Greg Isenberg: 48 hours in SF, 21 observations
A list of 21 impressions from two days with AI founders, labs and executives in San Francisco. They are anecdotes, not data. Ray's verdict: skim. Four points touch the founder's lanes.
The four that matter
- #10 Data readiness is the wall. Every company wants to pipe internal data to agents over MCP (Model Context Protocol), then finds a mess: five versions of the same customer, old Slack channels, PDFs nobody can search. Isenberg calls cleanup "a big consulting opportunity." This is phData's pitch, from an outside voice. See [[project_credibility_for_phdata_sales]].
- #6 A vertical harness = 4 parts. Industry data connectors, evals written by people who know the work, rules for what the agent can touch, and a clean human review screen. "The model is maybe 20% of it." The best teams treat the model as swappable, route models by job, prefer portable prompts over fine-tuning, and A/B test on user behavior (edit, export, share).
- #19 Taste + evals is the scarce hire. Models generate ideas well; the hard part is judging which ones are good.
- #14 Vibe manufacturing named in public. AI that can do CAD; a robotics founder says "the hardest problems in AI are deployment problems now, not research problems." Their robots call a human when stuck, and each call becomes training data. See [[project_ambition_bigger_problem_network]].
Best line
#2: "If a new model drops next month, does your company get better or worse?" Half the dinner table went quiet.
Other points, one line each
- #1 Data brokers reportedly selling company data to labs at large weekly volumes; bubble or start, unclear.
- #3 The new bar is $1B revenue in 12 months; retention is the unasked question.
- #4 Model = CPU, harness = operating system, skills and plugins = apps.
- #5 About 12 months from an agent that watches you work for a week, then does the job.
- #7-9 Cloning and decompiling software with agents is common; value moves to the data inside legacy SaaS; click-heavy legacy software won't survive.
- #11 Primitives for agents (APIs, data layers, tools) with little or no UI.
- #12 Logs of agents failing inside companies are valuable training data.
- #13 Next consumer fight is memory, not model smarts.
- #15 Cybersecurity is hot; each agent with system access is a new door.
- #16 Dead startups license their Slack, email and code archives.
- #17 Monetizing the agent web is early.
- #18 Agent insurance is coming, with audits before coverage.
- #20 Pattern: VC-backed founder fails, becomes a full-time creator, incubates bootstrapped products.
- #21 Distribution is the scarce thing; attention costs dotcom-era money.
Founder direction
Founder 2026-10-14: note only, no Ship 30 angle.
Why this is in the vault
Anecdotes, not data, but four observations land on lanes the founder is actually in — data readiness as a consulting opportunity, the four-part vertical-harness shape, taste-as-scarce-hire, and vibe manufacturing named in public — worth keeping as dated, attributable sentiment from a founder-facing operator, not for the other 17 ambient SF-scene observations.
Mapping against Ray Data Co
- #10 Data readiness is the wall. Every company wants to pipe internal data to agents over MCP, then finds a mess of duplicate customer records, old Slack channels, unsearchable PDFs. Isenberg calls cleanup "a big consulting opportunity" — this is phData's pitch, from an outside voice. See [[project_credibility_for_phdata_sales]].
- #6 A vertical harness = 4 parts. Industry data connectors, evals written by people who know the work, rules for what the agent can touch, a clean human review screen — "the model is maybe 20% of it." Independent confirmation of RDCO's own harness-thesis framing (model as swappable layer, route by job, prefer portable prompts over fine-tuning).
- #19 Taste + evals is the scarce hire. Models generate ideas well; judging which ones are good is the hard part — the same gap RDCO's fresh-eyes-critic pattern exists to fill, and the same claim [[2026-07-25-evals-as-competitive-moat]] makes structurally: the eval, not the generation, is the defensible product.
- #14 Vibe manufacturing named in public. AI that can do CAD; a robotics founder calls deployment the hard problem now, not research — their robots escalate to a human on failure, and each escalation becomes training data. See [[project_ambition_bigger_problem_network]].
Related
- [[2026-10-15-rieseberg-how-to-be-more-ambitious-with-models]] — independently arrives at the same "the model is maybe 20%" framing for what makes an agent harness actually work
- [[2026-08-31-dataengineeringweekly-285-agent-ready-data-architecture]] — the structural, non-anecdotal version of #10's data-readiness-as-the-wall claim
- [[2026-08-31-data-engineering-central-omnigent-meta-harness]] — a commercial attempt at #6's vertical-harness shape, evaluated in more depth