Why Gamma Would Rather Own a Design Lab Than Another Template Pack

Gamma didn’t buy a feature here, it bought an opinion about where AI content generation goes next, and that opinion is that the slide is a legacy output format.

The facts are thin, so let me lay them out before I start reasoning on top of them. In 2026, Gamma acquired Lica, an Accel-backed design startup, to stand up a new AI design research lab. Lica’s co-founders will run that lab, with a stated focus on new presentation methods. Lica’s product turned screenshots and recordings into presentations and videos. Terms were undisclosed.

That’s it. No headcount, no price, no roadmap. But for anyone building generation pipelines, the shape of the deal says plenty.

Research lab is the tell

Companies acquire teams for two reasons: to fill a gap in the product, or to buy time thinking about a problem they haven’t solved yet. Calling the result a design research lab, and putting the acquired founders in charge of it, points squarely at the second.

If Gamma wanted better slide themes, it would hire designers and ship them. A research division implies open questions. And the open question in AI presentation tools right now isn’t rendering, it’s what the output should even be.

I run into this constantly when I wire up bots that produce documents or reports. Getting a model to generate content is the easy half. Deciding on the container, and whether that container should be fixed at all, is where things get hard. A deck assumes a linear walkthrough with a human narrator. A video assumes passive viewing. An interactive page assumes the reader steers. Same underlying content, three completely different information architectures.

Input modality is the part builders should copy

The detail I keep chewing on is that Lica worked from screenshots and recordings. Not outlines. Not prompts. Artifacts.

That’s a meaningfully different ingestion problem, and it’s one most of us dodge. Text prompts are cheap to parse and cheap to be wrong about. Screen recordings are messy, temporal, and full of implicit structure: someone clicked here, then waited, then scrolled to the thing that actually mattered. Extracting intent from that is closer to behavioral inference than summarization.

If you’re building bots in this space, the practical takeaway is that your input surface determines your ceiling. A pipeline fed by prompts can only ever reflect what the user could articulate. A pipeline fed by artifacts can reflect what the user actually did, which is usually more honest and always more detailed.

What that means for architecture

The pattern I’d pull from this deal, and the one I’ve settled on in my own builds, is a hard separation between three layers:

• Extraction. Turn raw artifacts into structured claims, steps, and assets. This is where multimodal models earn their keep, and where you should be spending your eval budget.
• Intermediate representation. A format-agnostic content model. Sections, hierarchy, emphasis, supporting media, dependencies between ideas. No pixel decisions live here.
• Rendering. Deterministic transforms from that model into whatever the audience needs. Deck, video, scrollable page, one-pager.

Most tools I see collapse layers two and three, generating a deck directly because a deck is what the user asked for. It works, right up until you want a second output format and discover the layout logic is tangled through the content logic. Pulling them apart later is a rewrite.

A design research lab is exactly the kind of team you’d staff to figure out what belongs in that middle layer. That’s not a rendering question or a model question. It’s a question about how ideas are structured before anyone decides how they look.

The evaluation problem nobody wants

Here’s where I’d temper the excitement. Design quality is brutally hard to measure programmatically. You can score whether text fits a box. You can’t easily score whether a layout carried an argument.

Any team pushing past conventional slides runs into this immediately. New output formats mean new failure modes, and new failure modes mean your existing evals go quiet exactly when you need them loudest. If Gamma’s lab produces anything genuinely useful to the rest of us, I’d bet it’s less about novel formats and more about how you judge them at scale. That’s the unglamorous work, and it’s the work that transfers.

What I’d watch

With undisclosed terms and no public roadmap, judging the deal on the announcement alone would be guesswork. What’s checkable over time is narrower: whether Gamma’s outputs start diverging from the deck format, whether artifact-based inputs become a first-class path instead of a side door, and whether the lab publishes anything about how it measures design quality.

Those three signals would tell you whether this was a real research bet or a talent acquisition with a nice title attached. For those of us assembling generation pipelines, the second and third matter more than the product itself. I’d rather borrow a solid content model and a working eval use than a template pack.

🕒 Published: August 25, 2026