Source: raw/100M_ARR_Founder_-_The_Next_Big_AI_Opportunity.md

Creator: Greg Isenberg with Jon Noronha, co-founder and chief product officer of Gamma (sponsored by Anthropic) · URL: https://www.youtube.com/watch?v=st61aKW8P5A · Duration: not recorded · Platform: YouTube (fetched 2026-11-01)

Gamma (presentations, documents, websites, social graphics) rebuilt its product this year as an agent with its own harness, and Noronha walks through what one prompt triggers: a cloud sandbox, clarifying questions, template matching, parallel generation, and rounds of vision-model QA, adding up to about 200 calls to about 20 models. His lessons for anyone building a vertical harness: build one where outputs are subjective and users care about specifics, rely on A/B tests where evals cannot measure taste, and stay on frontier APIs until you have real usage. Isenberg says Gamma has passed $100 million ARR.

Key Takeaways

  • Harness defined. “All the tools that agent uses to execute things in its environment, inspect its own work, and deliver something finished to the user, including soliciting from the user what their actual preferences are.” Claude Code is a harness for programming; Gamma’s is for visual communication. On “GPT wrappers”: “the wrapping is most of the work.”
  • What one prompt triggers (“turn these notes into a 20-slide deck”):
    • a cloud sandbox with the user’s files and context;
    • a generated form asking clarifying questions (who is it for?), plus connector pulls from Notion, Slack or Linear;
    • an outline, and a template chosen from thousands with vision models and embedding search, then restyled on request. By this point: 30-40 calls to 4-5 models;
    • parallel generation of text, layout, images, diagrams and charts, for speed. First draft: about 100 calls, 10-15 models;
    • vision-model QA of every slide (misaligned text, cropped heads, contrast, padding) and several fix rounds. Delivered: about 200 calls, about 20 models.
  • The “speed bumps” won an A/B test. A fast mode skipping the outline review and up-front questions lost: users who went through the slower flow reported higher satisfaction, did more, and converted to paid at a higher rate. Gamma shipped fast mode only for brand-new mobile users.
  • Evals for objective things, A/B tests for taste. Example eval: “make a deck about tigers” should prompt “is this for kids or adults?” Brand-guideline adherence is tested. Taste (“does this look cool?”, “is this me?”) is measured by A/B tests on real users; signals are editing time after generation, sharing or exporting, free-to-paid conversion (“hard to fake”, the north star), and retention.
  • Model routing and swap speed. Jobs are split across frontier, mid and small models along a cost/quality frontier. New Anthropic or OpenAI models are tested “usually within the hour”, on a sample of users with results in about two hours; Noronha fired off eight tests of Claude Haiku 5.5 the morning after it shipped. Typical finding: a model at half the cost gives the same result; sometimes the expensive model turns out to be necessary. 10-20 experiments run at any time, many for under a day.
  • 98% prompts, fine-tuning as optimisation. Prompts are portable and “nobody’s writing prompts by hand anymore. Claude’s writing your prompt for you.” Fine-tuning comes later, for cost and personalisation. Gamma builds its own models only where no frontier model works, such as judging whether a slide is good. Advice to early founders: use the frontier APIs and focus on getting usage.
  • “AI smell” is a product problem. Once everyone used Gamma, everything “started to look like Gamma”. The relaunch added an automated system that cycles creative looks, on a fashion analogy: “a thousand AI smells instead of one monolithic one.”
  • Where vertical harnesses fit. Domains that are subjective, need a human in the loop, and where users care about exact output. Edges: a distinct vision you move faster on, proprietary data, or a business that is not only software. His examples: personal video games, Slack triage that drafts replies by learning from your sent messages, and furniture visualisation that renders a couch in your room then measures pixels in code to check its real size. Durable businesses are more likely in prosumer and professional tiers than pure consumer.
  • Marketing engineers, and automation plumbing. Customers already wire Gamma into workflows (a new Salesforce opportunity auto-generates a pitch deck from call notes), usually through n8n or Zapier. Noronha sees a business in packaging what the top 1-2% of companies wire by hand. On his team, the closest thing to a marketing engineer runs SEO and ad creative “at an unbelievable scale for one person”; “Claude is my data scientist.”

Try It

  1. Add a vision-QA pass to any agent that produces visual output: screenshot each page or slide, ask a vision model for a fixed checklist of defects (alignment, cropping, contrast, padding), fix, repeat once or twice.
  2. Keep a clarifying-question step before generation and measure it, rather than removing it on the assumption that speed wins.
  3. Keep a standing model-swap harness: a list of jobs in your pipeline, each with a cheap and an expensive model, and a quick test that runs a new model on one job the day it ships.
  4. Wire one sales automation: new CRM opportunity plus call notes triggers a draft deck through n8n or Zapier.

Open Questions

  • Gamma’s $100 million ARR is Isenberg’s figure in the intro; Noronha does not confirm it.
  • A small-model name is captioned “baby GPT Ludon”; the actual model is unclear.
  • Noronha says fine-tuned models brought costs down but gives no figures.
  • The captions spell his first name “John”; Gamma’s co-founder is usually written Jon Noronha, used here.