Scott Belsky: Models Are Commoditizing. Controls, Data, and Taste Are the Moat

Jacob Effron asks Scott Belsky for a piece of AI-generated work that is a window into where creative tools are going — not the pop-culture Gucci-jacket stuff, but something that actually moves professionals. Belsky, founder of Behance, prolific investor, and Adobe’s chief product officer and chief strategy officer, points to a mash-up Adobe did not expect. The company shipped Project Neo, a 3D illustration program for rendering structures, aimed at illustrators and unrelated to image generation.

Creators figured out they could take a Neo structure and feed it to Firefly, Adobe’s image model, as a structure reference. Firefly suddenly had extreme precision in how a prompt was directed. The workflow unlocked on social: people showing their audience and flexing “look what I discovered.” It popped. Belsky’s takeaway is that obsession with the next-best model is probably a little passé. The era now is not the next-best model. It is the controls one can apply on top of the model.

Adobe, Belsky says, is focused on interfaces, models, and the data underneath. The Firefly family is homegrown, trained on licensed content, with a compensation program for contributors and an aim to indemnify customers. They care how the models were trained. On the interface side they are integrating those models into Photoshop-class products, opening them to third parties and big companies that need complicated at-scale workflows instead of individuals doing the work piecemeal, and building controls like structure reference and Style Match.

On the data side the push is custom models: imagine Nickelodeon training a version of Firefly on SpongeBob SquarePants so it can generate, without concern, tons of material to ideate new character development and storylines. LLMs they do not build; those come through partnerships, surfaced in Acrobat’s AI assistant and in digital-experience products for marketing analytics. Effron recalls early Firefly: people prompted Spider-Man and got nothing. Belsky treated that as a feature, not a bug. Enterprise customers’ top concern is legal risk, and Adobe hears it constantly.

What separates people who mess around from people who make this a routine? Belsky goes back to the Behance years. The greatest creative professionals cover the full surface area of possibility: they pursue little mistakes of the eye, go down a track, revert, try another path. When they finally hand a client two or three solutions, that is the end of a process that probably started at twelve. The constraint they always fight is time. They have only so many cycles. The consistent theme across Adobe’s franchise is people using the tools to explore far more surface area, far more quickly.

Generative Recolor in Illustrator lets you take hundreds or thousands of palettes from a prompt, apply them to vector work instantly, and pick the best combinations. That used to take days and dedicated teams. Now those people are unlocked for higher-order exploration of the vector objects themselves, not just color variations. Startups, Effron notes, obsess over the blank-canvas problem: powerful features, a blank screen, no idea how to start. Superpowers do not mean people know how to use them.

Belsky reaches for the camera. When it was invented, artists hired for family portraits were deeply offended that a click could capture the same thing. It was a disaster for that industry. Photography became an art anyway, because you still had to choose how to capture the image, which lens — almost like which prompt — and lighting, and then, from hundreds or thousands of frames, which ones were best. That last act is taste, a skill more important than a lot of the rudimentary craft now offloaded to compute. Taste, he argues, is more important than ever. Adobe’s job is to let professionals flex it: prompt augmentation, onboarding, taking what is in the mind’s eye and making the better vision produce the better outcome.

Effron asks whether the industry converges on one creative model or a bunch with different aesthetics. Belsky thinks we will have to choose among models. A lot will be commodities; a lot will be trained on a specific slice of data and tuned for niche reasons. Adobe’s tools should let people pick different models for different use cases. Commercially safe, licensed models matter for shipping work. Mood boards are another matter: they are one of Behance’s most popular features, and no one has ever complained that people make them from other people’s images, because that is how every creative process starts. Why shouldn’t you use models trained on other content for ideation, so long as you are not using the output commercially? He wants Adobe to lean into an ecosystem of models for those jobs. Compliance needs differ by stage of the process, so the stack should be a fleet, not one model.

Is owning a model a moment-in-time problem, because nothing compliant existed for Disney, or a forever need? In areas where Adobe is a world expert, Belsky says, they will deliver a better end-to-end experience with their own models. Media generation is not only training data and a massive parameter count. It is fine-tune controls, custom-model capabilities, and data about how people actually use the tools. Look at prompts online: people specify the lens, the lighting, contrast, density, tints. Those are Lightroom dials and filters. Adobe knows that mapping better than anyone in the world, and not using that dataset would be throwing away the family jewels.

The litmus test for his team: if it is something they believe they can do better than anyone else in the world, it is their responsibility to do it. If it is commoditized, or some other company will definitely invest more and always have something better, partner. Today everyone is focused on a command-line, prompt-and-get-it experience. The next generation is that you ask for something, the model gives it back, and it also gives you custom UI to fine-tune what you just got — UI on demand for the use case of the moment, then it disappears. That is not one standard Photoshop chrome. Unlocking it, he thinks, means building their own models. Where are we on that journey? It is coming. It also has to be professional grade, which is the tricky part.

Startups keep announcing a video model; there are already half a dozen or a dozen. Adobe is building its own. When the focus is commercial-grade media, the bar is high — it cannot look weird. Belsky analogizes to self-driving: people talked about it fifteen years ago, and only now can you get in a Waymo in San Francisco. He expects a big gap in some areas between the cool happy-path demos and pragmatic, everyday, commercially viable quality. All the video models have great demos; you prompt them yourself and they do not quite look like the demo.

The milestone he cares about is pragmatic, because he is focused on what customers will actually value. You are editing in Premiere Pro. The scene is three seconds short of the song mapping, or the director wants two more seconds of mood, and you do not have the footage. If you could extend by a few seconds, would that change everything? Would it save going back on set and reshooting the whole scene? One hundred percent. Double down horsepower on that, he says, rather than generating 30 seconds of 4K aerial b-roll of any geography in the world. Some of those things are further off. Compute should go where customers will pay.

Does he worry OpenAI or Pika get so far ahead on video that they capture the market before open source arrives? Adobe is working with Pika and Sora in the product. As long as people still need an interface to tell a story, data to customize it, and tools to make it look like whatever is in their mind’s eye, Adobe wins when every model gets better. These are platforms, the way a better Apple chip makes Adobe’s products better. They are also building the models that unlock things only they can do with a full-stack advantage. In many ways he is rooting for the frontier labs.

Is AI just an incumbent bonus for companies with workflow products, or is there room for net-new companies? Internally and in the market, Belsky’s bet in creative and marketing is differentiation at the interface and data layer. Unless you absolutely have to build your own model, they become further commoditized, and most startups are on third-party models anyway. The industry, he thinks, is going toward hyper-personalized digital experiences at scale: websites that speak as if they know you, emails with a video featuring the stuff you bought. To do that you need customer data — Adobe has a CDP for big brands — marketer workflows to deploy, measure, and optimize campaigns, and a creative stack with guardrails, including brand-check AI, because one bad asset can screw up a brand. That is not models. It is interface, workflow, and data. When startups say they do that, he asks how they really have all the pieces.

There is still a ton of opportunity in work that was extraordinarily cumbersome. He has started calling this an era of small businesses operating as huge businesses: the equivalent of a hundred-person marketing team because of what the tools enable. Big companies do not focus on helping small businesses look huge. Law, government, and other ridiculous old-school processes are also open. Hyper-personalized media, though, has a pendulum. Brands are about to flood the zone. The next generation will wonder how websites used to ask gender or shoe size. Content will be free to generate, vary, culturalize, and translate. Humans will be inundated. Every time something becomes commoditized, we seek more scarcity and meaning in that category: expensive shoes that flex cultural sensibility, luxury cars. He has been thinking about luxury software, even handcrafted software “not prepared by AI coders.” Superhuman manually onboarding you is a luxury software experience — an interface on Gmail that you pay for as a flex that your time is worth more than the average person’s.

That craving for scarcity, meaning, and craft will reboot the role of humans in content creation. He agrees we will not want entertainment that features us; we will want a shared experience to talk about with friends. Music startups, focused on the AI’s role, forget why we listen to a song over and over: often we know the story. This is one of Taylor’s breakups, and everyone wonders who. That folklore drives the value of the experience. Lower the floor so more people can make music, but also let them tell a very human story behind it — this is the song from a profound moment in my life — so the listener leans into the words. That relationship between artist and consumer is what a lot of AI tools are still missing.

Customers are climbing a ladder. The base of generative AI is glorified Q&A. One step up is doing things for you: skip learning Photoshop tools, hop through a command-bar experience. The next tier they are asking about is proactive suggestion — explore this way, that would not perform in that geo, make a variation for this version. That is a superpower. They also want speed and quality; a lot of unlocks are still latency. Cost matters for going faster, and there is a parallel track of making models cheaper, but when Belsky talks to design and product he never has them work inside cost constraints. Unlock whatever is in the mind’s eye. Adobe’s business model is an economy of generative credits. If something is extremely intensive but saves you days of work, you will pay, and they should deliver it rather than withhold it because it is expensive.

Pricing for AI features is still unsettled across the industry. Adobe shipped a basic allotment of generative credits with existing plans. A subset of customers broke through; Adobe made new plans with more credits. There is little risk in the approach: as they unlock generative video or other intensive capabilities, they can change how many credits a job costs. The mission is to keep it as cheap as possible, accessible, part of everyone’s workflow. He thinks we are very close to a time when everything anyone makes has some degree of generative AI in it. Pricing correlated with cost lets new features drop in without rebuilding the whole scheme.

Biggest surprise building AI features? He quotes his friend Dave Morin’s classic: the devil’s in the defaults. When they put generative fill in Photoshop in a new default generation bar that people saw as soon as they came into the product, it unlocked utilization they could never have forecasted. The question now is how AI becomes a default part of how people use every product, in a way that gives them superpowers. That is a different playbook from debuting a feature and hiding it in a panel on the third tab. What has he changed his mind about since the ChatGPT craze? He used to think a few models would win, the next Googles. Now he thinks there will be thousands, and a lot will happen on the edge. Capability is rising so fast that models displace each other every day, but most real use cases sit below the frontier, so teams will get less choosy and more focused on cost, especially with local and open-source models.

Outside creative AI, the startup he is most excited about is KoBold, using AI to identify mineral deposits. He had no prior passion for mining; meeting the company showed him how antiquated the world is. R&D departments are dilapidated. When a mine comes online it is $10 billion of value and billions to build, massive and inefficient. If AI can increase identification of deposits four to ten times in speed and reliability, and reduce cost, you have a global-changing company — minerals for batteries and EVs. If he left Adobe to build again, not in mineral deposits, he would go after personalization in the physical world. Every website will be a conversation with a brand; in every physical space we are still strangers. Being remembered at a restaurant for your favorite drink, or having a personal shopper you have worked with for years, is magical. When Garrett Camp and Travis first pitched Uber it was “everyone’s personal driver,” a sensation few people were entitled to, suddenly available to everyone. The hospitality version of feeling known wherever you go is very interesting to him.

His side hustle is Implications, a monthly newsletter that is an accountability mechanism: connecting the dots for his teams at Adobe, the founders he works with, and people he respects. It is usually written at 30,000 feet, trying to be provocative and direct, and he is not known for brevity. Does he use any AI in the writing of it? He does not. He tries to scrape his own intelligence for that one.