Altman scrolling TikTok for 3 hours: It's not gossip, it's the ultimate challenge of AI productization

If even the smartest people in the AI world can't escape algorithm traps, can AI products truly serve humanity? This isn't a joke; it's a product logic gap spanning a galaxy between Sora and TikTok.

Sam Altman himself admitted that to understand how to build the Sora app, he downloaded TikTok to study it, ended up lounging on the sofa scrolling for three hours every weekend. Finally, because it was too addictive, he deleted it. This story is interesting; it exposes the huge chasm from technology to product for AI companies.

Addiction isn't a BUG, it's a Business Model

From an asset allocation perspective, TikTok's valuation is built on the core metric of "user time." Short video platforms are essentially efficient harvesters of the attention economy. Their recommendation algorithms are validated by hundreds of millions of users; every swipe trains the model, every second of dwell time optimizes matching. This isn't evil design; it's business logic.

But Altman's experience reveals a cruel fact: AI companies are learning how to make "addictive" products, but they haven't yet figured out if this addiction is worth replicating.

Honestly, after using Hugging Face and the latest large models for three weeks, retention rates for most AI products are dismal. Users open ChatGPT to solve problems and leave once solved; average session duration might be under 10 minutes. This differs by two orders of magnitude from TikTok's daily average usage time.

From an investment perspective, this is a fundamental valuation difference. Short video platforms can be valued based on DAU and user time; AI products currently can only be priced by "problem-solving capability." Which model is more sustainable? Short term, the former; long term, the latter might be healthier, but Wall Street's patience is limited.

Product Logic Behind Sora's Shutdown

The material mentions that in March 2026, Altman personally shut down Sora. On the same day, ByteDance's AI comic drama track broke through 70 million yuan in single-day ad spend. This timing is meaningful.

Sora had technical capability, but its productization didn't work. Why? I guess there are three reasons:

1. Failure in managing user expectations. TikTok users just want to watch 15-second funny videos; Sora users want to generate a movie. The former is "entertainment," the latter is "creation." Entertainment is a rigid demand; creation is a niche need.

2. Mismatched cost structure. TikTok's recommendation algorithm completes matching in milliseconds; Sora generating a video segment might take minutes. Users are willing to pay for "fast" and "fun," but not for "slow" and "uncertain." This is the cost structure issue I mentioned when analyzing MediaTek's semi-annual report.

3. Niche misalignment. How to put it, TikTok is a consumption platform; Sora is a production tool. Consumption platforms earn money via ads; production tools can only rely on subscriptions. Companies making production tools wanting to become consumption platforms are missing a massive content ecosystem in between.

In contrast, ByteDance uses AI to make short dramas, essentially building a "content consumption platform," not an "AI creation tool." This positioning difference determines victory or defeat.

Two Paths, Whoever Runs Through First Wins

From an investment perspective, two clear routes have emerged in the AI video track:

Route A: AI as a Tool, Helping People Produce Content

Representatives: Sora, Runway, Pika

Features: High technical barrier, clear business model but low ceiling (subscription revenue)

Risks: Poor user retention, high customer acquisition costs

Route B: AI Creates Content, Direct Consumption

Representatives: ByteDance AI short dramas, Kuaishou AI effects

Features: Low technical barrier, but large user scale, short monetization path (ads + in-app purchases)

Risks: Unstable content quality, regulatory risks

My judgment is: In the medium-to-long term, Route B has a higher win rate.

The reason is simple: User time is finite. AI video generation tools require users to actively "create," while AI video consumption platforms only require users to "lie back and watch." The latter better fits human nature.

Altman spending three hours on TikTok is essentially a failed "user behavior study." He thought he was studying competitors, but found himself becoming a user first. What does this show? It shows AI product managers still don't truly understand why people spend three hours on short videos instead of three minutes on AI tools.

From an asset allocation perspective, future investment opportunities in the AI field may lie not in "the smartest technology," but in "the most addictive products." The story of ByteDance piecing together a Sora competitor in 40 days proves that technical barriers are relative in this field; product barriers are absolute.

One-sentence summary: The core competitiveness of AI products lies not in how smart they are, but in how well they understand human nature.