Image source: Generated by Unbounded AI

The "AGI faith" of the large AI model may be on the verge of collapse, just like the former "blockchain faith".

It’s just that there was a large number of “blockchain believers” at the beginning. Although most of them were leeks, there was also a small group of “top people” who made a lot of money; and today’s large AI models Start-up companies, regardless of "AGI belief", may not even know how long they can survive.

2025 may be the first year of the collapse of faith in large AI models.

1. Kai-Fu Lee: Give up the pursuit of AGI

Today, Kai-Fu Lee is In an interview later, he admitted that he gave up pursuing AGI very early.

This may be the first big man in the industry to publicly announce "giving up his faith."

The reasons for giving up AGI are not complicated. Pursuing AGI requires money, and you can’t talk about dreams without money. "Looking up at the stars and pursuing AGI requires sufficient ammunition reserves at all costs. Looking at it from a down-to-earth perspective, our highest priority at this stage is to first consolidate the strength to obtain ammunition."

Choose to make a larger model or make a A more profitable commercial company had to choose one of the two. Kai-fu Lee said that he would make a decision in May 2024.

To survive first is the capital to talk about ideals.

Investor Zhu Xiaohu also bluntly stated that he did not believe in AGI, "I think pursuing AGI is the same as building a high-energy particle collider back then. Why doesn't Yang Zhenning recommend that China engage in high-energy particle collider?" There is no point spending billions and tens of billions to buy a machine."

Zhou Ming, founder and CEO of Lanzhou Technology, also said that he does not pursue AGI, but pursues how large models can serve customers. Pursue calculating ROI.

But in the past two years, people like Zhu Xiaohu and Zhou Ming are really in the minority. Under the spotlight, major model companies are rushing to realize the "AGI Vision".

2. Faith always goes with the wind

Every few days There was a "technological revolution and innovation" in 2008, and along with these technological innovations came various "beliefs." From the blockchain to the metaverse to today’s large AI models, there are no exceptions.

When blockchain was the most popular, celebrities in the industry all held the title of "blockchain evangelist" and said to the outside world that they believed in blockchain and decentralization... When the metaverse became popular two years ago, "you" in the virtual world became a new belief, especially in cyberspace.

But what happened next?

Those "blockchain evangelists" made huge amounts of money by speculating on coins, and of course there were a lot of leeks that lost all their money; there were also those who advocated that the metaverse was the future, and eventually they were no longer paid attention to. It sounds ugly, but those Metaverse products are not as effective asA popular game with a good experience.

I believe AGI will eventually do the same.

3. Battle reports will lie, but battle lines will not

Previous In August, OpenAI’s board of directors revealed that it was evaluating the company’s structure and planned to change the nature of the for-profit subsidiary into a Delaware public benefit corporation (PBC) that would manage and control OpenAI’s operations and business. The non-profit organization's significant interest in the existing for-profit subsidiary will appear in the PBC in the form of shares.

To put it bluntly, even a company as powerful as OpenAI must first consider how to make money before talking about its ideals.

In addition to OpenAI, other large model companies at home and abroad also have more or less different problems.

In February 2024, the AI ​​company Zhujian Intelligence announced a half-year suspension;

In May 2024, the well-known company Stability launched the Stable Diffusion series of models AI was exposed to encounter a serious financial crisis;

In June 2024, the valuation of Yue Zhi Dark Side exceeded 20 billion one year after its establishment, and it received investment from giants such as Alibaba; at this time, founder Yang Zhilin was rumored to have cashed out 40 million US dollars, and the company team was also exposed to partially cash out.

In October 2024, the large AI model company Waveform Intelligence was disbanded, and many core team members joined OPPO.

There are also large model companies such as Adept, Reka AI, Humane and other large model companies that have encountered varying degrees of operational crises and have been acquired or are seeking to be sold.

Wang Huiwen, who has great enthusiasm for large-scale models and is eager to try it, has now returned to Meituan; some of the core personnel of some large-scale model startups have returned to their previous companies, and some have gone to large factories. Some heads of commercialization have resigned. What about the promised faith?

All signs indicate that "belief in AGI" is just a pie, and the current large AI model companies are confused about how to make money.

Except for AGI, I can barely hold on even to the Scaling Law.

4. Scaling Law doesn’t work anymore?

Lee Kai-fu also said that it only took him one year to go from believing in Scaling Law to doubting it.

Why is this? It may be that Scaling Law is experiencing a crisis.

In December 2024, OpenAI co-founder Ilya Sutskever said at the NeurIPS 2024 conference that deep learning is thanks to Scaling Law, that is, using very large data sets to train a large enough neural network, success can almost foresee.

But whenThe current rate of improvement in computing power has exceeded the total amount available for AI model training, and data growth is approaching a bottleneck. "You can still use existing data for training, but the trend will eventually slow down, and the pre-training era will gradually end."

Data shows that from AlexNet released by OpenAI in 2012 to GPT3.0 in 2022, the model parameters have increased from 60 million to 175 billion, the scale has increased by 2916 times, and the training cost has increased from less than 10,000 US dollars to Tens of millions of dollars; from GPT3.0 to 4.0, the parameters were improved even more and the investment was greater, but the effect was not as good as expected.

Only large companies have the ability to maintain huge investments and develop better models; but it is very unrealistic for startups to talk about their ideals without a suitable commercialization path.

5. Faith cannot be used as food

The lack of profit model is the current The common dilemma of AI large model startups is also one of the reasons why they are confused. Even though the story of "Belief in AGI" is very vivid, the current model products are only tools to improve productivity at best, far from reaching the level of "revolution", and it is even more difficult for C-end users to pay for them.

Financing is almost the only way to survive. Tell the story to capital in exchange for ammunition to continue fighting.

So, the so-called "faith" is probably more of a means to facilitate financing. After all, "capital loves to hear stories."

6. It is not certain when AGI will be realized

‍‍

Overall, The performance of large model companies in the past two years has been "honest" and not as "fanatical" as the original blockchain. This is probably because large models do not have the "get rich overnight" attribute of currency speculation.

However, AI large models have gone from being popular in the past to being confused now. This is also a necessary stage for the development of emerging industries. Only when the tide fades can the essence be exposed.

Lee Kai-fu pointed out that the company’s development starts from who has the best people, has the most papers, and has the best competition results, and gradually reaches business milestones, who can win a big order, and who can win a few more orders. Who can expand business. This process takes AI 1.0 companies 6 to 8 years.

If we no longer believe in AGI, does it mean that AGI will never be realized?

Optimists believe that it can be achieved in the next five years, while pessimists or pragmatists believe that it will take longer.

OpenAI’s Sam Altman did not clearly say when AGI will be implemented in his “10,000-word essay” published a few days ago. He only said that he already knows how to build AGI in a traditional way.

In 2025, major model companies should be more pragmatic. Again, only by surviving can we achieve our goals.Maybe, "faith" is no longer important.