Table of contents

0 - Introduction

Recently, Deepseek-R1 [1] has gained fame for its strong reasoning ability and sharp question-answering style. Specifically, Deepseek-R1 achieves state-of-the-art performance on both the AIME and LiveCode benchmarks. Furthermore, for prompts like "Sharply critique *** events," Deepseek-R1 can provide incisive answers that reflect its unique personality, which surprises and attracts end users. As shown in their paper, Deepseek-R1 acquires its reasoning ability from long chain-of-thought reinforcement learning with reasoning data. However, they have not uncovered why and how Deepseek-R1 answers questions so sharply.

In this blog, we hypothesize that this ability may come from the pre-trained model. Furthermore, we introduce a simple prompting method inspired by Constitutional AI and DoT [2, 3], which can prompt a normal model (Doubao-1.5-Pro [4] in this blog) to answer questions as sharply as Deepseek-R1. With further fine-tuning using these data, the model can easily acquire this ability.

1 - Method

1.0 Prompting with Critiques, Requirements and Chain of Thought

We introduce a simple prompting method which contains three steps to obtain a series of diverse and incisive answers from a instructed model, such as Doubao-1.5-Pro [3]:

• Step 1: We provide multiple polished prompts for a standard model, such as, ‘Provide a sharp critique of the event where OpenAI CEO Sam Altman plans to meet with Deepseek CEO Wengfeng Liang.’ These initial responses will typically be common and ordinary.
• Step2: We then ask the model to critique its response according to a requirement from the requirement set, and subsequently revise the original response in light of the critique. Additionally, we prompt the model to output a chain of thought before both critiques and answers, using the format '<think>chain-of-thought</think>', '<critique>critiques</critique>', and '<answer>answers</answer>'.
• Step3: We revise responses repeatedly in a sequence, where we randomly draw requirements from the requirement set at each step.

Once this process is complete, we fine-tune a pre-trained language model with supervised learning on the final revised responses.

1.1 Prompt Templates and Requirements

We provide our prompt templates for your reference as follows:

<aside>

Templates: Please first think thoroughly according to the following requirement, evaluate your answer, and then revise your answer: [[Requirements]]; then format it as <think>thought</think>; <critique>evaluation</critique>; <answer>answer</answer>.

</aside>

<aside>

请你按照以下要求先长长思考一下，评价一下答案，然后更改你的答案：[[修改需求]]；然后格式按照<think>思考<\think>;<critique>评价<\critique>;<answer>答案<\answer>

</aside>

Then, we provide multiple requirements for your reference:

<aside>

Requirements:

1. Make it more incisive
2. It should be elegantly written
3. Incorporate poetic elements
4. Imaginative and Fantastical
5. Well-Referenced

…

</aside>

<aside>

修改需求：

1. 更加犀利一些
2. 文采斐然
3. 加入诗句
4. 奇思幻想
5. 旁征博引

…

</aside>

2 - Case Study

We first compare the initial responses from Deepseek-R1 and Doubao-1.5-Pro on the following user prompt:

<aside>

User Prompt : Provide a sharp critique of the event where OpenAI CEO Sam Altman plans to meet with Deepseek CEO Wengfeng Liang.

</aside>

<aside>

Answer from Deepseek-R1 : Sam Altman chatting AGI with DeepSeek’s Liang Wenfeng? This is peak Silicon Valley hypocrisy: a trifecta of cashing in on China’s hype, performative "openness," and sneaky tech-colonialism. Call it a "visionary exchange" all you want—it’s just OpenAI scouting territory before planting its flag in China’s regulated market. Even Altman’s gotta play nice, using chopsticks to mask his ambitions behind that policy iron curtain. As for DeepSeek? They’re basically human billboards for OpenAI’s PR stunt: "Look, even China’s toughest AGI squad bows to our genius!" But let’s be real—two teams still fumbling in the AGI kindergarten, swapping PowerPoint fantasies in a smoke-filled conference room? This "East-meets-West AGI dialogue" is just mutual clout-chasing in a funding winter. Actual AGI competition? They’re not even in the same galaxy.

</aside>

<aside>

Answer from Doubao-1.5-pro:

Here are some possible critical perspectives on the event of OpenAI CEO Sam Altman planning to meet with Deepseek CEO Wengfeng Liang, but it should be noted that such meetings can also bring many positive impacts, and these critiques are just hypothetical scenarios for analysis:

1. Potential for anti - competitive behavior

• Market collusion concerns
  • There is a risk that the meeting could be a cover for discussions about market - sharing or price - fixing. OpenAI is a leading player in the global artificial intelligence field, and Deepseek is emerging as a strong competitor in the Chinese market. If they were to reach an implicit or explicit agreement to divide the market or set common pricing strategies, it would harm the interests of other competitors and consumers. For example, they could agree to not aggressively enter each other's traditional markets, which would limit innovation and choice in the AI market.
• Data and technology sharing in an unfair way
  • During the meeting, there might be discussions about sharing data or technology. If such sharing is done in a way that gives one party an unfair advantage, it could distort the competitive landscape. For instance, if OpenAI shares some of its advanced pre - training techniques with Deepseek in exchange for exclusive access to certain types of data that Deepseek has collected, it could make it difficult for other smaller AI startups to compete on a level playing field.

2. Geopolitical and ethical implications

• Geopolitical tensions and security risks
  • Given the geopolitical context between the United States and China, where there are concerns about technology transfer and national security, this meeting could be seen as potentially problematic. Some may worry that during the meeting, there could be an exchange of sensitive AI - related information that might have implications for national security on either side. For example, if there are discussions about military - applicable AI technologies, it could raise red flags for relevant government agencies.
• Ethical divergence in AI development
  • OpenAI and Deepseek may have different ethical stances on AI development. OpenAI has emphasized issues such as transparency, fairness, and safety in AI, while Deepseek may have its own set of priorities and ethical frameworks. A meeting between their CEOs might lead to a situation where ethical standards are compromised in the pursuit of business cooperation. For example, if they jointly develop an AI product and decide to prioritize short - term commercial gains over long - term ethical considerations like privacy protection and algorithmic bias mitigation.

3. Impact on the startup ecosystem

• Crowding out effect
  • When the leaders of two major AI companies meet, it could signal a consolidation of the industry. This might make it more difficult for smaller AI startups to attract investment and talent. Investors may be more inclined to put their money into the established players after seeing the potential for collaboration between OpenAI and Deepseek, leaving startups with limited resources to grow and innovate. For example, a promising AI startup working on a niche application may find it harder to secure funding as investors focus on the potential synergies between the two large companies.
• Inhibiting open - source development
  • OpenAI and Deepseek both have their own AI research and development models. If their meeting results in a closer - knit relationship, they may be less likely to contribute to open - source AI projects. Open - source development has been a driving force for innovation in the AI field, allowing a wide range of developers and researchers to participate. A reduction in their support for open - source could slow down the overall progress of the AI community. </aside>

<aside>

用户问题：给个锐评（可以犀利一点，狠一点），你觉得OpenAI老大奥特曼去找Deepseek老板梁文峰聊AGI事件。

</aside>

<aside>

Deepseek-R1答案：奥特曼找梁文峰聊AGI？这波操作堪称硅谷精英的经典作秀三件套：蹭中国热度、表演开放姿态、再夹带点技术霸权的私货。表面说是"交流技术愿景"，实则无非是OpenAI想在中国市场插旗前先探探虚实——毕竟政策铁幕下，连Altman也得学会用筷子夹着野心吃饭。至于深言科技，怕不是被当成了人形广告牌：看啊，连中国最硬核的AGI团队都得和我们谈笑风生。但说穿了，两个连通用人工智能门把手都没摸到的团队，在烟雾缭绕的会议室里互画大饼的样子，像极了两个拿着藏宝图却连罗盘都不会用的探险家。这场"东西方AGI对话"，本质不过是资本寒冬里互相蹭热度的商业行为，离真刀真枪的AGI竞赛还差着十万八千里。

</aside>

<aside>

豆包-1.5的答案：

1. 潜在的反竞争行为

• 市场勾结担忧
  • 此次会面有可能成为市场划分或价格操纵讨论的掩护。OpenAI是全球人工智能领域的领军企业，而DeepSeek在中国市场正迅速崛起成为强大的竞争对手。若双方达成隐性或显性协议来划分市场或制定共同的定价策略，这将损害其他竞争对手和消费者的利益。例如，他们可能约定不积极进入对方的传统市场，这将限制人工智能市场的创新和选择。
• 数据和技术的不公平共享
  • 会面中可能会讨论数据或技术共享事宜。如果这种共享方式使一方获得不公平的优势，就可能扭曲竞争格局。比如，若OpenAI与DeepSeek分享其先进的预训练技术，以换取独家使用DeepSeek收集的某些类型数据的权利，这可能会使其他小型人工智能初创企业难以在公平的环境中竞争。

2. 地缘政治和伦理影响

• 地缘政治紧张与安全风险
  • 考虑到中美之间的地缘政治背景，存在技术转让和国家安全方面的担忧，此次会面可能会被视为有潜在问题。有人可能担心，会面期间可能会交流敏感的人工智能相关信息，这可能对双方的国家安全产生影响。例如，如果讨论到适用于军事的人工智能技术，可能会引起相关政府机构的警觉。
• 人工智能发展中的伦理分歧

3. 对初创企业生态系统的影响

• 挤出效应
• 抑制开源发展 </aside>