At a time when the domestic large-scale model industry has finished rolling up the application of the model volume, the application volume ecology, and fighting each other to compete with each other, the basic model has finally ushered in a breakthrough after the release of Claude 3 yesterday
Overall, of course, there are many surprising points of Claude 3, such as extremely fast response time, powerful visual capabilities, 200k context length, and various data sets and improved accuracy on tasksBut unlike the promotional video (not the Gemini manual dog head) when Gemini was first released, which brought people not only "beyond GPT-4" but also "beyond the GPT-4 paradigm", GLUDE 3 is more like a tinkering breakthrough on the development path that GPT-4 has already charted for the big model
And the interesting point is that even though the Claude series has actually maintained good performance since its inception, it is also the second and third in the major evaluation lists, and even this time the Claude 3 "turned over the serf and sang", surpassing the boss GPT-4 in one fell swoop,However, the hottest topic in everyone's discussion is still OpenAI
As long as you are searching for content and inquiries related to Claude 3, almost every Claude 3 must bring OpenAI or GPT-4Of course, this is also because the biggest selling point of Claude 3 is to "beat GPT-4", but for a company's popularity in the market, empathy is if OpenAI releases GPT-45 or GPT-5, the title may never have claude.
It's not just the title of **, when everyone in the community sees Claude 3 performing like this, and even their first reaction is "I'm looking forward to GPT-5 more"?
And this expectation of netizens does not seem to be groundless, not to mention that there were rumors about GPT-4 three months ago5 Suspect** (GPT-4.)5 suspicious**! Three model models, multimodal capabilities have been greatly upgraded, and pricing has been leaked), just said in Google's Gemini 1 last monthOn the same day that the 5 Pro was released, OpenAI released Sora, which immediately seized the heat and occupied the traffic highland, Gemini 1The 5 Pro hasn't been in the headlines for a full day, and the information is all about Sora right away.
And this situation was summarized by Zhang Junlin not long ago as the "suppression chain theory".Specifically, in the entire field of large models, OpenAI has become the undoubted "industry leader", at the top of the chain, and OpenAI, which occupies a first-mover advantage, "has the ability to suppress competitors who have the potential to catch up with it", such as Google and Anthropic
In this chain of repression, the theory envisions a boss at the top of the chain, OpenAI, and a group of followers at the bottom of the chain. Because of its first-mover advantage, OpenAI has the ability to build a "technology reserve" that stores many excellent models that are "secret". In order to suppress competitors and maintain their "dominant position", OpenAI will not choose to directly release the new models it has developed to seize market opportunities, but will wait for competitors to release models, and select models in its "technology pool" according to the advanced degree of its models to "suppress" competitors.
Due to the high degree of "homogeneity" of the market product of the large model (in fact, it is also the core reason why this suppression chain can be played), when OpenAI implements the suppression strategy, it often makes the high R&D costs invested by competitors in the past unable to be converted into effective market share, and becomes a sunk cost, so as to achieve OpenAI's means of "creating competitive pressure" and even "forcing them to withdraw from the market" for the followers of the market
This tactic can be seen when OpenAI launched ChatGPT to suppress Claude, and we can also see it when SORA cracked down on Gemini this year, and everyone in the community is even looking forward to seeing the moment when GPT-5 cracks down on Claude 3. However, 24 hours have passed, and the king bomb of GPT-5 has not yet appeared, and this "where is GPT-5" emoji has also ushered in an evolution:
In fact, although OpenAI seems to be a good "sniper" under the suppression chain theory, this time it seems to be a "counterattack" by Claude 3.
As we reported, in the past two days OpenAI is in the "quagmire" of being sued by Elon Musk accusing it of not being open source (Musk officially sued OpenAI and Ultraman!). ）。
And according to the analysis of netizens,In fact, Claude 3 seems to have been deliberately chosen to be launched quickly at such a time when OpenAI may not be able to quickly implement "suppression", and it was taken by surprise to catch OpenAI off guard
Perhaps Matt Schumer, CEO of Anthropic, or any member of the large model community, probably everyone knows that there must be a more advanced large model within OpenAI, and GPT-5 is only a matter of time. And at this time, Musk's prosecution, whether it can be successful or not, may buy valuable time for "latecomers" like Anthropic.
So, under OpenAI's suppression chain, are latecomers really no chance at all?
In the top journal of management sciencei don’t “recall”: the decision to delay innovation launch to **oid costly product failureThe team from Harvard Business School sets up a scenario in which every company is in a hurry to bring innovative products to market in a competitive market. However, the downside of releasing innovative products quickly may be that these products have not been rigorously tested (e.g., security risk review and alignment in large models), and products that have not been rigorously tested will be at risk of "recalls" in the future.
In this context,Through a dynamic game model, this article found that even if the risk of "recall" is high when a product is launched quickly, when competitors are expected to launch innovative products in advance, the best course of action may still be to release quickly to seize the opportunity. Because this model is more likely to make competitors go bankrupt in the event of a "recall", thus gaining the possibility of "winner-takes-all".
And against the background of the suppression chain,Although Anthropic's launch of Claude 3 may be "suppressed" by OpenAI, the launch of Claude 3 is in turn forcing OpenAI to use GPT-4The rapid launch of 5 or GPT-5 forces OpenAI to take the huge potential risk of not working well with their "star product" (similar to Google's Bard debut demo rollover).
Moreover, it is also due to OpenAI's "reputation" that the "recall" cost caused by OpenAI's overturn will be far greater than the "recall" cost that Anthropic may bear, and at this time, everyone's expectations for OpenAI GPT-5 will in turn become a real double-edged sword for OpenAI.
And if we turn our attention to OpenAI's suppression chain itself, the fundamental reason why this logic takes effect is also the high degree of "homogeneity" of market products.
The large models of each family, whether it is OpenAI's ChatGPT, Anthropic's Claude, or Google's Gemini, have not shown obvious "irreplaceability" at present, and even in the domestic large model "100 model war", most of the functions that the model can provide are on the way to catch up with GPT, at this time, a more advanced GPT-like model will inevitably suppress the remaining 100 or 1,000 models.
However, if various large models seek differentiated competition, such as vertical domain models and small terminal models, OpenAI's suppression will not necessarily take effect
In fact,Although OpenAI still firmly occupies the "first-mover advantage", this position of "industry leader" is not unbreakable, although OpenAI has the identity of "the top of the suppression chain", but this suppression strategy can not always be effective
For market leaders, in addition to the "first-mover advantage", there is also a "late-mover advantage", just like many domestic large-scale model companies have done, on the road to openAI's success through trial and error, the "trial and error cost" paid by late-mover companies will be greatly reduced.
And just as Musk sued OpenAI, mature bosses often face more unexpected and even potential threats, and these "opportunities" also make it possible to catch up with OpenAI.
Well, in fact, outside of the release of Claude 3,We may be able to find that under the praise and expectation of OpenAI in ** and the community, the "world's strongest model change hands" behind Claude 3 may not be as simple as a percentage increase in a few datasets!