Several Judgments on DeepSeek and the Second Half of Intelligence

From the beginning of 2Q3, we have been talking about the second half of AI. Although OpenAI o1 proposed the RL narrative, it did not break through due to various reasons. DeepSeek R1 solved the RL puzzle, advancing the entire industry into a new paradigm and truly entering the second half of intelligence.

There has been a lot of discussion about what DeepSeek is and why it is being discussed. The more valuable discussion now is how the AI race should proceed. Summarizing the thoughts of the past half month, I hope to provide a roadmap for exploring the second half, to be reviewed periodically. Here are a few of the most curious questions:

• Where will the next intelligent breakthrough Aha moment come from?

• If you have sufficient exploration resources, where would you invest them?

• For example, the next generation of Transformer architecture, breakthroughs in synthetic data, and more efficient online learning methods, what are your bets?

1. Has DeepSeek Surpassed OpenAI?

DeepSeek has undoubtedly surpassed Meta Llama, but compared to first-tier players like OpenAI, Anthropic, and Google, there is still a gap. For example, Gemini 2.0 Flash has a lower cost than DeepSeek, and its capabilities are also strong, and it is fully modular. The outside world has underestimated the capabilities of the first-tier players represented by Gemini 2.0, just because they did not open-source and achieved such explosive results.

DeepSeek is very exciting, but it cannot yet be called a paradigm-level innovation. A more accurate description is that it has opened up the previous OpenAI o1 semi-enclosed范式, pushing the entire ecosystem towards high transparency.

From a first-principles perspective, it is difficult to surpass the first-tier model manufacturers within the Transformer architecture. It is also difficult to achieve overtaking on the same path. Today, I am more looking forward to someone exploring the next generation of intelligent architecture and paradigm.

2. Has DeepSeek Started a New Paradigm?

As mentioned earlier, strictly speaking, DeepSeek did not invent a new paradigm. However, the significant meaning of DeepSeek is to help the RL and test time compute paradigm truly emerge. If OpenAI's initial release of o1 was a riddle to the industry, DeepSeek was the first to publicly solve the riddle.

Before DeepSeek released R1 and R1-zero, only a few people in the industry were exploring RL and reasoning models. DeepSeek provided a roadmap for the industry to truly believe that doing this can enhance intelligence, which is of great help in boosting confidence and attracting more AI researchers to study the new paradigm.

With talent, there can be paradigm innovation. With open-source, there can be more computing resources invested. After DeepSeek, the originally planned non-open-source models by OpenAI were released as o3mini, and they plan to continue releasing o3, even considering open-source models. Anthropic and Google will also accelerate RL research. The industry's advancement of the new paradigm has been accelerated by DeepSeek, and small and medium-sized teams can also try RL in different domains.

Furthermore, the enhancement of reasoning models will further help the landing of agents. AI researchers are now more confident in the research and exploration of agents. It can also be said that DeepSeek's open-source reasoning model has promoted the industry's further exploration of agents.

Therefore, although DeepSeek did not invent a new paradigm, it has promoted the entire industry into a new paradigm.

3. What is the Difference Between Anthropic's Technical Route and R1?

From Dario's interview, it can be seen that Anthropic's understanding of R1/reasoning models differs from the O series. Dario believes that base models and reasoning models should be a continuous spectrum, rather than the independent model series of OpenAI. If only the O series is done, it will quickly reach the ceiling.

I have always wondered why Sonnet 3.5's coding, reasoning, and agent capabilities suddenly became so strong, but 4o has not been released yet.

They did a lot of RL work in the pre-training base model stage, with the core being to improve the base model. Otherwise, relying solely on RL to improve reasoning models may easily lead to diminishing returns.

4. The Necessity and Contingency of DeepSeek's Surge

Two former OpenAI researchers described DeepSeek as fitting the title of "Why Greatness Cannot Be Planned."

Technically, DeepSeek has the following highlights:

• Open-source: Open-source is very important. After OpenAI started to close its technology details from GPT-3, the first-tier players no longer publicly disclosed technical details, leaving an open-source position vacant. Meta and Mistral did not take up this position, but DeepSeek has taken the lead in this race.

• Cheap: "Your margin is my opportunity." This phrase's significance is still rising.

• Internet + Open CoT: For users, these two points can bring a great user experience. DeepSeek has combined these two points, which can be said to be a bomb, giving C-end users an experience completely different from other chatbots. Especially the transparency of CoT, which makes users more trustworthy of AI, promoting the breakthrough. However, Perplexity's service was unstable, and the Perplexity team quickly launched R-1, which instead承接了大量DeepSeek R-1溢出的用户.

• RL Generalization: Although RL was first proposed by OpenAI o1, due to various operational issues, the penetration rate was not high. DeepSeek R-1 greatly accelerated the progress of the reasoning model paradigm, and the ecosystem's acceptance has significantly increased.

DeepSeek's technical exploration is an intelligent achievement worth more attention and discussion. However, the timing of DeepSeek R1's release also made this surge somewhat accidental:

• The past US has always claimed to be leading in basic technical research, but DeepSeek, originating from China, is also a highlight. In this process, many US tech bloggers have begun to promote DeepSeek, challenging the dominant position of US tech giants.

• DeepSeek was passively drawn into the debate.

• DeepSeek R1 was released just as the OpenAI Sargate $500B event began to ferment. This huge investment and DeepSeek's team's intelligent output efficiency contrast sharply, making it hard not to attract attention and discussion.

• DeepSeek caused a significant drop in NVIDIA's stock price, further fueling the debate. They probably didn't expect to become the first black swan of 2025's US stock market.

• The Spring Festival is the testing ground for products. Many super apps in the mobile internet era have erupted during the Spring Festival. The AI era is no exception. DeepSeek R1 was released just before the Spring Festival. The public was surprised by its literary creation ability, not the coding and mathematical learning emphasized during training. This content is more likely to be shared by the public, making it more viral.

5. Who is Hurt? Who Benefits?

The players in this field can be divided into three categories: ToC, To Developer, and To Enterprise (to Government):

1. ToC: Chatbots are definitely the most impacted. The mindshare and brand attention of ChatGPT have been taken away by DeepSeek, and ChatGPT is no exception.

2. To Developer: The impact on developers is limited. We see users commenting that r1 is not as good as sonnet, and Cursor's CEO also said that Sonnet still outperforms. Users surprisingly choose Sonnet in a high proportion, and there has not been a large-scale migration.

3. To Enterprise and Government: The business of enterprises and governments depends on trust and demand understanding. Large organizations' decision-making involves complex interests and is not as easy to migrate as C-end users.

From another perspective, considering closed-source, open-source, and computing power:

• In the short term, everyone will feel that closed-source companies like OpenAI/Anthropic/Google are more impacted.

  • The technological mystery of AI has been broken by open-source, and the most important mystery premium in AI hype has been shattered.

  • More realistically, the market believes that the potential customers and market size of these closed-source companies have been taken away, and the payback period for GPU investment has been extended.

  • As a leader, OpenAI is the most "eaten" among them. The dream of hoping to maintain multiple technological premiums through closed-source and semi-open-source has been shattered.

• In the medium to long term, companies with sufficient GPU resources will benefit. On one hand, the second-tier players like Meta can quickly follow up with new methods, and Capex is more efficient. Meta may be a big beneficiary. On the other hand, intelligent enhancement still requires more exploration, and DeepSeek's open-source has raised everyone's level, entering a new exploration phase that requires 10 times or even more GPU investment.

From a first-principles perspective, for the AI intelligent industry, whether it is developing intelligent technology or applying intelligent technology, it must consume a huge amount of computing power. This is a fundamental law determined by physics, and it cannot be completely avoided by technological excellence.

Therefore, whether it is exploring intelligence or applying intelligence, even if there are doubts in the short term, the demand for computing power in the medium to long term will explode. This also explains why Musk, starting from a first-principles perspective, insists on expanding clusters. The underlying logic of xAI and Sargate may be the same. Amazon and other cloud vendors have also announced an increase in Capex.

Assuming that the global AI research level and cognition are aligned, with more GPUs, more experiments and explorations can be done. In the end, it may still come down to competition in computing power.

DeepSeek is not afraid of being naked, has no commercialization concerns, focuses on AGI intelligent technology exploration, and the open-source action has significant implications for promoting AGI progress, intensifying competition, and promoting openness, with a鲶鱼 effect.

6. Can Steamed Buns Surpass SOTA?

There is an uncertain detail point. If DeepSeek used a large amount of steamed bun CoT data from the pre-training stage, and the current effect is not surprising, it is still based on the huge pre-training foundation of the first-tier players, and then open-sourced. But if DeepSeek did pre-training from scratch to achieve the current effect, that would be amazing.

Additionally, it is unlikely for steamed buns to surpass SOTA in the base model. However, DeepSeek R-1 is very strong, and it is speculated that the Reward model is doing very well. If the R-1 Zero path is reliable, there is a chance to surpass SOTA.

7. No Moat!

Google's previous evaluation of OpenAI: "We Have No Moat, And Neither Does OpenAI." This phrase is also very fitting here.

DeepSeek's wave of chatbot users has seen a significant migration, giving the market an important insight: technological progress is very iterative, and stage products are hard to form absolute moats.

Whether it is ChatGPT/Sonnet/Perplexity, which have formed mindshare and reputation, or Cursor, Windsurf, and other developer tools, once there is a more intelligent product, users have no loyalty to the "previous generation" products. Today, it is difficult to build a moat at both the model and application levels.

DeepSeek has also verified this: model = application. DeepSeek has not innovated in product form, and the core is intelligent + open-source. I can't help but think: in the AI era, any product and business model innovation is not as important as technological innovation.

8. Should DeepSeek Take Over This Chatbot Traffic and Do It Well?

From the explosion of chatbots to the present, through the response of the DeepSeek team, it can be clearly felt that DeepSeek has not yet figured out how to use this wave of traffic.

The essence of whether to take over and actively operate this traffic is whether great commercial companies and great research labs can coexist within an organization.

This is a great test of energy and resource allocation, organizational ability, and strategic choice. If it were a large company like ByteDance or Meta, their first reaction would be to take it all, and they have the organizational foundation to take it. But DeepSeek, as a research lab organization, the pressure to take over this huge traffic must be very great.

But at the same time, we also need to think whether this wave of chatbots will be a stage traffic? Is the chatbot not on the main line of future intelligent exploration? It seems that every intelligent stage has corresponding product forms, and chatbots are just one of the early forms unlocked.

For DeepSeek, from a 3-5 year perspective, if it does not take over the chatbot traffic today, will it be a miss? What if the sky shows a scale effect? If AGI is ultimately realized, what will be the carrier?

9. Where Will the Next Intelligent Breakthrough Aha Moment Come From?

On the one hand, the next-generation models of the first-tier players are very important, but today we are also at the edge of the Transformer's limits. It is uncertain whether the first-tier players can release models with generational improvement. Even if OpenAI, Anthropic, and Google release models that are 30% to 50% better, they may not be able to reverse the situation because their resources are 10 to 30 times more.

On the other hand, the landing of agents is more critical because agents need to perform long-distance multi-step reasoning. If the model is 5% to 10% better, the leading effect will be magnified many times. Therefore, OpenAI, Anthropic, and Google need to do full-stack integrated models + agent products, like Windows + Office. At the same time, they need to show more powerful next-generation models, such as the complete version of o3, Sonnet 4/3.5 opus.

Under technological uncertainty, the most precious thing is talented AI researchers. Any organization that wants to explore AGI needs to invest resources in more aggressive bets on the next paradigm, especially under the current pre-training stage where things have been aligned. With good talent and sufficient resources, exploring a paradigm with an Aha moment is the most important.

10. DeepSeek Makes Me More Confident in China's AI Talents, Very Encouraging

Finally, I hope that technology knows no borders.