Jon Noronha
Related people
Episodes
Notes
The Startup Ideas Podcast · Stop building AI wrappers. Start building agent harnesses.
-
What Is a Marketing Engineer?
Greg Isenberg says a marketing engineer runs dozens of experiments a day with AI agents. He traces the job from one big commercial a year in the Mad Men era, to the digital marketer, to the growth hacker running a few experiments a week, to an AI-assisted marketer, and now to this role. Jon Noronha buys the thesis. At Optimizely, before Gamma, the hope was a hundred tiny versions of a website so each visitor saw a different page. That mostly did not happen. Sites and apps are still mostly one version, except feeds such as TikTok and Instagram Reels, where companies spend heavily to personalize. A backlash risk, including the bad feeling of surge pricing, is part of why wild personalization s
-
Should Startups Fine-tune Their Own AI Models?
Jon Noronha says early startups should not fine-tune or train their own models as the plan. Gamma has built on other companies' models from the start, moving among Anthropic, OpenAI, and Google. Training costs more than an early company can sensibly spend, and the models on the market keep getting better. He would rarely want to compete with them. The first job is real usage, and the models already out there are how a team gets it.
Fine-tuning comes after that usage exists. Gamma has used it to cut cost and, in some cases, to personalize. Judging whether a slide looks right has not gone well on the big models, so Gamma is doing its own research for that job. He still says to use the strong
-
Why Does Taste Beat Evals in Creative AI?
Jon Noronha says taste beats evals when the question is creative. An eval is a fixed question used like a software test, and it works when the answer is objective. One Gamma eval says that "make a deck about tigers" should make the agent ask whether it is for kids or for adults. Scores compare prompts and models. Whether something looks cool, or sounds like a particular person, has no single objective test.
Gamma still checks proxies, such as whether the output follows instructions or sticks to a brand guideline. Founders compete by bringing their own sense of what looks good and by listening to users. The less verifiable the field, the more room there is for a point of view. He thinks clot
-
What Is AI Smell?
Jon Noronha says AI smell is the tell that work was made with AI. It shows up most in writing, when an email uses a word the person would not use and does not sound like them. He also sees it in images. Walking around San Francisco, he can spot cafe and restaurant posters that look like Nano Banana or a GPT image, and more people are starting to notice as that look spreads.
Gamma hit the same wall. When few people used it, a deck stood out. Once everybody used it, everything looked like Gamma, and people did not want to be recognized as using AI. The launch they just shipped was meant to widen the range, so there can be a thousand looks instead of one, and a person can move between them. Th
-
When Should a Startup Build Its Own Agent Harness?
Jon Noronha says a startup should build its own agent harness when the job looks very different from plain coding. For Gamma the work is subjective: two people disagree about what good looks like, a human stays in the loop, and the person using the tool cares about the specifics. He treats those traits as the test.
A personal video game is one opening he is bullish on. Coding agents already make games, and they are weaker at taking in what the player wants, the way non-coders shape Roblox or Minecraft. He can imagine a harness that runs inside a 3D world. The practical rule is to build something that looks nothing like what everyone else is shipping, then keep moving, because copies trail b
-
What Is an AI Agent Harness?
Jon Noronha says an agent harness is the set of tools an agent uses to do work, check that work, and hand the user something finished, including asking what the user wants. Gamma, a tool for presentations, documents, websites, and social graphics, started in 2020 and took off in 2023. Early on the team made one-off model calls, such as writing one slide or an outline, and wired them by hand into a preset path that could build a 10-slide deck. The change in the last year or two is a higher-level loop: one model does some work, inspects it, uses tools, and keeps going across many calls. Rebuilding Gamma as an agent this year meant building that harness. Gamma is a creative agent, not a coding
-
Why Can One Niche Grow Into a Big Company?
Jon Noronha says a big company is a thousand little niches piled on each other, so a product can start with one niche and then look for the adjacencies. It does not have to start broad.
Less than a quarter of Gamma's revenue is in the United States. The user base is very broad because many kinds of people make presentations, and revenue is spread across job roles, from consultants to teachers to marketers to salespeople to lawyers. Lawyers make presentations. Where that work leads is unknown at the start. A line of work built for lawyers can become a huge part of the business. A founder does not have to know which niche will grow.
-
Why Can Slowing Users Down Raise Paid Conversion?
Jon Noronha says slowing people down raised paid conversion. Gamma tested a fast mode that skipped the outline and the upfront questions, on the idea that less friction would get people to a useful result sooner. He wanted that test to win. It lost. Seeing the outline, choosing the high-level arc of the story, and answering questions led to much higher satisfaction ratings. People who got through the slower path felt more invested in the output, did more with it, and converted to paying at a higher rate. A test is not only one answer. Gamma still shipped a version of fast mode for brand-new users on mobile, where typing a stack of answers is a poor fit.
The same testing habit is constant. G
-
What Happens When You Paste Notes Into Gamma?
Jon Noronha says that pasting messy notes, a document, or a PDF into Gamma and asking for a 20-slide presentation starts an entire computer in the cloud. The sandbox holds the prompt, the file, and other context about the user.
Gamma reads the files and usually asks what else it needs. A form asks who the deck is for, such as school, adults, or work. Connectors to Notion, Slack, or Linear pull in related context. The first 10 seconds, just sourcing information and asking those questions, already use three or four models, and a generated interface shows the choices. Gamma then writes an outline and styles a template. Embeddings search thousands of visual templates, vision models judge the fi