September 21, 2026
SIG-P4B 2026-09-21
Participants: Sachin, rafa (UTC+1), hbando, Spencer Norwick, Rhea Myers, Max Hodge UTC-8, Andres Guevara, Mike Travers, doxxed_as_zak, Eric Alston (UTC - 6), jorge vega matos, Kannen123
This was the opening session of an experimental, multi-day AI workshop facilitated by Rafa and Sachin of the Protocol Institute. It combined theory — protocols as "engineered arguments," a car/tractor "kit" analogy, and a maturity model for AI adoption — with a live exercise asking participants to have their own AI assistants inventory the tools and "kits" they use. The core thesis: AI use is currently highly individualized ("solipsistic"), and the hard problem is making that private work legible and shareable so people can collaborate. Much of the transcript shows overlapping side-conversations (Eric, Max, doxxed_as_zak, Rhea) running in parallel with the main presentation.
- Protocols as engineered arguments (Sachin): A conventional argument relies on persuasion, memory, status, or trust; a protocol externalizes the argument into something durable in the world that diffuses as more people use it. Example: a traffic light replaces reliance on other drivers' affordances. Externalizing our AI workflows lets us communicate them and possibly "protocolize" useful discoveries.
- The car "kit" analogy (Sachin): Early cars were treated as general-purpose portable power sources; farmers built kits to attach plows, butter churns, washing machines. Specific products (Ford's tractor, ~1917–18) eventually replaced the kits by the 1950s. Sachin flagged the analogy's limit: cars arrived in a low-protocolized world (pre-Federal Highway Act, 1955/56), whereas AI arrives in an already heavily "productized" environment — so a better metaphor might be new car tech landing in the 1970s–80s (evoking Ballard's Crash / Concrete Island).
- AI CMM — two axes (Sachin/Rafa): "Governed" (top-down, org-sanctioned tools, token budgets) vs. "Cultivated" (bottom-up, individual/small-group adaptation). Their claim: organizations are roughly at the second governed level (sanctioned access but no protocolization of use), while individuals are at the kit level — bespoke, one-to-one, hard-to-translate personal workflows.
- Kit → Bridge/Factory progression: The workshop's arc is to move from personal "kits" to "bridges/factories" that let other people's systems talk to yours without direct conversation. A "factory" is legible/runnable without the original author's context; a "kit" is more like a private journal.
- Three affordances of AI (Sachin): automation (repetitive tasks), "spookification"/bespoke one-to-one tools built fast and thrown away (e.g. his RAG agent for one book), and "pensification" — sharing raw outputs/interfaces within trusted small groups.
- Pre-commodification (Rafa): Everyone queries a shared model/substrate, so people independently converge on the same answers without collaborating — "common knowledge" forms invisibly. The model is "like an index fund": you can't see how many others reached the same conclusion, which is disorienting (cf. disputes over who "really" solved a math conjecture).
Session Recording Summary · 17m 38s · Full notes ↗
Reading: Not clearly identified in the discussion. This transcript is from a breakout session focused on a hands-on project idea (building an AI "factory" for job searching) rather than a discussion of a specific shared text. There are passing references to a main "Kafka" room, but no reading is named.
This breakout conversation was a working brainstorm about designing an AI-powered pipeline ("factory") to assist with job searching. Participants shared their personal job-hunting strategies and mapped them onto a possible automated workflow: gathering context (CV, LinkedIn, contacts), searching/reducing the field of jobs, tailoring CVs and outreach, and identifying the highest-leverage next actions. The tone was practical and speculative, ending with the group agreeing this could be a good project to "one-shot" before returning to the main session.
- rafa — closed-loop job search design: Inputs include your CV, target job type, and your network (exporting LinkedIn/email contacts to see where people work, industries, and roles). Emphasized finding "weak network ties." It's a closed loop: once you target open jobs/companies in your network, you go back and update your CV — and possibly even do a project over the next month to strengthen your portfolio for a target role.
- rafa — avoid over-optimizing / define the MVP: A common pitfall in AI projects is over-engineering. The MVP is essentially what "Launch series" already does (auto-tailoring a CV); improvements (finding better-suited companies, checking what past coworkers did) come as later iterations.
- rafa — archetype-based approach: Rather than going job-by-job, an LLM could surface ~4 job "archetypes" most related to your network and CV, pull benchmarks (example pay, example companies), and let the human choose a direction. Then it could recommend actions per archetype: auto-apply, reach out to specific people, or flag where you're under-qualified.
- rafa — outreach tactic (from an Instagram video he made): Search jobs, have the AI find the hiring manager, read their last three posts, and draft a personalized message based on those posts — a way to get outside the job portal.
- Rhea — differentiation strategy: Starts by "panicking," then calms down and looks for a "statistically improbable phrase" in her resume — unusual skills rather than common titles (PHP developer, web designer). She projects where she differs from the field to find jobs she's uniquely qualified for, then rewrites her CV to lead with what each specific job wants.
- Rhea — a three-stage mental model: (1) research/mapping stage (gathering details), (2) reduce stage (narrowing to the jobs you actually want — she mentioned ~200), (3) reformat CV. She framed this as expressing a personal brand / unique selling point to stand out — both to a human you hope to meet and to the LLM that screens applications first.
Questions & Disagreements: - Not a sharp disagreement, but a contrast in approach: rafa's design leans toward mass-applying (at least initially) and automation, whereas Rhea explicitly does not mass-apply — she's careful and selective. rafa acknowledged this difference and suggested a "human check" within the protocol to reconcile the two. - Open technical question (Sachin): Can applications be submitted …
Participants: rafa (UTC+1), Rhea Myers, Sachin, PAtwater