98% of Users Froze on a Blank Page. Then Gamma Let AI Finish 90% of the Deck

Nathan Labenz opens with a reversal that still surprises him. A few years ago the consensus was that AI would automate manual labor first, then knowledge work, and only last, if ever, creativity. Generative models hit creative tools first. Talking with Gamma co-founder Jon Noronha, he names three reasons. The blank page is real: people run out of ideas, and even a GPT-3-class model can brainstorm concepts fast enough to get someone unstuck. Second, many people would rather not learn another interface at all, yet they will say what they want in plain language if the tool is smart enough to catch them. Third, in a creative setting it is acceptable for the model to be wrong on the first try.

Labenz generated a first draft of his own essay with Claude 2 and then rewrote the entire thing. That did not make the detour pointless. Getting an approximate structure onto the page, the slides, or the video still makes the end goal anywhere from two to ten times faster. What that means in practice, he argues, is that AI empowers people to make things they simply could not and would not make otherwise. The people being changed are not the ones who already lived in PowerPoint. They are the ones who never started.

He invited Noronha on after a live workshop he calls savvy shopping for AI products, run with about a hundred executive assistants he is training to tell a real product from a thin wrapper. In front of the group he asked Perplexity for the highest-rated AI slide makers, then tested several of them on the spot. The room was essentially unanimous: Gamma was the one they would actually adopt. Noronha laughs that he will take the assist from Perplexity and from an army of EAs, then agrees it is time to unpack how the product was built, because the AI layer was not the original plan.

They started the company in 2020, in the depth of the pandemic, when the models were not good enough to bet on. Noronha had tried GPT-3 that same year and concluded the technology was not there. The meat of the idea was the anti-PowerPoint. If someone tells you that tomorrow you have a high-stakes deck to present, do you feel a thrill or a cringe? For most people it is the latter. Public speaking and being judged by the cover of the work are part of that dread, but so is the format. Visual communication is barely taught. You drag boxes in an unfamiliar tool, force a story into same-sized rectangles, cut text to fit, and go hunting for pictures.

The first product bet was a writing experience inspired by Notion: type on a page, then pull in multimedia blocks, instead of assembling a collage of rectangles. They also built for mobile, because a presentation is not only a thing you stand up and click through. It is something you send before the meeting and comment on afterward, so it had to look right on a TV and still reflow on a phone. The irony, Noronha says, is that they almost built a website builder for slides. It was more limited than a typical deck precisely because it had to be responsive. They stripped out freeform dragging to keep the medium simple and writing-based.

That constraint later looked like foresight. Large language models are tools that can write almost anything and turn writing into almost anything. Gamma had already built an interface in which a human makes a presentation the way they write a document. When the models arrived, the two halves snapped together. Labenz hears a parallel in Waymark, the AI video company he founded: amateurs mostly do not make video because the craft feels insurmountable, and years of UI polish still left people unsure what they were even supposed to do. Both products were fighting a format problem before they had a model to throw at it.

Gamma launched in August 2022 with none of the AI features. The motto at the time was "write like a doc, present like a deck." It went better than Noronha expected in one narrow sense: they won early adopters who already hated PowerPoint's format and immediately understood the grievance. They did not break out beyond that core. A new person could not be taught what a Gamma was. It was like a presentation without feature parity, plus extra ideas that only landed if you already got them. Users watched a short video and were dumped onto a blank page. About 2% took the first step. The other 98% did not.

AI entered the conversation the following fall as an onboarding patch. Instead of making users build their first Gamma, the team asked whether a model could spit one out so people would actually see the power of the building blocks. They expected a rough skeleton that mixed in various features as a tour. It did much better than that. It generated a full presentation, found images, and created layouts, suddenly showcasing three years of blocks. Then came the second surprise: many people treated the output as a finished product. They would finesse a bit, but they said the tool had taken a task that would have taken 10 or 20 hours and "done 90% of it." That had always been the mission, Noronha says. It had never occurred to them that AI could solve so many of the pieces at once.

Labenz had tried to beat the same blank page at Waymark with templates, always starting from a pre-filled video rather than an empty project. It was better than showing people nothing, and still a hurdle, because users did not know how to project their own content onto someone else's layout. In retrospect the stuckness was often not about buttons. It was conceptual: what is this for, what can I make, why am I even here. People describe Gamma as what you would get "if Notion and Canva had a baby." Both of those companies compounded great products with enormous template marketplaces. Canva might have 10,000 or 50,000 templates, accumulated over years, which is a brutal asset for a small startup to match.

Gamma started using AI to manufacture templates, then asked the obvious next question: why have the model make templates at all? Let it make a perfectly tailored first draft and cut out the middleman. In doing so they also hoped to sidestep the template libraries that protect established tools. The onboarding hack had become a go-to-market strategy. It changed how they prioritized the rest of the product, because AI could sit at the core of almost every bit of the workflow rather than at the welcome screen.

Labenz's own test used an "AI Scouting Report" he had already written as notes: bullet outlines mixed with paragraphs, dropped in with an instruction to make a deck from this. Two things separated Gamma from the other tools in the workshop. Every card looked immediately professional, with theme, color, and layout that passed what he calls a sub-second "does this look Pro" test. He is not, he admits, the aesthetic authority on his own team, but he can tell when something is worth his time, and Gamma was the only one that cleared that bar. The other win was transformation with restraint. Other tools either rewrote his story or printed his paragraphs as giant text. Gamma made the notes presentable without distorting them.

Noronha's unlock came in his first week with ChatGPT, watching people prompt a story about a grilled-cheese sandwich stuck in a VCR, written in the style of biblical verses. The creativity was startling. The deeper lesson was transformation: models are extraordinarily good at turning one shape of information into another. Hip-hop lyrics are a toy. Taking written bullets and messy notes and turning them into a structured, visual, persuasive presentation is the actual job. Both founders agree that, at current quality, converting A into B is where most of the value is. Asking a model to invent the scouting report from scratch would have been useless. Asking it to reshape existing notes was not.

Getting from that idea to a working system was, Noronha says, the hardest problem they faced. Plain text produced bullets and no layout. JSON as structured data reached about 85% reliability, then a curly brace or semicolon in the wrong place collapsed the whole output. YAML was the same failure in a different costume. The PowerPoint instinct, "draw a rectangle here and a rectangle there," was worse, because the model only seems to see. It is doing text in and text out, and generating on a grid almost never works. He still remembers trying to get a three-level pyramid of the kind every corporate deck uses. "It could not draw a pyramid of triangles to save its life."

HTML was the format the model already knew, because these systems were trained by scraping the web. Gamma had already built the editor as stacked, mobile-responsive blocks, so they could generate HTML, then convert it into their own representation. Bold, tables, and lists came for free. The team only had to teach custom tags and elements: side-by-side layouts, columns, timelines. The same lingua franca powers editing. They feed the current card back as HTML and prompt something close to: you are the world's best web designer, your client handed you this markup, change this one thing. The vocabulary is semantic rather than pixel-level. The model is not told how many pixels from the top. It is given the building blocks the interface already understands, plus precious tokens of examples, and asked what it can do with this problem. Waymark, by contrast, keeps templates rigid so the model mostly fills copy and images. Gamma accepted more degrees of freedom, then fenced them with blocks.

Gamma has mostly ridden OpenAI's GPT family. Early on they had no vein of data, so few-shot prompting beat fine-tuning. Timing helped: they launched the AI features in early 2023 just as OpenAI jumped from GPT-3 to GPT-3.5, which changed the cost curve by making everything about ten times cheaper and, at the time, offered no fine-tuning. That made the decision for them. Was fine-tuning going to be ten times more effective? Then, no. Noronha is now watching open-source models and other foundation providers, and he thinks the story is about to invert. The company has generated a huge pile of input-output pairs. Scale, which used to be a prompting problem, could become a training advantage.

GPT-3.5 is remarkable, cheap, fast, and easy to parallelize. GPT-4 is simply better, and also hugely expensive and slow. The dream is GPT-4-level performance with 3.5's economics. Fine-tuning is one path; so is taking a model that is not a chatbot and pointing it at document generation. "GPT-3.5 is really trained to be a chatbot," Noronha says, and many jobs, including generating an entire presentation, do not fit that schema. Labenz is still on a fine-tuned OpenAI model, historically unbothered by token cost because Waymark sells licenses into large companies and quality dominates the P&L. He is now tempted by 3.5-class economics for a different reason: if generation is cheap, you can show three options instead of one. His worry with a single frozen prompt is that it will not match a fine-tuned specialist unless the prompt becomes dynamic, pulling in relevant examples at run time.

The chatbot personality shows up in the feature Noronha thinks is both the coolest and the trickiest in the product. After the first draft, users talk to a card, Gamma's word for a slide, and ask it to change the layout, make it more visual, tighten the copy, or translate it into Spanish. There are so many degrees of freedom that he does not claim it works every time. Honestly, it works well about half the time. One failure mode is that the chat interface wants to be so helpful it will not follow instructions. A user working on a rock-music card asks for a picture of a panda. The model replies that a panda is a bad idea and suggests a guitar. The user has to insist: I want the panda, I am in charge. Fine-tuning, or an open-source model that was never pushed toward companionship, is the hoped-for release valve.

Labenz singles out a design he cannot believe more products skip. After a chat edit, Gamma puts the original card beside the suggested update and lets the user click back and forth before choosing. Even if the edit only works half the time, the choice itself is a clean preference signal. Noronha says the pattern was inspired by Midjourney, which he still considers perhaps the most impressive AI product despite an interface he finds enormously frustrating. So much of Midjourney's product, from variations to upscales to liking parts of an image, is tuned to gather the data you need to make the system better. The landscape of how to use that data is not fully formed. Without the compass, he says, you have no idea where to focus.

Prompt evaluation is still homegrown. The team built a small internal studio for editing a prompt and watching it against different examples. Given how high-stakes those prompts are to the business, Noronha compares the current state to running a huge codebase with no automated tests, shipping on vibes. That itch is biographical. Much of Gamma's team comes from Optimizely, the company that tried to make A/B testing standard practice in marketing and software, so gradual rollouts and measurement are in the DNA. The tooling on the AI side has not caught up. Labenz points him at HumanLoop, whose CEO Raza Habib had been on the show, as infrastructure aimed at application developers rather than researchers.

Labenz's production worry with open-source models is not the fine-tune. It is bursty inference: big-company demos where everyone clicks at once, or jobs that kick off a hundred images and need them back immediately. Auto-scaling on hosted endpoints has not felt ready. Noronha agrees that this is one reason Gamma held off on AI image generation even though the product is full of pictures. The bigger reason is blunter: most generated images still looked like crap, though quality is improving fast. In the decks Labenz tried, he detected no generated images at all, only library search. For image models, open source is often the leader among the weights you can actually run, which makes the serving problem inescapable. He does a little reverse business-development and recommends Baseten, which packages fine-tuning with autoscaling for bursty traffic. Gamma was about to roll out its own image generation on top of it.

Noronha thinks in thresholds. A capability is not good enough until it is, and then it goes super wide, super quick. Midjourney had already crossed the line. Open-source models, including Stable Diffusion XL, had just crossed it. Video generation felt similar, and voice was having its own week, with Labenz citing an ElevenLabs clone on *The AI Breakdown* that he could barely distinguish from the host, plus a strong PlayHT drop. The job of an AI-enabled product, as Noronha describes it, is to leave a hole for cool images so that when cool images are ready you can flip a switch. Thinking a year ahead, with 3D and voice and whatever comes next, is almost staggering. The practical move is to stay in position for sudden sparks rather than to forecast each one.

They are still too early, he says, to make generated images the default. Libraries have a known quality bar; AI images still feel like a toy much of the time, and Gamma's aspiration is not to be a toy. The magic, and the control, will matter for a specific deck problem: twelve images throughout a presentation that are stylistically and color coordinated, a job models will go from bad at to good at in months. Waymark has been equally cautious, because its videos often run as television commercials, just as Gamma decks get presented to audiences that matter. Their stack prefers the user's own photos, uses Salesforce's BLIP family to pick which ones belong, and when it needs more, queries Shutterstock's computer-vision API for visual neighbors, then filters on metadata so similar-looking does not mean wrong content. Two sieves, and the quality bar holds.

Beyond models, Gamma is still chasing the original medium. To go up against PowerPoint they need unglamorous features, from PDF export to speaker notes. The surprise pull from users is websites: personal pages and digital faces on the internet. The accidental slide website-builder turned out to be a website builder, in part because it lacks the complexity of a full site tool. The promise is write like a doc, share like a website. They are adding custom domains, headers, footers, and forms. Being great at presentations, sites, and documents at once is hard. AI, Noronha says, makes it more feasible than the old assumption that a company cannot compete in several surfaces, because the model changes how much unique machinery each surface needs. They are trying to forge a thing that is not a slide, not a site, and not a doc, and that still fits how people actually work.

The question that keeps him up is Microsoft and Google. Familiarity and distribution are real advantages, and any fundraising conversation would lead with what happens when they ship an AI slide maker. Gamma's strategy, even before AI, was not to compete head-on with tools that are already good at being PowerPoint. They wanted a different object, with AI as an accelerant. Every month that Google Slides AI and PowerPoint AI fail to appear, the temptation to compete head-on grows. He does not know how good those tools will be, or how accessible they will be to small business and education, versus Microsoft's enterprise core. Google he calls a fierce competitor, an oil tanker that does not turn quickly and then moves relentlessly. A startup's bet is to stay nimble and a little bit elsewhere. The website market, he notes, has incumbents that have shown much less savvy with AI, and that may be the larger opening.

Labenz finds Gmail's writing assistant still not useful, in part because it offers a few modes, shorten or elaborate, and never lets you just say what to do. Enterprises are shipping, and some of it looks half-baked. Noronha used to work at Microsoft, and from the inside the scale is easy to underestimate. Gmail and Google Slides have hundreds of millions of users, each with contracts, languages, and privacy requirements. The number of humans you must convince internally to make even a low-risk change is large. AI adds risks the industry is only beginning to understand, from hallucination to misuse. Incumbents' hands are far more tied than a startup's. Their rational move is to let smaller companies establish patterns, then incorporate them. Distribution means they will not be doomed by taking their time. Behavior that looks odd from a fast-paced startup is, on that view, perfectly understandable. If you want the cutting edge of AI interfaces, you almost have to look outside those ecosystems. That is a preview of what will be in PowerPoint in a year or two.

Labenz offers two opposing stories about bloated suites. Either all that complexity makes AI harder to apply, or the model becomes a natural-language layer that hides the mess. Noronha will not pick a winner. So far the AI tools that work have picked off narrow, specific workflows and carefully crushed prompts and examples against them. He would be a fool, he says, to predict a year of model evolution. The trend has been surprising generalization: one model for translation, one for search, one for summarization, and then "god models" that do it all, a trend that might only accelerate. He will not dare say how the story evolves. What he will say is the URL: gamma.app, which Labenz calls, for his money, the market-leading AI presentation maker of the moment.