When anyone can ship an AI SaaS in 48 hours, real differentiation is no longer a feature. It is judgment, taste, speed, and identity.
Tseka Luk11 min read
Continue reading
Research Notes
In 1997, not long after Steve Jobs returned to Apple, TBWA\Chiat\Day creative director Ken Segall put the line "Think Different" in front of him. Grammarians immediately objected: after "think," the adverb should be "differently." That is middle-school English. But Jobs insisted on the adjective "different." In an explanation that has been widely quoted over the years, he said the point was not to think the same way, but to think different: think a little different, think a lot different, think different. He treated "different" as a noun, a target, not merely an adverb modifying the act of thinking.
That ungrammatical line won the second Primetime Emmy Award for Outstanding Commercial from the Academy of Television Arts & Sciences in 1998. The first winner, in 1997, had been HBO's "Chimps."
This story is worth revisiting today, in 2026, when AI has turned "building software" into a weekend assignment. The whole commercial world is now trapped by the same question:
When everyone can use Cursor, Claude Code, Lovable, Bolt, and v0 to ship a working SaaS in 48 hours, where does differentiation come from?
The answer is not hidden in the action of "doing something differently." It is hidden in the adjective Jobs insisted on: whether the product itself is truly a different thing. To answer that, you first need independent thinking.
First, the problem is not imaginary.
In March 2025, Y Combinator CEO Garry Tan told CNBC that roughly a quarter of YC's Winter 2025 batch had codebases where 95% of the code was AI-generated, and the batch was averaging 10% week-over-week growth. "You don't need a team of 50 or 100 engineers," he said. In mid-March 2025, Pieter Levels launched the Vibe Code Game Jam, requiring more than 80% of each game's code to be AI-generated. He wrote on X that the deadline to enter was 25 March 2025, giving participants just seven days. That seven-day window still produced 1,170 games.
The other side of exploding supply is product sameness. In February 2026, National Business Daily cited a comparative analysis by Tian Feng of the Zhongguancun Digital Intelligence AI Research Institute: chatbot interfaces behind different AI products had a similarity as high as 92%, and 80% of new AI companies were crowding into red oceans such as intelligent customer service, AI image generation, and voice assistants. China's average data-center compute utilization was below 20%, yet the market had formed a loop of homogenization, low retention, high acquisition cost, high compute spend, and low ROI.
Across the Pacific, Marc Andrusko, an a16z partner focused on B2B AI applications and fintech, defined the coming year directly in his December 2024 prediction for "Big Ideas in Tech for 2025": "2025 will be the year of AI companies turning differentiation into lasting defensibility." He emphasized that "Differentiation and defensibility are distinct, and startups that conflate the two risk being outflanked by more strategic competitors."
In other words, "we built a different feature" does not equal "we built a moat." That is the lesson everyone is relearning.
In China, the debate peaked on March 31, 2025, when Zhu Xiaohu spoke at the Zhongguancun Forum. His line was blunt:
"All AI applications are wrapper applications. Saying they have barriers is fooling people. You need to build barriers on non-AI capabilities."
The logic that followed was even harsher: grand visions eventually become red oceans; the dirty and difficult work becomes the moat. He advised startups not to waste a single cent training foundation models. Instead, they should embrace open-source models and APIs, and put their energy into four things: deep workflow and editor integration, proprietary hardware and IP, proprietary data, and human-heavy operational work that must be delivered manually. At a Shanghai venture forum in late August, he added another cut: GPT-5 had arrived after huge anticipation, but everyone was disappointed, and the upper bound of AGI under the Transformer architecture had basically become visible. His conclusion: all moats must be found outside AI itself.
Zhu is not alone. Ji Yichao, co-founder and chief scientist of Manus, openly admitted the product was a "wrapper," then added that the ultimate wrapper is victory. After DeepSeek's rise, Kai-Fu Lee redirected 01.AI toward turning DeepSeek's strong foundation into enterprise deployment and customization solutions, comparing his company's role in AI 2.0 to Windows, with DeepSeek as the kernel.
Dai Yusen, managing partner at ZhenFund, went further in his 2025 mid-year reflection. He defended the "wrapper":
"So-called wrapper products, meaning applications that call APIs, will not necessarily be crushed by model-native products... especially because agents need more context and tools, and much of that depends on the wrapper and the application environment itself."
His example was pointed. When Cursor first launched, the models were not strong enough to support its vision. Only after Claude 3.5 Sonnet arrived did Cursor become a truly useful product. "Good application companies need to design for the models that will exist 6 to 12 months from now."
That is the Chinese consensus of 2025-2026: model capability will be democratized; product capability will not. Technology itself is no longer the barrier. The thing you build around the technology is.
a16z's January 2025 AI applications report put it plainly: the golden rule of the past two decades of SaaS, turning manual workflows into tools and charging per seat, has broken down. The new logic is that software is eating labor: AI products directly deliver outcomes. But to win that fight, a16z's June 2025 essay "Trading Margin for Moat" named the work clearly. Stop optimizing for 80% gross margins. Do the messy work of forward deployed engineers: data ingestion, workflow integration, customer implementation. In one sentence: trade margin for moat.
In June 2025, a16z partner Bryan Kim gave the consumer-AI version: "In consumer AI, momentum is the moat." Models update weekly. Infrastructure refreshes monthly. You do not have time to move slowly and polish for years as in the mobile-internet era. Speed, distribution, and mindshare become the only things that matter.
Garry Tan explained it most clearly in his May 2025 conversation with Vanta CEO Christina Cacioppo. In the age of "intelligence on tap," he said:
"The two things that are most important when intelligence on tap is available is actually agency and taste."
He then added: "It's not clear to me [taste] can be solved... That's actually turning out to be the moat for many startups." Tan also gave a simpler version: in the AI era, reasons for startup failure are collapsing. The remaining constraint is whether the founder can get inside the customer's head, understand what they actually want, and build something worth paying for.
This echoes Perplexity CEO Aravind Srinivas. In October 2025 at Berkeley Haas, when asked what Perplexity's moat was, he gave one word: speed. At YC AI Startup School, he added: "Live with that fear. You have to embrace it. Your moat comes from moving fast and building your own identity." His sharper warning to students was: "You should assume that if you have a big hit, a model company will copy it."
Cursor, or Anysphere, is the sample path. It started with an $8 million seed round in 2023, then reached a $2 billion annualized revenue run rate by February 2026 according to TechCrunch's reporting from two people familiar with the matter. It passed one million paid users, reached 70% of the Fortune 1000, and moved toward a valuation around $50 billion, with talks reportedly pushing the range toward $50-60 billion. Cursor has deeply integrated Claude, GPT, Gemini, and other models, while also launching its own Composer model in November 2025. Its differentiation was not reinventing the IDE. It was betting, under the familiar skin of VS Code, on the agentic coding workflow that would only fully mature 6 to 12 months later.
The most interesting argument, and the one closest to "independent thinking equals differentiation," comes from two of the world's most studied indie founders.
In multiple 2025 interviews later compiled by tldev.co, Pieter Levels repeated one lesson: AI startups have thin technical barriers, high churn, and volatile GPU and API costs. Sell first, do it manually, automate later. In the early days of AvatarAI, he processed orders manually through the night. Only after validating the workflow and demand did he spend a week automating the backend. A KOL video once took his MRR from $12,000 to $40,000-$50,000. His philosophy is often called Lindy: choose technologies that have survived reality and can keep running for five or ten years, not whatever framework is newest.
Marc Lou, founder of ShipFast and a $50,000-per-month indie builder, launched TrustMRR.com on October 30, 2025 and reached $10,000 in revenue within 36 hours. This was not a miracle. It was a perfect example. He identified the attention-arbitrage window hidden inside James Potter's tweet complaining about fake MRR screenshots, amplified by Pieter Levels, then used the simplest possible Stripe API, leaderboard, and verification badge to turn community anxiety into a product. His repeatedly summarized rule is: "You're not selling features. You're selling outcomes, status, or access to attention."
Vietnamese indie developer Tony Dinh's TypingMind, a better ChatGPT client, reached $130,000-$160,000 per month in October 2025, with the B2B Team version contributing half the revenue. He was not technically harder to copy. Anyone can build a client on top of the ChatGPT API. But he first grew an audience of more than 130,000 followers on Twitter, and every build-in-public update became both marketing and feedback collection. Distribution comes before product. That is the indie-founder truth repeatedly verified by the market.
A sharper summary came from designer Andrés Max in his early-2026 essay "Taste Is the New Moat":
"Code is cheap. Taste is the product. You can't prompt your way to taste. You either have people with real judgment and standards building your product, or you don't. And users can tell the difference in about three seconds."
At a16z's January 2026 LP meeting, later discussed publicly on The a16z Show, Marc Andreessen pulled this observation down to the level of economic history: "This is the biggest technological revolution of my life... clearly bigger than the internet. The comps on this are things like the microprocessor and the steam engine and electricity." He used Cursor to explain how application-layer companies can integrate backward. They start by calling one model, then quickly use 12, 50, or more, because they understand the customer context better than anyone else. Models are replicated every quarter. DeepSeek reproduced GPT-level reasoning in January 2025, and within weeks Tongyi, Hunyuan, Wenxin, and Kimi followed. But the winners did not become poorer. They started selling $200 and $300 monthly subscriptions. The moat is not the model. It is everything you build around the model.
Put these voices together, and the core argument is surprisingly consistent:
They all point to the same thing: once AI makes building a thing cheap, differentiation no longer comes from "we built a feature others did not." It comes from a form of judgment. More concretely:
Return to the ungrammatical adjective in "Think Different." Jobs was not insisting on "think differently," meaning a different manner of performing the act of thinking. He was insisting on "think different," meaning that the state itself is different. The former is a pose. The latter is an outcome.
In the AI-homogenized 2026, the winning paths of Cursor, Lovable, Marc Lou's ShipFast, Pieter Levels's Photo AI, Manus, and DeepSeek were not about deliberately looking unlike everyone else. They came from a person, or a few people, who actually saw something others did not, then put everything into making it real while accepting the repeated grinding of taste, dirty work, distribution, and community trust.
Independent thinking is not about being different for the sake of being different. It begins with conviction, and only then becomes differentiation.
So the next time you open Cursor to vibe code another "AI for X," pause first and ask the same question Zhu Xiaohu, Tan, and Levels would ask:
What have I seen that others have not? If I have not seen anything, then this thing I am building is just another noise inside homogenization.
That is the real homework for every product person and founder in 2026.
0
Discussion
Join with your Nebutra account. New comments enter moderation first.