Silicon Valley AI scandal exposed! Directly hitting the unfinished GPT-4.5, Meta imitating DeepSeek with issues, NVIDIA and Apple have long-standing beef.

Renowned semiconductor analyst Dylan Patel was interviewed on July 1 to analyze the current competitive landscape among AI giants. He not only discussed the reasons behind the development failures of GPT-4.5, the acquisition actions of Meta and Scale AI, and the insider information on Apple lagging in the AI competition due to its fallout with NVIDIA, but also shared why AMD still cannot compete with NVIDIA's powerful CUDA ecosystem. At the end of the interview, Patel candidly expressed his optimism about OpenAI's future development in superintelligence (Superintelligence).

Super intelligence is the way to go; if you don't keep up, you're destined to lose.

Patel was the first to state that the entire AI industry has shifted from the previous pursuit of General Artificial Intelligence (AGI) to aiming for Super Intelligence (Super Intelligence). Since the co-founder of OpenAI, Illya Sutskever, established Safe Super Intelligence (SSI), it has been seen as a critical turning point, and it was from that time that Meta, OpenAI, and xAI began to pivot. He pointed out that if one does not catch up now, they will ultimately be the loser.

The image shows OpenAI co-founder Illya Sutskever discussing why GPT-4.5 may fail.

When discussing why OpenAI GPT-4.5 failed, Patel believes the problem lies in the fact that although the model's parameters have become larger and smarter, in reality, it is too slow, too costly, and has a very low user adoption rate. The main reason for the failure is that "the amount of data is insufficient," which has led to the "overparameterization" of GPT-4.5, that is:

"The larger the model, the greater its thirst for data. If it is only allowed to hold limited information, it cannot truly understand."

Patel stated that looking back, the key to the failure of GPT-4.5 was not the size of the model, but rather that the data and reasoning framework did not keep up.

(Is GPT becoming a bit bizarre? Three major events reveal the potential risks of AI going out of control)

Why does Meta's AI keep losing?

When it comes to Meta's Llama 4 and the delayed Behemoth model, Patel believes that the problem lies in the overall technical decision-making issues. Although Meta has strong researchers and a bunch of GPUs, without a leader who truly understands the technology and can make decisions, it is easy to go in the wrong direction.

He cited an example where, during the development process of Meta's AI models, some teams attempted to mimic China's DeepSeek, but due to improper training design, certain "expert modules" within the model were not effectively utilized, resulting in a waste of training resources. The issue lies in the absence of a leader who understands technology and can make decisions to guide the direction.

(How is Meta rebuilding market competitiveness through large-scale acquisition of AI companies from Scale AI to NFDG?)

Why are OpenAI and Microsoft gradually drifting apart?

Patel believes that the collaboration between OpenAI and Microsoft (Microsoft) has already passed the honeymoon phase. Although Microsoft has invested over 10 billion dollars in OpenAI, they have only received revenue sharing and profit distribution, without holding any shares of OpenAI, which makes the situation somewhat awkward. Additionally, Microsoft originally had exclusive computing rights with OpenAI, but that has been canceled this year.

OpenAI has now begun collaborating with suppliers such as Oracle (, CoreWeave, and others to build data centers, freeing itself from dependence on Microsoft. Patel also revealed that OpenAI is the most cash-burning startup on the planet, with no plans for profitability in the coming years, only aiming to continue expanding its scale and valuation.

)Microsoft and OpenAI are discussing cooperation terms again: reducing shareholding in exchange for access to technology, with the $13 billion cooperation project facing restructuring pressure(

Apple's lagging is not without reason; it's conservative and refuses to cooperate with Nvidia.

Patel pointed out that Apple’s culture is too conservative, leading to a lack of major breakthroughs in the AI field, as they only make acquisitions of smaller startups that are not influential, lacking the "guts" to make a one-time acquisition of an entire top team.

In addition, Apple previously had the "BumpGate" incident with NVIDIA ), where a generation of NVIDIA GPUs failed due to poor soldering, leading to both parties blaming each other and creating "beef". Later, due to patent litigation issues, Apple stopped using NVIDIA chips.

Apple's emphasized "device-side AI" strategy, while having advantages in security and response speed, still leads Patel to believe that overall market demand will lean towards cloud services. He bluntly stated:

"Everyone talks about valuing privacy, but in reality, they still prefer free and easy-to-use services."

Is Grok substantial? At least it's bolder when discussing certain issues.

Patel's evaluation of Grok is better than expected, especially when querying real-time news and discussing topics like geography and history, the responses are more natural and less afraid to touch on sensitive subjects. He mentioned that when asking about changes in the white and black population ratio in the southern United States, the history of slavery, or the history of oil monopolies, Grok is willing to provide more in-depth background rather than just the safe and harmless version.

As for Musk's claim that Grok is "the smartest AI in the world," Patel remains cautious. He said that Grok might indeed be good, and Musk is indeed very effective in building data centers and buying power plants, but he feels that further observation is needed.

(Musk xAI is set to retrain Grok, experts criticize: deliberately distorting history)

AMD is making a strong comeback, but Nvidia still maintains its leading advantage.

Patel finally analyzed the current chip battle between NVIDIA and AMD. He agrees with AMD's aggressive approach to catch up, and indeed, they perform well in certain situations, but overall, NVIDIA is still ahead.

Patel cited that NVIDIA's NVLink allows for high-speed interconnection of 72 GPUs, while AMD can only achieve 8 at this stage. Additionally, NVIDIA's software and development ecosystem is quite complete, enabling users to execute models with just a few clicks, whereas AMD still requires manual adjustment of a bunch of parameters.

However, recently NVIDIA launched its own cloud rental platform Lepton and DGX Cloud Lepton AI, which made many cloud providers that rely on selling NVIDIA chips feel backstabbed, and also gave AMD an opportunity to capture some market share.

( AMD's latest MI355 performance skyrockets 35 times! Su Zifeng: AI market will exceed 500 billion in three years )

Who is most likely to seize the throne of super intelligence?

Patel ultimately believes that OpenAI is the most likely contender to seize the throne of superintelligence, as all major breakthroughs have come from them. Following closely are the steadily advancing Anthropic, which has recently become more open, and then there are Meta and xAI, which are still making efforts.

This article exposes the dark side of Silicon Valley AI! Directly hitting the GPT-4.5 fiasco, Meta's imitation of DeepSeek goes wrong, and the long-standing beef between Nvidia and Apple first appeared in Chain News ABMedia.