Heart Capital: After Generating 500 Videos with Sora, Here's What We See | Voice VOICE

The recent launch of Sora 2 has once again pushed AI video generation to a new technological inflection point. It's not merely an upgrade in parameter scale, but a major leap forward in coherence, physical plausibility, and creative controllability. At Heart Capital, our partner Bingjian Wu spent the holiday immersed in hands-on experience with Sora 2, personally generating nearly 500 videos. He admits that to some extent, he even developed a bit of a "generation addiction," gradually building up genuine intuition through the process.

Drawing from this practical experience, Heart Capital would like to share some of his observations: what truly noteworthy opportunities emerge from Sora's capability evolution, and what it actually means for entrepreneurs today.

First-Hand Experience with Sora

1/ The biggest change Sora brings is that AI video content now has consumable value. A nine-second video can contain a conflict or twist — this is the foundation of watchability.

**2/ **Previously, making video models produce consumable content was rather cumbersome: long, precise prompts, generate images first, then image-to-video, then editing. Sora essentially compresses this into one step: text → multi-shot video.

The first consumer of AI video is yourself. The demand is self-gratification; the act of generation itself is entertainment. This suits those with creative desire and imagination.

**3/ **In practice, Sora is mainly addictive to generate. Bingjian kept creating through the holiday, remarking that it has the thrill of opening blind boxes. Below are two videos Bingjian made.

$2

$2 **4/ **It's somewhat difficult to binge-watch Sora videos long-term. AI video is a bit like a single dish — eating only this gets tiresome. You still need to mix it up with variety on Douyin and WeChat Channels.

**5/ **Cameo, where you record your own likeness upfront, was a smart move that solves the cold-start problem for IP. Only two types of IP have viral potential: either celebrity IP, like Altman or Mickey Mouse, or familiar IP, like yourself or people around you.

**6/ **Sora's power lies primarily in model capability and tool attributes. There's an opportunity to build a content community, similar to Instagram,切入 through memes and viral clips.

It has little to do with social relationships. In its current form, it's difficult to build strong ties — or perhaps Sora simply doesn't aim to. Otherwise, why would its growth mechanism be "four invite codes" rather than "import your contacts"?

Opportunities Seen in Sora

7/ When the entire internet is flooded with Altman videos, it shows there's an IP shortage. Currently, celebrity IP is limited to Altman alone. Due to copyright reasons, other celebrity IPs cannot appear. How to work with IP, even aggressively so, is an opportunity for startups.

**8/ **With Sora's emergence, we should have some basic assumptions about AI video models:

• Within a few short months, there will be models that catch up to or even surpass Sora. Each will have its moment in the sun for two or three months. The model layer is insanely competitive — we've heard Veo 3.1 is coming soon, capable of generating one-minute videos.
• There will be excellent open-source models available to anyone: technological democratization, quality democratization.
• Some startup will succeed with a wrapper play, riding another Manus-like narrative to gain more chips at the table.
• The Cursor vs. Claude Code competition will play out again in video applications. Startups that validate good PMF will face strong competition from model companies. (But there's no choice — you still have to do it.)
• Entrepreneurs in video applications already have a history book. Reviewing the old to understand the new: much of what happened with LLMs will replay in video. Understanding the cases, opportunities, and challenges in LLMs provides a prescient perspective.

9/ AI video startups have some opportunity to build moats — content accumulation moats. This is somewhat better than with LLMs.

10/ OpenAI has a mountain-climbing map, L1-L5 for LLMs. There should be such a map for video models too; today we're probably at L1.

L5 is the endgame. What I can see is AI interactive video: real-time generation, both video and game, an infinite game with unlimited narrative flexibility. It would achieve one goal: using operation costs close to "scrolling Douyin" to deliver entertainment value close to "playing Genshin Impact" — rich, stimulating, immersive.

Exactly how this will be achieved, I don't know. But when it emerges, it will definitely be a new content format, a killer app. Haoyu Cai, founder of miHoYo, has a new venture called Anuttacon that seems aimed in this direction. Their first game, Whispers From The Star, is a trial run.

As for what L2-L4 look like, that's difficult to predict.

**11/ **What are the opportunities for startups post-Sora? Here are several angles we can think of:

• What do users want to generate that Sora cannot? That's the opportunity. Due to copyright issues, about half of generation attempts get rejected by Sora — prompts involving IP like One Piece, Elon Musk, or the cinematic styles of Christopher Nolan or Wong Kar-wai. What Sora rejects is the startup's opportunity.
• IP will be a key word for startups. Be aggressive at cold start: which IPs can you ride for traction? Once on the path to legitimacy, which compliant IPs can you accumulate?
• Startups shouldn't only generate video — images, games, and AI coding are also part of the experience. Mix video with images and coding to create new experiences for certain vertical user segments. If it's just video, that format already runs on major platforms, making "content migration" easy. When you mix content formats, "content migration" only works on the product you've built.
• Which niche users can you capture? What unique content can you form? Can the per-video view count on your platform be 10x that on Sora? Opportunities for user identification and community atmosphere.
• For pure video tools, how can you make the tool 10x better? How can you charge premium prices while users still want to pay?
• Video ads, motion comics, creative shorts, viral clips, various "Guan Gong vs. Qin Qiong" mashup videos — Sora can already do these. Nailing the last mile is the opportunity.

**12/ **Getting views on Sora itself isn't easy currently. When videos are moved to WeChat Channels, views easily pass a thousand. Bingjian had one video unexpectedly hit 100,000+ views: Altman imitating a silly roe deer. The video took just minutes to make. Recommendation algorithms truly defy prediction.

Analysis and Speculation on Sora

13/ Sora's strength is creativity over precise control.

Using brief prompts and letting the model improvise tends to yield more satisfying videos. For example: "1990s Zhongguancun, Altman starting a New Oriental-style study-abroad business, Wong Kar-wai style." Sora will arrange its own dialogue and shots — better than writing a long prompt with precise descriptions.

This indicates two things. First, Sora's controllability is relatively poor; third-party applications should leverage its creativity, not its controllability.

Second, we suspect Sora was post-trained extensively on classic works — it learned them well. You could say many artistic styles are compressed within it; the right prompt decompresses that style.

Sora likely trained heavily on Douyin and Kuaishou videos too — the memes and shot-switching rhythms feel very similar.

**14/ **Post-Sora, many features will become standard for video models, such as:

• Character consistency. Digital humans are a killer app in video.
• Audio-visual sync: sound comes with the image, lip-synced.
• Understanding of physical properties: a ball bounces off a wall, doesn't bounce off a sponge.
• Shot transitions: how many cuts a video can contain will also become a metric.

**15/ **Referencing LLM progress, ChatGPT had many problems at launch; how fast it improved, we all saw. Video model issues are merely a matter of time, such as:

• Duration limitations. Nine-second video will become the minimum baseline; soon it'll compete to 30 seconds. Reference how "long context" competed in LLMs.
• Characters looking too raw, not beautiful enough. Don't worry — look at Meitu. It can make you beautiful enough that your own mother wouldn't recognize you.
• Audio-visual inconsistency: only a matter of time.

16/ Don't build on video models' weaknesses; build on their strengths.

This is like when LLM reasoning was insufficient, developers wrote lengthy prompts to guide reasoning — proven useless. Once reasoning models emerged, CoT solved it directly.

In other words, leverage video models' creativity; don't focus on precise control. Product design should emphasize the joy of generation; don't worry yet about whether it produces consumable value for general audiences.

**17/ **Only OpenAI or major tech companies can offer Sora for free. Startups definitely have to charge — video inference costs are too high, a bottomless pit. So how to build chargeable scenarios becomes the startup's challenge.

18/ Over past decades, three new industries formed — programming, film/television, and games: technically demanding, yet also requiring massive human labor.

These three industries will be reshaped by AI. AI coding's reshaping of programming is already visible. AI's reshaping of film/television and games is coming too — not from a single model, but from LLM + image + video + coding, the full technology stack.

Ultimately, these three industries will no longer require such heavy labor costs. Premium content will become even more premium, while personalized content will become more effortlessly accessible. The boundaries between film/television and games will blur further.

Questions Sora Leaves for Entrepreneurs

1/ How long to replicate Sora 2? When will video's DeepSeek moment arrive — a SOTA model going open source?

2/ What was Sora 2's training cost? How large are its parameters?

3/ How to predict model trajectories and wrap ahead of time?

The same analogy: what kind of wrapper has value? If we see model capability as a waterline, some wrappers are pillars — when model capability rises, the pillar is submerged. Some wrappers are boats — when model capability rises, the boat rises with it.

Sora has raised the ceiling for AI video application and AI entertainment application entrepreneurship significantly.

Entrepreneurship is about seizing opportunities — capturing the openings created by external change. Let's see in a few months which smart people have seized them.

---

The Heart Capital team continues to monitor the complete AI investment ecosystem, including infrastructure, model layer, application layer, and applications with self-built models.

If you're also working in this industry, also following AI frontier technology, we welcome your pitch: bp@soulcapital.com. We look forward to connecting with you.

Founded in 2022, Heart Capital is an early-stage venture capital fund focused on technology and digitalization in China. The Heart Capital team is primarily composed of Yan Han, founding partner of Lightspeed; core investors; a CFO; and senior investors from industry. The team's past investments include Series A investments in Xpeng Motors (NYSE: XPEV, 09868.HK), Full Truck Alliance (NYSE: YMM), as well as FinVolution (NYSE: FINV), RoboSense (02498.HK), Baichuan, Yunmanman Cold Chain, Dedao, World Logistics, Micro-Nano Star, LandSpace, Lanhu, Starfield, and others. Rooted in China with a global outlook, Heart Capital is committed to finding true value in non-consensus. Heart Capital respects the value of "people" and advocates for the potential of the "heart," looking forward to accompanying more young Chinese entrepreneurs to strengthen China and go global.