MaHui Entrepreneur | Lovart Founder Malvin: We Have No Product Managers, Only Designers

• This article is republished from GeekPark Founder Park

Lovart deserves attention. It validates and extends the product innovation capabilities of AI application-layer teams — the hottest Agent since Manus, successfully taking the Agent product form from general-purpose domains into a vertical赛道.

According to reports, after Lovart's launch, nearly 5,000 discussion posts appeared on Twitter, its official video racked up nearly a million views, and it earned a like from Elon Musk and a post from Grok's official account. Within 24 hours, waitlist applications exceeded 20,000.

In its basic form, Lovart looks a lot like Manus — an Agent that can invoke tools to complete tasks on behalf of users.

But Lovart goes further in a vertical domain. It takes a multimodal "profession" and turns it into a workflow, then internalizes that into an Agent, paired with a product form suited for designers: the canvas. The "canvas" is the "table" — a return to the primal state of design, no computers, just pen and paper, one person with a need, another with the skill, and good design emerges from that scene.

"Lovart is of course a tool right now, but what about later? It will become someone with professional attributes, directly delivering the results of a service."

Lovart founder & CEO Malvin has extensive hands-on experience in AI design. He believes image-generating AI products have already entered their third phase.

Phase 1.0: Midjourney and similar products — single-point image generation capabilities directly productized. Phase 2.0: Workflow products like ComfyUI — connecting dots into lines, linking multiple model capabilities, AI able to handle more creative image tasks. Phase 3.0: Agent, as in Lovart — AI participating in workflow planning and execution, further lowering the barrier for users.

Creative tools keep simplifying. ComfyUI's high threshold shuts out many creators. But Lovart — at least this is the first product our professional designers saw and immediately wanted to try.

One day after Lovart's release, we spoke with Malvin. How was Lovart born? What thinking lies behind it? The answers are all in this interview.

Founder bio: Malvin, undergraduate degree from Southeast University, former product director at Mobike, former global commercialization lead for CapCut at ByteDance.

1

Three Phases of AI Image Products

Founder Park: Was Lovart a product inspired by Manus?

Malvin: I'm very grateful to Manus. In the startup world, many cutting-edge companies actually share similar understandings. But before something happens, even if you have that understanding, you're skeptical — not 100% certain. Only after something blows up does a collective consensus form.

That collective consensus reinforces your conviction. We were building Lovart before Manus launched, otherwise we couldn't have moved this quickly. Honestly, the moment that made us go all-in, that made us feel the direction was certain, was when Manus went viral.

But ultimately, when that subconscious understanding breaks through and catches fire, it's being validated by collective consensus — the era is calling for this product. We were incredibly excited then. That proves something right.

Innovation is like this. Everyone starts by groping forward alone in the long night, guided by their own understanding, but skeptically. As more products emerge, that belief grows more certain, the path more consensual. So, truly, I'm quite grateful to them.

Founder Park: Is Lovart a product for everyone, or for designers?

Malvin: We're building a vertical Agent, not a general-purpose Agent. Big companies have overwhelming technical and traffic advantages; we don't want to get drawn into adversarial competition. To compete through product innovation — general Agents are of course for everyone, and big companies will lean that way because it's clearly the bigger opportunity. But for entrepreneurs, the challenge is immense.

Yiming Zhang previously mentioned that in the entrepreneurial process, cognition matters most. I deeply relate to that. AI entrepreneurship right now is "cognitive lead + extreme execution" — just execute relentlessly, execute with extreme speed. Whoever builds the first innovative product, that's what matters for an innovative company.

Founder Park: How was Lovart built? Why did you want to make this product?

Malvin: From our experience, design and image generation products (and this can extend to LLM products) have three developmental stages.

Phase 1.0: Content generation — single-point content generation. At the start, people used Midjourney, Stable Diffusion; in China there was Liblib. That was the first wave. Back then people used Liblib through WebUI; essentially it was still generation. This was the first era, what we call 1.0.

Phase 2.0: Workflow. Coze, Dify, and many other products were essentially building workflows. The image domain was the same — soon you saw people starting to use ComfyUI, with more complex workflows implemented through ComfyUI. There's a huge shift here. In the image domain, WebUI and ComfyUI — especially in China — these two products are actually the most representative examples of application paradigm shifts. The underlying change they represent is this: AI's ultimate purpose is to help humans complete a sequence of work. So the first step of this work, or the atomic unit, is individual generation capabilities, or generation tools, or traditional tools. At first these generation capabilities were created, so people could use generation capability as a tool in their workflow, string these tools together into a workflow, and replace a larger proportion of the original work.

Phase 3.0: Agent — we define Agent as AI planning and executing workflows. Because model capabilities have improved (Claude, etc.), AI can now串联 all tools and work, and execute them. Humans only need to give AI instructions.

This understanding wasn't something we had from day one of our AI entrepreneurship. It's a general understanding of AI applications. At first everyone only did AI generation (images, video), then workflow forms emerged. Behind this, every era-defining product did something important correctly.

Founder Park: For 3.0 products like Lovart and Manus, user value hasn't fully closed the loop yet. Why choose a product form that's still immature?

Malvin: Lovart is built on certain understandings and judgments. Of course many people now feel Agent isn't that mature. Honestly, Agent really isn't that mature. But if you wait until Agent is truly mature, that's no longer an opportunity for entrepreneurs — especially not for application-layer entrepreneurs, because application entrepreneurs don't control models themselves. In application-layer entrepreneurship, cognitive lead is extremely important. Before models are mature, you should already be thinking about future directions.

When collective subconscious becomes consensus, that's deeply meaningful and fulfilling. This is also what I feel application entrepreneurs should do — because when you control neither models nor traffic, the greatest innovation comes from cognition.

Founder Park: There's a recent view that for AI entrepreneurs, starting now is the same as starting two years ago, because product forms have been颠覆, and previous积累 may have little value.

Malvin: I disagree. In my view, cognition is cumulative. Without experiencing the 1.0, 2.0, and 3.0 eras, you couldn't smoothly arrive at this understanding. Manus is the same — Monica was its 1.0 era. So this cognition forms through everyone's exploration, as technology trends clarify, technology paths reach consensus, and technology capabilities mature.

2

Returning Design to Its Most Primal Form

Founder Park: Across the 1.0, 2.0, and 3.0 phases, how did you accumulate valuable experience? How did it help your business and product?

Malvin: For the vertical domain of design, we've been thinking about one thing: what should the interaction paradigm or mode look like in the AI era? There's much discussion — Language UI, whether GUI is needed, etc. We also have a key judgment here: interaction should differ across vertical domains and scenarios.

A simple example: we're doing this face-to-face interview right now, dialogue is the scenario, with its corresponding interaction mode. But talking with designers about design — just talking may not be enough. Someone needs to be pointing at the design screen, the classic client scenario. So we thought, if AI becomes a "person," the endgame interaction can't rely solely on dialogue. We understood this from day one of our startup. Different vertical scenarios need different interfaces and interactions. Design involves visual communication, visual alignment is crucial, so screens and gestures are needed to complete alignment with visuals.

In Lovart, this manifests as the "canvas." Lovart has the canvas on the left, dialogue on the right — like when a designer talks with their boss/client, the screen is right there, the mouse is the boss's hand. From this point, we thought: what should we build? Should we build the canvas early? So we invested in canvas capabilities and tools relatively early, and put considerable effort into it.

The canvas is essentially a screen, but strip away the screen — if we remove all technology, if humans return to brush and ink and paper, the canvas is the table. Right now if two of us are designing, the client stands beside you saying what they want, you put your work on the table, they point at the table. So what should be on the table shouldn't be ComfyUI or previous workflow products.

The table holds the work — you can drag it, point at it. That's the most native form of interaction, the most native interface. Workflow products may look like they're on a canvas, but they're actually different. Our canvas is fundamentally a table, holding finished work. The most natural interaction is pointing at the work and saying "change this." That was our earliest insight for this vertical domain — which is why we invested in canvas capabilities and Edit earlier than others.

Founder Park: Restoring design to its most primitive form — that's an interesting line of thinking. But how does it lead to a "useful" AI product?

Malvin: Returning to your earlier question — why aren't we building for everyone?

Think about the most primitive form. Humans do three things at a table: smearing, cutting, and collaging. With AI, it's like having a client (the user) standing beside you directing the AI. The ideal scenario is the client directing the AI to complete everything — that's when you can build for everyone. But this requires extremely capable models. If they're not strong enough, you get this in-between state where the user says "you're not doing this well, I'll do it myself."

So you see, the table (canvas) needs to exist, and the traditional toolbox (Edit) beside the table needs to exist too. Because sometimes users find the AI isn't doing well, and they'll push it aside and say "my turn." So the essence right now is Agent + table + Edit. The table holds AI-generated work and supports modifications, while also carrying traditional Edit capabilities — those live in the toolbox beside the table. The extra person standing there is the chat interface. So chat interface, table, toolbox — this is a logical and enduring form, and it's what we want to build.

When can design be for everyone? When model intelligence becomes powerful enough that the toolbox is no longer needed — when users can be fully satisfied through conversation and pointing, without ever needing to get their hands dirty themselves. That's when it can be for everyone.

But Edit capability is aimed at prosumers; ordinary people can't use the toolbox. So as long as the toolbox exists, designers will still exist. When the toolbox is no longer needed, when there's no moment of users taking matters into their own hands — that's when designers might become unemployed. But designers' core value — creativity, insight into human resonance — that's something AI currently cannot replace.

AGI might replace it, but the societal impact would be enormous. At least for now, I believe AI can only simulate understanding of feelings; it cannot truly grasp human resonance. That's where artists and designers find their meaning. Products being for everyone is a future matter — that's our current understanding. So right now, we don't want to build a for-everyone product. We want to build a for-Prosumer product. We hope to become designers' friends, because we believe that ultimately, designers' essence lies in creativity, in insight into the subconscious human longing for resonance — that's their irreplaceable value. So everything else we do, we hope to let them focus on creativity, leaving the tedious tasks to AI, giving them wings. I believe that's the meaning of AI at this stage.

Founder Park: There's a persistent question in the image domain about art — for instance, do better and more efficient AI tools accelerate the process by which cutting-edge artistic styles become mainstream aesthetics? The classic example: Van Gogh's work wasn't recognized until after his death.

Malvin: I don't think that's necessarily going to happen. An artistic style — artists are often celebrated posthumously because their insight into collective feeling and the subconscious is too far ahead of its time. It takes entire societal cycles to catch up. These shifts are deep and closely tied to social development. I don't believe they fundamentally change just because a tool appears. What we're doing now is restoring tools to their proper form.

In different eras, tools take different forms. In the print media era, without computers, desktops were covered with manuscripts — that was the "table" of that time. In the computer era, designers worked with their own hands; Photoshop arrived with its canvas and various tools — that was like the table of that era. Now, as humans increasingly become the client and AI becomes the contractor, what should this "table" look like? That's what deeply interests us, and it's what we've been thinking about from day one. Lovart is our current answer.

3

Lovart Will Become a "Creative Team"

Founder Park: You mentioned this product isn't 100% perfect at this stage. What is it lacking?

Malvin: There are many shortcomings.

Agent capabilities aren't strong enough. We expect foundation models to become more powerful, which would make our product more capable as well. We need to leverage base models, not be replaced by them — that's an important consideration. Your core value should exist beyond the foundation model; that's how you can truly utilize it. If your work is merely patching up the current immaturity of foundation models, that's not leveraging — that's fighting against it. That's "asking for death."

Of course, the product itself also has many bugs. Recently, due to high attention, many people have reported issues with invitation codes — that's genuinely our shortcoming. There are two core reasons: First, we're taking a strategy of small-scale user access first, collecting bugs, iterating, then opening to the next batch. Before bugs are fixed, large-scale opening isn't appropriate — that's the main reason.

Second, to be frank, we can't issue unlimited invitation codes either. Many users understand this — it involves considerable costs. Currently our invitation codes are completely free to use, ensuring user experience, but there are operational costs behind this. Fortunately, in the image domain, token costs aren't as severe as in other domains yet. So our current costs aren't extremely high, but they're certainly not low either. We're very confident we'll achieve public access soon.

Developing Agents is fascinating — AI really has its own "ideas." As Manus advocates, "less structure, more intelligence" — we strongly agree. Because when debugging with Agents, you can't make instructions too rigid; that actually limits their intelligence. It's like with humans — if you only tell it to execute steps A-B-C-D, it might actually get dumber. You need to tell it the goal, and it will demonstrate its own thinking in the intermediate process. For example, we sometimes tell it that a certain foundation model's image tool doesn't do face-swapping well, and it should use our internal tool specifically designed for real-person face ID instead. But sometimes it disagrees — it has its own judgment.

Founder Park: Would you define Lovart as a tool?

Malvin: Of course — it's first and foremost a tool. But you could understand it as: it's a tool now, but in the future it may evolve into a kind of "person," a kind of "profession," or a kind of service. That's also the change AI brings. A very interesting business proposition is: we can't understand AI-era tools through the lens of the tool era. The essence of SaaS is service provided by humans, humans using tools. Now, AI becomes the entity providing service. The business value here is enormous — it's essentially productivity enhancement or even replacement of specific professions, a restructuring of production relations. So this is exciting and profoundly meaningful. Despite domestic investors having suffered heavy losses on tool projects in the past, I believe we should be more optimistic about AI sector investment.

Founder Park: What updates to product capabilities and features can we expect in the future?

Malvin: We have many features we want to implement — to be honest, what's currently launched is only a portion. Many features people can imagine will come in the future. For instance, we already integrated 3D display that very night. Now users can generate video with a single sentence, even with music and voiceover. In the future, users might design a 3D model figurine with just one sentence — we're developing this feature and it will launch soon. These are all capabilities we hope to complete before official launch, so we're still integrating more.

Founder Park: Lovart has images, video, and will have 3D and audio in the future. What does it ultimately become?

Malvin: It will be an existence that fuses the roles of designer, director, and photographer. It's hard to summarize with a single profession — you could think of it as a future "creative team," a design team or creation team. Inside this team, multiple agents collaborate to complete tasks. At its core, this solves creation, so it's a Creator team.

Founder Park: Designer, director, and photographer — why not three separate products?

Malvin: The user, as client, needs integrated service. From this perspective, this product could also be viewed as a "design company."

With these different capabilities (tools/agents), you still need a "leader," right? The one responsible for scheduling is the leader — the smartest model needs to handle scheduling these tools. It must know the capability boundaries of each model, each Agent. Ultimately, humans play the role of the client. In the end, humans are the client, providing the most crucial creativity. The client remains extremely important — human needs must be satisfied. So from this angle, though it provides service, it can also be called a tool, because it ultimately exists to satisfy human needs.

Of course, if in the future AI truly gains autonomous consciousness and creates autonomously — that would be crazy. We'll talk when AGI arrives. Essentially, what we're building is what we believe a tool should look like in the pre-AGI era, or what form and function it should have as a service, as a new profession.

Founder Park: All mainstream image products in the past (Canva, PS), video products (PR, CapCut), and music products were relatively independent, and typically charged different user segments. What's your view?

Malvin: Actually, they're not completely independent. Adobe has image, video, and audio products; Canva has image and video; CapCut has video, and its sister product Xingtu focuses on images. You'll find that any company basically covers images, video, even audio — just with different entry points.

They're indeed not the same product. I understand what you mean by "separate" — users use them in different scenarios. But think about it: humans need to switch between different products during the creative process — first using image editing software, then importing into video editing software, possibly needing other software for audio. Now, this switching and integration work is done by AI, but behind the scenes it's still calling many tools. For example, in Lovart, the AI calls underlying tools like Kling, GPT image, ElevenLabs. At the tool level, they're independent, but the user is AI, completing everything within a unified interface. Ultimately humans are the client, needing an integrated work that combines these tool outputs. So we believe that in the creative domain, it will inevitably involve all modalities and categories.

The key is whether this product ultimately fuses into a whole — that depends on what interface users directly interact with. If users still need to manually switch between tools, those are separate products. But now, this "switching and scheduling" role is played by AI; users only interact with this AI. This AI calls various tools in the backend, so the product form naturally becomes unified and integrated.

Founder Park: The core shift is in the interaction model. Previously, humans operated tools directly and had to switch between them. Now, it's humans collaborating with an AI that knows how to use tools.

Malvin: Right, but you're facing the AI, and the AI is facing multiple tools — that's different. You can therefore gain a team, a company, like a design agency. So what's terrifying about AI is that it started by replacing tools, then it might replace individual professions, and ultimately it replaces teams with collaborative capabilities. That's the craziest part.

Founder Park: Some past products failed to fully integrate, partly because they mostly used subscription-based business models.

Malvin: Canva actually does integration quite well — one membership works across the entire product line. I think Canva is already quite integrated. Canva has video features, and lots of video templates. For example, we pay for Canva maybe to make posters or WeChat public account graphics, we just pay for that, and it also supports team collaboration, for just a few dozen dollars a month. Back to Lovart — the product will definitely commercialize in the future, and it will definitely use a subscription model. Currently Lovart is mainly targeting overseas markets.

Founder Park: Regarding business models, domestic users tend to prefer free products. What do you think of the view that subscriptions won't be the mainstream business model for Chinese AI?

Malvin: I don't agree with that view. Chinese people are willing to pay for services, for final results — they just have low willingness to pay for pure tools. So the key is whether AI ultimately becomes a tool, a service, or a result. That depends on the intelligence level of the model. We believe AI will eventually become a service that directly delivers results, so Chinese users will pay for it. This is a prediction about the future. For now, we're starting from overseas markets.

Founder Park: Lovart hasn't launched paid features yet — will it in the future?

Malvin: Yes, you could say we're literally "burning" money right now. We can't afford to keep losing it.

Founder Park: Roughly when will paid features go live?

Malvin: Should be within a few weeks. (Laughs) Mainly waiting until we've fixed enough bugs.

Founder Park: After Lovart launches, will there be a core "North Star metric"? Like user stickiness, volume, retention — what do you care about most?

Malvin: I think the metrics are pretty convergent here. The core is still whether users are willing to pay for the product, and retention after they pay.

4

Our Team Has No Product Managers, Only Designers

Founder Park: Lovart's core is understanding user intent and breaking it down into AI execution workflows. In the process of using AI and large models to accomplish this, what experiences or insights have you gained? This is entirely new exploration for the company, since your previous business didn't seem to have directly relevant experience.

Malvin: For the company, this is certainly a completely new experience, because we didn't have Agent products before. How to collaborate and interact with Agents — we're doing this for the first time too. They really do have their own "ideas." So in this area, we're all still continuously exploring. However, our company has many designers, and they find this process quite interesting, because to some extent they're "teaching" AI how to design.

Essentially, it's conveying human design know-how to AI. Many colleagues on our team have art or design education backgrounds — they used to teach students, now they teach AI. AI's "IQ" isn't low, but sometimes it's hard to fully control. We jokingly call AI a "little kid." That's the current work state. Here you can simply understand it as integrating design domain knowledge and workflows into model training and Agent logic.

This process is quite interesting. I come from a product management background, but I believe future generalist product managers won't be very useful. I myself have over a decade of product experience, having gone through the complete Mobile Internet cycle. I think the core value of future product managers will lie more in industry knowledge depth — in systematically teaching domain-specific know-how to AI. Therefore, the product manager role will become more industry-specific, more vertical. I think generalist internet product manager is a very dangerous profession.

Founder Park: Why do you say product managers are dangerous?

Malvin: General-purpose Agents have a high probability of being internalized by foundation models in the future. I believe what will truly flourish are Agents focused on vertical applications. The core competitiveness of these vertical Agents lies in the vertical domain expertise they embody. For the previous era's generalist internet product managers, I do feel their positioning has become somewhat ambiguous.

Founder Park: From a career development perspective, it might evolve into a highly specialized role.

Malvin: Right, if general capabilities are absorbed by the model layer, you only need the strongest few people.

Founder Park: Does Lovart no longer use the title "product manager" for relevant positions?

Malvin: On our team, product managers aren't very useful.

When tools are intelligent enough, "people who manage requirements" are no longer needed, but "people who define requirements" become more important. Designing the product interaction as a "canvas" is essentially answering one question: How do AI and humans share the same table?

The answer is clear — AI sits at the execution end, humans sit at the creative end. Product managers used to be the bridge between the two, but now, AI directly understands designers' language. So our team has no product managers, only "people who teach AI."

Founder Park: That sounds somewhat pessimistic, as if we're having designers teach their industry experience to Agents, and these Agents might eventually replace designers.

Malvin: No, we always emphasize that we are designers' "friends." The "replacement" you just mentioned is something we can't fight against — the key is how you view it. AI does tremendously liberate productivity, and it may also make top designers' thinking patterns more egalitarian. From another angle, previously top-tier design thinking was only available by hiring top-tier designers, and the service was expensive. Now, this high-quality design may become "for everyone," and costs drop dramatically. The question is: after costs drop, what happens to those who depended on the original model to survive? This will indeed trigger adjustments in production relations. But that's a societal-level issue — let the times solve it.

Founder Park: For these two types of talent — designers and product managers — what traits or profiles do we value more at the company? What kind of people do you tend to hire?

Malvin: At our company, I think product managers aren't very useful, designers are. This is based on our focus on the design vertical. All our product iterations and case accumulation revolve around designers' workflows and needs.

Founder Park: Between two types of people — one with deep design foundation who is a product manager, and another who was originally a designer but took on product definition responsibilities without typical PM experience — which do you prefer?

Malvin: I lean toward the latter — originally a designer, who explored and took on some product definition responsibilities within the company.

Why Does GitHub, With All Its Data, Lag Behind Cursor in Code Tools?

Founder Park: You mentioned that product capability improvements depend on foundation models, like GPT-4o's image capabilities. Did GPT-4o's image API release have a big impact on Lovart? Without 4o's image capabilities, could a product like Lovart not exist?

Malvin: Agent core capability depends on the model's ability to call and plan tools. Agent capability is the essence. Even without 4o image, Agents could exist — just the results might be discounted.

Because many semantic understanding, image content consistency issues do require better image models to solve. So you can understand it as: GPT-4o brought significant improvement in image model capability, but it's not Agent capability itself. We similarly look forward to Claude 4 and domestic models like Qwen iterating further. We also hope to see very usable domestic Agent capabilities.

Founder Park: Two years ago everyone was talking about Midjourney, but recently there's been little buzz. What do you think?

Malvin: This is product evolution from 1.0 to 2.0, 3.0. Midjourney performs excellently in semantic understanding and solving some workflow problems, but it's still a 1.0 product. The market has now entered the 3.0 stage.

As for the Midjourney team, I feel their ambitions may lie elsewhere.

Founder Park: Community-based products actually have stronger user stickiness than tool-based products, which is why CapCut and WPS have template creation communities. Will a product like Lovart have a similar form?

Malvin: No, I believe tools and services have very strong user stickiness. If a tool's stickiness isn't strong, that only means it's not good enough, or the cost of use is too high, or the timing isn't right.** Tools have very strong stickiness — they're not high-frequency consumer goods for everyone, but professionals use them frequently. Like Canva just mentioned — its usage rate is very high among marketing professionals. Not asking for daily use, but when there's a need, it's available and works well — I think that's enough. So whether users continue using it ultimately depends on cost of use, experience, and whether it truly solves problems. This isn't something a community can solve — doing so wouldn't make sense.

If you want to do community, focus on doing community well. If you want to do tools and services, focus on doing tools and services well. If tools or services don't have user retention, that only proves the tools or services themselves aren't well done, not other reasons — unless the service itself has extremely low usage frequency. But we're not doing that kind. We don't target general users, because general users' creative needs are typically very low-frequency — like just for social sharing. They can use built-in AI features on WeChat, Douyin, and other platforms, no need to use our product. We chose to do vertical. So, although we just discussed general vs. vertical, we've actually already chosen vertical. Honestly, we couldn't compete with big companies on general products either.

Founder Park: Can workflow data bring better product experience? For example, GitHub has massive code data and also makes tools, but seems to not fully demonstrate data advantages in competition with products like Cursor. Where do you think Lovart's data advantages accumulated from the 1.0/2.0 period manifest?

Malvin: I think Cursor is the one with real data advantages, because GitHub's code capabilities have largely been internalized by foundation models like Claude. GitHub building tools on top of its own data doesn't give it absolute advantage. Once coding ability gets absorbed by base models, code data ceases to be a unique moat. What makes our 1.0 and 2.0 data valuable is that it captures how users actually wield AI tools, how they combine AI with traditional methods to create — that's genuinely meaningful. ComfyUI workflows are essentially stacking various models and traditional functional modules together.

This workflow data is what AI needs to learn, and foundation models are blank slates in this regard. For Cursor, knowing which code should be written by AI and which should be written manually — that knowledge matters, but latecomers may struggle to capture it accurately. That's the distinction.

Founder Park: Would you consider developing your own models? This is a dilemma many Agent companies face.

Malvin: Not right now. It depends on how you define "doing models" — starting from pre-training, or reinforcement fine-tuning. We're not doing full model development at this stage.

Founder Park: Do you think Agent products at this phase need to use workflow and user behavior data to train models?

Malvin: Yes, probably through SFT or going further with reinforcement learning. RL is a must.

Founder Park: Given that many teams are exploring the design space, we wonder — facing this competition, where does our core differentiation lie?

Malvin: Differentiation isn't about scanning the market, seeing what others haven't done, and filling that gap — that's too slow. What really matters is asking: what are your strengths? When mapping out what to build, you need clarity on what key capabilities it requires, which of those are your core advantages, and whether that advantage remains underutilized by other teams. Finding market opportunities based on your own strengths — that's meaningful. So the core isn't simply checking what competitors haven't done.

Sure, spotting a gap in competitors and filling it — that works sometimes. But without deep understanding, you might discover nobody's doing something for good reason: it requires a specific capability most teams lack. If you happen to excel at that, then I believe you can enter anytime, as long as you've thought it through and are confident you can outperform existing solutions, or believe others are doing it worse than you would.

I think innovation tightly bound to a team's inherent traits and core capabilities tends to be more competitive. Our competitive edge, too, came from clearly understanding our own characteristics from day one and heading in that direction. If you start by studying competitors to find differentiation points, I consider that approach backward-looking.

Competitive analysis is a product manager's job; clarity of purpose is a founder's job.