Yunqi's Chen Yu: ChatGPT's Qualitative Leap Is More Than a Technical Shift | The Paper X Yunqi Capital ChatGPT Special

With ChatGPT's explosive debut and the ensuing public frenzy, large language model startups are heating up, and AGI (artificial general intelligence) seems closer than ever.

As early as late 2021, Yunqi Capital participated as the sole early-stage investor in the angel round of MiniMax, a domestic multimodal AI large-model startup. Its first product, "Glow," has since amassed nearly 5 million users, making MiniMax possibly the first domestic startup to simultaneously wield large-model capabilities across three modalities — voice, image, and text. Several other Yunqi portfolio companies have also explored related directions.

"Yunqi Tech π" launches its ChatGPT special. In this installment, The Paper's tech desk interviews Yunqi Capital partner Chen Yu on the qualitative shift behind ChatGPT and the future of AI's march toward general artificial intelligence.

Source | The Paper

Reporter | Chengtian Meng

➤➤➤ Chen has long focused on frontier technology and intelligent driving investments. Before joining Yunqi, he worked as a Google engineer and as CTO at a domestic publicly listed company.

**He argues that the qualitative leap behind ChatGPT was not a technical breakthrough but an engineering one — a phase change driven by scaling up model parameters, with human feedback as the root of ChatGPT's standout performance. As commercial use cases for ChatGPT come into view, the path toward AGI is gradually clarifying, though the journey to the destination remains long.

Human Feedback Is the Root of ChatGPT's Breakthrough

"For every bit of artificial intelligence, there's a bit of human intelligence," Chen says. "That's no longer an industry secret." Training large language models requires massive human effort for data cleaning and labeling, as well as human feedback to align the model with human linguistic habits and sound values.

Natural language processing (NLP) entered scientists' purview as early as the 1950s. After 2010, surging computing power and deep learning advances set NLP on its current trajectory. "Once model parameters scaled up, the phase change arrived suddenly," Chen explains. Beyond parameter growth, the technique of reinforcement learning from human feedback (RLHF) played a critical role in elevating language model quality, enabling models to learn human modes of expression.

Human feedback is the root of ChatGPT's breakthrough. Using reinforcement learning to iteratively refine an AI model is like teaching a child — rewards for correct answers, criticism for mistakes. After countless iterations of positive and negative feedback, the model converges toward human-desired outcomes, all of which demands enormous human labor. Consequently, ChatGPT turned to Africa and Southeast Asia, where labor costs are low, hiring local workers for data annotation. Kenyan data annotators, working under intense conditions for just $2 an hour, drew international scrutiny.

Current Costs Are 5-10x Traditional Search

"Large AI models are a track built on mountains of cash. From buying compute to hiring top AI engineers to collecting, labeling, and evaluating data — every step demands massive capital," Chen says. "$50 million is enough to get started, but in the long run, you need $300–500 million in funding to truly sit at the table." He adds, "It's a lot like autonomous driving back in the day — the bar for fundraising is exceptionally high."

The large models behind ChatGPT carry not only steep training costs but also formidable inference costs. "Right now, each ChatGPT query costs 5–10 times what traditional search costs. Whether that math works is still an open question," Chen notes.

Yet Chen also sees a path to cost reduction. First, Moore's Law applies to large-model compute as well. "NVIDIA's GPUs — from V100 to A100 to the upcoming H100 — improve overall efficiency by an order of magnitude each generation, so costs will keep falling."

The industry is already experimenting with various cost-cutting approaches, Chen says, "from chip-level compute to algorithmic improvements, and even questioning whether such massive parameter counts are truly necessary. Maybe 10 billion or 50 billion parameters will work just fine." He understands that Microsoft's integration with ChatGPT uses not the largest trillion-parameter model but a medium-sized one with only hundreds of billions of parameters — all of which will significantly reduce inference costs.

Tech Giants vs. Startups

In Chen's view, data volume matters enormously for large-model companies, but startups still have a shot at challenging the incumbents. OpenAI's achievements in challenging Google's search business model have become something of a legend.

"Internet data is equally available to all players. The difference lies in the human feedback behind that data," Chen explains. "Startups certainly don't have the data volume of tech giants, but they can innovate in how they collect data and feedback. ChatGPT's chat service, for instance, records every user conversation — and learns from that conversational feedback."

Chen also believes startup agility gives them an edge over big companies in large language model development. "As a department inside a large company, you suffer from all the typical corporate maladies. If a big company tried to launch something similar, someone internally would surely object — the model isn't perfect, it'll damage the brand, it'll cannibalize our existing business model... endless reasons to kill it. At a startup, I have no such fears."

As early as late 2021, Yunqi Capital participated as the sole early-stage investor in the angel round of MiniMax, a domestic multimodal AI large-model startup. Chen reveals that MiniMax's first to-C product, Glow, now has nearly 5 million users. "On the commercial front, MiniMax may be the first domestic company to make large models work based on user feedback." Several other Yunqi portfolio companies — including Jina AI, PingCAP, Sobot, and Bailing Intelligence — have also explored related products.

The Commercial Path Gradually Clarifies

In Chen's eyes, ChatGPT already has abundant commercial applications, especially in domains related to human language capabilities.

"In to-C, there's emotional companionship; in to-B, intelligent customer service, grammar correction, article summarization, financial analysis — these are all things large language models excel at," Chen says. "For now, though, they're mainly about augmenting human productivity. Current AI, including robots, can't fully replace humans. For a long time to come, it'll be human-machine collaboration."**

Artificial general intelligence (AGI) — the ultimate goal of AI development, capable of general human-level intelligence and performing tasks requiring human intellect — remains a distant concept. ChatGPT is an intermediate milestone on that research path.

"Progress is coming faster than expected, and it's given everyone a sudden sense of direction toward AGI. We don't know how far this path ultimately leads, but it produces staged, commercially viable results," Chen says. "AGI is still a remote concept — whether it's 5, 10, or 100 years away, nobody knows. So beyond technical capability, teams really need ideals and conviction." As a VC, Chen adds, "you need rational idealism too. Beyond capital support, you have to genuinely back those grander visions. Otherwise, neither ChatGPT nor any other breakthrough could ever happen."