A Guide for AI Product Managers: Who I Am, Where I Come From, Where I'm Going — A Conversation with ByteDance AI Product Lead Vanessa: Lessons from Interviewing 100 AI Product Managers

Not long ago, Crossing invited our good friend Vanessa, who leads AI product at ByteDance, onto our podcast. We talked about how she interviews AI product managers, and what new insights she's gained after conducting over 100 interviews.

AI is transforming every industry, and perhaps the first group to feel its impact are those closest to it — product managers.

Vanessa leads AI product work at ByteDance. Starting from her reflections on "interviewing 100 AI product managers," we discussed what exactly an "AI product manager" is. Vanessa, a seasoned PM and now AI product lead, also shared her view of the day-to-day reality and darkest moments of AI product management.

We also asked her to share stories of standout candidates from her interviews, along with tips on resumes and interviews from her perspective as an interviewer. Finally, Vanessa's own career transition may offer lessons for many: her background was in advertising, then she did a master's at CMU, became a designer, then a PM, then an AI PM. She says: switching isn't hard — in fact, diverse experiences become assets.

Beyond these takeaways and stories, we also hope to pass along some of Vanessa's personal philosophy from the podcast:

Follow your interests. Love life first, then build products.

Whether or not you're a product manager, staying sharp and open-minded toward new things is the fundamental engine of career growth. We hope this episode offers something for everyone.

🟢 This interview first appeared on the Crossing podcast. You can read this text version, or listen on WeChat or Xiaoyuzhou.

Reflections from Interviewing 100+ AI Product Managers

The Typical Candidate Profile

🚥 Koji

You've interviewed over 100 AI product managers at ByteDance. We're curious — what have you noticed and felt during this process? Compared to traditional PMs, what do you see as the key differences?

👩🏻 Vanessa

I also transitioned from user-facing product work. Previously, I'd typically assess whether you were data-sensitive, your ability to design user interactions, your communication and execution skills with other teams, and so on. For AI PM interviews, there's an additional hurdle — on top of user PM skills, you need deeper understanding of AI technology to enter this field.

Typical AI PMs come from a few backgrounds:

One is traditional PMs transitioning over, which is very common. Another is machine learning engineers moving to PM, though there aren't that many of these. A third is computer science or HCI students who were already interested in AI, accumulated some experience, and entered directly as AI PMs.

🚥 Koji

It sounds like the biggest difference for AI PMs is whether they've grasped generative AI technology. Beyond that, in terms of identifying user needs and understanding market competition, do you think today's AI PMs are different?

Given that AI isn't a settled technology — past PMs also needed some technical knowledge — the biggest difference now is that the technical boundary itself is fluid, which brings different competency requirements. Have you noticed these differences when reviewing AI PM resumes?

👩🏻 Vanessa

Let me start with what kind of resumes pass the AI PM screening.

One type is candidates whose past experience has nothing to do with AI, and they've submitted a generic resume they used for other major companies without tailoring it for AI. When I see that, I likely conclude they're not a strong industry or experience match, and they won't get far. Another type is candidates with some AI experience, plus generalizable skills that transfer to AI. For example, someone from industry analysis or strategy who previously covered non-AI products but could deeply analyze product thinking, market positioning, and business models — these insights transfer completely to AI products, and the barrier isn't that high.

So if I'm hiring a PM just to provide insights on AI industry or strategy, deep AI knowledge or skills aren't that critical. If the need is urgent and I need someone to ramp up quickly, having tangentially relevant experience is a plus.

🚥 Koji

So could we understand it as: PMs who excelled in the past all have a shot at becoming good AI PMs? They just need to learn some knowledge — and this knowledge is factual, not something more complex.

👩🏻 Vanessa

Yes. I think the reason it currently leans factual is that the industry is changing so fast, it's new for everyone. So no one can yet understand things at a deeper, more philosophical level — most people are still absorbing the new, largely factual knowledge the industry is producing.

This is actually the advantage of entering AI now — the barrier isn't as high as people imagine. Once you've accumulated more knowledge, you already have an edge. You don't need to reach a level of distinct superiority in user psychology insight or interaction design like before. The true giants emerging now are still few, so newcomers still have major opportunities.

Reading Papers Isn't the Only Path

🚥 Koji

Previously, another AI PM, hidecloud, joined our podcast. He encouraged all AI PMs to read papers daily to keep up with the latest AI developments. But many listeners felt stressed out after hearing that.

👩🏻 Vanessa

Haha, it felt pretty discouraging ~

🚥 Koji

So what's your take — in 2024, is reading papers daily, like hidecloud does, the only path to becoming an excellent AI PM?

👩🏻 Vanessa

Reading papers daily is a very hardcore approach, but it depends on your stage. If you're a student, especially in CS, it's hugely beneficial.

One question in our outline was "an AI product that recently impressed you" — and I actually couldn't think of a particularly stunning product. Instead, many paper ideas struck me as genius. So if you're a CS student, reading papers is absolutely a major plus. It expands your thinking and gives you deeper understanding of AI, rather than staying superficial and just looking at what products others have built.

AI has gone through multiple stages from technology layer to product layer. AI products are the final output. It's like making bread — only tasting good bread doesn't help you understand how flour became bread step by step. So understanding AI's technical principles is especially important.

But if you already have some foundational knowledge or experience in visual algorithms, you're not limited to grinding through papers. A PM who doesn't read papers can still become one who better understands users, markets, and application scenarios. In large companies with clear division of labor, professional algorithm engineers get earlier access to cutting-edge papers — it's not necessary for PMs to out-algorithm the algorithms. For PMs, reading papers isn't a must-have, it's a nice-to-have. Don't stress too much about it.

A Product Manager's Guide to the Post-AI Era

Four Types of AI Product Managers

🚥 Koji

When we talk about PMs, we typically categorize them as to B or to C, across different application scenarios like text, image, video, and productivity. So for AI PMs, what are the categories?

👩🏻 Vanessa

Since generative AI emerged — including visual and language large models; I'm using this loosely to distinguish from traditional AI, not as a rigorous concept — AI PMs have been categorized by their role in the AI value chain.

One type specializes in data production. The essence of large models is algorithms, compute, and data. There's an overseas product called Scale AI whose team excels at data cleaning and categorization, converting unstructured data into algorithm-usable data. These PMs form the foundation of the entire large model stack.

The second type works on large model strategy and evaluation. A large model may have different versions, and training on different data yields different results. The strategy PM's role is evaluating training outcomes and defining optimization strategies. For example, after multiple iterations, a language model gets a text test set with various questions to assess its capabilities — including even bizarre questions to test its limits, like posts from the "Dumb Questions Bar" forum: "Why is sashimi called dead fish slices?" or "How is three and a half hours one and a half hours?"

The remaining two types are closer to our first impression of "AI PMs."

The third type is PMs like me in AI-native products, especially those focused on effects. There aren't many of these PMs yet, since few companies are investing in pure AI products. They need to figure out how PMs should work and what workflows look like in the generative AI era.

The last category is product managers responsible for designing AI features within existing products, and they're also the easiest to transition into. For example, some friends work at Google, handling Google Calendar and Gmail. These products have AI assistants — your calendar acts like a little secretary, summarizing who you're meeting today, and if there's a video recording after a meeting, it can jot down follow-up items for you. These PMs are essentially adding AI elements on top of existing products, gradually moving closer to the AI track.

🚥 Koji

When you mention the fourth type of PM, you're mostly talking about to-C, but I think the same applies to to-B. Recently, looking at the 260 AI companies Y Combinator has invested in over the past year, we found that roughly 70-80% of them are to-B in the United States. Any to-B sector that previously had SaaS can now be converted to AI SaaS. For instance, in fields like veterinary medicine and dentistry, there were companies doing CRM well in the past, and now there are companies using AI to rebuild CRM from scratch.

I think this is another PM profile — using AI to improve work efficiency and empower existing products.

🚥 Ronghui

Do these people coming in for interviews have a clear sense of which category of PM they're applying for?

👩🏻 Vanessa

Probably not. Because the AI industry is so new, and during hiring there may be confidentiality considerations, job descriptions are often written very vaguely without clearly describing the day-to-day specifics. So many job seekers don't have a clear self-positioning either.

I've encountered candidates in interviews who, when they say "I want to be an AI PM," don't actually understand the challenges they'll face. Their understanding is limited to having used ChatGPT or Miaoya in daily life, or having read related news, thinking these things are incredibly cool.

Just as many people previously taught how to transition into product management, the current trend is teaching how to transition into AI product management. The AI PM title is becoming increasingly prestigious, so many people want to flood into this field. I think what they're saying isn't realistic enough, or they're deliberately hiding a lot of things.

The Darkest Hours and Sense of Value for Product Managers

🚥 Koji

What's the reality?

👩🏻 Vanessa

When teaching people to transition, sometimes these instructors only talk about the advantages, earning traffic and money without any after-sales support.

There's also "only talking about the 'why' without the 'how'" — only talking about how important AI is, without explaining how you actually become an AI PM, what the day-to-day looks like specifically.

Some people think everyone's starting point is roughly the same; actually, the paths for current students or people with zero experience, people with some AI product experience, and people with extensive non-AI product experience should be quite different when entering the AI PM field.

So what exactly are AI PMs doing behind the glamorous title? Here are a few "darkest hours" for AI PMs —

One is that AI PMs are constantly pressed with the question: "Where exactly is my value?"

From the outside, AI industry technical capabilities are already very strong, and AI PMs may not be able to add much on top of that; actually, quite the opposite — everyone assumes current technical capabilities are strong, but in practice they're full of holes.

🚥 Ronghui

This is an outside assumption, right?

👩🏻 Vanessa

I think it's a misunderstanding. AI still has major capability limitations, and these limitations come more from the models themselves than from product strategy. The entire AI industry is technology-driven. For example, when Moonshot AI's large model unlocked very long context limits, it dramatically expanded long-form application scenarios — you could throw a legal or medical textbook at it and immediately unlock more use cases. Therefore, what an AI product can do is heavily dependent on the capabilities of the underlying model.

Sometimes people question, "As an AI PM, am I just doing ornamental work?" This is a self-worth question that constantly troubles AI PMs.

I think PMs still have very high value. Models are evolving rapidly, and we expect them to become increasingly powerful, but there's a long road before that expectation is realized. We need to layer on large amounts of strategy, the so-called ornamental work, but this isn't entirely about compensating for technical shortcomings — it's about deeply understanding what users actually want, what their scenarios are, whether their interactions are smooth. The more mature the model, the simpler it is to layer on strategy. However, current models are far from mature and still require large numbers of PMs to pile on strategy.

Another major challenge is finding Product-Market Fit.

I saw an example of an overseas product for AI-generated comics. Products like Midjourney can already generate all kinds of stunning images, but applied to comics and picture books, it's still missing many things — like maintaining character consistency, fixed scenes, helping creators expand from one idea into a complete story outline or even detailed scripts. These still require extensive manual processing. That product targeted the comic creation scenario and built many useful features.

AI PMs need to deeply understand the user flow — what steps users go through in doing this thing, what their needs are — and string together all AI capabilities.

Another dark hour is that AI PMs wake up every morning to a constant stream of market news flooding their feeds, and when you stay up late, some conference is happening across the ocean yet again, with yet another new thing.

We're in a state of being disrupted at any moment every day. The ornamental work you carve today may become unusable tomorrow when the model improves.

👩🏻 Vanessa

There's another point: most people think AI PM work is very advanced, tweaking models every day, without understanding how much dirty work this job involves. I once posted on Moments, quoting an internet meme:

"Between city walk and dirty talk, chose dirty work."

Product managers also do a lot of very basic stuff, the so-called dirty and tedious work.

🚥 Koji

Can you give some examples?

👩🏻 Vanessa

I work with visual models. Currently there are two different UI interfaces for visual models, one called Web UI and another called Comfy UI. Both can interact with Stable Diffusion, guiding it to exercise its model capabilities and output images. However, there are many model options within these UI interfaces, and each model has many parameters — for example, each parameter can be adjusted from 0.1 to 0.99... This parameter-tuning work is extremely tedious. Even with automation tools to assist, it still consumes massive amounts of human effort.

🚥 Koji

I'm also curious — having one good image doesn't prove my parameters are correct; you probably need many images of different styles and characteristics to prove that. With so many models, plugins, and parameters to choose from, doing this well sounds like a pipe dream.

👩🏻 Vanessa

Someone once said that tuning large models is like alchemy, with a certain mystical quality. For example, you have two parameters that may interact with each other. When you raise one parameter and lower another, its effect on the overall image may be very subtle. Beyond understanding models and parameter effects, you also need to constantly explore and develop a feel for it — for instance, when you notice the image looks a bit greasy, which parameter should you adjust or what node should you add? These are things that require accumulated experience.

👩🏻 Vanessa

There's other dirty work too. Like I mentioned, getting one good image doesn't count for much — you need relatively stable results across hundreds or thousands of images in bulk. Batch image generation is very basic work without particularly high technical barriers, but it's what AI PMs need to do every day.

🚥 Koji

Current automation tools aren't very good either, right?

👩🏻 Vanessa

We're developing our own, but there's currently no universal automation tool on the market, though I believe one will emerge. Just as there initially wasn't Airtable or multidimensional tables to help PMs organize requirements and ideas, these tools appeared over time. We need to give these tools time to be built.

👩🏻 Vanessa

Another dark hour is that most users don't understand AI — they still use the mental model from interacting with non-AI products to reason about what should happen when interacting with AI products. However, previous products weren't probabilistic models — they were step-by-step executions set by programs. For example, when you use a coffee machine, if you order a latte, it will definitely give you a latte. But generative AI is a probabilistic model; there's a certain probability it won't give you a latte, even if that probability is as low as one in a thousand. Users can't understand this — they think AI should understand and respond to their needs 100% of the time, when in reality AI can't do this.

Users also don't understand that AI itself has capability ceilings. Even the most advanced Midjourney or Stable Diffusion XL/3 versions can still produce unnatural limbs, inaccurate hands, blurriness, and other issues. There's a prompt called "yoga girl" — if you input this into all image generation tools, the probability of bad images is quite high because yoga has so many poses, making this prompt particularly challenging for large models. Users can't understand this — as soon as they spot an extra finger or something off with the eyes, they think the product is bad. In reality, this is always a probabilistic model; we can strive for 99.99% accuracy, but 100% accuracy is impossible.

🚥 Koji

Do you have any ways to address this? For example, having a preview or double-check step before image generation, so problems can be fixed in time. Would that work?

👩🏻 Vanessa

That's another form of ornamental work. But scenarios can't all be enumerated — if you add a check step to all image generation, the generation process won't seem as smooth and simple. This is also a trade-off to face.

🚥 Koji

So you probably receive a lot of user feedback every day about being unsatisfied with generated images, but sometimes there's nothing you can do, because this is just the hallucination or limitation of large models.

👩🏻 Vanessa

We can work through various approaches to get closer to the model's ceiling, but breaking through the ceiling requires the model itself to iterate, or using a better model to replace the current one. However, even the most advanced models on the market (SOTA) will inevitably have some flaws.

🚥 Koji

I really understand this kind of dark hour — even when you receive a bug, there's no solution.

👩🏻 Vanessa

This also relates to how users interact with bots. In the past, GUIs (traditional graphical interfaces) presented users with four buttons, making it crystal clear they could only do those four things. Now users get a chat box. They don't know what's possible, and might throw out bizarre requests like "do my laundry" — completely unfulfillable prompts. This happens because AI is still so new; the market's understanding and education around AI hasn't caught up yet.

🚥 Koji

Any more dark hours?

👩🏻 Vanessa

Not for AI product managers, but for AI itself — I think we're being too harsh on it. If we treat AI as a person, what would "his" dark hour be? We're not being fair in our judgment of him.

Take autonomous driving. Even people who've passed driving school make mistakes. Human driving skills vary widely. AI might actually be safer than some drivers, yet we hold AI to an extremely strict standard, allowing no errors whatsoever. Or consider medical diagnosis, where AI still serves as Copilot — it can assist doctors but not replace them entirely. Even if AI's misdiagnosis rate is lower than doctors', we still feel uneasy entrusting decisions to it.

I think this might stem from human responsibility — when we delegate life-or-death matters to machines, we feel compelled to be demanding. Or perhaps from human superiority — the belief that machines and humans shouldn't be conflated. That is AI's dark hour. There's an entire field of tech ethics devoted to discussing these issues, and it can get very serious and academic; I'm just scratching the surface here.

The Fulfillment of Product Managers: Worth Getting In

🚥 Ronghui

I want to ask about fulfillment. You mentioned so many people applying for AI product manager roles — it seems many are bullish on this field. This reminds me of something Facebook COO Sheryl Sandberg once said:

"If you're offered a seat on a rocket ship, don't ask what seat. Just get on."

👩🏻 Vanessa

Yes, from my own judgment, becoming an AI product manager now is still worth it. There have been other emerging technologies that everyone rushed into, which in retrospect look somewhat bubbly.

🚥 Koji

In this wave, what makes you feel it's worth trying?

👩🏻 Vanessa

First, the technology is no longer an urban legend or armchair speculation. By using ChatGPT and other products, you can genuinely feel this technology is real and capable of changing your daily life. Plus, prices are gradually coming down. More and more people will benefit from AI's impact, and the day AI truly enters every household isn't far off.

I've always believed in a saying:

"The future is already here — it's just not evenly distributed."

For those who frequently interact with AI, especially people at internet companies or in related professions, they have earlier access to understanding and applying AI. So they know AI is no longer a bubble, and they're more likely to have the opportunity to jump on this rocket ship — why not give it a shot?

---

Standout Stories from Interviewing 100 AI Product Managers

Stories That Dazzled and Disappointed

🚥 Koji

Among the 100+ AI product managers you've interviewed, have you met any candidates who truly impressed you, whom you wanted to offer on the spot?

👩🏻 Vanessa

I think such people exist, but it depends on fit with the open role. Some candidates who wowed me had excellent educational backgrounds and were very smart, but needed significant landing time — not quite matching the immediate needs of the position, so we had to pass with regret.

In terms of hiring expectations, we hope candidates have sufficient knowledge and understanding of AI technology, even relevant work experience, so they can ramp up quickly. They should have their own perspective and understanding of AI, and original ideas about which scenarios AI can be applied to. Those are the people we consider truly outstanding.

🚥 Koji

Any memorable interview stories? Especially about strong candidates.

👩🏻 Vanessa

I recall one interviewee. He worked at a major internet company, where his day job had little to do with AI. In his personal life, he was a photographer with high standards for aesthetics, photography, and imagery. When he saw AI, he started having ideas: Could he skip location shoots? Could AI transform a photographer's workflow? So he started from a real problem, gathered some people, and built a side project.

I felt this demonstrated his self-learning ability and passion for AI. Under the intense pace of a big tech company, he still managed to complete an AI product project in his spare time, driven by passion. Such people are worth paying attention to and giving more opportunities.

🚥 Koji

Any stories that made your jaw drop — in a bad way?

👩🏻 Vanessa

There were candidates with very impressive backgrounds whose resumes mentioned AI-related content, and we expected them to hit the ground running. Yet when we asked very basic questions in the interview — like one of my favorites, "Have you seen any AI-related news recently that you found interesting?" or "What AI products have you used or heard of recently that you think are well done or poorly done, that you want to complain about?" — they'd think for a long time and come up empty.

🚥 Koji

You'd think they'd anticipate questions like that in an interview.

👩🏻 Vanessa

Yes, you could prepare a little.

🚥 Koji

Beyond what you mentioned, are there other questions you always ask in interviews?

👩🏻 Vanessa

The questions above test whether you follow the AI field, keep up with the latest market developments, and whether you genuinely have passion and interest for AI.

I also ask more fundamental questions. For practitioners who've worked with graphics and images, I might ask, "How did you use these algorithms in your projects? What difficulties did you encounter and how did you ultimately solve them?"

If candidates lack relevant experience, I might ask, "For industry-standard model architectures like Stable Diffusion, what's the underlying principle? Can you explain your understanding in non-technical, accessible terms?"

These questions assess whether they have basic cognition of the AI field. If you want to interview for an AI product manager role but lack this foundational understanding, I'd interpret that as insufficient preparation, insufficient motivation for the job, not enough commitment.

Who Finds It Easier to Transition to AI

🚥 Koji

Among the people you've interviewed, there are both fresh graduates and experienced professionals. In your view, who finds it easier to transition to AI now?

👩🏻 Vanessa

I think it depends on your hiring needs. Fresh graduates don't have much experience making trade-offs, but if you don't require deep product accumulation and instead need someone to ramp up quickly and bring fresh eyes to do something original, fresh graduates actually have greater advantages. People with extensive experience in non-AI fields may develop mental inertia, doing things along established patterns.

AI is a field where you need to learn new things every day. If you have many years of rich experience, maintaining this beginner's mindset is actually a challenge. People who are both experienced and capable of keeping a beginner's mindset are extremely sought after.

Vision and Language Models Don't Transfer

👩🏻 Vanessa

One more thing: HR might push up any resume with AI or large model keywords. We work on vision-related products, and we've seen people with LLM large language model fine-tuning experience, even those who've done LLM projects. Do you think these people have an advantage when applying for vision large model roles?

🚥 Koji

I think there's some advantage, but it's hard to judge how much.

👩🏻 Vanessa

Actually the advantage isn't that significant — or rather, smaller than people imagine. Many assume large models are basically either vision large models or language large models, and that capabilities should transfer between them. But in terms of technical essence, they're different.

Vision diffusion models, unlike language large models, aren't intelligence models or AGI models. The underlying architectures they correspond to — diffusion and transformer — are fundamentally different. At a higher level, they can inspire each other, but at the practical level, the help may not be as great as imagined. (Update: This was recorded relatively early, using SD models as representative of vision models; this doesn't mean there's no trend toward fusion — for example, DiT models use transformer networks.)

🚥 Koji

For product managers working on vision large models versus language large models, are their boundaries with technology very blurry? Or have clear divisions of labor also formed?

👩🏻 Vanessa

This depends on how AI-native the company is.

I previously read an article about Perplexity AI. Their talent DNA is that everyone can do everything — with AI assistance, one person handles operations, product, and technology, so division of labor becomes very blurry. I think this is partly because each person's skill set can expand with AI assistance, and partly because limited manpower forces expansion.

Conversely, having clear division of labor where everyone focuses on their strengths isn't an outdated way of working. I think both approaches have their merits.

Who Stands Out in Resumes and Interviews

🚥 Koji

In your view, after interviewing so many product managers, who stands out more easily?

👩🏻 Vanessa

This breaks into two stages. First you have to pass the resume screen to get into the interview, and only then can you make the interviewer want you.

First, regarding the resume screen, we look at candidates' fundamental qualities and the relevance of their past experience to AI. If a resume doesn't mention AI at all, unless the candidate is exceptionally outstanding, they'll be tagged as "industry mismatch" or "experience mismatch."

If your resume isn't closely tied to AI, you haven't worked on a complete AI project, yet you want to apply for AI product manager roles, I suggest rewriting your resume to incorporate AI elements — you can even have ChatGPT or Moonshot AI do it. You can tell AI, "Assume you're a senior HR professional reviewing this resume; how would you rewrite it to make the work appear more AI-related?" Let the large model help you brainstorm directions for rewriting your resume.

Of course, I'm not saying you should make things up. The idea is to highlight the parts of your work that involved interacting with AI. For instance, if you've chatted with ChatGPT and used it for brainstorming, that's your intersection with AI. Going deeper, if AI has become a fixed step in your workflow and you've spent time tuning or refining it, those are elements you can add to your resume.

Even deeper, if your work experience touches on AI-related domains — whether today's AGI or earlier machine learning, recommendation algorithms, search algorithms, or even classical computer vision like face recognition or facial landmark detection — that knowledge is transferable. Including this experience signals to interviewers that you're more likely to successfully transition into AI.

The best scenario is that you got into AI early. For example, Xiaohongshu has many AI effect templates, and if you previously owned that module and deeply understood the technical principles, input/output formats, and model debugging methods, you'd be a rare find in the market — exactly the type of talent we're especially eager to recruit.

Second, the interview stage has different requirements depending on the role.

For intern interviews, the main focus is interest in AI, enthusiasm, and learning ability. Take the "new AI products you've been following" question I mentioned earlier — if a candidate can name some AI products and offer their own analysis of strengths and weaknesses, that's already great. Learning ability is hard to assess in a short interview. Say your current work has nothing to do with image generation, but you've deployed related models through cloud platforms or locally — that would be excellent proof of learning ability. I remember hidecloud published an article on how to train a GPT-SoVITS voice-cloning model, and I trained a similar one on my own machine. Don't set too many boundaries for yourself. Current technology doesn't require you to write code; it's mostly about problem-solving and prompt engineering. The self-learning cost isn't that high — it's worth trying.

👩🏻 Vanessa

For non-intern roles, I ask: "Do you understand the principles behind these models? For recent research, do you grasp its core ideas and what makes it stand out?"

For example, Xiaohongshu previously launched InstantID — with just one image and zero training, it could generate images while maintaining strong facial similarity. The breakthrough was in how it fixed the face: it extracted facial landmark keypoints and passed them to the model, essentially the "three courts and five eyes" principle in traditional Chinese facial aesthetics. Slight shifts in the middle court or individual positions are perceptible to the naked eye, so this approach better anchored facial information. If a candidate can articulate one or two points like this, their technical understanding is solid. Even if they can't lead an entire team, being able to quickly grasp technical ideas when interfacing with algorithm teams and push projects forward is crucial.

🚥 Koji

From what I'm hearing, the essence of being an AI product manager isn't fundamentally different from other product management directions — you still need to understand the technical principles behind the product.

Switching to Product Manager: Start with a 5% Change

🚥 Koji

Our "Crossing" was inspired by teacher Li Songwei's concept of "5% Change." He believes:

Asking others to make a 100% change is impossible. But a small 5% change — many people might take that tiny step after hearing it, get a small positive feedback, and slowly start rolling a snowball that grows into bigger feedback.

So I want to ask you: for someone who wants to switch to product management today, what's your "5% suggestion"?

👩🏻 Vanessa

I think you need to immerse yourself in the environment, whether that's cutting-edge AI papers or more tangible — the AI products and news emerging in the market. When you see an interesting product, you'll speculate about the principles behind it, and that sparks curiosity. It's like tasting delicious bread and wanting to explore how it's made, and the interesting secrets in its process. This exploration is spontaneous — it won't leave you feeling at a loss or intimidated by esoteric knowledge.

Starting from interest is crucial. First, know that something exists, which generates interest; then, work backward from that interest to what you can do.

I encountered Stable Diffusion quite early, tried a few img2img generations in the Web UI, then put it down. I reconnected with it in 2023 when I watched Creation of the Gods I and really liked the actor Yu Shi. On Xiaohongshu, I saw many AI-generated fan art posters. That was when I first realized: so this is what AI can do — it preserved the face shape, features, and expression so vividly, with such refined image quality. I started wondering: what steps would I need to replicate this?

🚥 Koji

So your earliest interest in AI image generation came from wanting to make Yu Shi fan art posters.

👩🏻 Vanessa

Yes. I wanted to create a stylized poster, imagining what he'd look like in a wuxia film or a historical romance drama. Starting from this idea, I began working backward: "How can I fix his face in the image?"

So I searched and found I'd need to train an AI LoRA. How do you train a LoRA? It takes N steps. First, collect a training dataset — what kind of images work best? Should include close-ups, medium shots, and so on. After collecting, what resolution to use? Along the way, a chain of questions emerges. Ultimately, to make good bread, you need to understand many things behind it. This 5% change is just about letting you first see that bread you love.

🚥 Koji

I still remember you once said something golden in a chat: "You have to love life before you can make products." You fell for Yu Shi first, and everything else followed.

👩🏻 Vanessa

If you stay in an ivory tower, doing vertical work in a closed space, you lose certain sensitivity to the outside world. You won't encounter these soft, fluffy breads, and they won't be able to influence what you're doing right now. So I think you must stay sharp to the outside world and love life.

Projects Should Start from Real User Needs

🚥 Koji

What you just said was excellent — what makes a candidate stand out, on one hand the resume, on the other the interview. Is there anything else you particularly care about in interviews?

👩🏻 Vanessa

Beyond some basic threshold requirements, another thing is long-term沉淀 and accumulation. I look for candidates with genuine AI project experience, whether from work products or side projects. What matters is that the project starts from a real pain point or need.

A good interviewer will definitely press: "Why did you want to do this project?" and "What problem is this project fundamentally solving?" These are soul-searching questions for product managers — every good product manager should be grilled on these two.

🚥 Koji

Have you encountered any memorable projects?

👩🏻 Vanessa

The AI photography workflow project I mentioned earlier is a good example combining hobby with actual product. Let me share a few more ideas I haven't had time to execute:

For instance, when scrolling Xiaohongshu, I see many people posting chat screenshots with their crush, asking netizens for advice on whether their crush likes them back. So I thought of building an LLM-based "crush analyzer" — collecting all such posts and comment interactions on Xiaohongshu, using OCR to extract text conversations, and building an LLM expert knowledge base on top of that, then designing an interaction flow. Say I'm pursuing someone: I could feed the conversation screenshot to the model, which extracts text from the image, combines it with the knowledge base, and helps analyze whether I have a shot — and if so, what to do next.

This is a concrete project born from observing real user scenarios, not some simple exercise from a training bootcamp.

Going through projects like these, you can demonstrate your ability to spot user pain points and uncover user scenarios. A spherical chicken in a vacuum is perfect, but it's a theoretical scenario that doesn't exist in reality — don't do this kind of "spherical chicken in a vacuum" project.

🚥 Koji

Find a real user need, then roll up your sleeves and build it yourself.

Personal Journey: Transition and Interest

From Designer to AI Product Manager: Transition Isn't Hard

🚥 Ronghui

You've gone from designer to product manager to AI product manager — a series of transitions. You've said transitioning isn't hard. Let's talk about your personal experience: what were these two key transition points like?

👩🏻 Vanessa

Actually, I originally studied advertising. From advertising to designer to product manager, I don't think the boundaries are that rigid — there's a lot of overlap. I've always preferred entertainment-heavy to-C products, so after graduation I consistently worked in to-C spaces like short video and image-related fields. Fundamentally, it wasn't a huge leap.

Also, things I learned earlier could transfer to later domains, giving me many advantages. For example, in advertising, besides image-related skills like Photoshop, I had to develop aesthetic sense for posters and videos, and even other more interesting content — skills that go beyond typical UI/UX designer territory.

For instance, advertising requires marketing thinking. The goal of advertising is to sell, so it emphasizes user orientation and extreme sensitivity to commercial opportunities. Also, advertising is fundamentally about capturing attention, so people in advertising have strong presentation skills.

These skills might be uncommon for UI/UX designers, but when I transitioned roles, I could build on my existing tools and aesthetic foundation while retaining the advantages from my advertising background — giving me more strengths than a typical UI/UX designer. I don't think career switching is difficult; on the contrary, it added some highlights for me.

🚥 Koji

Did you switch careers through further education?

👩🏻 Vanessa

I started in advertising, worked as a designer after graduation; later studied product design, but didn't pursue product design — I became a product manager instead. I don't believe you must do what you studied. Your major may relate to your work direction somewhat, but they don't need to fully align.

🚥 Ronghui

When did you start wanting to transition to AI product manager? What attempts did you make at the time that proved useful in retrospect?

👩🏻 Vanessa

I chose this industry entirely driven by interest. I'm not a particularly ambitious person — interest weighs very heavily in my value system.

Initially, I worked on content safety for Douyin and TikTok, in a review team. The team was very small in the early days, so I simultaneously handled review policy design, review product building, and review model training — giving me a comprehensive perspective that provided many insights.

One realization was that I wasn't cut out for content moderation — I found it boring. I like doing interesting things. Moderation demands rigor, not creativity. So I eventually left moderation and pivoted to creation tools.

Another insight came from the exposure to models I gained during my moderation days. The job required analyzing similar videos to deliver appropriate recommendations. We also used extensive visual models to automate judgments about policy violations — pornography detection, violence detection, and so on. Training these models gave me substantial knowledge about image-related technology.

With this foundation, I moved into creation tools. I started with products like Duet, which had little to do with algorithms. Then I began working on effects, where algorithms are inseparable from the product. My work became almost entirely algorithm-centric. Effects are also entertainment-driven consumer products — a perfect match for my interests. My personal passions and my work aligned, which made the job much more fulfilling.

In the early days, we used traditional algorithms for effects. One was image segmentation, designed to replace green screens by identifying and separating people from their backgrounds. We created an effect called Green Screen, which remains the most-used effect on TikTok. It's simple: it replaces whatever's behind you with any image you choose. Users deploy it to photoshop themselves into group photos, share chat screenshots, even explain concepts. I once saw a video where someone was explaining the differences between N species of octopus, with octopus photos cycling behind them like a PowerPoint presentation.

Later our team built the "Living Photos" and "Teenager GAN" effects, also deeply algorithmic. Living Photos animates still images — making figures blink, smile, and so on. The Teenager effect renders what someone looked like around age ten or so, trained on extensive datasets of youthful faces. All of this laid deep groundwork for our later work on visual foundation models like Stable Diffusion.

Around 2022, when DALL-E 2 emerged, everyone was stunned. We discovered that AI could generate, say, an astronaut riding a horse on the moon — something human illustrators might struggle to match. So we quickly pivoted in this direction.

Initially, someone on our team built an effect called "AI Dreamscape." The name came from the fact that our models weren't yet mature; what they generated didn't quite make sense, more like a hallucinatory dream. Users could input text and have the model paint the described scene.

As model capabilities improved, and as capabilities and papers like DreamBooth became available, we realized faces could be fixed while generating varied artistic renderings. We launched AI Portrait and related effects — you capture a selfie, and the model paints you in parallel universes: your courtly portrait, your watercolor incarnation. This was what I described as the fourth type of AI product manager — someone adding AI elements to existing products. Though I later jumped out to become the third type.

🚥 Ronghui

Is there anything you wish you had known earlier?

👩🏻 Vanessa

I think I knew what I needed to know at the time. I'm not someone who does much second-guessing or "if only" thinking. If I didn't know something then, it simply wasn't my moment to know it.

🚥 Ronghui

That echoes what we concluded from our conversation with hidecloud — follow your curiosity and interests, and let them guide you toward longer-term directions.

👩🏻 Vanessa

Many podcasts brand themselves as shattering knowledge gaps, promising to rapidly inject information into you. But whether you truly understand what you hear is another question. It's a shortcut, or perhaps just reveals that a path exists — but you still have to walk it yourself.

🚥 Koji

One more thought: the common thread between pre-AI and AI-era PMs is understanding user needs and pursuing extreme user experience — this fundamental capability never changes, and PMs across industries need it. The difference now is that with relatively mature technology, there's greater emphasis on demand discovery and competitive positioning.

👩🏻 Vanessa

You could say PMs used to be carving decorative patterns on wood; now they're carving decorative patterns on top of models.

🚥 Koji

Now you need to understand the model in order to carve decorative patterns. This raises the bar for PMs — technical understanding isn't a one-time acquisition but continuous learning, because change happens daily, occasionally in leaps. It sounds like being a PM is harder in the AI era.

👩🏻 Vanessa

You asked what makes a top-tier product manager. Actually, the hardest thing for a top-tier PM is prophesying what technological changes will occur and what levels they might elevate us to.

Doraemon in the AI Era

🚥 Koji

At Crossing we have a column called "Doraemon in the AI Era." Why Doraemon? Because we feel that pioneers in AI today all believe the future will be better — they're people like you who "see that the future has already arrived, it's just not evenly distributed." Suppose you were Doraemon, traveling back from the future to this exact moment in 2024. What would you most want to bring back with you? It could be a product, a service, or an idea.

👩🏻 Vanessa

My Doraemon probably wouldn't fly too far into the future. Distant futures aren't something we can prophesy now — let me speak of just a few years ahead.

When I watched TV as a child, I'd think: "Why can't I be the protagonist?" I'd even close my door and imagine how I'd act the scene if it were me. Nowadays, male and female leads in many dramas generate controversy. Some shows have even sparked face-swap trends on Bilibili and Xiaohongshu, where viewers feel a different actress would have been better. It reflects a user need: "I want my designated person to play the lead, whether it's my favorite actor or myself." AI can fulfill this — through more advanced, natural face-swapping, designating any actor as lead, even letting users become the protagonist themselves.

TV drama plots could also be freely AI-generated rather than fixed works where viewers remain passive spectators unable to participate in creation. We could become players in a TV-drama-style interactive game, placing ourselves or our favorite actors in the story, controlling their branching choices, or even rewriting the plot.

🚥 Ronghui

That's fascinating. I remember the Hollywood writers' strike not long ago, with concerns about AI taking their jobs.

🚥 Koji

Sam Altman also said:

The future of film is games; the future of games is unimaginable.

👩🏻 Vanessa

Yes, it might become highly interactive.

"Emergent" Learning

🚥 Koji

We just discussed a crucial capability for PMs in the AI context — learning ability, the capacity for sustained, rapid learning in this era of discontinuous progress. How has your learning approach changed over the past year or two?

👩🏻 Vanessa

My learning style is what I call "emergence," similar to AI "emergence." When you accumulate massive amounts of information in your mind, sometimes they connect with each other and form new ideas. For instance, you're stuck on a problem, then you read five books, ten WeChat articles, and eight podcast episodes, sleep on it, and suddenly you figure it out.

I call it "human brain emergence" — like the saying "read a hundred times and its meaning will naturally appear." When you've input enough relevant information, perhaps one day you'll naturally figure it out.

🚥 Koji

So don't expect immediate payoff or answers — wait for the "emergence." Not waiting idly, but consuming related and even unrelated material, and the answer may come naturally.