and the reader will break their rhythm with excessive time spent in saccadian movement.

System diagram showing the optimization problem of reading comfort. An increase in line length leads to a decrease in saccadian effort, but an increase in line tracking effort. An increase in line spacing leads to an increase in sccadian effort, but a decrease in line tracking effort. An increase in distance to text results in a decrease in saccadian effort, but an increase in line tracking effort.

What’s the optimal line length for the web then?

Bernard et al. (2003) studied this very question. Their study looked at three different layouts presented on what would be considered a regular desktop screen.

Examples of the three layouts from the Bernard et al. (2003) paper are presented below and are illustrative of what the test subjects saw.

From the paper (emphasis mine):

No differences were found for either reading time or efficiency for either adults or children. However, adults preferred shorter line lengths to full-screen line lengths. In examining perception of reading efficiency, the results were mixed. For adults, the full text lengths were perceived as providing the optimal amount of scrolling in comparison to the two other narrower line length conditions. The narrowest line length condition was perceived as promoting the highest amount of reader concentration, while the medium line-length condition was considered to be the most optimally presented length for reading.

– Bernard et al. (2003)

Beymer et al. (2005) performed an eye-tracking study along these same lines.

From the paper:

Comparing the wide and narrow formatting conditions, our analysis shows that for narrow formatting, subjects (a) read slightly faster, (b) have fewer regressions, (c) retain more information in a post-test of the material, but (d) tend to abandon the ends of longer paragraphs.

– Beymer et al. (2005)

In other words, there’s a slight preference for narrower columns, but there’s no consensus on whether there’s a significant improvement in reading speed. Some evidence exists to suggest that comprehension is slightly better with narrower columns.

Bernard, M. L., Fernandez, M., Hull, S., & Chaparro, B. S. (2003). The effects of line length on children and adults’ perceived and actual online reading performance. Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 47, 1375–1379.

Beymer, D., Russell, D. M., & Orton, P. Z. (2005). Wide vs. Narrow paragraphs: An eye tracking analysis. In M. F. Costabile & F. Paternò (Eds.), Human-computer interaction - INTERACT 2005 (pp. 741–752). Springer Berlin Heidelberg.

An Intuition on Shannon Entropy

Sat, 27 Sep 2025 00:00:00 GMT

This post skips over an in-depth explanation of Shannon entropy and only addresses the intuition surrounding it and how it can be misleading. Googling for “information theory” should come up with plenty of options for a more thorough treatment of information theory.

In information theory, entropy is a way to measure “uncertainty” or alternatively a measure of “information”. The field of information theory itself was first formalized by Claude Shannon who’s name comes up often in the literature.

Conceptually information and uncertainty are opposing forces. More of one and you have less of the other.

Lack of information leads to uncertainty. In the absence of any information we are in total uncertainty.

In practice in any given situation where we are concerned with information begins with a well defined foundation which lays out a set of possible outcomes. Each piece of information we receive should narrow down the set of possible outcomes until we reach an unambiguously certain outcome.

If we only have one possible outcome, then we are certain of that outcome. The measure of uncertainty could be thought of as a way of measuring how far we have to go to get there.

Shannon entropy is a measure of uncertainty

A good example of this is the classic game “Guess who”, where you have to guess a person from a set of pictures based on a few pieces of information.

We want to pinpoint a person of interest; which in this case translates to narrowing down the list of possibilities to 1. Each person has an equal probability of being the person of interest. Therefore at the beginning our uncertainty is 1 in 24 — or equivalently \frac{1}{24}.

Let’s look at two possible pieces of information:

wears glasses – matches 5 persons
presents as a man – matches 18 persons

We started with 24 possibilities. Once we know that the person of interest is wearing glasses, we can narrow down the field to 5. On the other hand, knowing that the person presents as a man only narrows down the field to 18. The first piece of information is clearly more valuable than the second. By “valuable” we mean “reduces uncertainty”.

Entropy is a way to quantify –or measure– this valuableness.

Choices of measure for a single outcome

For the sake of discussion, let’s say the probability of a single outcome (call it A) is \Pr(A). And the measure of information learned from knowing this outcome is M(A), where M is some function which we will figure out here.

Adding independent pieces of information should correspond to adding entropy

Channeling some high school math, we know that if the probability of event A happenning is p, and the probability of event B happening is q, and if these two events are independent, then the probability of both happening is going to be pq.

For example, the event person is in the first row (A) has a probability of p = \frac{1}{3}, while the event person is in the third column (B) has a probability of q = \frac{1}{8}. Meanwhile, both of these happenning at the same time (A \wedge B ; where you can read the \wedge as and) has a probability of pq = \frac{1}{3} \cdot \frac{1}{8} = \frac{1}{24}. I.e.

A \text{ and } B \text{ are independent events} \implies \Pr(A \wedge B) = \Pr(A) \cdot P(B) = pq

A good measure should be intuitive. Which is to say, it should behave in ways that we expect. So for example, if we were to take two independent signals, adding them should yield a measure equal to the sum of each measured separately.

\begin{equation} A \text{ and } B \text{ are independent events } \implies M(A \wedge B) = M(A) + M(B) \end{equation}

The simplest way to convert a multiplication like pq to a sum is to … drumroll … take the logarithm. So we could start with M(p) being something like \log(p).

More uncertainty should correspond to a larger entropy

To further intuit, an outcome with a lower probability is more surprising and hence strictly more informative than an outcome with a higher probability, we need:

\begin{equation} \Pr(A) \lt \Pr(B) \implies M(A) \gt M(B) \end{equation}

No uncertainty should correspond to zero entropy

In other words,

\begin{equation} \Pr(A) = 1 \implies M(A) = 0 \end{equation}

because the outcome A is going to happen with certainty.

Putting it all together

The simplest way to make entropy larger for outcomes with lower probability is to negate the logarithm. Doing so has the additional desirable effect that probabilities –which are in the range [0,1]– will yield positive values.

So finally, we have the measure:

M(A) = - \log\left( \Pr(A) \right)

Since Claude Shannon was interested in measuring the information content of digital signals at the time, it made sense to use a more familiar base for the logarithm like 2 instead of e.

So the final equation is:

M \left( p \right) = - \log_2\left( p \right)

The choice of 2 as the base has another nice feature that we can now interpret the information quantity as a number of bits. Because, for example, if we have 5 bits of information where each of the 32 values is equally likely, then any specific outcome has a probability of \frac{1}{2^5} of occurring.

So knowing the outcome gets us M(\frac{1}{2^5}) amount of information; which happens to be 5 — the number of bits we started with.

This measure of information for a single event is called the pointwise entropy, or self information. “Pointwise” because it is measuring a single outcome in the probability space.

Choice of measure for a probability space

Now that we have a measure of the information content of a single outcome, how do we measure the information content of a probability space?

For this, we take another simple formula we already know; which is the expectation. Hence we define the information content of a probability space to be the expected pointwise entropy.

This entropy of a probability space is often denoted H(P) where P is the probability space (defined formally below). For simplicity we can think of P as being a function of some outcome to the probability of that outcome.

The expected pointwise entropy is therefore:

\begin{equation*} \begin{split} H(P) &= \sum_{\omega \in \Omega} P(\omega) \cdot M(\omega) \\ &= - \sum_{\omega \in \Omega} P(\omega) \cdot \log_2\left( P(\omega) \right) \end{split} \end{equation*}

Where \Omega is the set of all possible outcomes.

Formalities

A common critique is that Shannon entropy can be misleading. Let’s see why.

Let’s start with a probability space. It consists of:

A set of possible outcomes. Usually designated as \Omega.

For example, the set of possible person-of-interest consists of the 24 individuals who are pictured above.

Note that here we are dealing with a discrete probability space in that we have a countable set of possible outcomes. This isn’t always the case.
A set of subsets of possible outcomes. Usually designated as \mathcal{F}.

In our example, it could be people with moustaches, people wearing hats, people with glasses, etc. The set doesn’t need to have any semantic meaning. So any group of people from the 24 possible people is a valid subset.

Typically there would be one element in \mathcal{F} which contains all the possible outcomes. I.e. we say \Omega \in \mathcal{F}.
A probability function. Usually designated as P, that when applied to an element in \mathcal{F} yields the probability of that set of outcomes occurring.

For example:

P(\text{wears glasses}) = \frac{5}{24} P(\text{presents as a man}) = \frac{18}{24}

Also, when applied to \Omega (i.e. the set of all possible outcomes), it yields 1 as one would expect since otherwise it would not be the set of all possible outcomes. I.e.:

P(\Omega) = 1

Let (\Omega, \mathcal{F}, P) be a probability space.

So, \mathcal{F} is the set of all possible subsets of \Omega, which is also written as 2^{\Omega}.

\mathcal{F} = 2^{\Omega}

… and …

P is a function that maps from \mathcal{F} to the real numbers in the range 0 thru 1 inclusive.

P : \mathcal{F} \mapsto [0,1]

The Shannon entropy of the probability function P is given by:

H(P) = - \sum_{\omega \in \Omega} P(\omega) \cdot \log_2\left( P(\omega) \right)

Incentives

Tue, 12 Aug 2025 00:00:00 GMT

Teams can fail in many ways¹. But when they succeed, the most critical roles are those that form the communication structures within and across teams. Forming these structures require teammates to have some level of autonomy. In turn this requires a team stucture that allows the team the freedom of movement to reconfigure itself to serve the teams’ goals.

If the conditions are right, a healthy team will fill in the gaps by rebalancing their responsibilities. Such a rebalancing will inevitably lead to an uneven distribution of burden. With time the burden will become increasingly visceral, which will lead to burnout, resentment, and attrition. For an uneven burden to be sustained, that burden must be commesurately rewarded.

Accurately measuring the contributions of each team member is a hitherto unsolved problem. Managers are often left with gut instincts to fall back on, which comes with a laundry list of shortcomings like bias and favoritism. Those who come up with the right formula against all odds are celebrated as good managers.

People write books and give speeches about their favorite good manager. But often the actual skills involved aren’t articulated. The list of skills isn’t long. Primarily these skills revolve around the ability to understand what each person is doing. Secondairly they revolve around the ability to understand how those actions affect the rest of the team. Both require a manager to have visibility into how the team interacts – either by direct observation or by frequent and extensive communication with each team member.

All of this still won’t do much if the team desn’t have psychological safety. Without this, team members may not be willing to recognize and reward their peers. Nor would they be liklely to disclose bad news in a timely manner.

Bad incentive structures lead to people ridigly performing just their roles and letting things fail around them (“it’s not my job”).²

To summarize, good things happen when the incentives for a team:

Never pit employees against each other. One person’s success cannot be another’s downfall.
Assume that the leadership hasn’t figured out exactly what needs to be done. The team should be empowered to course-correct as they learn.
Reward those who pull the team together and help the team communicate internally and externally.

Ode to Glue People

Tue, 12 Aug 2025 00:00:00 GMT

The concept of glue work isn’t new. Team building, mentoring, running meetings, and many other mundane tasks fall under this category of often uncompensated work. How glue work establishes unofficial communication structures within and across teams is, however, discussed less often.

Not every manager is a great communicator; and that’s okay. This is specially true in large tech companies where the people management ladders are the natural next step for individual contributors who are seeking career growth. When important information that affects people’s careers isn’t communicated in a fair and timely manner, trust collapses taking down the team with it.

Bound by needs for confidentiality and mandatory reporting, managers are often not the best people to confide in or trust. Meanwhile peers with the right set of social skills become speakeasies: people with whom others feel psychologically safe. The level of trust between a speakeasy and other team members can be much higher than with managers.

Consequently, these speakeasies become social hubs where both work and non-work related information flow freely. Some might call it gossip. But when the information is factual and timely, it can help teammates made better decisions, at least for themselves if not for the company. In times of turmoil – like the increasingly frequent “once-in-a-lifetime” recessions that keep happening in the US – the glue people take on a dual role: as a pressure release valve for individuals, and as a vehicle for breaking their sense of isolation.

No surprise that such roles cannot be written down in a normal job description. Also of no surprise that such roles aren’t recognized nor compensated.

Soft Roles as Organizational Cross Beams

Tue, 12 Aug 2025 00:00:00 GMT

Beyond the obvious team scoped function of soft roles, they also perform important functions at the organizational level. Of these two common types are:

Glue-people who are necessary to tie teams together, build cohesion, connect the dots between silos that are not effectively communicating with each other, and establish an unofficial foundation of trust that prevent teams from collapsing when the trust between leadership and their underlings inevitably crumbles.
Immune Systems are the people necessary to stop the company from shooting themselves in their corporate feet. These are the people who internally oppose measures that would cause widespread collapse of user-trust, violate privacy policies, open up the organization to litigation, irreversibly misuse customer data, unnecessarily jeopardize the operational stability or security of critical systems, or commit war crimes and human rights violations.

Emergence of Soft Roles

Sat, 02 Aug 2025 00:00:00 GMT

Google’s early success is partly attributed to the use of commodity hardware that were expected to break occasionally. Rather than demanding perfection from complex systems it is far more effective to expect and design around imperfections¹.

Similarly when we design products, we design them for fallible users who are prone to making occasional mistakes, and beholden to some number of misconceptions.

As we design products, so we must design our internal processes. We once again assume that our teammates are fallible, prone to making occasional mistakes, and operating under some number of misconceptions².

Fallibility runs both ways up and down the org chart. Leaders operate under vague, incomplete, and sometimes wildly inaccurate information³. Further complicating matters, it is impossible to understand the capabilities and shortcomings of each individual person or understand every nuance of what must constitute a role.

The result is that team roles are neither comprehensive nor precise. In any given team structure people are expected to adapt to a broad range of situations that are unpredictable, and frankly unknowable. Some (most?) of those situations are directly due to other team members.

As teams mature, the actual role each person takes on will gradually anneal around individual strengths and weaknesses forged by need and cemented through consistency.

So in reality, what is intended to be a neat and well circumscribed role profile becomes irregular, illdefined, and dynamic.

In short: Following the spirit of designing with imperfections, team organization must always account for this fluidity of roles.

Roles in a team will rarely be circumscribed by the corresponding job description. People will always do work outside their scope.
Roles will overlap in ways that will run counter to the intents of the official role allocations. People will do other’s work occasionally.
Roles will emerge that are not mentioned in anyone’s job description. Sometimes called soft roles.

In this series we will look at a few of these soft roles that emerge often and look into the critical work that these roles perform.

A former Googler, Avery Penwarr wrote an influential document while they were at Google called View from Above about the perspective of executives. Fortunately he also wrote a public version called What do executives do anyway?. The basic premise is that when it comes to any big decision, the executive highest up the ladder is the least informed about the pertinent details – but that’s okay because knowing the details isn’t their job.↩︎
See Postmortem Culture: Learning from Failure for how anticipating mistakes lends itself to a healthy culture where people proactively report and learn from mistakes.↩︎
A former Googler, Avery Penwarr wrote an influential document while they were at Google called View from Above about the perspective of executives. Fortunately he also wrote a public version called What do executives do anyway?. The basic premise is that when it comes to any big decision, the executive highest up the ladder is the least informed about the pertinent details – but that’s okay because knowing the details isn’t their job.↩︎

Draining Our Intellectual Catchment Areas

Fri, 04 Jul 2025 00:00:00 GMT

It is July 4, 2025, and many major universities (Alan Blinder, n.d.) in the US are struggling to deal with threats and funding cuts from the Trump administration.

A major impact of funding cuts is the loss of enrollments of promising students due to fewer programs. International students now face an even more uncertain future where they may not be able to enroll at all despite being accepted (Harvard Staff, 2025), their student status may be invalidated without notice (Big Immigration Law Blog, 2025), they are detained without recourse for exercising their rights (Charlotte Alter, 2024), or they are detained and denied entry to the US (Melissa Hanson, 2025).

In addition, funding cuts have resulted in universities having to freeze academic programs or renege on accepted students (BestColleges Staff, 2025), jeopardizing ongoing research programs. Also as a side-effect, promising academics both in the US and abroad are looking towards universities outside of the US (Sequoia Carrillo, 2025).

Let’s have a look at the role these academic institutions are playing in the US’s industrial dominance, and the long-term effects of hampering their progress.

Migrations of Intellectual Capital

Throughout history a number of events have contributed to the present distribution of intellectual capital. As of this writing (July of 2025) the intellectual capital – especially when it comes to information technology and biotech – is heavily concentrated in the east and west coasts of the United States.

Use the interactive timeline below to explore important events in the migration of intellectual capital to and within the United States.

For a more detailed treatement, see The Accumulation and Migration of Intellectual Human Capital in the United States: A Historical Analysis.

When:

Where:

Decline

Just as a reservoir relies on its hydro-catchment area to replenish their water supply, intellectual endeavors are similarly dependent on “intellectual catchment areas” —-regions that gather and channel information, ideas, and people.

As a hydro-catchment area’s health dictates the long-term viability of its reservoir; so does the richness and connectivity of an intellectual catchment area directly influence the vitality of intellectual pursuits.

Without these intellectual catchment areas, the wellspring of creativity and innovation for a particular field risks drying up.

Over many decades, particularly starting towards the end of the second world war, the United States has became one of the largest reservoirs of intellectual capital¹. While the rise of fascism in Europe drove some precipitation of knowledge into the US in the 1930s and 1940s (L{“a}ssig, Simone, 2017), the intellectual dominance that followed in the decades since the world wars was driven by a positive feedback loop which made the US increasingly attractive to the best minds of the world.

The groundswells of industry established around major universities as well as billions of dollars of research funding from the NSF, NIH, and DARPA created prolific centers of intellectual activity. The Immigration and Nationality Act of 1965 opened the floodgates that were arguably already partially open to international students and skilled labor from all over the world. Out of all this emerged ecosystems of collaboration with regional specializations.

These ecosystems built between industry, government, and academia are collectively the US’s intellectual catchment areas. It took centuries to build what we have now.

What’s more important than what it is, is how the system sustains itself.

And the short answer is that it works because it was working before. We were in a positive feedback loop.

Specifically, the fact that the whole world thinks of the United States as the leader in technology means that all the best minds want to live and work in the United States. The fact that the conditions were favorable for skilled immigrants to study, live, and work in the United States so far means that this dream becomes a reality to millions of people.

It’s not just skilled immigration.

The universities need to continue fostering the thought and collaboration that breeds innovation. Until this year, this positive feedback loop has remained stable. But positive feedback loops can break – especially if one were to violate several of the reasons why the loop exists in the first place.

The ongoing threats against these elite universities at the centers of intellectual catchment areas and the ongoing threats against immigration will undo what took a century to build.

Alan Blinder. (n.d.). Trump Has Targeted These Universities. Why? Retrieved July 4, 2025, from https://www.nytimes.com/article/trump-university-college.html

BestColleges Staff. (2025, June 10). These Colleges and Universities Are Cutting Costs Over Federal Funding Freezes. https://www.bestcolleges.com/news/colleges-university-cut-costs-federal-funding-freeze/

Big Immigration Law Blog. (2025, April 10). The New Risk for Global Talent: F-1 SEVIS Terminations and Student Visa Revocations. https://www.bigimmigrationlawblog.com/2025/04/the-new-risk-for-global-talent-f-1-sevis-terminations-and-student-visa-revocations/

Charlotte Alter. (2024, April 2). These Are the Students Targeted by Trump’s Immigration Enforcement Over Campus Activism. https://time.com/7272060/international-students-targeted-trump-ice-detention-deport-campus-palestinian-activism/

Harvard Staff. (2025, June 12). Federal judge blocks Trump plan to ban international students at Harvard. https://news.harvard.edu/gazette/story/2025/06/federal-judge-blocks-trump-plan-to-ban-international-students-at-harvard/

L{“a}ssig, Simone. (2017). Strategies and Mechanisms of Scholar Rescue: The Intellectual Migration of the 1930s Reconsidered. 84, 769–807.

Melissa Hanson. (2025, June 25). Foreign Harvard students were detained at Logan Airport, denied visas after Trump proclamation. https://www.masslive.com/news/2025/06/foreign-harvard-students-were-detained-at-logan-airport-denied-visas-after-trump-proclamation.html

Sequoia Carrillo. (2025, June 5). International students look to the U.K. instead of the U.S., as Trump squeezes visas. https://www.npr.org/2025/06/05/nx-s1-5415740/trump-international-student-visas-uk-universities

Act III: Flame

Wed, 07 May 2025 00:00:00 GMT

We are prone to anthropomorphizing anything that looks or acts vaguely lifelike. All the way back in the 1960s, when people were confronted with a rudimentary chatbot, they were convinced that it embodied empathy and understanding (ELIZA effect)¹.

Roll forward to the 2020’s and chatbots are substantially more complicated. Unsurprisingly, people attribute more and more human qualities to bots. We saw people stake their reputations on the claim that some AI systems are sentient and therefore have feelings that might be hurt by humans’ relentless experimentation.

There are obvious ethical considerations that follow from such claims. Anthropic recently announced (April 24, 2025) that this is a question they wish to explore in earnest.

Anthropic’s work is very forward looking. Currently, we are nowhere near an answer to the question of whether LLMs are conscious. However there are good reasons why they aren’t.

Before answering the question of whether something is conscious, one must know what consciousness is. We don’t know what that is yet (see The hard problem of consciousness). Our near term goal is to establish what’s called neural correlates of consciousness – observable phenomena in the human brain that are highly correlated with the subject’s state of consciousness. We haven’t yet (as of this writing on May 5, 2025) achieved this near term goal.

Short of knowing what brings about subjective experience, people resort to convoluted experiments like the Chinese room to eliminate situations where consciousness is absent and to point out that nothing along the path we are traversing really reaches any meaningful self-awareness.

Is this picture of a flame hot? There never was an actual flame in this instance because this image was generated by AI. The AI may have seen pictures of many real candles, but it has never felt the warmth of a candle. It is therefore safe to assume that the sense of heat played no role in this picture.

What if we start with a more realistic looking model of a flame. Then we can calculate how the flame should behave² with time. No matter how detailed this model is, the model does not constitute a flame. We can crank up the fidelity and accuracy of our calculations as far up as technology allows, but it will still not burst into flame³.

This is where we are with AI and consciousness. We have descriptions of how to determine the next word in a sequence of words. These descriptions are extremely detailed and based on unfathomable amounts of information. They are so complicated that nobody knows how they behave, so the only way we can find out what they will do is to run them against some input and see what happens. We know these models can be made even more complicated than their current, already incredible, complexity.

But no matter how far we go, we still only have a description and a thing that can interpret this description. We can increase the complexity of the description or increase the speed of the interpreter, but there is no setting on this continuum where information suddenly becomes self-aware. It is intuitive that the description will not achieve any physical accomplishment no matter where we take it. So the mechanism of consciousness must depend on the implementation of the interpreter. Currently, that too has no “ignite consciousness” threshold.

So this post will also end the way Anthropic’s announcement does:

“There’s no scientific consensus on how to even approach these questions or make progress on them. In light of this, we’re approaching the topic with humility and with as few assumptions as possible. We recognize that we’ll need to regularly revise our ideas as the field develops.”

— Anthropic

Act I: Meaning

Sun, 04 May 2025 00:00:00 GMT

It doesn’t look like much because I haven’t told you what it is yet. That’s going to come in the following paragraphs. All I’ve done is toss an abstract concept up in the air. It seems pointless, but it gives us a canvas to paint with meaning later.

For example, if I tell you that this meaning concerns a feeling – something very abstract – it suddenly looks a tiny bit different. Now look at it.

Not all the words I speak will add meaning equally. Some are digressions that you have learned to ignore like this sentence that isn’t talking about our thing at all. In one ear and out the other. You won’t pay much attention to those.

Still others, like the sentence that is about to follow this one, are in fact talking about our thing, so you will pay attention to those. Our meaning, the one we are talking about, is going to be about a feeling – a calm feeling. Something that you feel when you suddenly run out of ideas or your mind stops racing and you become aware of the environment around you.

The feeling you get when you are in the middle of a forest. A calm forest where you are not worried about any creepy-crawlies. It’s the afternoon, but not a lot of sunshine is making it through the canopy. You can smell the moss – it feels wet and earthy even though you know you can’t smell wet.

Unlike two steps ago, these last few sentences were a bit special. You see, up till now I tossed out a very plain and boring thing in the air and painted it with a couple of distinct and well understood – if a bit vague. But in the last paragraph, I made you dip a brush into your own subjective experience and paint our thing in your mind. The meaning is now yours and yours alone.

And that’s special because your experiences are not things that can be communicated to you. We don’t have language for that – because those experiences cannot be constructed innately. They must be communicated to you through your senses. Those bodies have ways of constructing feelings for you that mere abstract meanings cannot. However once you have those feelings in yourself, our abstract words can conjure them out at will; just like we did when we spoke about the forest.

If we had a succinct way to communicate this meaning in one go, we would have.

But our language and our facility of communicating ideas which we stimulate with our language doesn’t afford us such shortcuts. So our language tends to be very abstract; and sometimes it takes a little shortcut and conjures up a meaning from the back of the subjects’ mind like we did with our allusions. Again, these are not always shortcuts. These are the only ways in which we can communicate certain things – things that can only be constructed in one’s head with one’s senses.

So to understand these ideas, one must have the right ingredients in their minds. To create these things in their heads one must experience things with their senses.

And if we can’t conjure meaning then we can’t understand each other. And if we can’t understand each other then we can’t have compassion and empathy.

Interestingly, this is how we fundamentally differ from the current LLMs. You see, they have another problem. They don’t have any experiences to conjure because they don’t have feelings. So when they encounter language, they can’t understand what those words mean. But one thing that they can do is to reach into their vast embeddings and see how someone who can feel would have responded. And then they can respond accordingly. What they conjure isn’t life experience or feeling, but evidence of others having reacted a certain way to a certain sequence of ideas.

Because they can’t construct new feelings, their universe of feeling will always be limited by what people have felt and reacted to. That way they can look at those reactions and imitate them. The meanings in your head are yours alone.

No one can see your dreams. Especially not the robots.

Intermission - Q1 of 2025

Sat, 03 May 2025 00:00:00 GMT

It’s been a while since I’ve published anything here. This is partly because every time I try to write there is something much more important and terrible happenning in the world that I’m not qualified to write about. Everything else including whatever it was that I was trying to write about seems trivial and inconsequential by comparison.

Many years ago I’d have said that there is very little evil in the world and that most people you will meet are good. I still sort of believe that people are good, but I’m beginning to think that the word “most” may not be readily applicable.

School children has this would-you-rather question about a big red button.

“Hypothetically, if there were a big red button that when pushed gives you a million dollars, but also kills a random person, would you press it?”

It is safe to say that the answer to this question is predominently yes.

More worryingly, the reward doesn’t need to be nearly that high. It seems, for some, there doesn’t need to be a reward at all as long as the person who would be harmed is of a specific race, sexual orientation, body dysmorphia, where they were born, or religion. There’s enough hate to take the place of a reward.

People aren’t born this way. Decades of being conditioned to cast out a group of people gradually dehumanizes them. With enough rhetoric, likemindedness, and reinforcement anything can be normalized. Anyone disagreeing – and is not a member of the outcasts – can be easily disregarded as being the enemy or at least misguided. Hate becomes self-sustaining, and can last generations.

We can win battles or wars, but we are not going to really move forward as a species unless we figure this one out – how to teach compassion, and how to reason someone away from hate?

Contrary to popular opinion I think pushing the hateful away is not a sustainable solution in the long run. Suppressing hateful behavior without corrective guidance has historically just made the hateful group more resentful and even more entrenched in their beliefs. Ultimately they will – as evidenced in recent events – have enough momentum and intensity to turn the tables.

Preface

Thu, 01 May 2025 00:00:00 GMT

Large language models’ ability to mimic human language and show signs of intelligence is quite strange yet makes sense at the same time. This series is a very gentle introduction to how they almost capture human experience through language and how language is perhaps the weirdest invention of all.

It’s going to be a bit of a long ride. Perhaps by the time you read it I haven’t quite finished the whole series. So you’ll see what looks like an abrupt ending or missing episodes. They are coming. Just slowly.

Onwards …

Intrinsic and Extrinsic Motivation: An anecdote

Sat, 22 Mar 2025 00:00:00 GMT

A long long time ago in a totally hypothetical company that still exists but only in name, there was a little bit of discontent¹. You see, one of the core values of the company was its commitment to diversity and inclusion. Rather than lowering the bar, like how some people think inclusion works, they had many outreach programs that encouraged women and minorities to take up STEM. These outreach programs almost necessarily fell on the shoulders of the women and minorities who were already employed at the company – after all representation matters. Those who worked on these did so out of their own passion. They wanted to help others like them.

Anyone who has worked in tech will tell you it’s a lot of work. This company’s performance review system attempted to measure the amount and quality of said work so that they can reward the right people for the right reasons. Since the company deemed that having a diverse workforce was in their best interest, ideally they would reward people working towards that goal. However, the performance review system didn’t measure anything about volunteer work like what the people doing outreach were doing.

So those volunteers were unhappy. All that extra work, often done at the expense of work that is measured in the performance review system, didn’t count for anything when career progress was concerned. They wanted change.

The company listened.

“Community contributions” got added to the job descriptions for people at all levels. Naturally the higher the level, the broader in scope those contributions were required to be. It was now part of everyone’s job to make contributions that made the company’s culture better – be it recruitment, outreach, or whatever.

The changes to the performance review system was an improvement. But unintentionally it had other side-effects. Notably, those higher up the organizational ladder found that their community contributions need to be commesurate with their status. But starting up an effort that is at the company or region level was a lot of work. Lucky for them, those volunteers who were passionate about helping others had already established programs that were of the right scale.

So as career focused people are wont to do, they threw their weight around and effectively took over these outreach programs. The volunteers were upset. Some left. Others kept working but at a smaller capacity than what they were before. The career focused people were happy, but the volunteers were not.

The outreach programs still existed² but were run by people who were no longer doing it just to help other people. They were run by people who were trying to help themselves. The programs faltered; quality went down; the effectiveness of the programs went downhill. It wasn’t fun or rewarding as it was before.

Buffer Bloat

Fri, 03 Jan 2025 00:00:00 GMT

From Here To There On The Internet

Devices on the internet including your phone, laptop, gaming console, refridgerator, etc. communicate by exchanging data “packets”. They are of course not physical packets. It’s just a useful way to describe a unit of data sent over a wire or some other network connection. A packet could be thought of as, say, a single frame in a video or a small slice of audio, or a couple of paragraphs of text in a webpage. Devices on the internet will take a big chunk of data (e.g. a video clip) and break it down into smaller pieces (e.g. a single frame in the video clip) and then sends each of these pieces along the network. The receiving device puts the pieces back together to reconstruct the original thing (e.g. the video clip again)¹.

Each packet has little pieces of extra information to help it get to where it needs to go. One piece is where it comes from (a source “address”). Another is where it is going (a target “address”). Devices along the network (like your internet provider’s networking equipment, satellites, etc.) use this extra information to correctly route the packets to their intended destination.

The network is large and constantly changing. So each device focuses only on getting the packet at least one step closer to its destination. The specific journey taken by a packet is unkonwn until it is on its way.

Faster and Faster

So let’s say a video streaming service wants to send your mobile phone a video. First the streaming service breaks up the video into small packets. Then it throws all the packets into the network so that the network can deliver them to your phone.

If the streaming service pushes data out at exactly the right speed and the network doesn’t break, then everything works out great. All the packets make it to the destination without hassle. But neither the streaming service nor your phone know what that right speed is due to the aforementioned unknowability of the network.

Sending too much data can overwhelm some devices or connections in the network. If a device gets more data than it can handle, it’s likely that it will start dropping packets until it can catch up.

Sending too little data isn’t great either. The speed might be too slow for streaming video. The receiving device will underperform. Users will be confused why their expensive broadband connection doesn’t let them watch streaming video.

So how does the device know how much data to send out?

Flow Control

The mechanisms we use to regulate the flow of data is called “flow control.” We are going to look at a method based on coordination between the sender and the receiver because they are the only two actors in this scenario that we consider to be reliable and stable.

Different packets can take different paths See the network connections between C-D and C-E in the diagram above. Either of these paths can take a packet from C to F. Packets going down different paths will arrive at different times. Also, if the network gets overwhelmed or breaks, then some of the packets may not make it at all.

The reciever needs to know how to put the packets back in order and tell if any are missing.

To keep track of which packet is which, we can add some new information to the data packet: let’s call this the sequence. As you might imagine, the sequence can start at 1 and count each packet until the transmission is over. Each packet now has a unique position in the list of packets.

Great. But how do we know which packets made it to the other side? The recipient will need to let us know this information in some sort of reply. This reply is also a data packet called an acknowledgement. It’s sent from the original recipient back to the sender.

Ok so now we have all the information we need to perform some flow control.

We can now figure out the optimal rate of data transfer experimentally using some method like this:

Send data as fast as possible.
If the recipient claims some packets were missing, then resend those packets and send packets slower than before.
If the recipient hasn’t complained of missing packets in a while, try sending them faster.

Stuck On The Wire

Now imagine you sending data out - a rate of 100 packets per second. And let’s say it takes one second for a packet to get to the other side. Just before the end of the first second, the destination hasn’t received the first packet yet. But the sender has sent 100 packets.

Where are the packets?

We call this “data on the wire’. Some data is indeed on the wire literally, but that tends to be a tiny amount. Most of the data on the wire is actually at each of the devices along the way either being received, or waiting to be sent.

So once we factor in the equipment in the middle, your path to the destination looks a little like this:

The device sees some data come in, stores it, figures out where to send it next, and sends it on its way. There could be other stuff these boxes could be doing, but that’s something for another post.

So from the perspective of a single box the world looks a bit like this:

If this is the limit for the network and the sender keeps sending more packets, then there’s no room for the packet. So it gets “dropped”.

It is common for a device to be at the border of a “good” network and a “not”good” network. The goodness here is the amount of data that can be stuffed into it.

When a burst of data comes along, not much data can be stuffed into the small pipe before no more can fit.

As mentioned before, this means that the sender will need to re-send the dropped packets and also slow down the rate of data. Given enough time the rate of data transfer will approach the throughput of the smallest pipe along the way.

Buffer Bloat

Some folks looked at this setup and decided to improve things. There are some valid reasons for doing what they did – which is why so many people did this. But it didn’t quite have the intended effect.

So, what was the thing they did? They added buffering. Too much of it.

A buffer is a resevoir where excess packets can be kept until the next link in the network is ready to accept more. What this means is that every device along the way is trying to do flow control by storing and resending packets instead of letting the sender know that packets are getting dropped.

Device B is dropping all excess packets into a buffer.

As before, let’s say the sender is sending 200 packets every second. The network still can only transfer 100. The excess packets – another 100 of them – gets tossed into the buffer.

What happens just one second later?

The very next packet – let’s call it Bob – that the sender tosses into the network will end up in the buffer since B is still waiting for some of the first 100 packets to get past C.

Bob is the last packet in the buffer and must wait for all the other packets in the buffer to be send before it will get queued to sent out.

In another second, the first 100 packets make it to the target, and now the 100 packets that were in the buffer are back on the wire.

Bob is still in the buffer but it’s the next one in line to get back on the network.

Once back on the network it will take another second for Bob to reach the target.

Even under ideal conditions eventually the system settles on a new equilibrium with one difference.

The length of time it takes for the data packets to make their way to the destination is much larger. Packets accumulate in buffers along the network. Each buffer adds an additional delay for the packet.

Overall the throughput of the network is the same. But latency – the time it takes for a packet to get to the destination – is much larger.

The effect is even worse in practice because this setup resists any attempt at flow control. It delays the discovery of actual throughput.

This phenomenon is called buffer bloat.

Fourteen Years

Tue, 31 Dec 2024 00:00:00 GMT

It’s been fourteen years since I joined Google. Much has happened since then, and people think the old Google is dead. This is somewhat correct.

One way to understand what is happening is to compare and contrast the Google of yore and the present (as of late 2024) Google.

From Illegal Flower Tribute

Perhaps the most defining moment for Google around 2010 was Google Search’s exit from China. The reasons for the exit and the aftermath are many. The high level reason was due to non-capitulance to the Chinese government’s demands for censorship and compromising user data.

In addtion, Google’s hand was forced by Operation Aurora, a cyber attack on multiple US based companies including Google orchestrated by Chinese state actors.

The fascinating accounts of the events surrounding Operation Aurora is covered in a documentary called “Hacking Google”.

Hacking Google - YouTube

Amazon and Microsoft were in a similar situation as Google. But they were unwilling to turn away from access to the Chinese market. There was a palpable sense of pride that we were working for the right company.

To Project Nimbus

Time changes everything.

Eight years later Google reneged on its stance against censorship by attempting to build a censored search engine called Project Dragonfly. The seeds of capitulance was well under way by this point despite the fact that the censored search engine project was ultimately dropped.

Fast-forward a few more years and we get to Project Nimbus. Google signed the contract despite warnings by its lawyers that “Google Cloud services could be used for, or linked to, the facilitation of human rights violations…”.

One could argue that tools like Google Cloud services should be value-neutral. But even then, what matters here is that Google intentionally changed the terms-of-use for this contract to allow uses that were explicitly called out as going against Google’s values.

That’s Where We Are

And that’s where we are right now.

Google has had some rough years during which its relationship with the rest of the tech ecosystem and the world in general has changed. The old “focus on the user and all else shall follow” matra no longer applies. The company’s focus is now exclusively inwards where the fact that something benefits everyone is no longer justification to continue working on it. For example, Google laid off most of the core Python team. While it may seem like a move towards cheaper labor, it is more than that. The previous Python team were also industry veterans who were effectively in the service of the broader Python community far beyond Google. The replacements are people who are tasked with maintaining internal tooling and not much else.

Google is no longer the giant benefactor of technology for everyone. However it’s still better than the other big tech companies. I keep telling myself that.

I Moved This Blog To Eleventy

Fri, 13 Dec 2024 00:00:00 GMT

Astute observers of this blog may have noticed that the blog looks suspiciously different lately. That’s because I changed the static site generator I was using.

Previously the static site generator I used was Hugo. This arrangement worked for a while but it had some shortcomings.

Growing Pains

I was using Pandoc to convert Markdown to HTML. Hugo’s support for Pandoc had some limitations. For example, you can’t specify options for Pandoc or use a bibliography. When I moved the blog to Hugo, I lost some of these features. Attempting to close the gap I authored a patch that added support for configuring Pandoc.

Pandoc is extensible via filtering the AST of a document through a series of programs which can modify it. There is no restrction on the kinds of programs that you can use.

If Hugo allowed configuring all this via frontmatter, that may introduce some tricky security issues.

Say you download a Hugo theme from the web and try to build your site. A theme could include frontmatter that causes Pandoc to run an unsafe binary. That’s not great. So it is understandable that this kind of feature isn’t a part of Hugo proper¹. While my patch worked, it was safe only because I controlled every aspect of the blog.

While tinkering on this was fun, it created an additional headache. Namely, whenever and whereever I wanted to build my blog, I had to maintain a custom build of Hugo. After moving my build a couple of times maintaining a custom build became a painful chore.

The Exodus

So started the process of looking for an easy-to-use static site generator (hereafter SSG). My requirements were:

The SSG should support Markdown.
Markdown should support some of the Pandoc features I was using, or at least have the flexibility to extend support for those features.
Making this work should not require any custom patches.

Eleventy was almost a no-brainer. People had high praise for it, including developer relations folks at Google. So I felt like that would be an easy move.

Eleventy’s Mardown support is based on Mardown-It. One thing that I liked about Markdown-It from the start is that there appeared to be people who are extending it to support various features of Pandoc. In fact I found plugins for basically all Pandoc features that I used in this blog. Everything was looking quite rosy.

Templating

WebC + Eleventy Is Broken

Eleventy supports a templating language called WebC. It follows the web-component way of defining layouts, but can avoid much of the custom element overhead by opportunistically generating plain HTML. I had already accepted that switching the SSG meant rewriting layout templates, and using web components to do that seemed like a good idea².

A couple of standalone WebC based templates worked fine. But the moment I had nested templates, Eleventy hung and eventually crashed. I didn’t want to go on a side quest to debug it, and ultimately that would’ve likely led to be violating the “no local patches” rule. So that was that for WebC. If WebC was more robust or just worked, it would have been my choice of template engine.

One cool feature of using WebC with Eleventy is that bundling works trivially. For example, if you have a <script> tag or a <style> tag in a component, WebC automatically pulls the content into the relevant bundles. All other templating engines supported in Eleventy (save for third-party engines that I didn’t try) require manual labor to get stuff into the bundles.

Nunjucks + Eleventy Works

The next templating language I looked at was Nunjucks. I had briefly looked at Liquid as well, but for my purposes Nunjucks and Liquid seemed to be quite similar. Nunjucks won out because it had safer defaults³.

Handrolling

Pandoc Again

I tried to rely only on Markdown-It for this blog. But very soon I had a big stack of extensions to support the various features I needed, like supporting <div>s, bibliography, and attributes. While the Markdown-It ecosystem is extensive, it also introduces a lot of confusion. For example, there are multiple packages that all provide the same functionality and also based on the same base package. Keeping track of all this was starting to be a pain. Also, I had trouble getting block level containers working.

After a while I decided not to waste time trying to replicate features that already existed in Pandoc. So I wrote a little library that invoked Pandoc within Eleventy to convert Markdown to HTML. It was a little slower, but worked out great.

Bundling Files

While it is possible to add content to bundles (like CSS and JS), curiously it was not possible to stuff files into said bundles. The suggested method was to do something like this:

  {% css %}
    {% include "some-file.css" %}

    body {
        // other CSS.
    }
  {% endcss %}

This seems like a good idea, but there is one crucial issue: The contents of the included file (some-file.css in the example above) is also processed by the same templating engine. If the file doesn’t contain any sequences that look like template tokens, then we are good. But this is asking for trouble.

So I wrote a custom tag called bundleFile that can take the contents of a file and stuff it directly into a bundle. The earlier example now looks like this:

  {% bundleFile "some-file.css" %}

  {% css %}
    body {
        // other CSS
    }
  {% endcss %}

SASSy

Eleventy doesn’t support SASS out of the box, but there are examples in the documentation for how you can set it up. But all of the examples assume that you have SASS or SCSS files that you are going to conver to CSS, and then serve without bundling. I wanted to be able to compile SASS and stuff the resulting CSS into a bundle.

So alongside the bundleFile shortcode, I added a bundleScss shortcode so I can do stuff like this:

   {% bundleScss "my-styles.scss" %}

Katex

I like Katex for my math typesetting. Pandoc supports it out of the box. But the rendered HTML needs to invoke Katex in the client to render the math. I wanted to avoid this step entirely and try to render Katex on the server.

This was a pain to do with Hugo, but since Eleventy was alredy runing in Node, server-side rendering is a breeze.

Conclusion

Overall I found Eleventy fun to work with. My gripes are mostly with the documentation which lacks a reference. It’s a pain to go through tutorial-like documents to find out what exactly gets passed to a shortcode or what to do when a shortcode adds a dependency. I made way too many trips to the 11ty GitHub repo. The right number of required visits should be zero.

In the coming weeks I’ll package Eleventy Pandoc support and some of the shortcodes so others can use them.

The Letter අ

Welcome, our AI User Agents

User-Agents with AI superpowers …

… need to feed on your data …

… but can’t swear any allegiance.

They are opaque, even to those who make them.

They are deceptively trustworthy …

… and difficult to align.

In summary,

Something Happened On The Way To Perfect History

False Precision and Vibe Coding

The Increasingly Inaccurately Named User-Agent

The User-Agent is supposed to be your buddy.

Modern User-Agents are technical marvels.

But they are bound by the same implicit social contract …

That the User-Agent must act ONLY on behalf of the user.

But how do we know how the UA balances conflicting interests?

It matters who owns the User-Agent.

What's your cog ratio?

Optimal line-length for publishing on the web

An Intuition on Shannon Entropy

Shannon entropy is a measure of uncertainty

Choices of measure for a single outcome

Adding independent pieces of information should correspond to adding entropy

More uncertainty should correspond to a larger entropy

No uncertainty should correspond to zero entropy

Putting it all together

Choice of measure for a probability space

Formalities

Incentives

Ode to Glue People

Soft Roles as Organizational Cross Beams

Emergence of Soft Roles

Draining Our Intellectual Catchment Areas

Migrations of Intellectual Capital

Decline

Act III: Flame

Act I: Meaning

Intermission - Q1 of 2025

Preface

Intrinsic and Extrinsic Motivation: An anecdote

Buffer Bloat

From Here To There On The Internet

Faster and Faster

Flow Control

Stuck On The Wire

Buffer Bloat

Fourteen Years

From Illegal Flower Tribute

To Project Nimbus

That’s Where We Are

I Moved This Blog To Eleventy

Growing Pains

The Exodus

Templating

WebC + Eleventy Is Broken

Nunjucks + Eleventy Works

Handrolling

Pandoc Again

Bundling Files

SASSy

Katex

Conclusion