Jia Zijian has built three AI products: Talkme, for practising spoken languages; ListenLeap, for learning languages through podcasts and YouTube; and Stances.ai, which pulls first-hand sources out of podcasts for research.
They look unrelated. Underneath is one judgment: voice is an underrated medium. It carries more than text, he says, with far less noise than video.
Talkme isn't competing with Duolingo
In early 2024 he made a simple AI conversation demo and posted it to Bilibili and a few overseas directories. It went unexpectedly viral and tens of thousands of users arrived. This was the GPT-3.5 Turbo era, and he took it as proof that AI speaking practice had product-market fit.
He's clear it isn't another Duolingo. Duolingo is about habit-building for beginners: five minutes a day, streaks and points. Talkme is for people who already have the basics but don't dare to speak or don't know how to use what they know.
What about voice mode in the big chat apps? His analogy: a large model is like an American friend who knows everything and will talk about anything, but isn't a teacher. It won't correct your pronunciation, design a progression for you, or stop you the third time you make the same mistake. Talkme has a full curriculum and has collected close to 100 million voice recordings from users around the world, data the general models don't have.
AI also makes languages cheap to add, so the same product can be rolled out market by market and language by language. Taiwan is its largest market, Japan has good returns, and mainland China is growing fast.
ListenLeap: listening fits more of life than speaking
Language learning has input and output. He noticed that listening fits far more moments than speaking: commuting, cooking, running. You can listen in those gaps, but you rarely speak.
ListenLeap turns podcasts and YouTube into immersive listening practice. After launching on Xiaohongshu, its growth went up more than tenfold and it became the top language-learning product there. Notably, the team got smaller during this period while revenue grew.
Stances.ai: retrieve, don't generate
Stances is the newest and most distinctive. It started as a feature inside ListenLeap: they were already processing huge amounts of podcast content, so why not extract what's valuable in it?
The selling point is first-hand sources. Ask a question and Stances doesn't generate an answer the way a chatbot would. It tells you which human expert said what, on which episode; whether different experts agree or disagree; how a view has shifted over time. Every claim links to the timestamp in the original audio. Search for DeepSeek and you see what people on a16z, YC and Nvidia's AI podcast each said.
The key technical choice is that Stances retrieves rather than generates. They trained their own models, in the billions of parameters, on more than a year of annotated podcast data from ListenLeap. If nobody has talked about something, it says so. Ask how to make Hainanese chicken rice and, if no podcast has covered it, you get nothing rather than an invented recipe.
That's also the difference from Perplexity. Perplexity searches the web and can answer almost anything, but web pages are second-hand at best. Stances only does first-hand sources, aiming for depth over breadth, for people whose decisions are expensive: investors, analysts, founders, researchers, journalists. They don't need an AI that answers everything. They need to know what the people who really understand the topic have said.
Not judging for the user
I asked whether Stances would warn users when new evidence overturns an earlier view.
His answer: they don't decide who's right. They lay out how views changed over time and let the user judge. A physicist and a chemist reason about the same question differently and reach different conclusions, and neither is necessarily wrong.
That restraint is unusual. Most AI products compete to give you the best answer. Stances holds back. Its value isn't thinking for you, it's showing you how everyone else thinks.
Three stages of building in AI
He's been a founder for six years, and building on large models for more than two. He describes three stages. Early on, the rush to ship something and prove yourself. Then, after some validation, the rush to scale while the business model is still shaky. Now, a steadier state of mind: knowing his limits, not over-promising to himself, and saying what he really thinks to investors and other CEOs instead of telling them what they want to hear.
He's frank that 80% of the time a founder is under heavy pressure and feels low. Each time a new opportunity works out, he feels capable again.
He also described how the products get built: young engineers from top universities in China and the US, with senior engineers guiding architecture and security. Fresh creativity plus experienced judgment, and it moves very fast.
Building is easy. The 10,000th paying user isn't.
The line I remember most:
"With vibe coding, building something close to our product in a weekend isn't impossible. Getting to the 1,000th or the 10,000th paying user is the hard part."
Cold-starting several products to 10,000 paying users and the first US$100,000 in revenue taught his team things that apply both globally and in China, and much of it can't be written down. You only know what to do when you reach that point and hit the problem.
The technical barrier keeps falling. The operating barrier doesn't. That may be the most honest description of where AI startups are now.
*This article draws from a podcast conversation with Jia Zijian, founder of Talkme, ListenLeap and Stances.ai.