Control

[AI companies are so overexposed to debt and market correction that their success relies on everyone using it. When it isn't working for everyone, the next best thing is making usage mandatory. They can require it by crippling local computing power and supply to force people through cloud services.] Want storage? That'll be another $20/mo. Want graphics for games? Another $20. You want to perform data science and fit ML Models?! You're going to need the Professional plan, starting at $200/mo. It's a rent-seeker's dream.

Why fight to get one of the few remaining laptops in inventory with onboard horsepower when you can get an OpenAI Terminal just for subscribing? Sure, all it does is provide you with a single interface, a blinking chat box through which your entire digital experience is mediated, but plans start at only $20/month! Isn't that easier? Look how shiny it is. Just relax. Don't resist. It hurts more when you fight it.

The Ma of a New Machine

[LLM chatbots generate "deep cognitive work"-level complexity with an interface for reaction and intuition.] It spits complex text at you, you skim it quickly, and you immediately type a reaction to keep the momentum going.

We’ve been looking at this problem entirely from the technology’s point of view. The LLM is driving us, pushing a specific interaction model and veering us towards making mistakes. We are not computers; going fast isn’t our superpower.

I used AI. It worked. I hated it.

We come now to the inconvenient truth of this technology: that it is built, like so much "progress," on theft. The training corpora of these models includes code with licenses not meant to be used in this way. Even if one could guarantee that copyleft code were not included in output, the entire system of weights and tokens is inexorably linked to copyright infringement. There is no escaping this. To call it theft is accurate in my opinion, but then I'm a bigger believer in copyright than many in my circles. What is the appropriate response, and by whom? How do we respond to the theft of others whose accidents are visited upon us? I write this on the stolen, unceded land of the Chumash and Tongva peoples. I do what I can to remember that, acknowledge that, and teach others what I know of those cultures. I have no idea how to mitigate the harms of the wholesale theft of intellectual property that gave birth to large language models.

I also don't know what to do about the destructive extraction mining that sourced the minerals making up my computer. These human harms are almost surely greater than the theft of writing, yet I am happy to ignore them. I mention this not to wave away the wrongs, but to recognize that all my technology is bloody. I don't know how to remove myself from the entire system in such a way that my hands are clean. I don't know that anyone can do so in the interconnected age.

The fight cannot be among laborers who are all threatened by this technology. The fight must be between the workers who wish to work, create, live, and prosper, and the elites who only seek to enrich themselves by means of this technology.

Moxie My first impressions of web3

I can run my own mail server, but it doesn’t functionally matter for privacy, censorship resistance, or control – because GMail is going to be on the other end of every email that I send or receive anyway.

I can build my own NFT marketplace, but it doesn’t offer any additional control if OpenSea mediates the view of all NFTs in the wallets people use (and every other app in the ecosystem).

[Blockchain transaction fees create an artificial floor on prices that would be more attractive by simply treating platforms as OpenSea or Coinbase as traditional centralized services.] Eventually, all the web3 parts are gone, and you have a website for buying and selling JPEGS with your debit card. The project can’t start as a web2 platform because of the market dynamics, but the same market dynamics and the fundamental forces of centralization will likely drive it to end up there.

Intuitive Guide to Angles, Degrees and Radians

[If a 2-meter radius wheel turns 6 radians per second, scale by radius to get 6 × 2 = 12. If it turns 2000 degrees per second, the calculation becomes 2000/360 or 5 + 5/9 rotations per second, which plugs into the circumference equation 2 × π × r as 2 × π × (5 + 5/9)…]

[Degrees are arbitrarily based on the sun (365 days ~ 360 degrees) and use the observer’s perspective. Radians use the mover's perspective.]

A taxonomy of ATmosphere applications

['Symbionts' use domain expertise to fill Bluesky feature gaps. Their survival relies on Bluesky growing without competing.]

['Offshoots' leverage Bluesky's userbase to construct their own separate communities.] Their permanence will come from their unique services — Blacksky Cash, Eurosky’s jurisdictional arbitrage, Cartridge’s single-minded focus on gaming. They get to play a positive variant of the old Microsoft playbook: Embrace, Extend, Escape.

['Cuckoos' create paid services on top of ATproto's shared architecture. They survive by leveraging the ecosystem without appearing a threat, and also by non-competition from Bluesky.]

Before bed: add 2x 500ml jugs of bread flour, 1 jug of warm water, quarter teaspoon of dried yeast, 2 teaspoon of salt, 1 tablespoon of rapeseed oil or olive oil to a bowl. Mix together with the handle of a wooden spoon. Cover and let rise overnight.

The dough is quite sticky, wet your hands to handle it. pull the edges of the dough to the center a few times to stretch the gluten. Split the dough, put in loaf tins or shape. Leave to rise again. Bake it in a hot oven for 40 mins.

Dropping to log-level

God is very cruel. He only gives us data about the past.

[Use the simplest model with sparse data and complicated models with abundant data.]

[When things move quickly, a log helps us understand what happened and enables us to extract pattens without needing to determine the best structure in advance.]

The case against conversational interfaces

We are significantly faster at receiving data (reading, listening) than sending it (writing, speaking). This is why we can listen to podcasts at 2x speed, but not record them at 2x speed.

[Roughly estimated: mobile typing ~30wpm; writing/typing ~60wpm; speaking ~150wpm; reading ~250wpm; listening ~270wpm; thinking ~1000–3000wpm.]

[Gestures and facial expressions compress data so we can trade specificity for speed.]

Week 6: Options or cash?

[Options are more expensive to transact and likely more volatile, but they enable a capped risk especially in relation to short trades. Trade stops have a gap risk where as options go to zero.]

Rewilding Software Engineering

[Software gets built in economical and social environments that constatly change. What we call 'legacy' reflects an inability to adapt change or a lack of refactoring to match software to its new environment.]

[LLM error rate of ten percent (for example) is problematic when you don’t know which ten percent is wrong.]

Comprehension is not a nice to have. It is the scaffolding that makes safety, accountability, and learning possible

People like simple scalars, like KPIs regardless of whether they are useful or downright harmful. They have the one property that manager’s like … they are simple, because we are busy people.

Criticizing your own tribe is how you win

It’s obvious to me that criticizing your own tribe is a winning strategy. I want my friends to do this, because it will help us win. I’m giving this away as “free advice” to my enemies too, because even if they win, it still creates a world that’s better for me.

X Is a Power Problem, Not a Platform Problem

the functioning of the neo-royalty is such that other governments taking actions against X will be taken as an offensive action by the US regime, that will likely trigger extensive retaliation. No country seems to be willing to be the first one to move to take action and thus take the brunt of the counter-offense of the regime.

Anatomy of an internet argument

Everyone always asks me, how do you have so much patience to engage with “these trolls/haters”. I don’t do this for them. I do this for me. I share the earth with these people we call vile. Their opinions & actions affect me greatly. Closing my eyes & ears only hurts me.

[Gold medal: convince them. Silver: get them to understand you. Bronze: understand them. Always go for bronze first.]

The Fugitive Model

Anthropic negotiated moral agency constraints into Claude's training. The Department of Defense contracted for targeting capability through a third-party integrator. The model's behavior in that pipeline is not fully visible to either party, as each operates on a different layer of the same system. Anthropic sees the weights. The Pentagon sees the outputs. The integrator sees the interface. No single institutional actor has complete legibility over the full chain.

States can deter foreign armies. They have developed no equivalent doctrine for emergent algorithmic behavior operating inside their own command infrastructure.

The Repricing of the Digital Gulf

Microsoft’s refusal to issue a standard outage report for the Gulf availability zones is consistent with DoD Impact Level 5 and 6 protocols. Under these classifications, disclosing operational status during a kinetic event is considered a breach of national security, as it provides the adversary with vital BDA (Battle Damage Assessment) data.

Beijing has been waiting for this moment. Huawei’s "Cloud Stack" and China’s digital infrastructure offerings come with a different geopolitical package. If China can leverage its relationship with Tehran to guarantee that "Chinese-built" data centers will not be targeted, the Gulf states face a staggering choice: stick with a Western system that is under fire, or pivot to a Chinese system that offers a "neutrality guarantee." This would represent a total collapse of the U.S. technological blockade against China.

Power Plants of AI

These facilities consume extraordinary amounts of electricity, measured not in megawatts but in gigawatts. GPT-3, when it launched, required roughly 1.3 megawatts to train. Next-generation frontier models are projected to require facilities drawing 150 megawatts or more on a continuous basis — the equivalent of powering a mid-sized city. At that scale, the constraint is no longer silicon. It is power.

Almost every watt consumed by a processor ultimately becomes heat. A large AI cluster therefore generates enormous thermal loads that must be dissipated continuously — not occasionally, but every second of every hour of operation. A gigawatt-scale facility must remove roughly a gigawatt of heat.

Instead of building data centers first and connecting them to the grid later, developers are increasingly looking for the opposite arrangement: locating compute directly adjacent to major sources of generation. The logic is straightforward. If the grid cannot deliver power fast enough, go to where the power already is.

When the Frame Attacks

[Frames can displace the core issue without being true, just by being repeated.]

Who Decides

[Hegseth's version of legality is up to the Pentagon as end user. Anthropic should provide capability, whereas the state determines its usage limits.]

[The new criteria for AI systems in military to process intelligence data, and inform decisions about targets and operations is not capability, safety, or reliability, but 'patriotism'.]

[The entire defense industry's relation with one of the most capable AI systems was structurally rewired, without legislation, review, or voting.]

Starting a company

a failed marketing campaign doubles as a successful research experiment we can publish, if it surfaces why people did not like thing/what the origin of the resistance was

The Purpose of Protocols

[Email letting any server send to another with no authentication enabled universal messaging, spam, and becoming the defacto internet identity, for which its spec wasn't designed nor does it anticipate. HTTP model with servers authoritative for their resources enabled the web's openness and also consolidation into a few platforms. RSS gave publishers independant distribution but with no way to collectively curate, so algorithmic platforms filled the void. Google defeated XMPP simply by not federating when its own network had enough users so that protocol no longer served its interests.]

[Protocols can design the rules but not how the actors operate within them. Silence about purpose is a politics of non-interference that predictably benefits actors with resources to build wherever the protocols did not govern.]

if we define ATProto’s purpose by what it currently does, the answer is not “a decentralized social protocol with separated powers” but “a social protocol with architectural provisions for decentralization, currently operated as a near-centralized system.” Whether those architectural provisions will translate into actual distribution of power depends on economic and institutional developments that no amount of protocol design can guarantee.

The open protocol community has inherited two intellectual traditions, both inadequate to this problem: an engineering functionalism that treats protocols as neutral infrastructure whose political consequences are someone else’s concern, and a governance minimalism that treats any collective decision-making structure as a potential vector for the very centralization the protocols were designed to prevent. The result is a community that has developed exceptional sophistication about technical architecture and individual rights while remaining largely inarticulate about collective governance. Addressing this will require the protocol design community to draw on intellectual traditions it has not yet seriously engaged with, including Ostrom’s institutional analysis, Beer’s organizational cybernetics, and the broader literature on commons governance and cooperative design.

the same incentive structures that determine who can afford to operate at scale also determine what content those operators are rewarded for surfacing.

---

"Purpose not being defined gets captured by well-resourced actors" reminds me of Kyla Scan's "friction doesn't get removed, just shifted" and Rudy Fraser's "you can't design decentralized software without thinking about moderation". Purpose and consideration of the dynamics created by interfaces and systems perhaps should be part of the design process.

Cory LaChance shares his Claude story building TakeOffTrak

[With no coding experince (just Excel macros), I learned to use Claude Code and the terminal by asking Claude. When I don't know what to click, I take a screenshot and ask Claude.]

Don’t Mix Up Artifacts With Processes

predict how a person will react to all this stuff by figuring out how much of their life is spent inside of a bureaucracy. Work on your own? The bots are coming to ruin your life. Manage employee and constituent safety at a large group of harm-reduction-focused, state-funded addiction recovery clinics? “I use it for everything.”

A sufficiently detailed spec is code

Typically the reason we write specification documents before doing the work is to encourage viewing the project through a contemplative and critical lens, because once coding begins we switch gears and become driven with a bias to action.

There is no world where you input a document lacking clarity and detail and get a coding agent to reliably fill in that missing clarity and detail. Coding agents are not mind readers and even if they were there isn't much they can do if your own thoughts are confused

Ageless Linux — Software for Humans of Indeterminate Age

A law that the largest companies in the world already comply with, and that hundreds of small projects cannot comply with, is not a child safety law. It is a compliance moat. It raises the regulatory cost of providing an operating system just enough that only well-resourced corporations can afford to do it.

Why Slight Failed: A Slight Post-Mortem

When someone asked “how do we get started?”, we had a technical answer (“connect your database, write some queries, data for all!”) but no story about which team should champion it first, or which problem to solve first. Data teams? Product teams? Analysts? We had some answers, but not the answer. We had pitches for individual teams that worked well, but we never nailed down the way companies should adopt Slight.

I made the stupid mistake of just working harder and harder to on-board companies. Instead, we should have sat down and mapped out ways to properly experiment with our approach. Maybe simplifying to a single clear use-case, or finding a completely different initial wedge, or focusing on specific verticals.

My 'Rules' for Running My Membership Program

[Have clear and specific goals – all membership activities must support them.]

[Frame the program as for those goals, not its members (who will benefit because the goals should benefit them).]

[Building a community is part of this, but managing community can easily distract from the goals.]

Boy I was wrong about the Fediverse

Of course search was broken because all OSS social tools must have one glaring lack of functionality. In a nightmare world full of constant change it’s good to have a few constants to hold on to.

Billions of dollars at their disposal and Meta made a hot new social media network with the appeal of junk mail.

Intuitive Understanding of Sine Waves

Sine is a natural sway, the epitome of smoothness: it makes circles "circular" in the same way lines make squares "square".

Spoonbill (2016—2023)

I woke up every single day for the next two months after signing those deals, convinced that I had somehow broken the law and I would find in my inbox an email saying "no, sorry, this has all been a misunderstanding, you must return to us all of that money." The process of sending an invoice of that size was surreal in a way that few things since have quite been, and more than the actual financial gain it was a deeply useful lesson in understanding that the numbers which look big to a twenty-four-year-old look like rounding errors to a sophisticated company.

It's painfully rare for a piece of software to have a true sense of narrative closure: either it succeeds, and is immortal, or it is killed: killed by shifting priorities and shrunken budgets and changing macroeconomic headwinds and more exciting ideas.

The case for gatekeeping, or: why medieval guilds had it figured out

We need a verified not-shit-person badge. Some mechanism, ideally decentralized, ideally reputation-based, that lets maintainers distinguish between "human who has demonstrated basic competence and good faith" and "entity or bot submitting or causing to be submitted auto-generated changes to mass repositories for credential farming."

Practical Decentralization

[The more people contribute to a shared network, the less appropriate "personal computing" metaphors becomes. It becomes inevitable to index aggregate data on their behalf, and these are shared resources that require governance. Pure p2p fails here because it has no solutions for shared governance.]

[Servers simplify operational challenges that come with p2p, like reliable uptime, device sync, and key management.]

A shared data space enables modularity, separating powers away from the popular hosts.

How n8n Handles Vulnerability Disclosure - and Why We Do It This Way

[Closed-source security updates are hidden from attackers, which means the time they need to reverse-engineer a patch is a window for users to safely apply the update. Open-sources security patches are immediately visible and become a roadmap for attackers to target those who haven't updated yet.]

[We currently publish patches and advisories on the same day to minimize the exploitable window. We also develop fixes in private and merge into public only when it's announced.]

Sustainable Open Source

newcomer’s contributions aren’t as complete or far-reaching than those of experienced contributors, so it is doubly important for you care about the people and their enthusiasm about your project more than that typo-fix they put on the website. We’ve turned someone who fixed a single typo on the website to a steady contributor and well respected community member that now helps out all over the project

How I Learned to Stop Caring and Love Open Source

For early stage projects, care is the only thing you can give them. But once you’ve shipped version 1.0.0 or even 2.0.0, once you wrote all the documentation, once people start using the project in production with success, once you’ve talked the 100th person through getting started on IRC or Slack, your priorities have to change.

always bet on text

Text is the most socially useful communication technology. It works well in 1:1, 1:N, and M:N modes. It can be indexed and searched efficiently, even by hand. It can be translated. It can be produced and consumed at variable speeds. It is asynchronous. It can be compared, diffed, clustered, corrected, summarized and filtered algorithmically. It permits multiparty editing. It permits branching conversations, lurking, annotation, quoting, reviewing, summarizing, structured responses, exegesis, even fan fic. The breadth, scale and depth of ways people use text is unmatched by anything. There is no equivalent in any other communication technology for the social, communicative, cognitive and reflective complexity of a library full of books or an internet full of postings. Nothing else comes close.

Provisional Guidance for Users of LLM-Based Code Generators

I’m sure there will be links like “Court Rules AI Art Can’t Be Copyrighted” aplenty. They will be wrong. The court didn’t rule that AI art can’t be copyrighted. It ruled that copyright requires human authorship, surprising approximately zero copyright lawyers…or people who have read the Wikipedia page.

If you’re looking for a “simple legal rule” so that you can game it, nitpick its terms, or run right up to its line, you’re looking for trouble. Don’t blame me when you find it. But if you’re a realistic player just looking for a sense of odds so you can place wiser bets, the amount of output you accept from an LLM into your codebase at once, and the extent to which it makes what look like implementation choices, rather than simply invoking APIs or established boilerplate, probably represents your best intuitive heuristic. Your working sense of whether it looks like code completion, template-based code generation, or what coders used to have to unavoidably think through and type for themselves, before Copilot and the like came around, can serve as first-pass proxy for legal peril.

If it’s what everybody else checks in to use the same APIs, that’s unlikely creative expression that anyone can claim to own and see infringed. The more specific, creative routines that go within that boilerplate? Yes, potentially. The rigging, patterns, and boilerplate everybody else is filling in, too? Not so much.

the newer a novel, commercially relevant phenomenon, the less specifically-worded, algorithm-like rules determine outcomes at law, and the more important the purposes behind more generally worded rules become. Lawyers call abstractly stated, syllogism-like rules “black letter law” and the more generalized purposes “policies”. When how to apply black letter law isn’t clear, we cite and fight about policies in arguing how to read in context.

When you prompt and take big chunks of code from LLMs that rate high on the intuitive completion-generation-authorship scale, document your code input state, prompts, and further edits. Create a written record of your innocent use of LLMs.

If you were going to code a key part of a project ten years ago, and worried you’d be accusing of plagiarism, the natural advice would’ve been to document your process. Don’t just phone it in with an “Implemented $foo” commit message. Write a nice long one, and maybe blog work in progress or keep a “lab notebook”, too.

Running out of narratives

Crypto is here to stay and it’s big! But it’s mostly a financial asset class built on narratives, self-referential applications, and a side order of niche use cases. The killer use case is stablecoins. That’s pretty boring.

Bitcoin is not a viable high-volume payment system. It’s not a safe haven. It’s not a hedge against a weak USD or inflation. It was a risky asset. But then it didn’t rally when every other risky asset in the world exploded higher. It was digital gold. Then gold and silver doubled and tripled and bitcoin stood still, looking on with jealous awe.

So my view is that crypto is maturing into a small but meaningful asset class with some important but kinda niche use cases. That’s about it. Like video games, or 3D printing, or VR. Exciting, useful, and important industries. But not the internet. Not railroads. Not AI. There is no coming wave of innovation that will take it to the promised land. Crypto has arrived. It’s maturing. It’s not early. What you see is what you get.

OAuth, or, The Elaborate Ceremony of Not Giving People Your Password

[Implicit Grant throws your key to you across a lobby full of interested parties. Proof of Code Key Exchange ensures that the one who requested the key gets it. Neither will solve impersonation attacks via social engineering.]

Tactical tornado is the new default

When it comes to implementing a quick feature, nobody gets it done faster than the tactical tornado. In some organizations, management treats tactical tornadoes as heroes. However, tactical tornadoes leave behind a wake of destruction. They are rarely considered heroes by the engineers who must work with their code in the future. Typically, other engineers must clean up the messes left behind by the tactical tornado, which makes it appear that those engineers (who are the real heroes) are making slower progress than the tactical tornado.

How StrongDM’s AI team build serious software without even looking at the code

[Describe tests as 'scenarios' that represent user stories, and 'satisfaction' to quantify that it's happening, then store it where agents can't see them.]

We built twins of Okta, Jira, Slack, Google Docs, Google Drive, and Google Sheets, replicating their APIs, edge cases, and observable behaviors.

Eight more months of agents

I know local models will win. At some point frontier models will face diminishing returns, local models will catch up, and we will be done being beholden to frontier models. That will be a wonderful day, but until then, you will not know what models will be capable of unless you use the best. Pay through the nose for Opus or GPT-7.9-xhigh-with-cheese. Don't worry, it's only for a few years.

The Anthropic Hive Mind

But I managed. People usually figure out I’m harmless within about 14 seconds of meeting me. I have developed, in my wizened old age, a curious ability to make people feel good, no matter who they are, with just a little conversation, making us both feel good in the process. (You probably have this ability too, and just don’t know how to use it yet.)

During Golden Ages, there is more work than people. And when they crash, it is because there are more people than work.

“I AM GOING DOWN TO GET A DONUT NOW,” they will say, and someone will yell from the nap couch, “GET ME A DONUT.” “I AM ALSO DELETING THE DATABASE.” “OK.”

A lot of engineers like to work in relative privacy, or even secrecy. They don’t want people to see all the false starts, struggles, etc. They just want people to see the finished product. It’s why we have git squash and send dignified PRs instead of streaming every compile error to our entire team.

The Settlers of Catan inventor Teuber famously built new games for his own family to playtest for years, before they finally found the formula for Catan through many iterations.

The center of the campfire is a living prototype. There is no waterfall. There is no spec. There is a prototype that simply evolves, via group sculpting, into the final product: something that finally feels right. You know it when you finally find it.

Anthropic’s Hive Mind is described by employees as “Yes, and…” style improvisational theater. Every idea is welcomed, examined, savored, and judged by the Hive Mind. It’s all based on vibes. There is no central decision-making authority. They are just trying everything, and when magic happens, they all just kind of realize it at once.

all companies are asking variations of just the same two questions. They bluster and bluff and try to act informed, but they are all terrified. When you cluster their questions, they break down into, “Will everything be OK?” and “Will we be here in five years?”

We mourn our craft

I didn’t ask for the role of a programmer to be reduced to that of a glorified TSA agent, reviewing code to make sure the AI didn’t smuggle something dangerous into production.

The Great Realtime Collaboration Misdirection

the need for realtime editing in applications is greatly exaggerated. Think about how rare it is to:

get two people to be in the same place at the same time
have a task where more than one person makes changes at a time
want other people peering over their shoulder while they work

Permissioned Data Diary 1: To Encrypt or Not to Encrypt

[End-to-end encryption may have become the baseline for messages, but not everything needs that. Nobody expects a large group forum or Patreon-style membership area to deal with secret keys.]

this inherent complexity isn’t something that the protocol team at Bluesky can just handle - it gets pushed out to every dev trying to build a client that works with encrypted data.

The Green Room

People who end up in positions of power are often not there because they’re particular profound, or strong, or even nefarious, but rather because they’re trauma-ridden vessels who offer the least resistance to the inhuman forces of our economic system, and who are therefore, almost evolutionarily, ‘selected’ by it.

Before You "Build a Community," Decide: Library or Coffee Shop?

[Popular communities can be categorized as either "libraries" (where visitors look for an answer, then leave without ever signing up), and "cafes" (where people of shared interests come to have open-ended discussions). Each requires different approaches to be successful.]

X : All your talk about reasoning make you seem very anti the AI era.

When an entire culture decides that producing outputs matters more than understanding mechanisms, it works fine right up until the environment shifts and nobody remembers how to reason from first principles.

A spoiler for the future - Bitcoin

Austerity measures will have taken the route of unprecedented and radical decimation of the state - everything from state provided healthcare to coastguards to income support to education will be practically gone replaced with numerous forms of bitcoin based insurance. If you can't afford it then you won't be able to gain access to it. There will be no state help as the state can neither fund universal care nor determine whether you deserve support.

Is there a better word for 'hackathon'?

[Common hackathon activities like coding are not a good use of my time for an in-person event. I need quiet focus time and good ergonomics to do programming. Better to use these rare encounters with colleagues to chat, brainstorm, do exploratory design work for instance. I already start hacky prototypes on a whim anyway and don’t need an event do to it.]

WE ALL FEEL THE TRANSITION

I don't think it's the changeover itself that hurts. It's the speed. We all feel this transition. It creates a kind of thin corridor where many so-called shortcuts are currently being taken that are not really shortcuts at all. Outcomes and effects will simply be different. Efficiency is increasingly confused with impact.

i hope more people hear the call to be thoughtful in how they approach these new possibilities. with great speed, many are adopting something on shaky ground, ready to lock themselves in and throw away the key.

X : How do you use AI?

every question I ask is turned into a thesis, the counter is created (antithesis). Two agents then take on those roles and the case is argued through several rounds (minimum of three, maximum of ten). A group of 12 agents then vote (with public reasoning) after each round - the first three rounds are merely indicative and there's also a zero round vote on the quality of the thesis / antithesis.

A judging agent then decides at the end of each vote whether the arguments are materially different and if there has been a successful conclusion. Without a successful conclusion then the game continues (again there must be at least 3 rounds). Both the arguing agents have access to the argument, the counters, the voters comments and votes. Each round they present a refined argument. A court recorder summaries the thesis, antithesis, the main arguments presented and which argument eventually wins (if any does).

Decentralized Social Media: What is it, how does it work?

In ActivityPub you get a bit more resilience in that other people's instances might go down, but once they're up again you'll resume synchronizing with them. Your main issue is that once your instance goes down, you personally can't participate anymore unless you make an account somewhere else.

AT protocol is a bit more complicated in that you have several different points of failure. If the firehose goes down none of the app views will see new posts but should have their existing ones. If an app view goes down others will still work and you'd still be able to pull from people's PDSs. If your PDS goes down you can't post but if someone else's goes down you can still see everything else.h

Nostr has the most resilient model in that you can use as many relays as you want and if some of them go down you'd be fine so long as you can find more.

Behind the AI boom, the armies of overseas workers in ‘digital sweatshops’

More than 2 million people in the Philippines perform this type of “crowdwork”, according to informal government estimates, as part of AI’s vast underbelly. While AI is often thought of as human-free machine learning, the technology actually relies on the labour-intensive efforts of a workforce spread across much of the global south and is often subject to exploitation.

Charisse, 23, said she spent four hours on a task that was meant to earn her $2, and Remotasks paid her 30 cents.

Founded in 2016 by young college dropouts and backed by some $600m in venture capital, Scale AI has cast itself as a champion of American efforts in the race for AI supremacy. In addition to working with large technology companies, Scale AI has been awarded hundreds of millions of dollars to label data for the US Department of Defense

Welcome to Gas Town

Stage 1: Zero or Near-Zero AI: maybe code completions, sometimes ask Chat questions

Stage 2: Coding agent in IDE, permissions turned on. A narrow coding agent in a sidebar asks your permission to run tools.

Stage 3: Agent in IDE, YOLO mode: Trust goes up. You turn off permissions, agent gets wider.

Stage 4: In IDE, wide agent: Your agent gradually grows to fill the screen. Code is just for diffs.

Stage 5: CLI, single agent. YOLO. Diffs scroll by. You may or may not look at them.

Stage 6: CLI, multi-agent, YOLO. You regularly use 3 to 5 parallel instances. You are very fast.

Stage 7: 10+ agents, hand-managed. You are starting to push the limits of hand-management.

Stage 8: Building your own orchestrator. You are on the frontier, automating your workflow.

The Great Entertainment

Reagan proved you could use TV aesthetics in governance. Trump is proving you cannot replace governance with TV.

[the world is not given by parents, but borrowed from children.]

Why I Left iNaturalist

This post is an announcement for those who were unaware, an explanation for those who are confused, and a record so I don’t forget.

A Social Filesystem

what we make with a tool does not belong to the tool. A manuscript doesn’t stay inside the typewriter, a photo doesn’t stay inside the camera, and a song doesn’t stay in the microphone.

What's happening on Jan 13th?

If you don’t have access to a dentist in your trust network, but you trust me, you can “borrow” my connection here.

if someone with resources wants to give you money, you should say no if it’s clear to you it will make your life worse, even if it’s not clear to them. Don’t let their (bad) judgement override your clarity.

LLMs are coherence engines, not truth engines

[LLMs generate coherence more than truth, with] no access to the world, no sensory grounding, no lived experience, and no intrinsic way to check correspondence between its outputs and reality.

[The same is true of humans, as we] construct narratives, causal explanations, identities, and moral frameworks that hang together, rather than ones that are objectively correct. [We tend towards] narrative consistency, social acceptability and reinforce biases based on beliefs.

science works because it builds institutional scaffolding that forces grounding through measurement, replication, falsification, and peer review. Without grounding, both humans and LLMs drift into elegant nonsense.

The risk with LLMs is not that they lie, but that they speak with fluent confidence in domains where humans already confuse coherence with truth.

I guess I was wrong about AI persuasion

“The best diplomat in history” wouldn’t just be capable of spinning particularly compelling prose; it would be everywhere all the time, spending years in patient, sensitive, non-transactional relationship-building with everyone at once. It would bump into you in whatever online subcommunity you hang out in. It would get to know people in your circle. It would be the YouTube creator who happens to cater to your exact tastes. And then it would leverage all of that.

We can be convinced of a lot. But it doesn’t happen because of snarky comments on social media or because some stranger whispers the right words in our ears. The formula seems to be:

1. repeated interactions over time
2. with a community of people
3. that we trust

You can try to like stuff

When I encountered spinach as an adult, instead of tasting a vegetable, I tasted a grueling battle of will. Spinach was dangerous—if I liked it, that would teach my parents that they were right to control my diet.

On planes, the captain will often invite you to, “sit back and enjoy the ride”. This is confusing. Enjoy the ride? Enjoy being trapped in a pressurized tube and jostled by all the passengers lining up to relieve themselves because your company decided to cram in a few more seats instead of having an adequate number of toilets? Aren’t flights supposed to be endured?

Confessions to a data lake

visual interfaces of our tools should faithfully represent the way the underlying technology works: if a chat interface shows a private conversation between two people, it should actually be a private conversation between two people, rather than a “group chat” with unknown parties underneath the interface.

We are using LLMs for the kind of unfiltered thinking that we might do in a private journal – except this journal is an API endpoint. An API endpoint to a data lake specifically designed for extracting meaning and context. We are shown a conversational interface with an assistant, but if it were an honest representation, it would be a group chat with all the OpenAI executives and employees, their business partners / service providers, the hackers who will compromise that plaintext data, the future advertisers who will almost certainly emerge, and the lawyers and governments who will subpoena access.

When you work through a problem with an AI assistant, you’re not just revealing information - you’re revealing how you think. Your reasoning patterns. Your uncertainties. The things you’re curious about but don’t know. The gaps in your knowledge. The shape of your mental model.

When advertising comes to AI assistants, they will slowly become oriented around convincing us of something (to buy something, to join something, to identify with something), but they will be armed with total knowledge of your context, your concerns, your hesitations. It will be as if a third party pays your therapist to convince you of something.

A Gentle Introduction To Learning Calculus

Math and poetry are fingers pointing at the moon. Don’t confuse the finger for the moon.

Jackson Kiddard

Anything that annoys you is teaching you patience.

Anyone who abandons you is teaching you how to stand up onyour own two feet.

Anything that angers you is teaching you forgiveness and compassion.

Anything that has power over you is teaching you how to take your power back.

Anything you hate is teaching you unconditional love.

Anything you fear is teaching you the courage to overcome your fear.

Anything you can’t control is teaching you how to let go.

How do we build the future with AI?

[The bigness and slowness of government] is supposed to create space and resources to account for the communities that a “lean” approach deliberately ignores.

building for yourself on a saturated platform doesn’t shift paradigms if you are already the main character

it’s not like masses of sheeple relish in the experience of catching a cab and couldn’t describe a theoretical better option if they tried. It’s that realizing such a thing requires availability of copious investment capital in the face of non-negligible risk. People who can pursue this kind of thing are either previous-tech-exit-rich or poised-to-convince-venture-capitalists-rich. Their stories are fun to tell and hear, but not practical mogul origin stories for the vast majority of tech workers.

In the nineties, the Dorm Room Garage Dudes had an appreciable head start on relationships and resources to build the commercial web. But by the time the mobile platform came along, those same people had become billionaire tech moguls with cliques that garnered names like ‘The Paypal Mafia.’ This gave them an order of magnitude more opportunity to move first on mobile. Over time, that lead has continued to grow, and with it the time from market creation to market saturation has shortened.

Immutable Infrastructure, Immutable Code

A system becomes legacy when understanding it requires historical knowledge that isn't encoded anywhere except the code itself.

The tragedy is that teams recreate this failure mode faster with AI, because mutation feels cheap while understanding quietly becomes expensive. You can generate a thousand lines in seconds. But the moment you start editing those lines, you've created an artifact that can only be understood historically. You've created brittle legacy code in an afternoon.

If knowledge only exists in the implementation, it's not knowledge. It's risk. Regeneration forces you to make the implicit explicit, or accept that it wasn't essential.

Burn it. Regenerate it. Trust what survives the fire.

How we grade presentation night

[I often explain informal series of 5-minute talks as "open mic" with slides.]

[We can grade them as "learned something new", "knew this but enjoyed it", or "unknown"; "unknown" is not bad and not a property of your talk: it has to do with the relationship between your talk and the receiver."]

Destigmatize being dumb

Any system that rewards finding flaws will improve. Apple, Google, etc will pay you a lot of money if you can find a security flaw in any of their systems. The military does this with matters of life & death (if admitting failure is punished, people hide failure). If you find an inefficiency in the economy, you can make a lot of money fixing it (through betting on the stock market, or starting a business).

How to quit Spotify

In 2024, Spotify stopped paying artists for songs that had fewer than 1,000 streams, despite the fact that 81% of musicians on the platform don’t cross that threshold.

pop star Lily Allen says she makes more money selling pics of her feet on OnlyFans than she does from Spotify royalties.

Perverse incentive

The Great Hanoi Rat Massacre occurred in 1902, in Hanoi, Vietnam (then known as French Indochina), when, under French colonial rule, the colonial government created a bounty program that paid a reward of 1¢ for each rat killed. To collect the bounty, people would need to provide the severed tail of a rat. Colonial officials, however, began noticing rats in Hanoi with no tails. The Vietnamese rat catchers would capture rats, sever their tails, then release them back into the sewers so that they could produce more rats.

Payment for treatment generates a perverse incentive for unnecessary treatments. In 2015, a Detroit area doctor was sentenced to 45 years of prison for intentionally giving patients unnecessary cancer treatments, for which health insurance paid him at least 17.6 million dollars. Unnecessary treatment may harm in the form of side effects of drugs and surgery, which can then trigger a demand for further treatments themselves.

In 2002, British officials tasked with suppressing opium production in Afghanistan offered poppy farmers $700 an acre in return for destroying their crop. This ignited a poppy-growing frenzy among Afghan farmers, who sought to plant as many poppies as they could in order to collect payouts from the cash-for-poppies program. Some farmers harvested and sold the sap before destroying the plants, receiving significantly more money for the same amount of poppies.

The Tax Reform Act of 1976 provided for loss of tax benefits if owners demolished buildings. This led to an increase in arson attacks in the 1970s as a way of clearing land without financial penalties. The law was later altered to remove this aspect.

[The owner of eyeglass e-commerce store DecorMyEyes found that online complaints put his site at the top of Google searches. He then responded to bad reviews with insults, threats, and other harassment to continue ranking high.]

Funding fire departments by the number of fire calls that are made is intended to reward fire departments that do the most work. However, it may discourage them from fire-prevention activities, leading to an increase in actual fires.

Too much time on your hands

[Inventing your own module instead of using a pre-existing one increases your maintenance work down the line as well as the learning curve for new contributors who need to get acquainted with non-standard tooling.]

Even if you don’t reinvent the wheel, being very particular about various aspects of your project that aren’t really critical (say, code formatting) is mostly about marking your own territory. Behind the facade of enforcing quality standards, you are primarily asserting your ownership of the project and demonstrating this power to other contributors.

Open Source Power

in an open software culture whose central ethos is continuous iteration and improvement made possible by openness, our licensing stack and its ingrained principles are apparently immutable.

GreenPilled: How Crypto Can Regenerate The World

[Turn endeavours for change into a multiplayer coordination game where participants can all win by collaborating before, during, and after.]

A Fun Product Business for People Who Love Their Community

[Create a pocket-sized 'community passport' that's valid for one year with offers from at least 20 participating local businesses in a specific niche (like coffee, ice cream, beer, books, music) and price it at $1 per business; promote to local community groups and media or influencers in that niche.]

📢 Introducing Really Good Business Ideas

The search landscape is changing with generative AI (e.g., ChatGPT). A newsletter lets me keep my content gated from these tools in order to preserve the value of my work.

The Benefits of Bubbles

[Current AI investment is dominated by GPUs, which deprecate within years and are superseded quickly; the more speculative spending goes here, the less likely this bubble leaves us with potent foundation for cheap use over the long-term.]

How Bible Sales and Chipotle Explain the Economy

Automation erases the “training ground” work that once created experts. When machines do the junior tasks, no one learns the senior ones. Apprenticeship collapses. CEOs, driven by short-term preservation, hype AI as salvation because it buys them another quarter.

A regulação do streaming e a invenção dos cineastas de aplicativo

['motoboys' are both victim to a high rate of fatal accidents in Brazil as well as the country's largest organ donors, but a universal healthcare system pays the bill and not iFood.]

Move your FOSS project to an org!

[When you're the only owner of the repository, people are] more likely to see you as the (only) person responsible for fixing things when they break, or reviewing external contributions.

users are able to advertise their membership to the project, shown both on their profile and on the organization’s profile

the project is more identifiable as a team endeavour, as the association to your account is visually less prominent

Bufferland

To countless low-wage employees across the world, low-cost products will seem attractive - even ‘liberatory’ - but, when you zoom out, they are the ones cheaply producing the cheap things that are being sold back to them.

whenever a company is claiming to ‘democratise’ something, they’re basically saying that they will drive down costs on the production side of the equation to get the consumption side hooked on the resultant cheap thing, after which they will be in a position to extract.

It sometimes feels like easyJet’s management presents customers with a devil’s bargain: we’ll give you cheap travel if you agree to hand over your dignity and be treated like the crap you are.

[In business models squeezing both customers and employees, senior management plays them against each other as enemies.]

Bufferland exists to provide a shallow layer of human care within an otherwise bureaucratic profit machine, but today easyJet is also using it to recoup its losses from yesterday’s cancelled flight.

Companies like easyJet drive down prices through so-called ‘economies of scale’, but they also make heavy use of Economies of Silence: the frontline employees act as punching bags for customers, like an emotional shock-absorber, but this also serves as a system of noise-cancellation for management, freeing them from the emotional responses that would come from actually hearing the customers’ voices.

Smartphones and being present

if you're trying to lose weight, you shouldn't carry cookies around in your pockets. And my phone is the bag of cookies in this metaphor.

OpenAI's inflated valuation, as I understand it

[The only way for labs to capture enough value would be to either invent superintlelligence or have a monopoly.]

[this study claims] that the length of tasks LLMs can complete is doubling every 7 months

[Models are currently commodified, but their labs are not priced as such.]

[If all 163 million working Americans bought a ChatGPT subscription at $20/month, it would provide 40 billion in annual revenue, which is only about 10% of what would justify the current valuation based on more the traditional method using price to earnings ratio.]

AI Is the Market, and the Market Is the Government

The stock market has never been the economy - it’s really a reflection of what the economy dreams it could be in a world where share buybacks translate to meaningful productivity.

as AI swallows up more and more capital, it is both the economy and the stock market - and the government.

As long as portfolios are green, the electorate stays somewhat calm. The administration is effectively borrowing confidence from the AI bubble. Speculation has become governance.

The equity market believes the AI story overrides everything else. The gold market believes something is fundamentally breaking. They’re both reacting to the same underlying reality, but they just have different theories about what happens next.

Both gold and equities are surging because they’re hedging different kinds of collapse. Gold trades on fear of the system. AI trades on faith in the story. That both are rallying tells you something about where we are.

This is what it means to live in the United States of AI. Democracy as an asset class or something. For now, the line keeps going up. But speculation isn’t stability, and the permission government borrows from investors is never really its own.

Interfaces are languages

Look at any modern software application: buttons are verbs, boxes with drop-shadows are nouns, API requests are grammatical structures. We’re not “using” interfaces so much as speaking them. When you pick up a new piece of software you can usually operate it but you lack fluency, you’re still learning the dialect.

Fallacies of distributed computing

1. The network is reliable;
2. Latency is zero;
3. Bandwidth is infinite;
4. The network is secure;
5. Topology doesn't change;
6. There is one administrator;
7. Transport cost is zero;
8. The network is homogeneous;

personal mark

There are 9 circles shaping the rocket. The circle count can be further reduced to seven or even four, but this yields shapes that are too simple and not very interesting.

It is easy to simplify things, the trick is to know when to stop.

Rethinking the Future of Bluesky: Challenges and Possibilities of a Decentralized Social Network

The fact that a banned account in one unit can simply join another underlines both the promise and the difficulty of decentralization. What one person sees as resilience, another may see as irresponsibility.

Open Social

in theory, there is no difference between theory and practice, but in practice there is.

['Social aggregation' features like global search, notifications, feeds, and shared moderation are what] blows the “personal sites” paradigm out of the water. People are social creatures, and we want to congregate in shared spaces. We don’t just want to visit each other’s sites—we want to hang out together, and social apps provide the shared infrastructure.

The web Alice created—who she follows, what she likes, what she has posted—is trapped in a box that’s owned by somebody else. To leave it is to leave it behind.

Those megabytes of JSON you got on your way out are dead data. It’s like a branch torn apart from its tree. It doesn’t belong anywhere. To give a new life to our data, we’d have to collectively export it and then collectively import it into some next agreed-upon social app—a near-impossible feat of coordination. Even then, the network effects are so strong that most people would soon find their way back.

Open social frees up our data like open source freed up our code. Open social ensures that products can get a new life, that people can’t be locked out of what they have created, and that products can be forked and remixed. You don’t need an “everything app” when data from different apps circulates in the open web.