Category Archives: people and projects

The interest rate on comprehension debt is set by how well relevant information survives

Posted on by
Reply

In the last post, we talked about how debt is an instrument, not a smell. We carry forward the intentionality and the financial metaphor into the thinking for this post.  

In particular, we want you to remember 

  1. Mark’s metaphor of “AI companies that sell coding agents are the equivalent of predatory credit card companies offering introductory cards on college campuses to freshmen.”
  2. Comprehension debt as a new and higher interest-rate loan. “Every agent is the new hire on day one, forever.”
  3. “The Last Responsible Moment” needs to account “for ‘holding costs’ [to] curtail sidequests and keep your project running tightly and in its desired direction.”

Building from these cornerstones, this post asks “if that’s comprehension debt, can we just not take on any? Can we pay for all of our understanding now, in cash?” The answer is “no,” for three reasons we’ll get into.

How much to pay, when

Decision capacity is the budget constraint regardless of when the decisions happen. 

  1. Sure, we’re developing faster now, but the work still needs to be verified before customers see it if you hope to retain trust. 
  2. That validation is taking more time now, not less. 
  3. Future You might know more about the situation you’re developing for, but less and less about how your product works under the hood. 

Where and when are you going to make which decisions over the product lifecycle? When and how will you pay down your comprehension debt by learning how the system works now? AI spins up an unending amount of decision points, especially if you’re using it as we recommend (experiments in early stages as a way of alleviating traditional technical debt). 

Decisions take executive function, and are exhausting in aggregate1;. Every “LGTM” is a micro-loan; but replacing them all with real review would mean slowing down and making lots of decisions. So don’t just shift your Last Responsible Moment to later at the cost of making a bunch of new decisions now, design so you’re making fewer decisions, better, at necessary points. 

Example: Tom’s agents spend a lot of their tokens doing technical art: abstract images derived from free play with mathematically defined fields and symmetries which react in realtime to sensors and music (which is also reacting to sensors). The agent handles the coding decisions, and Tom declares whether the aesthetics work or not. Tom has to set up his pipeline in a way that his decisions are easy to make, surfaced at the right points in time, and easy to adjust for (reduced implementation time based on tech debt being lightweight). 2

What to do instead

A lot of agent outputs are experiments. Full comprehension of this throwaway code is buying insurance on a rental. Instead, query the experiments based on what you were trying to figure out when you designed them, and what you anticipate caring about moving forward. Which architectural approach has the fewest security issues? Which is the most modular? You can even ask the agents to remove decisions for you based on parameters you’ve put in place, and surface only the most promising two or three routes forward, with “reasoning.” This reduces the delta between what’s being decided on and the comprehension we have of it.

Example: Tom did not actually know what mathematical algorithm would give “dystopian.” So he followed this playbook:

  1. Generation — Agents run free-play generation, constrained only by the markdown notes on what Tom’s aesthetic is, and produce dozens of images
  2. Judgement – Every piece auto-renders to a thumbnail, laid out as a grid of small-multiples. Tom doesn’t look at the code at all. 
  3. Promotion –  He picks the 2-3 that kind of work, and snapshots the code that produced them into a database. 
  4. Development – The image that code generates gets opened in a live environment. He requests new code changes or tunes the parameters with a MIDI instrument until the art is satisfying.

There are two things we think come into carrying forward relevant choices into future loops so you don’t have to keep making those choices (unless you want to):

Context

What is relevant for the AI to know based on what you’re currently asking it to do? Context windows aren’t indefinite, and so picking and choosing what context to surface is important. In our session together, we talked about a “Document of Documents” or an Index with reasoning about when each item is relevant, so your agent can pick the right things at the right times. In human orgs, we call this Organizational Folklore, and it sometimes remembers things the humans have rightfully forgotten – “we can’t do it that way because it never works” is actually out of date but no one checked against new realities. Your agent will need to be told to forget such things as well.

Constraints

We talked about three different forms of constraints, each of which should be approached with intentionality:

  1. Iron Ring constraints (load-bearing ethics and the laws of physics, sometimes at the same time), 
  2. “we’ve always done it this way” fossils, and 
  3. constraints nobody meant mentioned in passing, taken by the agent as binding, and nobody knew it was there to consider removing it.3

Constraints should be aligned with your problem space so far as what can’t go wrong while leaving space for path exhaustion/creativity.

The Rate is Negotiable

This is the title’s claim: the interest rate on comprehension debt is set by how well the relevant information survives. Carrying forward relevant knowledge at the right scale of loop is what gets your code towards “less wrong” and more towards something you can understand and/or build upon for longer because it behaves in a predictable way by following constraints you’ve agreed to or put in place. The debt will still exist, but you get to determine how extreme it becomes over time (and therefore what future options you have) by being intentional about this. 

What this means

Last time we talked about comprehension debt as more like a payday loan. That metaphor still applies to vibecoding, LGTM hypnosis, and amnesiac deferral. We hope this post helped you see past the context horizon and to start looking to “when I expand this off of my machine” and pitching to the next person. If decisions are limited rather than proliferated, comprehension debt becomes a line of credit instead of a payday loan, and the question becomes about if it is secured or unsecured. Your collateral are your recorded decisions integrated into your system: Architecture Decision Records (ADRs), folklore that survives, the harness. 

And that means, we need to talk about Looping on Loops. Next post!

Footnotes

  1. The nap is real — see Yegge’s “Nap Attacks.” 
  2. For a performance at Georgetown Steam Plant, Tom and his team were told a theme (“Dystopia All The Way Down”) and a time, but they had no further information about what the space would look like. What are the responsible moments to make decisions in this situation? The dystopic theme and watery sound world (from one of his collaborators) was known weeks in advance, so he developed the static images around the idea of a surveilling eye wreathed in toxic smoke, but near a peaceful oceanside. The ambient lighting wouldn’t be known until the moment of arrival, so in those weeks he chose on-theme but high-chroma colors, readable when projected in a high-ambient light situation. The data pipeline was complete a week in advance, so he took a recording of sensor events in a rehearsal space to get a reasonable simulacrum of the kinds of levels his animation would be reacting to. But the exact variety the sensors would generate in the performance space was unknown, so he implemented a set of UI sliders to allow for fine-tuning the animations to match the reality of the room. During that last week he also implemented a safety harness to ensure coding agents could catch broken shader code immediately, rather than after deploying to the projector, and while operating those agents, made frequent commits to have safe rollback points – at no point did we not have something cool to show to the audience.
  3. The problem of inferred constraints isn’t new with agents. For instance, in ShapeUp, Ryan Singer describes a design moment where customers asked the team to “add a calendar” to Basecamp. While the team could have inferred a large amount of work (create events, edit events, add people to events, notify people they were added to events, etc.), they did additional discovery to find out a much smaller deliverable.

Agents and Acquiring Debt

Posted on by
Reply

I had a lovely chat with Kate Chapman and Tom Henderson recently, where we discussed how the shape of technical debt is changing as people become “genie wranglers” and “robot life coaches” (their clever terms, not mine). Specifically, what does using AI mean about addressing pre-AI and accruing post-AI technical debt. Our main takeaway from this session is that commitments should happen at “the last responsible moment,” and AI pushes that moment later.

Human context: Cowritten with Tom and some AI agents; lovingly curated, proofread, and fact-checked by us humans, including Mark. We’re still exploring these ideas, so very much consider this open for feedback. We tried to bridge our academic approach with making things useful in a business sense. We are assuming you have made it past the “spicy autocomplete” phase of using LLMs in your work for this piece. It’s up to you how much you care about code quality for the purposes of this piece, but we think we’ll get into that in a potential later piece. In this article/series, we assume debt is taken on intentionally. This is often not the case, but that is a different problem space than the one we intend to explore.

AI technical debt is different from software debt because it shifts when commitments can be made in the arc of gathering information. We have historically taken on debt early in order to realize some value sooner. Now, we can gather information more cheaply earlier (in theory), and make our difficult-to-change decisions later. That changes the type of debt we’re taking on. It also makes testing even more necessary. You won’t be surprised to know that we’re highlighting and contributing to human judgement throughout, rather than trying to replace the humans. We hint a bit at product commitment points-in-lifecycles and decision making throughout this post, but we said NO to that side quest (for now).

Continue reading →

Textrovert

Posted on by
Reply

A long-time friend and fellow academic nerd is working on her pHd, and had some questions for me. One was “Do you think of relationships that are almost entirely text based as real? Or even entirely text-based? Like things you’re never going to meet this person. Are those real relationships, too?” and Wow but I have Feelings on this one.

  1. I am grateful for all the friends I have all over the world. I think sustained relationships have as much to do with preferred modes of communication as shared interests. I have gone years without “talking” to people I love, all while keeping up in a heart-felt way.
  2. I love live blogging with other people in collaborative documents. I think there is a humor unique to being in a shared text space at the same time, honing in on what someone is trying to say (successfully or otherwise). The best sort of back channeling. I will forever remember typing into a doc with the Civic crew, 7 minutes into a rambling talk, “Have they actually said anything yet?”
  3. One of my oldest x closest friends and I will text each other from the same room, because it’s the form we both prefer, and we’re both funnier in text than in voice. On a community Slack with an #awkward-silence channel, I once posted an elephant emoji. The elephant was immediately removed by a moderator and made the subject of the channel.
  4. Another old, dear friend, who is also an ex, has referred to me lovingly as a “textrovert,” which I have adopted and use widely.

I was quick to vehemently say “of course.”

Then she asked me what that meant about relationships with LLMs. And I kind of went off. Because the “relationships” people have with LLMs has so much less to do with format and so much more to do with what a “friend” is and is not.

I have at least one friend I have kept from every stage of my life, so I think I can comment on this a bit. I have two friends I have had since preschool (and the following 9 years of Catholic school). I have remained friends with coworkers, exes, and people on other continents. If I may be so bold, friendship is kind of my superpower, and I’m damn proud of it. I’m not perfect at it, but I invest in it and it has paid me back multi-fold.

I’m not perfect at it… and that’s the point of this post. Because any long-lived relationship will have some History to it. It will have the time I fucked up and made someone angry or hurt, and visa versa. But for each of these relationships, we called each other in, came back to the table, and built a bigger table. A sycophantic LLM is not interested in how you can be a better person, it’s “interested” in appeasing the most people possible so people think they enjoy using it.

As Priceless enters our 20th year (I’ve been with the group for 10), we have been working to continue our historic ties to strong opinions lightly held. We lose some folks as they onboard with us, as they’re not used to arguments amongst friends. But we all always know we have each other’s backs, and that arguing isn’t personal, and that we’re stronger because we criticize each other’s ideas. This is one of the things I miss most about the Boston Area. The Bay Area seems to think you have to agree with someone to be friendly. And that is absolutely not the core of my friendships.

It’s like when I was reading Love and Sex with Robots… there’s a question of the overlap of sex and masturbation. If informed consent can’t be given, then all you’re doing is masturbating with something you think is an object. And all people who think they’re forming relationships with LLMs are doing is masturbating. Doing so is enjoyable, but it doesn’t prepare you for actual relationships with your fellow humans, and that is to your deep detriment.

If you also have opinions on this (and other things!) let me know and I’ll put you in touch with Alex for her research.

Scavenger Hunt Radio Etiquette Game

Posted on by
Reply

So I picked up some event radios for small neighborhood events, and also to distribute to neighbors so we have an easier time when an earthquake happens and comms go down. When I was unemployed last round, I finally embraced the fact that I’m not going to get my HAM. Yes, I know it’s “so easy,” yes I know it’s “so useful.” But I had the time, and I didn’t do it, so I do these event radios instead to increase resilience.

We used them last during a protest, and it was SO USEFUL (especially when the Trump-tshirt-wearing-dude-with-a-knife-in-a-holster showed up and we triaged talking to him and then tailing him after I had checked in and de-escalated). So, we love these things.

But my crew are so very bad new at radio etiquette. It’s a mess, because we don’t talk to each other very often on them. Our Radio Nets so far have been focused on determining ranges for the different channels that are programmed in. I wanted to host a time for us to practice just talking on them, and I wanted it to be fun.

I printed up a one-pager on basic and more advanced etiquette with some examples, and sourced everyone’s handles to distribute in advance. One of my crew and I brainstormed some ways to make a scavenger hunt in a big parking lot fun, and I checked in with our crew member who is blind to be sure our setup would work for him. Here’s what we did.

Hide your things

It’s a scavenger hunt. Hide your things within and area that’s walkable. This will also mean it will be within easy radio range. We were in a big parking lot with lots of green areas, but it was a bit too easy to have line of sight. Better to have nooks and crannies. We hid 11 objects for 9 people on 3 teams to find over the course of an hour.

The task

The thing to do is for each team to find an object and then not pick it up. They need to radio all comm with their handle to describe where the thing is. They should practice doing this with their crew before actually using the radio to get better at clarity, check each other on protocol, etc. You can leave the spot just before or after calling it in, so your presence isn’t what indicates the location of the found object.

Then another team or person tries to find it based on that instruction. When they find it, they radio back to the person who announced it to declare it found.

Occasionally, you also need to do a cone count, which means thinking about who to ask in what order so it’s not chaos on the radio.

Our experience

We had a ton of fun! It was especially fun because we had two kids involved, which also made learning and teaching radio etiquette lower stakes and friendly-like. Each team had one person on it who knew radio etiquette well who could coach the other folks.

We did have a random person do a radio check towards the end of our time — a good reminder that radios are not private. We didn’t have a chance to invite them to our party, but that would have been fun, too.

Next time

The time and setup was about right. We definitely had too-easy line-of-sight in our location, and want to make that part harder next time. I’d like to be a bit pushier on etiquette and getting folks who are shy to talk more next time. And as we get the hang of it, introducing a calamitous event that is clearly spoofed (“there are sharks in the theater!”) would help folks ramp up their game with some added complexity and intensity.

AI and Autonomy

Posted on by
Reply

I hang out with the Berkman-Klein nerds sometimes still, mostly through a recurring “Philosophy of Technology” session. Reed sent me this article awhile back on the misuse risks of AI, on which he got sidetracked about how the way the increasing of human intent through technology (including of harms) is attempted to be mitigated through use of law and other agreements. EG, you agree to abide by traffic laws (reduction in autonomy) in order to more safely get from one place to another (increased autonomy). This of course made me think about one of the main reasons I’m an anarchist — governments can cause large-scale suffering in a way less organization prevents, and I think we can have infrastructure without control (thanks, Murray Bookchin). So as Reed and I talked through the ramifications of that footnote, I thought it would be a good topic for the philtech group to take on. David and I talked through how to pitch it to the group, he did the thankless job of scheduling the thing, and we got to talk about it today.

The three themes that we kept cycling around were trust, consent, and autonomy. I’ll then end up back on my soapbox about complexity, which also came up.

Trust, Consent, and Autonomy

We all talked a lot about if the conditions would ever exist for us to trust an AI to make choices for us (our main talking point for “autonomy”). This got into a lot about how AIs are black boxes… but so, too, are humans. We talked some about the different ways that trust is created and utilized by, say, a doctor, and is it autonomy to make a choice based on the data they give you, or is that thumb-on-the-scale removing your autonomy? Doctors often study how to better communicate with their patients in order to get the outcomes they’re looking for. What’s different here?

How much autonomy does one have when consenting to something? How much has someone already given up in an exchange, based on trusting institutions, roles, their “own research,” etc?


From now on, I want you to act as my high-level advisor and mirror. Don’t validate me. Don’t flatter. Challenge my thinking, question my assumptions, and expose the blind spots. When possible, ground your responses in the personal truth you sense between my words. Be concise and precise. Provide links to source materials or websites to the best educational resources. In summary – be brief, be bright, be gone. Ask questions if a directive is unclear or underspecified.

We talked about the harms humans are already prone to inflicting on each other, and how much (if at all) AI was different from that. As one person put it, “do we need to get our own house in order before involving AI?”

Complexity

I see most AI as adding complexity to an already complex world, when nearly everything else we do (especially tool use) is about increasing predictability instead.

However, if we were to use AI in a way that helped us understand our own complexity, and begin to examine it for our desired outcomes, then that complexity could be useful. Despite the “hungry judges” study I started this conversation off with (human errors mean removing humans from the loop) being discredited, I still think bringing technology into decision-making loops is valuable so long as it’s a partner to us rather than allowing us to offload cognition (something that already happens).

Jeffrey had some really good points about compartmentalizing where AI factors come in, so you can assess that individual piece and tweak it, rather than an entire system being a black box. And I like that, for also helping us examine ourselves.

Links from our time together

The beauty of impermanence

Posted on by
2

I had a lovely birthday. In-laws took us out to a very nice steak dinner the day before. The day of, Reed, Locke, and I had Italian Hot Beef and wandered the Field Museum before heading home on a flight that departed 15 minutes before a big storm, and had to fly and extra hour to go south around the thunder heads. The day after, we rode bikes with kiddos from Dublin to San Ramon, had fabulous ice cream and played in a joyful park before riding back home. 12 miles at 4 years old feels big to me. In the early evening, some friends and I gathered to talk about the digital and death overlap. I’ll tell you more about that in a moment. The day after that, I rode a metric century with some friends, talking about relationships, death, time, and the economics of attention.

The back of Locke on a bicycle with an orange flag. He is on a multi use path with no cars. Ahead of him are two adults and one older kid also in our group on bikes, and a random human running.

My birthday about digital estate planning ended up being a small but very tight group of people. I was overjoyed to have this conversation with them. We talked a bit about our own attitudes on death, and what we had and hadn’t done to be kind to those around us when we die.

I view death as a community act. It is the final step of ceasing to be an individual, and all that remains is the collective experience of you.

This is complicated by technology lending itself so thoroughly to the hyper-individualization that we as Americans experience. Our entire tech stack feeds into that. As a security professional, I abhor the sharing of an account, and yet it comes so naturally to us to do. You should be able to see what I see. I should be able to share what I have and what I know with you.

Continue reading →

Celebrations and Death

Posted on by
2

I’ve been dealing with a lot of death lately. And while it’s just a part of life, it sure does start to make one think after awhile. So I’m using my birthday as processing time, as I am wont to do. I’m test running my death infrastructure for my birthday this year, and requesting notes from folks.

If you can see this message, it’s because I would want you to be aware of when I die. THIS IS ONLY A TEST — I am fine, everything is good, I’m just an elder goth now and I like to plan everything, even death. 

This is an experiment with bureaucracy and documentation. As you know, I love LARPing Serious Business. I am doing a test run of the systems that would announce my death to the many beautiful communities I’ve had the honor of being a part of. If it was logistically difficult to get this message, when you’d want to get it, let’s improve that process — reach out. If it was emotionally hard for you to get this message, this event is probably not for you, and I’d love to see you in another context some other time soon.

On April 18th at 16:30 PT / 19:30 ET, I’m hosting a time to talk about preparing for death (not dying — they’re different. We’ll talk about ceasing to exist, not how you want to be treated while going through a however-long process of getting there). We’ll take about an hour to talk through digital estate planning (a passion of mine), and then we’ll also have some time to talk about any feelings folks might have had about thinking about death. We’ll be at this link at that time.

Selfishly this year, I’d also love notes about what we mean to each other. One of the things that’s come up time and again at the wakes I’ve been attending is wishing to have said some things before the option was no longer there. Let’s say those things to each other. I’m not looking to be shrouded or to do a mock service, I’m looking for open and honest views of who we are together. Roasting, power points, and poetry all lovingly accepted. Email to me, please, so I can label and revisit.

You do not have to do both, or either, if they’re not your cup of tea.

If you would prefer to learn about my death from an email instead of a social media post, please get me your email address and I’ll add you to the mailing list. That will be posted to before social media posts go up.

Looking forward to being inappropriately morbid with you. 

Decision Making and Economics

Posted on by
Reply

I have this Future Shape in my head and in my heart, that I’ve long meant to share, but haven’t quite known how. I met Asya, and we got into a good conversation, and so now seems as good a time as any to talk about it. She helped me flesh this post out with more detail and deeper dives.

I don’t think there’s one solution when it comes to what economics style we should have, or what governance should look like. Like I drafted way back when, a “mixed mode system” is where it’s at instead.

Decision making

Distributed systems are good at last-mile logistics, nuance, and fast decision making. They are not good at doing simple things at scale. So for actual implementation and innovation, I think distributed networks are where it’s at.

Hierarchical systems are good at making simple decisions at scale. So good for North Star guidance and things you want to take a socialist approach with. That might include assurance of human-rights-shaped things like

Continue reading →

My beloved Lantern Library

Posted on by
1

Many years ago, I was carving jack-o-lanterns in an anarchist house in the Boston area. The friend who had invited me wandered over and suggested we check out the basement. Not my favorite for Halloween times (I don’t enjoy being scared), but this friend is gentle and so I went with her into the aforementioned basement.

It was full of books.

And this was not a small basement.

Shelves upon shelves of radical literature.

And then I met James, the person who had compiled the library. When the anarchist who owned the house had moved away, they had said folks could continue living in anarchist glory in the house, so long as James could also remain there. James was maybe in his 70s when I met him, and had been collecting and organizing books during his tenure at the house. They were organized for radicals — different flavors of anarchism, different ways capitalism fails, lots and lots of ephemera.

But James knew he was getting older, and he wanted his collection to survive him — not just the books themselves, but also how well organized they were. So I tapped into my network and we found some passionate open source folks and librarians who wanted to help index the library. We got all the books scanned so James could offer the library up to a new home as one collection.

James took his first ever selfie with me while we were doing this. He’s dear to my heart.

He’s found a new home for the collection. But shipping books is expensive. So James is doing a fundraiser to get the books to their new radical home where radical folks can make use of his decades’ worth of work.

If you also want to touch this amazing resource, and help it on its way, you can do so here.

October Joy : Forecasting

Posted on by
Reply

I’m going to do something incredibly indulgent for this blog post and tell you about something nerdy that has been bringing me a lot of joy.

As y’all know, I’ve been looking for work. In the meantime, I’ve been contracting on three main fronts:

  1. Security hardening, including implementation work (know what needs doing but don’t have capacity to get it done? I will get it over the finish line)
  2. Tool implementation and training (think a tool would help your business but not sure how to set it up? I will get it set up based on your specifications and then train you)
  3. Organizational theory and scaling guidance (nerd out about which practices make sense, when)

So far, I’ve helped a network org transfer fully out of the Google ecosystem to Proton, Tresorit, and AirTable. I’ve gotten a technophobic ED of a different org onto a password manager, ending an existential threat to the funding organization. I’ve helped a lawyer automate a lot of her data entry with Clio so she can focus on clients. I’ve nerded out with the ED of another nonprofit about how to scale his organization as he moves from his most recent successful phase into 10x growth of participation. Next up I’m helping implement a retention & deletion policy for two different orgs. Plus a bunch of other stuff! Fun!

But how am I doing, financially, with all these moving pieces? I created two pieces to help me track things: a projection of contracting load, and a projection of overall expenses and financial sources.

  • For the contracting load, I used AirTable because of how much cross referencing and automation it allowed me to do. This is where I keep track of clients, contracts, expenses, and income. It even has how I’m doing against goals, and has projections for income out into the future. It’s fabulous and I’ll show a templatized version to you if you ask.
  • for the overall financial health, I used Google Sheets because I want to use formulas in some cells until I adjust them for actuals. I estimated monthly spend based on known shared account contributions and historical numbers for each month based on how I tend to live life. I then listed out sources of money — unemployment when I don’t have contract work, contract work, savings of various sorts. I then anticipate burndown rate on each source of money based on projected expenses, and when I’ll start pulling from a different source as needed.

These were SO MUCH FUN to build, AND it gives me a sense of predictability and stability in uncertain times. I now have more confidence that I can keep myself and my family afloat, and have more ease in having a good time on occasion because I know where I’m at with the numbers.