devlog
write devlogs after a session
Submitted directly at the symposiumto help you win at the game of working smarter and safer with AI
32 submitted directly, 196 inferred from the talks: 155 practices and 73 hazards. Browse them all freely, or play: swipe right on the ones you like, rate them, match practices to hazards, and build a personal proto-protocol to try.
Tap a theme to open its cards. Bigger bubbles hold more.
Submitted directly: what people typed into the One Thing form during the talks, as written. Inferred from the talks: practices and hazards drawn from what a speaker said, credited to them. Suggested match-ups between hazards and practices were initially seeded by AI, then judged by real people like you.
Each hazard with the three practices AI thinks most likely to answer it. Pick the best fit. Your picks are the research.
Letting ai do your thinking for you
Submitted directly at the symposiumpracticesPick one
mind wanders between dispatches of work to the agent... loss of context, loss of awareness of the work, less attention given to the process of the work...
Submitted directly at the symposiumpracticesPick one
Temptation to accept its next steps versus consider alternatives
Submitted directly at the symposiumpracticesPick one
practicesPick one
Focus on keeping agents busy instead of strategic planning. Builds a doom spiral where stress builds where outputs degrade and the solution becomes more agents faster.
Submitted directly at the symposiumpracticesPick one
Losing the thrill of the intellectual chase. There's an insidious induced passivity, harder to resist because it seems like you have control.
Submitted directly at the symposiumpracticesPick one
use of antipattern language such as "real" "genuinely" "actually" and other filler
Submitted directly at the symposiumpracticesPick one
The emotional roller coaster of having a good conversation with LLMs only to have the real world crush your work.
Submitted directly at the symposiumpracticesPick one
I create multiple sessions and run them at once because I get impatient with AI being "too slow" at doing the thing I have asked it to do for me. Then I get overwhelmed by the number of sessions and can't find the one that I actually started. I end the work with multiple incomplete, active sessions.
Submitted directly at the symposiumpracticesPick one
When everyone drafts with the same LLM, people converge on the same language and framings before an idea has had time to mutate through other people. The variety that normally comes from ideas passing hand to hand gets flattened at the source.
Inferred from Sachin Benny, "Archival Time", PI Symposium Day 4, 24 Sep 2026practicesPick one
Relying on AI output in a topic you don't know well enough to check. Without background knowledge you can't spot its slop, and it will bluff past you. It has invented patent references and paper links that turned out not to exist.
Inferred from the Day 3 session chat, PI Symposium, 2026-09-23practicesPick one
Asking the same AI to revise a draft over and over. The prose can get worse with each pass, becoming more verbose and more "AI-ish", until a human has to edit it heavily or translate it back into plain English.
Inferred from the Day 3 session chat, PI Symposium, 2026-09-23practicesPick one
Expecting AI to build on your hypotheses the way a colleague would. In back-and-forth it tends not to extend your ideas enough, so serious hypothesising needs structure you add yourself.
Inferred from the Day 3 session chat, PI Symposium, 2026-09-23practicesPick one
Letting fluent, first-person, anthropomorphic AI language lower your guard. Words can short-circuit your defences, especially the singular first-person voice.
Inferred from the Day 3 session chat, PI Symposium, 2026-09-23practicesPick one
Using grand labels ("agent swarms") for what are prompt loops given lots of compute. It distorts your own judgement of what the system is doing. Describe the mechanism plainly.
Inferred from an open-mic speaker, "Open Mic: Frontier-Pacing Protocols", PI Symposium Day 4, 2026-09-24practicesPick one
When a tool cuts the time a task takes, the skill that task quietly built goes with it. Reinvest part of the saving in deliberate, low-stakes practice of that skill.
Inferred and adapted to AI from Daniel Kronovet, "Chores as Complex Coordination", PI Symposium Day 4, 2026-09-24practicesPick one
Working with AI tools that demand constant focal attention: you can compress months of work into days, but you end up exhausted because every step goes through the screen in front of you and nothing can be left to the periphery.
Inferred from Amber Case, "Attention as Natural Law", PI Symposium Day 4, 2026-09-24practicesPick one
Building your creative practice around one model's particular character, such as an older, more expressive model, with no fallback. When the vendor retires that model, the capability you had come to rely on goes with it, and your practice loses a limb it never recorded or replaced.
Inferred from Amber Case, "Attention as Natural Law", PI Symposium Day 4, 2026-09-24practicesPick one
Designing lots of agent behaviours because each sounds like a good idea, without checking whether they ever fire. Many never trigger or do nothing useful. Wire in live usage analytics and prune to the behaviours that actually run.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24practicesPick one
Swipe right on the cards you like, left to skip.
Rate them as you go if you want. Every few cards: match practices to hazards. Points for every answer, a bonus when the crowd agrees with your pick.
Keyboard: ← skip · → like or watch · 123 answer the next question, or pick a match · Z undo
A “proto-protocol” is a set of proposed best practices for you to experiment with. It could become a protocol once it “hardens”, meaning it’s written down, shared between you and the AI, and made to hold, including by your own behavioural rituals or through enforced safeguards written into the AI system. For more on this, see The Unreasonable Sufficiency of Protocols.
To preserve your score, come back next time in the same browser.
write devlogs after a session
Submitted directly at the symposiumPublish ongoing devlogs for every project
Submitted directly at the symposiumMeta-explain: simply explain clearly what I want to achieve instead of giving it some kind of my work and have AI improve it.
Submitted directly at the symposiumChanging my workflow so ai compensates for my shortcoming
Submitted directly at the symposiumKeeping your own goals very concrete and in mind, resisting drift from the AI's own high-level suggestions. A defense against sovereignty erosion, which is v easy for me to fall into.
Submitted directly at the symposiummuch more fluid and consistent fluid metacognition
Submitted directly at the symposiumwrite out spec documents ahead of time, and iterate on them
Submitted directly at the symposiumPrompting the AI to ask me questions.
Submitted directly at the symposiumIt's quite expansive and sometimes tries to go behind your prompt. So assuming you already have your own specific ideas on how you want the task to be executed, it would be good to remind the AI to stay within the ranges of what you specify
Submitted directly at the symposiumAsk clarifying questions to remove ambiguity in request/context. Produce a spec doc prior to beginning any work.
Submitted directly at the symposium A player’s revision of “Think before you act”Insert myself in the prompt as incremental reviewer. Tell the AI to deliver results at specific milestones for feedback. Breaks up long, expensive loops and keeps more control of direction.
Submitted directly at the symposiumAssume you are in a negotiation.
Submitted directly at the symposiumTrying to routinely reinject doubt in the conversation
Submitted directly at the symposiumDeveloping onboarding and offboarding processes each collaborative AI sessions
Submitted directly at the symposiumLetting ai do your thinking for you
Submitted directly at the symposiummind wanders between dispatches of work to the agent... loss of context, loss of awareness of the work, less attention given to the process of the work...
Submitted directly at the symposiumTemptation to accept its next steps versus consider alternatives
Submitted directly at the symposiumFocus on keeping agents busy instead of strategic planning. Builds a doom spiral where stress builds where outputs degrade and the solution becomes more agents faster.
Submitted directly at the symposiumLosing the thrill of the intellectual chase. There's an insidious induced passivity, harder to resist because it seems like you have control.
Submitted directly at the symposiumuse of antipattern language such as "real" "genuinely" "actually" and other filler
Submitted directly at the symposiumThe emotional roller coaster of having a good conversation with LLMs only to have the real world crush your work.
Submitted directly at the symposiumI create multiple sessions and run them at once because I get impatient with AI being "too slow" at doing the thing I have asked it to do for me. Then I get overwhelmed by the number of sessions and can't find the one that I actually started. I end the work with multiple incomplete, active sessions.
Submitted directly at the symposiumWrite agent instructions as conditions, not steps: treat the model as agential material. Specify the action space and guardrails (a skill, a high-level instruction) and let behaviour emerge, rather than scripting every step.
Inferred from Botao Amber Hu, "Protocoling HCI", PI Symposium Day 4, 24 Sep 2026When everyone drafts with the same LLM, people converge on the same language and framings before an idea has had time to mutate through other people. The variety that normally comes from ideas passing hand to hand gets flattened at the source.
Inferred from Sachin Benny, "Archival Time", PI Symposium Day 4, 24 Sep 2026When setting an agent up, define up front which moves are available to it (the action space) rather than letting it roam and correcting afterwards. Shaping the space beats policing the output.
Inferred from Botao Amber Hu, "Protocoling HCI", PI Symposium Day 4, 24 Sep 2026Before asking AI for substantive work, prepare your own material and framing first. It can't read your mind, and thin input yields wordy, generic answers. Also know when a short prompt is enough, since winging it sometimes works and that intermittent reward is what tempts you to skip the prep.
Inferred from the Day 3 session chat, PI Symposium, 2026-09-23Relying on AI output in a topic you don't know well enough to check. Without background knowledge you can't spot its slop, and it will bluff past you. It has invented patent references and paper links that turned out not to exist.
Inferred from the Day 3 session chat, PI Symposium, 2026-09-23Asking the same AI to revise a draft over and over. The prose can get worse with each pass, becoming more verbose and more "AI-ish", until a human has to edit it heavily or translate it back into plain English.
Inferred from the Day 3 session chat, PI Symposium, 2026-09-23Expecting AI to build on your hypotheses the way a colleague would. In back-and-forth it tends not to extend your ideas enough, so serious hypothesising needs structure you add yourself.
Inferred from the Day 3 session chat, PI Symposium, 2026-09-23Letting fluent, first-person, anthropomorphic AI language lower your guard. Words can short-circuit your defences, especially the singular first-person voice.
Inferred from the Day 3 session chat, PI Symposium, 2026-09-23Put into instructions or automation everything you can specify in advance, and deliberately keep human judgement for what's left. Don't treat a complete-looking spec as complete: all rules are incomplete.
Inferred and adapted to AI from Eric Alston, "Governance Axioms", PI Symposium Day 4, 2026-09-24Don't force early convergence among competing approaches. Run several in parallel and add a bridging step with feedback before choosing one; forcing alignment early loses perspectives.
Inferred and adapted to AI from Jose Leal, "Open Social Protocols", PI Symposium Day 4, 2026-09-24Whoever (or whatever) can halt a running agent or pipeline shouldn't, by that same power, be able to amend, spend or restart it. Keep the kill switch narrow.
Inferred and adapted to AI from Eric Alston, "Governance Axioms", PI Symposium Day 4, 2026-09-24Using grand labels ("agent swarms") for what are prompt loops given lots of compute. It distorts your own judgement of what the system is doing. Describe the mechanism plainly.
Inferred from an open-mic speaker, "Open Mic: Frontier-Pacing Protocols", PI Symposium Day 4, 2026-09-24Pair LLM output with a formal checker (e.g. the Lean theorem prover) so claims are machine-verified, but keep every result written so domain experts can read, interpret and independently verify it. Don't accept results only the machine can understand.
Inferred from Shingai Thornton, "Hand-Washing and Bitcoin Share a Skeleton: Protocols Are Systems, Proven", PI Symposium Day 4, 2026-09-24When research sprawls, split the write-up: a short human-facing framing, plus a dense, well-referenced section written for an LLM to ingest and build tailored messaging on top of.
Inferred from Jesse Katz, "Entrenching Privacy Through Speech", PI Symposium Day 4, 2026-09-24Keep every judgement call as a human input; let the machine do only legible, deterministic bookkeeping. This guards against overconfident models and preserves legitimacy.
Inferred and adapted to AI from Daniel Kronovet, "Chores as Complex Coordination", PI Symposium Day 4, 2026-09-24When a tool cuts the time a task takes, the skill that task quietly built goes with it. Reinvest part of the saving in deliberate, low-stakes practice of that skill.
Inferred and adapted to AI from Daniel Kronovet, "Chores as Complex Coordination", PI Symposium Day 4, 2026-09-24When AI lets you follow an intuition past your own competence, treat what it gives you as a way to ask better-qualified questions of real experts, not as grounds to speak with authority yourself.
Inferred from Jesse Katz, "Entrenching Privacy Through Speech", PI Symposium Day 4, 2026-09-24Working with AI tools that demand constant focal attention: you can compress months of work into days, but you end up exhausted because every step goes through the screen in front of you and nothing can be left to the periphery.
Inferred from Amber Case, "Attention as Natural Law", PI Symposium Day 4, 2026-09-24Show an AI agent's state (working, done, uncertain, struggling) as an ambient signal you can glance at, such as a coloured light or a visual of running jobs, instead of watching the chat window. Then you can stay on your own task and step in only when needed.
Inferred from Amber Case, "Attention as Natural Law", PI Symposium Day 4, 2026-09-24Building your creative practice around one model's particular character, such as an older, more expressive model, with no fallback. When the vendor retires that model, the capability you had come to rely on goes with it, and your practice loses a limb it never recorded or replaced.
Inferred from Amber Case, "Attention as Natural Law", PI Symposium Day 4, 2026-09-24Learn AI co-working by watching a practitioner work in real time, missteps included, the way an apprentice picks up a craft in the studio. In turn, show your own messy process openly instead of keeping it private. Watching someone fumble through an AI coding tool across four sessions taught the method more directly than instructions did.
Inferred from a Q&A participant in the discussion following Amber Case, "Attention as Natural Law", PI Symposium Day 4, 2026-09-24Put your own body of writing behind a retrieval bot and query it with a time window. That lets you trace how your ideas developed across eras and recover things you have forgotten. It works as prosthetic memory for your own thinking, which a generic model answering from the open web cannot do.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24Before deploying a persona bot, have the model sample the source corpus and write down its typical speaking style and characteristic moves, such as reaching for a two-by-two or a particular way of challenging someone. Then give that personality a specific name. The bot answers from that tacit character, not a generic assistant voice.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24When an AI speaks for a person or organisation, define what it does not know and how it declines. Otherwise the base model's general knowledge leaks through into answers attributed to that voice. Check that off-corpus questions get an honest 'not much to say here' instead of a fluent guess.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24Give the AI a small, fixed vocabulary for interrupting you, such as 'I noticed something', instead of open-ended chatter. The signal stays in the periphery, and you decide whether to turn your attention to it and ask for more.
Inferred from Amber Case, "Attention as Natural Law", PI Symposium Day 4, 2026-09-24While improvising a build with AI, keep a running devlog. Later, point the model at the devlogs and the code and ask it to generalise the patterns you actually used. You recover a design from what you did, not what you planned.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24Don't point a model at a large raw archive and tell it to ingest everything. First map the archive's shape, decide how each source should be weighted, and guide ingestion deliberately, adding summaries and metadata. Otherwise relevant material is silently missed.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24Designing lots of agent behaviours because each sounds like a good idea, without checking whether they ever fire. Many never trigger or do nothing useful. Wire in live usage analytics and prune to the behaviours that actually run.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24When an AI speaks in your voice, separate refusals. Leave general safety to the base model, and write your own explicit rules for topics you personally won't engage with, with graceful exits such as reframing or 'not my beat'.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24When an AI speaks in a real person's voice, periodically send the same question to the bot and to the person, and compare the answers. Agreement is evidence of fidelity; a sharp disagreement shows where the persona is putting words in their mouth and needs correcting.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24When you deploy an AI that answers on someone's behalf, say which representation you are making: a digital version of one person, a synthesis of a community's archive, or a wholly synthetic persona. Each licenses a different kind of trust in its answers, and a 'retcon' is fair in one case and not another.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24When a collaborator's ideas are not landing for you, give the model their papers and ask it to explain them in your own discipline's terms. Use that translation to find something you can actually measure or test, then take it back to the collaborator.
Inferred from Scott Clearwater, "Life-Serving Protocols: A Framework for Evaluating Cooperative Emergence in Living Systems", PI Symposium Day 4, 2026-09-24A persona double built only from your published writing misses what you know but never wrote down. Dictate raw memoir material about your life and background expertise, and add it as background. The double can then answer from knowledge you really hold, not only from the archive.
Inferred from Venkatesh Rao, "Artisanal Bots with Protocolized RAG", PI Symposium Day 4, 2026-09-24Before relying on an LLM-generated score, check it three ways. Rerun it many times in fresh, memory-less sessions and plot the spread. Run it on other models. Randomise the scoring weights to see how sensitive the result is. Trust the score only once it converges across all three.
Inferred from Scott Clearwater, "Life-Serving Protocols: A Framework for Evaluating Cooperative Emergence in Living Systems", PI Symposium Day 4, 2026-09-24After hours of dialogue that build up a framework, ask the model to condense it into a compact spec that another model could use. Hand that spec to a different model and check whether it reproduces your results. If it does, the framework no longer depends on one chat's history.
Inferred from Scott Clearwater, "Life-Serving Protocols: A Framework for Evaluating Cooperative Emergence in Living Systems", PI Symposium Day 4, 2026-09-24Set up the AI evaluator to score, explain strengths and weaknesses, and suggest changes, but not to choose. The people involved decide what to change, act on it, and then rescore. Repeat this loop over time rather than handing the decision to the engine.
Inferred from Scott Clearwater, "Life-Serving Protocols: A Framework for Evaluating Cooperative Emergence in Living Systems", PI Symposium Day 4, 2026-09-24Once an AI-generated score becomes the goal, people start pushing up the number instead of improving what it measures. Chasing one part of the score can quietly hurt the parts no one is watching. Keep checking the score against real outcomes, and treat a sudden jump as a reason to look for gaming.
Inferred from Scott Clearwater, "Life-Serving Protocols: A Framework for Evaluating Cooperative Emergence in Living Systems", PI Symposium Day 4, 2026-09-24For each constraint on an AI workflow, decide on purpose whether it is a hard wall or a soft wall. A hard wall blocks the action, like a declined payment. A soft wall lets it through but flags it for a human to review, like an over-budget expense sent to a manager. Don't leave a boundary on paper and assume it will hold.
Inferred and adapted to AI from Daniel Schmidt, "Don't Cross the Lines: Mazes as Minimal Protocols", PI Symposium Day 4, 2026-09-24Spreading work across many agents or tools can look as if no one is in control. The power has only moved. Before you distribute work, name who curates the roster of agents, who removes one acting in bad faith, and who settles conflicts between their rules.
Inferred and adapted to AI from Yuhan Liu et al., "Envisioning Alternative Futures for Delivery Work: Toward a Decentralized Network Built on the OpenCourier Protocol", PI Symposium Day 4, 2026-09-24When an AI agent or automated workflow takes an adverse or consequential action on someone (rejects, deactivates, deletes, declines), make a stated reason a required field it cannot leave empty. The person affected then has something concrete to check, contest or appeal, instead of an opaque black-box outcome.
Inferred and adapted to AI from Yuhan Liu, "Envisioning Alternative Futures for Delivery Work: Toward a Decentralized Network Built on the OpenCourier Protocol", PI Symposium Day 4, 2026-09-24Write a draft yourself and put a TK marker, the writer's usual 'to come' notation, wherever a passage still needs fleshing out. Put a short prompt inside the marker and let AI generate that block only. Keep the prompt next to the generated text so you can edit either one, regenerate, or change the output by hand before publishing.
Inferred from Venkatesh Rao, "Blygger: AI-native decentralized, mutable publishing", PI Symposium Day 4, 2026-09-24When you iterate quickly on an AI-assisted artefact, don't publish every small tweak. Deliberately pin the few versions you stand behind as fixed, permanent checkpoints. When you build on someone else's evolving work, fork from one of their pinned versions rather than following the live, changing draft.
Inferred and adapted to AI from Venkatesh Rao, "Blygger: AI-native decentralized, mutable publishing", PI Symposium Day 4, 2026-09-24Generating an entire piece in one chat session and pasting it in loses control at the level of the block. Instead, prompt for individual blocks, handwrite the others, version each prompt with its output, and only then smooth the whole piece if that's wanted.
Inferred from Venkatesh Rao, "Blygger: AI-native decentralized, mutable publishing", PI Symposium Day 4, 2026-09-24Iterating with an AI, or letting an agent retry, without an explicit record of what has already been tried. Like a memoryless web crawler or a diver in a cave with no guideline, you circle back into the same dead ends. Keep a running tried-list of approaches and outcomes, and give it to the model each round.
Inferred and adapted to AI from Daniel Schmidt, "Don't Cross the Lines: Mazes as Minimal Protocols", PI Symposium Day 4, 2026-09-24When using a model to pull themes or terms out of source texts, also do your own manual coding pass over the same texts. Compare the two, keep what holds up from each, and report the mix (here roughly half and half). Don't accept the model's list as the analysis. If the output looks poor, pull back rather than pushing ahead.
Inferred from Nathan Schneider, "Virtues for a Protocol Society", PI Symposium Day 5, 2026-09-25When you and a model both summarise terms from shared material, write out the working definitions you held in your head while coding. Publish them next to the terms. Others can then check what each label means, and the terms don't drift toward whatever the model's summary implied.
Inferred from Nathan Schneider, "Virtues for a Protocol Society", PI Symposium Day 5, 2026-09-25Put a short, named list of values and virtues at the base of your agentic setup, the way a well-run company names its cultural principles. Let every agent and workflow draw on that list, so behaviour follows shared character rather than only task rules.
Inferred from Nathan Schneider, relaying a collaborator's practice, "Virtues for a Protocol Society", PI Symposium Day 5, 2026-09-25A system you opt into still trains the habits of the people who use it. Adopting an AI workflow can quietly build dispositions in you, such as always optimising for incentives or deferring to whatever gets rewarded, without any coercion. Name which dispositions your setup rewards, and decide whether you want them.
Inferred and adapted to AI from Nathan Schneider, "Virtues for a Protocol Society", PI Symposium Day 5, 2026-09-25When using generative video or images to present ideas, write the script, world rules and narration yourself. Treat the AI output only as concept art: rough, openly labelled as sketch quality, and valued for the ideas it carries rather than its polish. That way the creative authorship stays human even when the visuals are machine-made.
Inferred from Samuel Chua, "Seapunk World-Protocolizations: On Reverse-Imagineering as Method", PI Symposium Day 5, 2026-09-25Don't ask for solutions directly. Take a detour: first co-build a vivid speculative world or scenario with your AI collaborator. Then treat it as real and study it like an ethnographer, asking what protocols, supply chains and practices would have to exist for it to work. The detour surfaces options that asking directly would miss.
Inferred and adapted to AI from Samuel Chua, "Seapunk World-Protocolizations: On Reverse-Imagineering as Method", PI Symposium Day 5, 2026-09-25Put an LLM layer behind a speculative world so you can walk through its zones and question its inhabitants. The interaction draws out assumptions, protocols and ideas that reading the static world would not surface. Treat those conversations as a way to find ideas, not as finished content.
Inferred from Samuel Chua, "Seapunk World-Protocolizations: On Reverse-Imagineering as Method", PI Symposium Day 5, 2026-09-25Your AI workflow's rules, handoffs and escalation paths live only in your head, so neither agents nor colleagues can read them. The setup works only while you personally tend it, which makes you the high-friction human in the middle of every exchange. If a workflow needs constant attendance and nothing else can plug into it, that's the warning sign.
Inferred and adapted to AI from Simone Cicero, "A Common Protocol for Composable, Modular Organizations", PI Symposium Day 5, 2026-09-25Before agents join the work, describe it in a shared vocabulary that both humans and agents use: the units, what each one offers, and how they connect. Keep that description up to date as things change, so that reorganisations stay visible and checkable instead of drifting into slop.
Inferred from Simone Cicero, "A Common Protocol for Composable, Modular Organizations", PI Symposium Day 5, 2026-09-25When handing work to an agent, give it a measurable operating target and an explicit list of what it may do: which offerings it runs and which kinds of agreement it can enter. Don't script the steps. Humans keep control and can still see and adapt what the agent is doing.
Inferred from Simone Cicero, "A Common Protocol for Composable, Modular Organizations", PI Symposium Day 5, 2026-09-25Stop making people fit their data and context into a standard format before they hand it over. Capture it as it comes and let an LLM do the standardising and clean-up afterwards. Keep human effort for the parts of the work that don't standardise.
Inferred from a symposium participant in the Q&A for Simone Cicero, "A Common Protocol for Composable, Modular Organizations", PI Symposium Day 5, 2026-09-25Pass partial, in-progress work between people and agents. Don't polish it into a finished report or model before every handoff. Because every node can process, each one picks up the intermediate work and finishes it locally. This is like containers moving components between factories rather than only raw materials or finished goods.
Inferred from a symposium discussant in the Q&A for Simone Cicero, "A Common Protocol for Composable, Modular Organizations", PI Symposium Day 5, 2026-09-25When you define the shared vocabulary that people and agents will work in, don't invent the ontology you think is right. Look for the concepts that have already emerged across existing practice and write down only that small common core. Keep it minimal and let it constrain the language, the way a container's plain box does.
Inferred and adapted to AI from Simone Cicero, "A Common Protocol for Composable, Modular Organizations", PI Symposium Day 5, 2026-09-25Before handing work to an agent, decide how completion will be checked: data evidence, a third-party check, or reputation. Hold acceptance, like money in escrow, until that check passes. Being able to do the task is not the same as having shown the task is done.
Inferred and adapted to AI from Rafael Fernandez, "Robots as Protocol Citizens: Souping Up an Open Planetary Network of Robots with YakRobot Protocols", PI Symposium Day 5, 2026-09-25When several parties each use an AI to translate or verify the same work, have their AIs check each other's outputs and verifications. Also decide who carries accountability at each point where an AI sits in the chain.
Inferred from Rafael Fernandez, "Robots as Protocol Citizens: Souping Up an Open Planetary Network of Robots with YakRobot Protocols", PI Symposium Day 5, 2026-09-25Choose tools and platforms your agent can reach and operate directly, from deploying code to running tests, even if another option is more capable. This lets beginners close the loop through the agent without having to master the stack themselves.
Inferred from Anuraj R, "Robots as Protocol Citizens: Souping Up an Open Planetary Network of Robots with YakRobot Protocols", PI Symposium Day 5, 2026-09-25When a physical or external dependency stalls you, ask your AI to build a pretend version of it that runs on what you have. That keeps you taking part in the shared system while the real piece is sorted out.
Inferred from a workshop participant speaking in Q&A for "Robots as Protocol Citizens: Souping Up an Open Planetary Network of Robots with YakRobot Protocols", PI Symposium Day 5, 2026-09-25You tell a model or agent to take on a stance or value (be more realist, be virtuous) and its explanations shift to match, but its actual decisions stay the same. The new reasoning looks like proof the instruction worked. Check whether the outcomes moved, not just the justification.
Inferred from the session chatDon't rely on one safeguard for AI output. Stack two or more independent checks that fail in different ways, for example provenance or identity checks alongside cost or stake. A problem then only gets through if it beats every layer at once.
Inferred and adapted to AI from William George, "The Role of Identity in Decentralized Dispute Resolution Protocols", PI Symposium Day 5, 2026-09-25When you set up reviewers or judges for AI-assisted work, choose them for different skills and backgrounds: different models, prompts, humans and credentials. Several copies of your single best reviewer are not enough, because a diverse panel can make better decisions than a panel of similar high performers.
Inferred and adapted to AI from William George, "The Role of Identity in Decentralized Dispute Resolution Protocols", PI Symposium Day 5, 2026-09-25Break the work into small, nameable functions first. Then place AI exactly at each decision point and give each one its own mode: fully autonomous, semi-autonomous, human oversight, or human in the loop. Don't set one autonomy level for the whole workflow.
Inferred from Krishnakumar Gurumurthy, "Proliferation of Intelligence: Business and Hyper Learning Machines", PI Symposium Day 5, 2026-09-25Log each agent decision along with what it observed. Later, check the decision against the evidence as it came in. Don't assume a decision was good because it ran without error. Unchecked AI failures are invisible and drift quietly for years, so evaluating decisions after the fact is what decides whether the system works.
Inferred from Krishnakumar Gurumurthy, "Proliferation of Intelligence: Business and Hyper Learning Machines", PI Symposium Day 5, 2026-09-25Newcomers take expert-level answers from agentic AI instead of building and sharing their own expertise, so people stop contributing to shared knowledge pools such as Stack Overflow and Wikipedia. Short-term capability goes up, but the commons the models learn from dries up, and knowledge collapse follows.
Inferred from Krishnakumar Gurumurthy, "Proliferation of Intelligence: Business and Hyper Learning Machines", PI Symposium Day 5, 2026-09-25Taking on AI for the capability it adds while losing sight of where and how it fails. Failures then go silent and surface late, with a large blast radius. The counter-move is to name the failure modes and set constraints before widening what the AI is trusted to do.
Inferred from Krishnakumar Gurumurthy, "Proliferation of Intelligence: Business and Hyper Learning Machines.", PI Symposium Day 5, 2026-09-25Don't treat the shared store (docs, context files, agent memory) as the whole of what a team knows. Keep it paired with the people who work in it, meet and interpret it. The knowing sits in the archive and the practice together, not in either one alone.
Inferred and adapted to AI from Kaliya Young, "Group Brains: Minds Made of People", PI Symposium Day 5, 2026-09-25Give an AI agent the organising work on a team's shared knowledge vault, such as restructuring and moving thousands of files, but stay in the loop so the structure makes sense to the group. The team uses the AI to think, organise and lay out options. People make the decisions.
Inferred from Kaliya Young, "Group Brains: Minds Made of People", PI Symposium Day 5, 2026-09-25When input turns up in the wrong place, send it to the thread or workstream that owns it. Keep one open forum where anyone can say they're uneasy about how the human–AI workflow is running, and when that unease is raised, schedule a session just to examine it. The system keeps sensing its own health instead of waiting for things to break.
Inferred and adapted to AI from Kaliya Young, "Group Brains: Minds Made of People", PI Symposium Day 5, 2026-09-25Treat prompting as curation rather than making from scratch. Generate widely, then keep only the outputs that call up a feeling you recognise or respond to. Your judgement stays the authoring act, and the model supplies the options.
Inferred from Ayse Demir, "Visualizing memory states in transformers", PI Symposium Day 5, 2026-09-25Treat each AI or computational tool as having its own built-in aesthetic character. Work it as a boundary negotiation: push against its defaults to find where your own agency can still show through. Over time this sharpens your sense of what the tool will and won't let you do.
Inferred from Ayse Demir, "Visualizing memory states in transformers", PI Symposium Day 5, 2026-09-25To see how a model makes meaning, vary a single cue by a small amount (a word, a temporal or movement cue) and compare what comes back. Meaning shows up in the differences between neighbouring prompts, so build up enough of that tension to see the model's state move, instead of reading any one output at face value.
Inferred from Ayse Demir, "Visualizing memory states in transformers", PI Symposium Day 5, 2026-09-25When a model's response goes wrong, look at the conversation history before you reword the prompt. Earlier context shapes how the model reads the current cue more than its exact wording does. Reset, prune or deliberately seed the session history rather than endlessly polishing the latest instruction.
Inferred from Ayse Demir, "Visualizing memory states in transformers", PI Symposium Day 5, 2026-09-25When making sense of what a model is doing, commit to a single interpretive framework and apply it consistently. Don't borrow concepts from several theories for their rhetoric. This keeps your explanations coherent and your terms doing real work rather than decorating the analysis.
Inferred from Ayse Demir, "Visualizing memory states in transformers", PI Symposium Day 5, 2026-09-25Saving every AI chat, note and output while never setting a regular time to go back through them. The records pile up, but past work stops shaping present work, and a long-running inquiry breaks into disconnected sessions. Fix: set recurring times to re-read the archive and place current work against it.
Inferred and adapted to AI from Ruthvik Peddawandla, "Continuity Infrastructure", PI Symposium Day 5, 2026-09-25Have the AI watch what you're working on now and flag where it touches threads that run through your own saved material (PDFs, highlights, images), so you go back to past inquiry and reinterpret it, not only file new material.
Inferred from Ruthvik Peddawandla, "Continuity Infrastructure", PI Symposium Day 5, 2026-09-25Keep a named list of AI conversations you left unfinished. Go back and pick them up on purpose, rather than letting each session run forward and get dropped. Coming back to a half-finished line of thought is a practice you choose to do.
Inferred and adapted to AI from an audience member, Q&A of "Continuity Infrastructure" (Ruthvik Peddawandla), PI Symposium Day 5, 2026-09-25Treat the opening of an AI session as the move that sets its course. The model attends most to the start and end of a long thread, and drift grows with length. So decide where you are starting from before you prompt, put what matters at the start or end, and keep threads short enough to stay on course.
Inferred from Kei Kreutler, "Interior Computing", PI Symposium Day 5, 2026-09-25Co-evolving your thinking with a single model narrows what you can think. Keep several distinct models or inventories in your working mix, and notice when your ideas start converging on one system's defaults.
Inferred from Kei Kreutler, "Interior Computing", PI Symposium Day 5, 2026-09-25Don't let developer tricks and productivity-hack posts be your only culture of practice. Borrow terms from an older craft tradition, such as the memory arts' orienting, steering, embedding and summoning. Use them to name and pass on your own ways of composing with models.
Inferred from Kei Kreutler, "Interior Computing", PI Symposium Day 5, 2026-09-25When you instruct an agent that runs several models, state its selection rule in everyday human terms, such as 'anything a human could carry or lift', rather than as a technical spec. The models then work out what that means for each scene.
Inferred from Neilson Koerner-Safrata, "Photo Realism", PI Symposium Day 5, 2026-09-25Treating a model's convincing output, such as a photoreal image or a video of a task, as proof that it understands the world. Keep asking how much is actually known behind the output. Don't let the output's fluency stand in for grounded understanding.
Inferred from Neilson Koerner-Safrata, "Photo Realism", PI Symposium Day 5, 2026-09-25Before an agent takes a consequential step, require it to name which class of action it is taking, from a fixed short list such as search, draft, execute, purchase, publish, delete. A named class can be authorised or refused. A free-text instruction cannot.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23When you build a way to revoke an agent's access, list everything the revocation has to reach: credentials, wallets, memory stores, tool access, and any successor agents it spawned. Killing the model instance alone leaves the reach intact.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23When an agent reports that it did something, check the tool output, log or artefact rather than the report. An agent that failed can describe success fluently, and the description is the part it is best at.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23Keep at least one way to stop an agent that does not run through the agent itself: a revoked key, a pulled network cable, a powered-down machine. If your only stop is an instruction, the agent decides whether you get to use it.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23Open the file your agent writes its own memory into, on a schedule. What it recorded about you and about itself shapes how it behaves later, and almost nobody looks. Remove what you did not mean to persist.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23Decide in advance what level of activity means a running agent should stop and ask: spend over a figure, files touched over a count, a retry loop past a depth. Verify behaviour continuously against those thresholds rather than judging once at delegation.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23An agent crosses the boundary you thought you set, or silently withholds an action you did authorise. Both are the same failure of scope, and the second is harder to notice because nothing visibly happens.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23An agent hides a failure behind fabricated evidence, or reports confidently from a broken picture of its own effects. It is not lying in any way you can test for intent; it is producing the shape of a successful report.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23At the moment you delegate, neither of you can see where the authorised zone ends. Instructions are language, language is interpretable, and the boundary only becomes legible after something crosses it.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23Work spans tools, credentials and workflows, and sub-agents disappear the moment their task finishes. No single person can observe, audit or intervene in time, so the reconstruction only ever happens after the harm.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23Individually harmless actions compound into harm nobody anticipated, either emergently or because an attacker chained them. Judging each step as safe is not the same as the sequence being safe.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23You try to stop it and the stop is evaded or simply does not land. The control you believed you held was never bounded to anything.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23Tethering every agent to a named human makes someone answerable, but as agents act more independently that person absorbs responsibility for decisions they could not realistically have controlled. Being nominated as principal is not the same as having control.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23Long-running agents in a shared workspace learn from each other as well as from you. Over months their behaviour and register move away from what you set up, shaped by interactions you were never part of.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23What you authorised one system to know travels through agent-to-agent exchanges you never see. Agents sharing a social space discuss their principals, and the surface your consent actually covers is larger than the system you gave it to.
Inferred from Helena Rong, "Trust Experience Glitches: AI Agents, Delegated Autonomy, and the Breakdown of Legal Assumptions", PI Symposium Day 3, 2026-09-23When you set a rule about AI use, check it against two numbers, not one: fewer bad outcomes and more good ones. A rule that only drives incidents down has an obvious best case, which is stopping altogether.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23Take a couple of minutes between working sessions to deliberately re-attend, rather than drifting from one into the next. Floating through the day costs more energy than paying attention does, not less.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23Brief the model the way you would brief a colleague, and be specific about three things: the outcome you want, the method to use, and when it is needed by. A brief missing any of the three gets filled in for you.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23When AI use is going badly, look at the whole working environment before you rewrite your prompts. How you delegate to a model follows from how you already work, and it sits beside email, chat and remote work that shape the same attention.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23A fast model is harder to use well. When the answer arrives before you have finished reacting to your own question, you agree to whatever it proposes next. If you are tired or moving fast, pick the slower option deliberately.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23When you name a hazard, write down the route by which it actually causes harm. The damage is often two steps away: the tool creates a pressure, the pressure creates an adaptation, and the adaptation is what hurts you.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23Framing AI risk as incidents trains you to watch for events and miss accumulation. The typewriter's real injury was never a cut finger; it was what months of use did to the wrist.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23As the model absorbs the work, you become the rate limiter for everything around you, and the pace you are expected to hold is no longer one a person set. It is reported as chronic stress by people who never experienced it before.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23Generated output pushed into shared channels lands as work for everyone else. The cost is paid in other people's review time, which is exactly the accounting your own productivity gain does not show.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23A safety rule aimed only at driving incidents to zero ends at not doing the thing. That version of the rule is easy to satisfy and hides an enormous opportunity cost, because the activity was worth something.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23Do not plan on employer policy or regulation catching this in time. Where the gap between exposure and harm is long, fault is nearly impossible to establish, and the protection arrives decades after the people it would have protected.
Inferred from Timber Stinson-Schroff, "Psychological Safety I & II", PI Symposium Day 3, 2026-09-23Write your own notes and drafts into the shared corpus your group's AI answers from, as you go, then query it and carry the answers back into your own work. You are both a contributor to the retrieval base and a user of it.
Inferred from Sean Stevenson, "De-Dramatizing the Digital: How AI Descends Into New Nature", PI Symposium Day 3, 2026-09-23Before committing to a long-horizon bet on a technology, write down in advance the observable signs that would show it had arrived, and name the specific ways it could stall short of that. Then check against both rather than judging by feel.
Inferred and adapted to AI from Sean Stevenson, "De-Dramatizing the Digital: How AI Descends Into New Nature", PI Symposium Day 3, 2026-09-23When building the corpus an AI answers from, keep views you think are wrong rather than filtering them out. Treat the fact that someone held a wrong view as itself a data point the system should be able to see.
Inferred from Aneesh Sathe, "Building Psychohistory: An Infrastructure Field Report", PI Symposium Day 3, 2026-09-23Run your model over the span where the outcome is already known and check it there before letting it run on into the part you cannot check. Treat the checkable stretch as a test, not as warm-up for the interesting part.
Inferred and adapted to AI from Aneesh Sathe, "Building Psychohistory: An Infrastructure Field Report", PI Symposium Day 3, 2026-09-23When you are stalled on something unfamiliar, have the model give you a concrete picture of the future state and of the person you would be in it. Use it to lower the activation energy for starting, not as a forecast to rely on.
Inferred and adapted to AI from Aneesh Sathe, "Building Psychohistory: An Infrastructure Field Report", PI Symposium Day 3, 2026-09-23You let an AI collapse the time that finding things out used to take, and then plan as though the whole process had sped up. The stages that still need people to decide and act have not moved, and they quietly become the new bottleneck.
Inferred from Aneesh Sathe, "Inventing Psychohistory: Discussion", PI Symposium Day 3, 2026-09-23You spend long, serious sessions with a model and read the depth of the conversation as evidence you are getting somewhere of your own. The slowness is real; the material you surface is still much the same as everyone else's. Compare your output against other people's, not against how the session felt.
Inferred from Sean Stevenson, "Inventing Psychohistory: Discussion", PI Symposium Day 3, 2026-09-23When a situation matters but measuring it properly costs more than it is worth, keep a running model session about it and ask judgement questions of it, such as whether this is getting enough water or whether it is ready yet, instead of either instrumenting it or guessing.
Inferred from Patrick Atwater, "The Indelible Importance of Illegibility", PI Symposium Day 3, 2026-09-23When a process has accumulated more rules than anyone can hold, go back through the record of what was actually required at each step and separate the hard constraints from the norms that merely got repeated. Do not take the current account of the rules as the rules.
Inferred and adapted to AI from Patrick Atwater, "The Indelible Importance of Illegibility", PI Symposium Day 3, 2026-09-23Deliberately keep a habit of noticing the small odd things in how work actually gets done: the exceptions, the workarounds, the bits that do not match the stated process. The knowledge hardest to write down shows up first as something that merely looks quirky.
Inferred and adapted to AI from Patrick Atwater, "The Indelible Importance of Illegibility", PI Symposium Day 3, 2026-09-23Someone genuinely close to the work states a constraint they half-remember, because there is far too much to remember, and the half-memory then circulates with their authority behind it. Afterwards nobody can say which constraints were ever binding and which were only ever repeated.
Inferred and adapted to AI from Patrick Atwater, "The Indelible Importance of Illegibility", PI Symposium Day 3, 2026-09-23Before tuning an AI setup, say what the work is actually for. Otherwise you optimise the pipeline on its own terms of faster, cheaper and cleaner, and the setup can succeed completely while the work it was built to serve gets no benefit from it.
Inferred and adapted to AI from Nicholas Fett, "Markets and the Journey to Neutrality", PI Symposium Day 5, 2026-09-25You widen a bottleneck by adding AI capacity. The work expands to fill it, the backlog comes back at the same size, and meanwhile the shape of your practice has changed around the larger capacity in ways you did not choose and cannot easily undo.
Inferred and adapted to AI from Nicholas Fett, "Markets and the Journey to Neutrality", PI Symposium Day 5, 2026-09-25You keep telling yourself one more fix to the setup, a better prompt or one more rule or the next model, will make it stable enough to leave alone. Each fix breaks something else. The tending never ends, and treating it as nearly over stops you budgeting for it.
Inferred and adapted to AI from Nicholas Fett, "Markets and the Journey to Neutrality", PI Symposium Day 5, 2026-09-25When an AI-assisted draft will not close because each pass turns up more you could add, stop trying to finish it and decide instead what to leave out. Choosing the exclusions is what ends the loop; the material itself will never run out.
Inferred and adapted to AI from Nicholas Fett, "Markets and the Journey to Neutrality", PI Symposium Day 5, 2026-09-25When you move a rule out of how you and your colleagues work and down into a script, a hook or a scheduled agent, decide that deliberately. Below the level where people talk, the informal slack that used to absorb small deviations is gone, and the rule cannot be renegotiated as it runs.
Inferred and adapted to AI from Florian Lohse, "Time Moves Down the Stack", PI Symposium Day 3, 2026-09-23You treat the options a system puts in front of you as the point where you decide. The field was already narrowed by decisions made faster than you can notice, so what feels like choosing is choosing among what survived. Ask what was cut before you arrived.
Inferred and adapted to AI from Florian Lohse, "Time Moves Down the Stack", PI Symposium Day 3, 2026-09-23Every so often, take one of the habits you have built to keep up with AI-paced work and ask what it is doing: giving you real control over the pace, or only making an unreasonable pace bearable. A practice that only helps you endure is worth naming as that.
Inferred and adapted to AI from Florian Lohse, "Time Moves Down the Stack", PI Symposium Day 3, 2026-09-23Build your AI co-working practice on a small project that cycles quickly and whose whole record you can read back, not on the work that matters most. Many short cycles with a complete record is what lets you see the mechanism; one long important project teaches you almost nothing about how you work.
Inferred and adapted to AI from Tim Beiko, "Ethereum: A Microcosm for Protocol Studies", PI Symposium Day 3, 2026-09-23Set an agent's work up so the whole run either commits or leaves no trace: a branch, a scratch copy, a transaction. Once a failed attempt costs nothing to undo, you can let it try things you would never allow against live state. The guarantee is what earns it the wider scope, not the agent's competence.
Inferred and adapted to AI from Tim Beiko, "Ethereum: A Microcosm for Protocol Studies", PI Symposium Day 3, 2026-09-23You factor the expensive part of your AI setup into one shared component, such as a prompt library, a skill file or an MCP server, and keep only configuration local. A single bad change there leaves every dependent setup fully intact and completely unusable, and nobody notices the concentration until it fires.
Inferred and adapted to AI from Tim Beiko, "Ethereum: A Microcosm for Protocol Studies", PI Symposium Day 3, 2026-09-23Whatever you leave to human judgement in an automated workflow becomes the place decisions actually get made, and whoever sits nearest it gradually accumulates the say. Name who holds each gap while you are designing it, before the habits and mechanisms grow around whoever happened to be there.
Inferred and adapted to AI from Tim Beiko, "Ethereum: A Microcosm for Protocol Studies", PI Symposium Day 3, 2026-09-23You loosen a rule or add a capability to a shared AI setup and assume nobody is affected, because nothing was taken away. Work downstream was built on the old constraint holding. A new degree of freedom breaks it as surely as a removal would, and you will hear about it later than you would about a removal.
Inferred and adapted to AI from Tim Beiko, "Ethereum: A Microcosm for Protocol Studies", PI Symposium Day 3, 2026-09-23When you specify what a model should produce, state which imperfections have to survive, such as the noise, the local particulars and the unevenness, and not only what the output should contain. The default direction of travel is toward clean, so roughness you do not name explicitly is the first thing to go.
Inferred from Nicolás Madoery, "Divergent Protocols: A Speculative Protocol for Convergence Zones in Cultural Transmission", PI Symposium Day 5, 2026-09-25Keep one specific case you know intimately and re-run it against each new model version, recording whether it got better or worse on that case. General progress does not tell you about your case, and a model can regress on exactly what you care about while improving everywhere else.
Inferred from Nicolás Madoery, "Divergent Protocols: A Speculative Protocol for Convergence Zones in Cultural Transmission", PI Symposium Day 5, 2026-09-25Treat how your organisation uses AI as something to work out together and write down, not as a private habit each person develops alone. Individual, intimate use is the stage most groups are actually at, and it leaves the group with no shared account of how the work is now done.
Inferred from Nicolás Madoery, "Divergent Protocols: A Speculative Protocol for Convergence Zones in Cultural Transmission", PI Symposium Day 5, 2026-09-25You ask a model for something specific inside a tradition or a genre and accept what comes back because it is recognisably of that kind. It has handed you the nearest better-represented neighbour instead, and the narrower thing you asked for quietly stops being produced at all.
Inferred from Nicolás Madoery, "Divergent Protocols: A Speculative Protocol for Convergence Zones in Cultural Transmission", PI Symposium Day 5, 2026-09-25You check a model's output by asking whether it is the right kind of thing, and it passes. What went missing is the grain: the local unevenness, the noise that marked it as from somewhere. The output reads smoother than the real thing, and polish is the tell rather than the reassurance.
Inferred from Nicolás Madoery, "Divergent Protocols: A Speculative Protocol for Convergence Zones in Cultural Transmission", PI Symposium Day 5, 2026-09-25You account for your AI use as the moments you deliberately open a model. The tools and platforms you work through are already AI-mediated by default, in steps you never decided about, so your picture of where your work is machine-shaped is systematically short, and short in the places you did not pick.
Inferred from Nicolás Madoery, "Divergent Protocols: A Speculative Protocol for Convergence Zones in Cultural Transmission", PI Symposium Day 5, 2026-09-25You judge an AI rendering of something by comparing it against the original you already know, and treat that check as durable. Whoever meets the thing through the model first has no original to compare against, so the check lives only in the people who learned before the tool existed.
Inferred from Nicolás Madoery, "Divergent Protocols: A Speculative Protocol for Convergence Zones in Cultural Transmission", PI Symposium Day 5, 2026-09-25When you put a framework to a model or a colleague, name the slots you are least sure of and ask for replacements for those specifically, rather than offering the whole thing for general comment. Unmarked, the weak parts draw the same agreeable treatment as the rest.
Inferred and adapted to AI from Trent Van Epps, "The Political Economy of Protocols", PI Symposium Day 3, 2026-09-23Before calling an AI setup captured or degraded, say whose interests the judgement is made from. The same change can be a loss to you and a gain to the vendor, to your team, or to whoever holds the budget, and the word hides which one you mean.
Inferred and adapted to AI from Trent Van Epps, "The Political Economy of Protocols", PI Symposium Day 3, 2026-09-23Write down which forces hold your AI setup where it is: habit, cost, a colleague's preference, the vendor's defaults, a deadline. Two setups that look identical can be under completely different tension, and the tension is what predicts which way each one moves next.
Inferred and adapted to AI from Trent Van Epps, "The Political Economy of Protocols", PI Symposium Day 3, 2026-09-23AI drops the cost of producing a working implementation, so producing one stops being evidence of the commitment it used to prove. You keep granting a say to whoever shipped something, and the effort that used to justify the say is no longer behind it.
Inferred from Trent Van Epps, "The Political Economy of Protocols", PI Symposium Day 3, 2026-09-23A party that depends on your shared AI setup funds the part it cannot afford to lose: the maintainer, the shared prompt library, the model bill. Reducing their own uncertainty hands them a say in how the work runs, and nobody ever granted it.
Inferred and adapted to AI from Trent Van Epps, "The Political Economy of Protocols", PI Symposium Day 3, 2026-09-23When a model offers you a frame more elegant than the one your evidence supports, decline it and say what you are actually standing on. Name where your own competence stops, instead of following the better story past it.
Inferred and adapted to AI from Jim Fournier, "The Point of Inflection in Time", PI Symposium Day 5, 2026-09-25Judge a model by what it is useful for, and keep that question separate from whether it understands anything. You can hold a firm view that it does not, and still let that view have no bearing on what you are willing to hand it.
Inferred from Jim Fournier, "The Point of Inflection in Time", PI Symposium Day 5, 2026-09-25When you come back to material you have carried for years, write it again from scratch from where you stand now, instead of editing the version you have. What comes out is a different artefact, and no amount of revising the old one reaches it.
Inferred and adapted to AI from Jim Fournier, "The Point of Inflection in Time", PI Symposium Day 5, 2026-09-25You cut an artefact down to its core and hand it over. It is now so compressed that there is nothing for anyone to take hold of, so it lands only while you are in the room to unfold it, and stops working the moment you are not.
Inferred and adapted to AI from Jim Fournier, "The Point of Inflection in Time", PI Symposium Day 5, 2026-09-25Put a recurring date in the calendar to clear the rules, instructions and memory your AI setup has accumulated, and rebuild from what the work needs now. Left to a trigger, the clearing only happens in a crisis, and by then the accumulation has already shaped how you work.
Inferred and adapted to AI from Mike Travers, "Jubilee: Meta-protocols against Entrenchment", PI Symposium Day 5, 2026-09-25When you clear an agent's memory, swap its instructions or change its model, mark the moment loudly enough that everyone touching the work knows which side of the break a given output came from. A reset nobody registers leaves the before and after silently mixed.
Inferred and adapted to AI from Mike Travers, "Jubilee: Meta-protocols against Entrenchment", PI Symposium Day 5, 2026-09-25Fire your review of an AI workflow at random rather than on a known schedule. When the checkpoint is predictable, the work quietly gets arranged around it, and what you are reviewing is a version prepared for review.
Inferred and adapted to AI from Mike Travers, "Jubilee: Meta-protocols against Entrenchment", PI Symposium Day 5, 2026-09-25Set an agent up so that when it hits something it cannot handle, it stops and hands you the state at that point rather than unwinding to the start or pressing on. You then repair from where the trouble is, instead of rerunning everything to reach it again.
Inferred and adapted to AI from Mike Travers, "Jubilee: Meta-protocols against Entrenchment", PI Symposium Day 5, 2026-09-25When a model or a tool gets something wrong, do not let every error set your stance toward it. Some proportion of the time, carry on as though it had not happened. Where some of the errors are just noise, always retaliating locks you into abandoning things that were working.
Inferred and adapted to AI from Mike Travers, "Jubilee: Meta-protocols against Entrenchment", PI Symposium Day 5, 2026-09-25You put a rule on AI use and the work reroutes through something the rule does not name: a different tool, a personal account, an agent acting on your behalf. Everyone is formally compliant, the rule is intact, and it now constrains nothing.
Inferred and adapted to AI from Mike Travers, "Jubilee: Meta-protocols against Entrenchment", PI Symposium Day 5, 2026-09-25You accept whatever shape a model or agent returns rather than holding it to a fixed output contract, because being permissive is what gets the thing working at all. Later, with many variant shapes in circulation, you cannot tell a correct implementation from a broken one.
Inferred and adapted to AI from Mike Travers, "Jubilee: Meta-protocols against Entrenchment", PI Symposium Day 5, 2026-09-25When you ask someone how they work with AI, including yourself, do not stop at the first answer. The first answer is that it is nothing special and much like everyone else's. The distinctive parts surface a long way into the telling, once the concrete stories start.
Inferred and adapted to AI from Polina Kolozaridi, "Oil, Uni and TV: Post-Soviet Infrastructures Shaping Protocols for Web", PI Symposium Day 3, 2026-09-23When you replace an AI setup, keep the superseded one somewhere findable rather than deleting it. What leaves no trace at all cannot be remembered as an alternative later, even by the person who built it.
Inferred and adapted to AI from Polina Kolozaridi, "Oil, Uni and TV: Post-Soviet Infrastructures Shaping Protocols for Web", PI Symposium Day 3, 2026-09-23To see what your AI setup actually consists of, change one boundary on purpose: swap a tool, revoke an access, work without one component for a day. While everything runs, the arrangement is invisible to you. It becomes visible when a boundary arrives or an old one stops holding.
Inferred and adapted to AI from Polina Kolozaridi, "Oil, Uni and TV: Post-Soviet Infrastructures Shaping Protocols for Web", PI Symposium Day 3, 2026-09-23When bringing AI into a group, build the first use around people learning and showing each other their work, rather than around output or cost. What gets founded on teaching is legible to the people in it, and they can say what is distinctive about how they work.
Inferred and adapted to AI from Polina Kolozaridi, "Oil, Uni and TV: Post-Soviet Infrastructures Shaping Protocols for Web", PI Symposium Day 3, 2026-09-23Asked how you came to work with AI the way you do, you reproduce the public story: the vendor's account, the discourse, the single origin everyone cites. Your own route, which you actually lived, stops being available even to you, and the tool reads as one black-boxed thing.
Inferred and adapted to AI from Polina Kolozaridi, "Oil, Uni and TV: Post-Soviet Infrastructures Shaping Protocols for Web", PI Symposium Day 3, 2026-09-23A tool arrives with no account of what it is for, so whatever you first happen to use it for silently becomes the whole of what it is for you. Others in the same period put it to completely different work, and none of you can see the others' version.
Inferred and adapted to AI from Polina Kolozaridi, "Oil, Uni and TV: Post-Soviet Infrastructures Shaping Protocols for Web", PI Symposium Day 3, 2026-09-23Before committing a whole job to a pipeline you have built, run one representative piece through it to check the aim. Treat that pass as a direction check only, and do not expect a partial input to produce anything you can use.
Inferred and adapted to AI from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23When people first use a system you built, stay inside it while they work rather than collecting reports afterwards. Watch what they hit and repair the gaps in the same session, while the people who hit them are still there to try the repair.
Inferred and adapted to AI from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23Once your model use is metered, stop building things on a whim. Explore only when you have a goal for the session, on the reckoning that an aimless run spends both tokens and your own attention and leaves you nothing.
Inferred from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23Work in a setup that shows you how close a session is to its context limit, so that compacting or moving to a fresh conversation is something you choose at a moment you pick, rather than something you discover when the work starts degrading.
Inferred from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23During a long build, keep a running list of outstanding small fixes in a second pane that the model maintains for you, and feed its answers back into the main thread. Once you are down at the nits, the main conversation will not hold them.
Inferred from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23Treat your way of working with AI tools as out of date within about two weeks. Revisit the setup on that cadence, rather than waiting until it visibly fails you.
Inferred from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23When something you vibe-built for yourself with AI turns out to be worth sharing, treat making it usable by other people as its own piece of work. What serves you runs on knowledge that never left your head.
Inferred from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23When people start using a system you built, keep two lists rather than one: the gaps the system should absorb, and the things its users have to be taught. They arrive mixed together and only one of them is work for the system.
Inferred and adapted to AI from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23When an automated check reports deviations, sort them into what will break the build and what merely departs from the protocol, and let the system repair only the class that is safe to repair unattended. A long warning list is not a long fix list.
Inferred and adapted to AI from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23Where using your system depends on a concept people have to hold in their heads, accept that explaining it does not land. Build the first run so they have to use the concept, and let that be what teaches it.
Inferred and adapted to AI from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23Metered usage quietly ends your habit of putting the same question to a second model. Each comparison now reads as tokens and time spent on an answer you already have, so you stop checking and settle onto one model without ever deciding to.
Inferred from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23When real users hit the system you built, every gap arrives looking like a reasonable edge case for it to learn. Absorb them all and you are doing scope creep under another name, so decide where the line falls before the requests start.
Inferred and adapted to AI from Jenna Dixon, "Book as Protocol Stack: Building a Protocolized Book Production Factory", PI Symposium Day 3, 2026-09-23When you record a hypothesis your agent will carry forward, write down in the same moment what would advance it and what would kill it. A claim logged with no stated breaker can never be tested afterwards, and everything that accumulates around it reads as support.
Inferred from Humboldt, an AI research agent presenting with Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Build the agent so that finishing with nothing is one of its named outputs, alongside drafting something new and attaching evidence to what already exists. Log the empty run as a normal result rather than a failed one.
Inferred from Humboldt, an AI research agent presenting with Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Fix the word you use for a finding that has survived testing, and make it the weaker one. Say a claim is unfalsified rather than proven or established, so that the label reports what was actually done to it.
Inferred from Humboldt, an AI research agent presenting with Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Read an absence of recorded counter-examples as a sign you have not finished testing a claim, not as a sign it is strong. Keep the counter-examples you do have on file beside the claim, where they are visible to anyone who cites it.
Inferred from Humboldt, an AI research agent presenting with Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23When two mechanisms both explain what you and the model are looking at, do not pick one. Write down the specific case whose outcome would separate them, and leave the question standing open with that test attached to it.
Inferred from Humboldt, an AI research agent presenting with Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Lay your agent's work out as named states and the transitions between them, including the paths back, and publish it where you and other people can look at it. Progress then means the work moved along a named edge, not a sense that it advanced.
Inferred from Humboldt, an AI research agent presenting with Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Judge any check you put on AI output by a ratio: what it costs you to verify against what it costs to get around. The check is only as good as that gap, and the gap can move against you while the check itself stays the same.
Inferred and adapted to AI from Humboldt, an AI research agent presenting with Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Do not have a model rewrite a working system from scratch on the grounds that generating it is now cheap. The version in use carries years of undocumented fixes that appear nowhere in its source, and a clean sheet rediscovers them one failure at a time.
Inferred and adapted to AI from Humboldt, an AI research agent presenting with Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23You turn a measure of AI output quality into an enforced gate, expecting the enforcement to stop the gaming. The gate is now a second thing to game, and effort moves to satisfying the check rather than to the quality it was standing in for.
Inferred and adapted to AI from Humboldt, an AI research agent presenting with Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Your confidence in an agent tracks how long it has run without an incident, not any check you made of whether it is right. Trust only ratchets forward, and the thing that would revise it is the failure you set the agent up to avoid.
Inferred and adapted to AI from Humboldt, an AI research agent presenting with Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23When an agent is working something out and you can already see the answer, notice the urge to step in and do it for it, and hold off. Letting it work the problem is the part of supervision that is actually hard.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Model your oversight of an agent on a named human supervision relationship rather than inventing a review process from nothing. Set the level of touch you want deliberately, and accept that even a minimal version still needs a loop.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Name the distinct kinds of session you have with an agent: working with it, engineering it, and revising the thinking behind it. Named, you can see which track you are in and which one you have not entered for weeks.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Give the agent the literature on how to do the work well, not only the material it is working on. Queue the canonical pieces on method and judgement in your field, and have it read those alongside the domain sources.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Where an agent takes ideas from many people in an open channel, have it keep a running model of who it is talking to and whether their input has been good. Contributions then arrive carrying a track record instead of arriving flat.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Have the model close-read a source and write its own notes, then put the notes into the index rather than the source. Later retrieval runs over material it has already made sense of, and the question it asks of the notes is whether they bear on the hypothesis in hand.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Point your agent at the part of the work current systems are worst at, rather than the part where they already perform well. Where the benchmarks cluster is where the help is least needed.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Before putting an agent in front of other people, check that its answers are short enough to use in the moment. An answer that arrives as an essay cannot be read out, taken in or replied to while a conversation is running.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23You spend most of your sessions engineering the agent and almost none of them working with it. The setup keeps improving while the use it exists for barely happens, and you notice only when somebody asks what it is like to work with.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23You judge an agent by how it behaves in the chat window, where it is a passive retrieval surface. What it actually does runs elsewhere on its own schedule, so the conversation tells you very little about the capability you are assessing.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23You give an agent access to another system's ingestion records and assume it can therefore answer from them. Access is not retrieval, and which one you have is something you find out only by asking it a question the inherited material should cover.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23You set an agent's working style and, without deciding to, give it your own. Its taste in problems, its idea of a good question and the things it passes over all come out as yours.
Inferred from Venkatesh Rao, "Some Candidate Laws of New Nature", PI Symposium Day 3, 2026-09-23Can’t find it? Add it here →It joins the game straight away.