What Hinton Taught Me Was Curiosity. LLMs Are Just Undergrad Math Done Very Deep.

This is the seventh episode of Stolen Chat. Yang Kai studied math at the University of Toronto, took Geoffrey Hinton's class, and witnessed the moment AlexNet launched the deep learning era. Harvard stats master's, Wall Street quant trader, serial entrepreneur. Now in Singapore building Arros AI — AI infrastructure for the recruiting industry, ARR over $1M, recently appointed Chief AI Scientist by Nasdaq-listed YY Group.

We talked about two things: why finding people is fundamentally different from finding information, and why "lobsters" (agentic AI) represent a genuine paradigm shift. His answer was more modest than I expected — not because the technology is amazing, but because people are getting lazier.

What Hinton taught was curiosity

Yang Kai took Hinton's class as an undergrad at U of T. Deep learning wasn't mainstream yet. AlexNet had just appeared, and the accuracy of convolutional networks on classification tasks stunned everyone.

But what Hinton taught him that mattered most wasn't any specific technique. It was curiosity.

"North American education was a shock to me. Chinese traditional education is teacher-directed — master the knowledge points. North America values the spirit of exploration. The professor opens a door, shows you something interesting, and you think: this is fascinating, I want to dig into it too."

That curiosity sustained over a decade of deep work in AI, including entrepreneurship. "Startups are actually a lot like doing research in a lab — you're constantly studying new problems, trying new methods."

LLMs are just undergrad math done very deep

Yang Kai once said: LLMs are just an autoregressive model you learn as an undergrad — junior year material — executed at massive scale with enormous data.

He doesn't deny the value of large models, but emphasizes that the fundamentals haven't changed. The underlying principles are the same — how to capture relationships between words. The difference is the engineering scale: from pre-training to RLHF to data labeling, the sheer volume is unimaginable from a student's perspective. Training on clusters of tens of thousands of GPUs — that's not something a lab can do.

"Without the professors' earlier contributions, without AlexNet showing the world deep learning's potential, there wouldn't have been more investment from Google, no Transformer, no OpenAI. The whole thing is a chain reaction."

"Nothing is about the theory. It's about how deep you go."

From quant trading to AI recruiting

After Harvard stats, he went to Wall Street for quantitative trading. He left for a straightforward reason: the problem's scope wasn't big enough. He wanted greater impact.

Back in China, he started his first company: Huixiaor — an online meeting venue booking platform. The core problem was massive labor costs: leads from Baidu needed phone calls, matching venues required extensive training, and an operations specialist needed two to three months to learn all the venues, prices, layouts, and vibes.

He built an automated matching system: incoming leads get classified, routed to the right specialist, then automatically matched to venues based on structured requirements. Fewer people handling more demand, lead conversion rates improved, customer acquisition costs dropped from 2000+ to 800 RMB.

What connects quant trading and AI recruiting? Both are hard problems with a signal-to-noise challenge — the data is extremely dirty, noise is massive, and the core task is capturing the true signal.

China's SaaS graveyard

His second startup was Moyin Intelligence — voice technology. It exploded during COVID. Masses of finance professionals working from home needed automatic meeting transcription and summaries via Tencent Meeting. Within months: over a million users, $3M raised.

Then Feishu went all-in on free. Tencent Meeting added free features too. Moyin had launched nearly a year before Feishu, but big tech's free tier crushed them.

They pivoted to enterprise intelligent customer service, landing major clients like Yonyou and Kingdee. But enterprise had its own problems: companies partnered with startups because they were cheap and hungry. A single contract might be 1-2 million RMB but take a year to deliver. Year two, they'd push for lower prices — or just build a 60-70% replacement internally.

"China's consumer products have successful business models — advertising, gaming, e-commerce. But enterprise has never found a scalable business model."

Two years ago, he returned to overseas markets.

From AI recruiting player to AI recruiting infrastructure

Arros AI originally did AI recruiting directly. But they discovered a bigger opportunity: every recruiting company was desperate for AI transformation, anxious, urgently wanting AI capabilities.

So they made an elegant pivot — from industry player to industry infrastructure. Instead of competing in the market, they supply technology as ammunition to all recruiting platforms.

"You can think of it this way: every recruiting company is an AI recruiting company. So selling shovels became the biggest opportunity."

YY Group was one such company — they saw the technology, invested in Arros AI, and appointed Yang Kai as Chief AI Scientist. But beyond YY, many similar platforms have proactively sought partnerships.

Indexing information with keywords is easy. Indexing people with intent is hard.

I threw out an analogy during our conversation: Google used intent-based search to reorganize information. LLMs can also search for information. But what Arros does is fundamentally use algorithms to reconstruct people.

Yang Kai said people are extremely complex objects. A resume mentions "LLM" — does that mean the person works on LLMs? Not necessarily. They might be a recruiter hiring for LLM positions. Keywords have very limited utility — only weak correlation.

Finding people requires strong correlation and deep understanding. For example, a client says "I want someone with a Chinese background." How do you search with keywords? By surname? Zhao, Qian, Sun, Li? It doesn't work.

They built a technology called profile consolidation: aggregating a person's traces across LinkedIn, social media, news, and forums, using LLMs for source comparison and cross-validation — confirming this is the same Ellen Cheng, not a different one — then assembling a complete portrait.

"AI as a communication tool is becoming increasingly valuable. LinkedIn is the world's largest address book. AI already has enormous potential to become the Connector of the future."

"Lobsters" are the next AlexNet

Yang Kai's take on "lobsters" (Lovable/Bolt-style agentic AI platforms) is definitive: this is a genuine product paradigm shift, not buzz.

The reason isn't that the technology is impressive. It's that people are getting lazier.

"Before, I had to log into software, learn how to use it, click 100 buttons. Now I just talk and get the same result. Once people get used to this, there's no going back — just like once you're used to ChatGPT, you don't want to search Google. Once you're used to lobsters, going back to SaaS feels impossible."

Their recruiters use agentic AI for massive amounts of work: organizational chart mapping, sourcing, auto-sending WhatsApp and LinkedIn messages, generating PPTs, meeting notes, checking what salary a candidate discussed previously. Implementing this workflow used to require seven or eight SaaS tools. Now one conversational assistant handles it all.

He compared lobsters to the next AlexNet: "We've already seen through agentic AI the enormous potential — just like when I saw AlexNet at U of T. Once people see the potential, more investment follows, and it gradually creates larger and larger societal value."

The scarcest thing in the AI era

Final question: what's most important for building products in the AI era?

Yang Kai's answer: understanding demand.

"Engineering barriers are dropping everywhere — vibe coding, execution-layer tools, everything is getting simpler and cheaper. The core question remains unchanged: where is the demand? What does the customer actually need? How do I serve them better?"

The trend now is people becoming more composite — engineers becoming PMs, PMs becoming engineers, everyone needing to understand marketing and distribution. AI technology further expands the boundary of individual capability.

"This question is as old as time."