Field Notes From Fireside: Kerp On AI As Daily Workbench

Kerp's fireside treated AI as a daily workbench: something used while teaching, writing, judging, building course tools, testing product ideas, and carrying context between machines. The useful thread was practical rather than promotional. AI helped when the human still owned judgment, could inspect the source, and knew what good work should look like.

Session Frame

The conversation started from Kerp's work as a professor at the University of Virginia McIntire School of Commerce, where he teaches information systems and entrepreneurship and is studying applied AI in practice. He also connected that work back to older RaidGuild, startup, blockchain, marketing, and legal-tech context.

The examples moved quickly: Canvas and Course Tools for course management, Hypercontext for carrying work context across offices and machines, Juris as a revived legal-tech negotiation experiment, AI-assisted startup judging, grading workflows, pitch decks, podcast research, and book drafting from outlines and conversations.

What Broke: Proxy Collapse

One of the strongest session ideas was proxy collapse in education. Written submissions and reflection papers have often stood in for something harder to observe: whether a student actually read, thought, connected the material, and can explain it. AI makes those artifacts easier to produce, so the old proxy carries less signal.

Kerp pointed toward interview-style and oral-exam patterns as one response, including AI-assisted assessment that can ask follow-up questions and probe whether students understand the course material. The important shift is from detection to design: instead of only asking whether a student used AI, ask what interaction would actually reveal understanding.

The Patient Tutor Pattern

Kerp described Claude Code as more than an engineering shortcut. Used well, it can explain libraries, schemas, structure, security logic, and implementation decisions while the builder is doing the work. That makes the learning loop tighter: build, ask why, inspect the answer, and keep moving.

The caveat matters. The session did not frame AI as a replacement for mentors, teachers, or engineers. It framed AI as a patient explainer inside a workflow where the human still has to test the answer and develop taste.

Context, Memory, And Model Routing

The group kept returning to context management. Hypercontext came up as a response to a practical problem: work spread across multiple machines, chats, local clients, cloud sync, style documents, and recurring project notes. Context is becoming part of the workspace, not just a prompt pasted at the top of a chat.

Model orchestration showed up as another unresolved pain. The session named the cost and confidence problem around choosing the right model for the right task. OpenRouter was raised as a chat suggestion, but the verified session evidence does not treat it as part of Kerp's current stack.

Execution Gets Cheap

The closing thread was about what stays scarce when execution gets easier. If apps, decks, prototypes, and written drafts become faster to produce, the harder work shifts toward taste, sequencing, judgment, QA, and distribution. The value is less in producing more artifacts and more in knowing which artifacts are worth making.

That is the session's cleanest builder takeaway: keep sources close, keep review loops visible, and design AI workflows where the human can still inspect the work instead of only receiving a finished answer.

Topic Map For Follow-Up

Broad wiki candidates: assessment after proxy collapse; personal context portability for AI work; model orchestration for everyday builders; human-calibrated AI judging and grading; taste and judgment after execution scarcity.

Narrow blog candidates: AI as an infinitely patient tutor; when execution gets cheap, taste gets expensive; the first twenty tests as a human-calibrated AI grading loop; proxy collapse came for the reflection paper as a session-grounded education field note.

Editorial Notes

Exact quotes should be checked against the recording or cleaned transcript before publication. Course Tools, Hypercontext, and Juris were shared in-session and can be treated as session links, but their current public state was not externally audited for this session report. Broader wiki topics need external source scans before they become durable topic pages.