2025: The Year of the AGI Faith Collapse?

2025: The Year of the AGI Faith Collapse?The "AGI faith" surrounding large language models (LLMs) is teetering on the brink of collapse, following a trajectory strikingly similar to the former "blockchain faith." The difference lies in the outcome: while most "blockchain believers" were left with losses, a small percentage of "top players" reaped enormous profits

2025: The Year of the AGI Faith Collapse?

The "AGI faith" surrounding large language models (LLMs) is teetering on the brink of collapse, following a trajectory strikingly similar to the former "blockchain faith." The difference lies in the outcome: while most "blockchain believers" were left with losses, a small percentage of "top players" reaped enormous profits. In contrast, many AI LLM startups, stripped of the "AGI faith" fog, are struggling for mere survival. 2025 might mark the year of this faith's definitive collapse.

I. Tech Giants Abandoning the AGI Faith: Pragmatism Trumps Dreams

Kai-Fu Lee recently publicly declared his abandonment of the pursuit of AGI, becoming the first high-profile figure in the industry to openly "renounce his faith." He candidly admitted that pursuing AGI requires massive funding, and securing that funding is the current priority. The "sky's-the-limit" pursuit of AGI necessitates limitless capital reserves, but the reality is that consolidating and acquiring financial strength takes precedence. Faced with a choice between "building bigger models" and "building a more profitable business," Lee made his decision in May 2024: survival first, then ideals.

Investor Xiaozhu Zhu also frankly stated his disbelief in AGI, likening it to the construction of a high-energy particle collidera massive investment with limited significance, echoing physicist Yang Zhenning's opposition to China building such a collider. Zhou Ming, founder and CEO of Lanyu Technology, holds a similar view, focusing on how LLMs can serve clients and deliver a strong return on investment (ROI). However, over the past two years, pragmatic voices like Zhu and Zhou have been in the minority. In the whirlwind of hype, most model companies prioritized achieving their "AGI vision."

II. The Interplay of Technological Faith and Hype Cycles

Every few years, a so-called "technological revolution" or "innovation" emerges, spawning various "faiths." From blockchain to the metaverse, and now AI LLMs, the pattern remains consistent. During blockchain's peak, "blockchain evangelists" preached their "faith" in blockchain and decentralization. The result? Some profited handsomely from cryptocurrency speculation, while countless investors lost everything. The metaverse boom followed a similar trajectory, with the concept of virtual worlds becoming a new faith, only to become a spectacle for a select few, offering an experience often inferior to even obscure games. Will the "AGI faith" suffer the same fate?

III. The Brutal Reality of Business: OpenAI's Transformation and Industry Crisis

OpenAI's board recently revealed an evaluation of the company's structure, planning to transition its for-profit subsidiary into a Delaware Public Benefit Corporation (PBC) to improve profitability. This demonstrates that even a powerful company like OpenAI must prioritize generating revenue to continue pursuing its ideals.

Beyond OpenAI, other LLM companies globally face challenges. Bamboo AI announced a six-month shutdown; Stability AI is experiencing a severe financial crisis; while after securing significant funding, the founder of Moonlit Face is rumored to have cashed out, with parts of the team following suit. Waveform Intelligence even dissolved, with core members joining other companies. Many LLM startups' core personnel have returned to their original companies or large tech firms, and business development leaders have also left in droves. These signs suggest that "believing in AGI" might be a beautiful vision, but current LLM companies are struggling with profitability, and even their faith in Scaling Law is wavering.

IV. The Limitations of Scaling Law: Data Bottlenecks and Cost Pressures

Kai-Fu Lee confessed that his belief in Scaling Law shifted to skepticism within a year. This is largely due to the challenges facing Scaling Law. Ilya Sutskever, co-founder of OpenAI, noted at NeurIPS 2024 that while the success of deep learning is attributed to Scaling Law, the rate of improvement in computing power has surpassed the total amount of data available for training AI models, nearing a data bottleneck. While existing data can still be used for training, this trend will eventually slow, and the pre-training era will gradually end.

Data shows that from AlexNet to GPT-3.0, the model parameter scale increased 2916 times, and training costs soared from less than $10,000 to tens of millions of dollars. From GPT-3.0 to GPT-4.0, the parameter increase was even larger, with greater investment, but the results were not as expected. Only large companies can afford the massive investment to develop better models, while startups, lacking viable commercialization pathways, find the pursuit of ideals unrealistic.

V. Lack of Profitable Business Models: Faith Fails to Translate into Reality

The absence of profitable business models is a prevalent challenge for AI LLM startups, contributing significantly to their uncertainty. No matter how compelling the "AGI faith" narrative, current LLM products are merely improvements to productivity tools, far from being "revolutionary," and struggle to attract paying consumers. Funding becomes almost their sole lifeline, relying on captivating narratives to secure continued development. "Faith," therefore, often serves as a fundraising tool, as "capital loves a good story."

VI. The Future of AGI: A Clash of Optimism and Pessimism

LLM companies' performance over the past two years has been relatively "honest," lacking the "frenzy" of the initial blockchain boom. This might be because LLMs lack the "overnight riches" potential of cryptocurrency speculation. However, the shift from hype to uncertainty in AI LLMs is a natural phase in the development of emerging industries; the receding tide reveals the true nature of things. Kai-Fu Lee points out that company development goes through a shift from talent-focused to commercialization-focused, a process that took AI 1.0 companies 6-8 years.

Does abandoning the AGI faith mean AGI will never be realized? Optimists believe it will be achieved within five years, while pessimists or pragmatists foresee a longer timeline. Sam Altman of OpenAI hasn't specified a timeline for AGI, only stating that he knows how to build AGI using conventional methods. In 2025, LLM companies will likely adopt a more pragmatic approach; survival is paramount to achieving any goal, and "faith" will become less significant.