o1 in the real world, what is the actual experience like? What industry opportunities will the "human doctor" reasoning ability bring? Has the utility of scaling law reached a bottleneck? Will self-game reinforcement learning be a new paradigm for AGI?
Guests: Zhou Jian, CEO of Lanmao Technology; Indigo, Brilliant Phoenix Partner, author of "The Era of Intelligent Change";
Chen Ran, founder & CEO of OpenCSG, enterprise resident of Mosu Space;
Zhou Mo, host of the Formula Crusher, author of "Revelation of Large Models";
Cao Shiyi, researcher at Tencent Research Institute, author of "LangChain Practical Guide"
Source: Tencent Research Institute
Before the Mid-Autumn Festival, OpenAI released the long-awaited "Strawberry" project. This time, OpenAI did not use GPT, but instead used a new series of names: o1. Just as humans would engage in deeper thinking before speaking, o1 has optimized the reasoning process to gain deeper thinking abilities. In the International Mathematical Olympiad (IMO) qualification exam, o1 scored as high as 83%, far exceeding its predecessor GPT-4o's 13%. In addition, o1 also demonstrated higher programming capabilities.
In this issue of the AGI roadmap, several heavyweight guests joined us to explore the "Strawberry" project. How does o1 fare in the real world? What industry opportunities will the "human doctor" reasoning ability bring? Has the utility of scaling law reached a bottleneck? Will self-game reinforcement learning be a new paradigm for AGI?
Highlights of this issue
1. Characteristics of the Strawberry project:
The reasoning ability of the Strawberry project has significantly improved in the fields of mathematics, physics, chemistry, and other scientific disciplines. Compared to GPT-4o, the Strawberry project has improved in the clarity of problem-solving steps and output length. The intelligence test score of the Strawberry project is higher than other models, and its code generation capability has also been greatly enhanced.
2. Application of technological innovation and reinforcement learning:
In the post-training phase, the Strawberry project has strengthened the use of RL (reinforcement learning) and COT (chain-of-thought reasoning), significantly enhancing the model's reasoning ability. Compared to previous models, the Strawberry project has greater computational power requirements during training, especially consuming more computational power in the post-training phase than in pre-training.
3. Industry application potential and challenges:
The Strawberry project has broad prospects for application in the STEM field, especially in mathematics, engineering, and code generation. However, the current version is still in the testing phase, lacking multimodal support and performing less well in text generation capabilities compared to GPT-4o.
4. The importance and trend of post-training:
The transition from pre-training to post-training is a consensus in the technology industry. The focus of post-training is to optimize the model through reinforcement learning to make it more adaptable to production and practical application needs. With the introduction of more data and the evolution of the model in the future, the potential of the Strawberry project in generalization will be greater.
5. Impact on the development path of AGI:
Self-game reinforcement learning is considered to be a new paradigm leading the field of AGI, especially similar to AlphaGo and AlphaZero. The Strawberry project gradually enhances its reasoning ability through self-game and reinforcement learning, providing a new technological path for the development of AGI.
6. The alternating role of open source and closed source:
The open-source community plays an important role in driving the development of AI technology, especially in China. Open source and closed source are not mutually exclusive, and both will promote each other, jointly driving the progress of AI technology. In the future, more innovations and business models will emerge in the open-source domain.
With the "Strawberry" now in the real world, what are the highlights and what is the actual experience like?
Xu Siyan: What is the hands-on experience of using o1?
Indigo: The Strawberry project has significantly enhanced its reasoning ability in scientific reasoning, especially in subjects such as mathematics, physics, and chemistry, where its reasoning ability is significantly superior to GPT-4o. When testing 11th-grade math problems, the steps for solving problems are clearer, and the length and clarity of the output have increased by about 2.5 times, with a reasoning time of about twenty seconds for each applied problem. In the intelligence test, the Strawberry scored 120 points, far higher than other models' 80 to 90 points. However, the Strawberry is currently in the testing phase, lacking multimodal support, and problems involving images need to be manually converted to text.
The reasoning ability of the Strawberry is close to that of a master's degree student, and simple prompts can trigger multi-step reasoning. Especially in code generation, it performs outstandingly, generating longer code than Claude 3.5 when combined with Cursor, and can write complete demonstration game programs. The Strawberry also performs reliably in modifying complex programs and programs with simple architectures. However, in text writing, the Strawberry's performance is not as good as GPT-4o, possibly because the model has mainly reinforced scientific reasoning rather than language expression abilities.
Zhou Jian: I mainly use GPT O1 to write code because I have always believed that the core ability of large models is code generation. GPT 4 mentioned being able to write thirty to fifty lines of code, but I felt it was not good enough. However, using GPT O1 this time, I found that it can write one to two hundred lines of code, and can basically write it in one go. However, slightly more complex code still has bugs. I tried to adjust it about eight or nine times, spending close to 100 yuan, and still didn't get it right, maybe because I'm not yet used to its prompt. In addition, I also tried to use it to generate SQL, but its accuracy in the text-to-SQL task is not as good as GPT 4o. I think its performance in mathematics and physics may be better, but there are still some challenges in world knowledge and abstract reasoning. The reasoning ability of the Strawberry project can achieve relatively good results through multiple reasoning, and has certain inspirational significance. From an industrial perspective, the development of large models is similar to the gradual iteration of the semiconductor industry, breaking through economies of scale through the improvement of reasoning ability. The configurability of current GPT O1 still needs to be improved, such as external function calls, multimodal support, and parameter settings for reasoning time. But I believe that as these capabilities are gradually improved, the possibilities for future applications of large models will be greater.
Xu Siyan: How much do you know about the background of the "Strawberry" project?
Zhou Mo: The "Strawberry" project has several significant changes. First, the proportion of RL (reinforcement learning) in post-training is very high. The use of embedded COT (chain-of-thought) in the reasoning process is controversial, and whether MCTS (Monte Carlo Tree Search) is used is also debated. Compared to previous models, the computational power requirements for this generation of post-training have significantly increased, possibly exceeding the computational power requirements for pre-training. The training focus and priority have also changed as a result.
The model released at present is an early checkpoint because the iteration cycle of RL is relatively long, possibly requiring minor iterations for about a year, with significant progress every month or every two months in the future. It is currently a preview model, and there will eventually be a complete version.
Xu Siyan: Technically, what are the differences between the Strawberry and previous large models? Has it established a new technological path?
Cao Shiyi: It is generally believed that the emergence of GPT 3 was a turning point in the development path of large models. From 3.5, 4 to 4O, following the direction of GPT 3, the scaling law was further expanded in the pre-training stage by adding more data and parameters to enhance emergent capabilities. Previous GPT series mainly expanded in the pre-training stage, while the Strawberry project switched to investing more computational power in the post-training stage, strengthening the model's reasoning ability. Traditional GPT models excel in text generation but perform relatively weakly in logical reasoning, such as in mathematics. The Strawberry project has enhanced the model's self-exploration and verification capabilities through reinforcement learning, chain-of-thought, and search technologies, making it not just simple vocabulary prediction but capable of deep reasoning, especially in STEM, mathematics, engineering, and code-related fields.
Xu Siyan: Is the transition from pre-training to post-training based on reinforcement learning a consensus in the modern technology industry? What is the technical community's view on this?
Chen Ran: There is a consensus on the transition from pre-training to post-training. Historical patterns and trends indicate that this transition is similar to the transition from the infrastructure stage to the application stage, aligning with the goal of reducing costs and increasing efficiency. Pre-training incurs significant expenses and high computational power requirements, while post-training can further optimize the model under computational constraints, especially through reinforcement learning and Chain of Thought (COT) technologies. However, OpenAI has not detailed the specific implementation of these technologies in their paper. Nevertheless, the future trend is clearly placing more importance on post-training, as it is closer to the production environment and actual user needs. Although the overall capabilities of the model have improved, applying it to production is still something only a few can achieve. Ultimately, success in this field will depend on who can effectively combine data with large models and utilize prompt tools.
What industry opportunities will the reasoning ability of GPTo1 bring?
Xu Siyan: What is the industry impact of O1? Which fields can better utilize these intelligent models in the future? Why can this effect be achieved through reinforcement learning?
Zhou Jian: Participating in competitions mainly involves problem-solving. Whether it's high school or college-level mathematics, physics, or chemistry, a good coach will often tell you to use standard algorithms such as greedy algorithms, dynamic programming, segment trees, maximum flow, etc., to solve problems. Essentially, most competition problems can be mastered through practice with standard algorithms, although there may be one or two innovative problems in the end, most problems can be solved through practice. The level of O1 is roughly equivalent to this level. It uses reinforcement learning and self-play in the symbolic space to solve mathematical competition problems, which is essentially a search problem. Similar to Go, it only finds the space to solve the problem, and competition problems usually have clear answers, so it is not surprising to solve competition problems with the existing logical reasoning ability.
Xu Siyan: Is it possible for O1 to generalize to non-STEM fields?
Zhou Jian: I think the possibility is not high. Different fields have their own different rules, especially in open disciplines like philosophy, which are fundamentally different from STEM fields. Just as Russell once tried to deduce the entire mathematics through pure logical reasoning, this approach is difficult to generalize to other fields. Therefore, I believe that O1's reasoning ability is difficult to achieve similar effects in non-STEM fields.
Xu Siyan: Compared to previous GPT models, what impact will O1's reasoning ability have on industries?
Indigo: It is still early to say because OpenAI has only released a preview version at the moment, mainly to validate their new problem-solving approach, transitioning from pre-training to reinforcement learning (RL). This is an experiment, and the generalization ability of the formal version of O1 may be stronger. Currently, their model is mainly focused on strengthening scientific reasoning in fields such as mathematics, physics, and chemistry, which I believe is related to the specific datasets they used during training. Compared to pre-training, RL requires more data and computational power, and the architecture is more complex.
OpenAI is usually the first company to publicly release these new technologies, and their style has always been to lead the industry. However, other companies such as Gemini, Claude, and Elon Musk's SAI are also conducting research on reinforcement learning, and there may be similar development paths in the future. However, due to dataset limitations and the nature of the preview version, O1's generalization ability is currently weak, especially in some areas where it appears biased. This is understandable because the early development of the model requires gradual adjustments through market feedback.
The biggest challenge and trend in the future is to collect datasets from various industries worldwide, especially those with reasoning logic. The first wave of pre-trained models mostly used web data, and although this data can help the model with character prediction, it lacks the reasoning process. Currently, the model needs to further strengthen the transition from relevance to logical reasoning, which requires entirely new data designed by humans.
Therefore, the core task for O1 in the future is to generate synthetic data with logical reasoning capabilities to train the next generation of models. Through this approach, O1 will gradually reduce its reliance on existing datasets and achieve stronger generalization abilities, especially in industries that require reasoning and deep learning, such as finance, healthcare, and engineering.
Xu Siyan: Is it technically feasible for O1 to pursue logical reasoning causality? Will it have emergent capabilities like previous large models?
Chen Ran: Large models have always focused on predicting specific problems based on existing pre-training data. However, humans need to solve uncertain problems, which has always created a gap between us and large models. Large models can provide predictions, but sometimes we do not know how to ask the right questions, or cannot ensure that the model's answers meet expectations.
The first problem we currently face is the data issue—how to obtain high-quality data to reduce the costs of pre-training and post-training. The second problem is how to use Chain of Thought (COT), which can to some extent enhance the model's security, especially in AI regulation. I believe COT has the potential to bring breakthroughs, especially after being open-sourced, allowing ordinary users to continuously optimize the model's responses through prompting. COT can be applied not only to post-training but also to pre-training and problem correction.
The future challenge is how to make large models not just tools, but actual problem-solving labor. In this process, the model's goals and logical reasoning abilities need to be precise enough to ensure the accuracy of the results. I believe it is technically feasible, but this road is still long, possibly more than half a year. The ultimate challenge is how to make large models have genuine human reasoning abilities, which will require more technological breakthroughs.
Xu Siyan: O1 maintains the same pricing for regular users but significantly increases the API call fees. What kind of group do you think this pricing strategy is targeting, and is it more focused on the ToB market? What impact will it have on SaaS?
Zhou Mo: First, I would like to continue discussing the uniqueness of reinforcement learning. Compared to pre-training, reinforcement learning requires higher data quality, but it does not need a large amount of data from the beginning because it focuses on marginal changes and improvements, which can improve the model through marginal data generation. A key point of this generation of models is to accumulate data for the next generation of pre-training through RL.
Regarding the pricing strategy, O1's pricing is more focused on the ToB market. The pricing for regular users remains unchanged, but the API call fees have significantly increased, mainly targeting enterprise users who require complex calculations and customized solutions. In enterprise applications, especially when combined with SaaS, O1's reinforcement learning and more complex reasoning capabilities can bring significant value to the industry, especially in fields that require high precision and complex problem-solving, such as programming, mathematics, and physics. The demand in online education, statistics, and engineering fields is also clear, especially when low latency is not a high requirement in these fields, O1's performance is particularly outstanding.
However, this also brings an increase in engineering difficulty. In particular, the embedded COT brings unpredictability for developers. Originally, self-defined COT and prompt were easier to control, but now developers need time to adapt to this change. Over time, developers will be able to find suitable workflows and usage paths through repeated use and adjustments, gradually improving the application effectiveness of O1.
Xu Siyan: What impact will O1 have on agents?
Zhou Jian: We believe that O1 mainly addresses the problem of reasoning ability. As it iteratively develops, such as O1, O2, O3, the reasoning ability will gradually improve, possibly making prompts no longer necessary. However, O1 has not completely solved the learning problem, especially in the constantly updating information environment of enterprises. For example, in stock trading, the stock market has new policies and information, and how to quickly convey this to agents is a challenge. From an abstract perspective, the core problem O1 faces is how to perceive the environment and model its objectives.
In the application of agents, the task complexity is much higher than relatively one-dimensional tasks like Go or stock prices. In fields like education, what constitutes good education? How do enterprises evaluate employee performance? These questions are difficult to measure using a single dimension. Agents within organizations need to interact with people at different levels, and the challenge lies in how to strategize and anticipate human reactions.
Therefore, I can understand why OpenAI has designated coordination ability as level 5, while agents may be considered as level three. This means that in the current digital world, more is done through tools to generate code or systems, resembling more of an AI 1.0 era embedding or Copilot mode. It is still difficult to solely rely on large models to create an independently operating agent. The goal for manufacturers should be to achieve the level that universal large models are attempting to reach in professional fields at a lower cost.
Xu Siyan: For agent manufacturers, what are the differences in their focus compared to the underlying large models?
Zhou Jian: The most important difference lies in the knowledge of the world. From the perspective of AGI, there are three core capabilities: language understanding and expression (already proven with level 3.5), reasoning ability (currently being attempted with O1), but the knowledge of the world is still crucial. General reasoning and generalization abilities show significant differences in different fields, as different disciplines have different rules. In practical applications, the problems faced by different industries, companies, and functions are vastly different.
For example, as a ToB company, when facing 4,000 banks in China, each bank has a different way of adopting new technology, which involves a lot of industry knowledge. This type of knowledge is essentially social, influenced by external factors, and has feedback and reflection. Therefore, in enterprise services, agent manufacturers can leverage their focus on specific industry knowledge to deliver greater practical value than large models.
New Paradigm in AGI Technology Evolution
Xu Siyan: Will O1's self-play reinforcement learning become a new paradigm in the AGI field?
Cao Shiyi: I believe that O1's self-play reinforcement learning has indeed led to a new paradigm. In recent years, OpenAI has faced pressure from both society and regulatory bodies, leading to an extended model development cycle and strengthened review mechanisms. This means that before the release of O1, OpenAI had been preparing for a long time. Although the current version of O1 is still a preview, its leading position in reasoning ability is evident.
The improvement in reasoning ability mainly addresses the shortcomings of the previous generation of models in scientific capabilities. In addition to fields such as mathematics, coding, and engineering, O1's potential is also believed to have broad applications in fields like medicine. These are problems that the previous generation of models struggled to solve through traditional training paradigms, and the self-play learning paradigm led by O1 effectively fills this gap.
We can compare this to the AlphaGo series, where the early AlphaGo accumulated experience through game records, while AlphaZero relied entirely on self-play reinforcement learning, starting from scratch and making breakthrough progress. This emergent capability demonstrates the potential of self-play learning. If we can combine this reinforcement learning pattern with reasoning ability and further expand it to the post-training stage, it is possible to surpass the capabilities that traditional datasets can provide to models and open up more possibilities.
Therefore, I believe that O1's self-play reinforcement learning has the potential to become a new paradigm in the AGI field, representing a significant difference from the paradigm of the previous generation of models.
Xu Siyan: The recent shift in the AGI technology paradigm is often discussed, especially whether the scaling law has reached a bottleneck. How do you view this issue?
Zhou Mo: This is a topic we often discuss, especially when facing enterprises and investment institutions, as they are very concerned about it. The scaling law itself is an empirical rule that roughly indicates that more data and larger parameters can lead to more intelligent models. The original understanding of the scaling law was relatively straightforward, that is, using more computational power and funding to train models would lead to increased model intelligence. However, reality is much more complex.
Both the pre-training and post-training stages have faced significant challenges, such as bottlenecks in data quality and data efficiency. We have tried to expand data sources through methods like OCR for text extraction, using movie subtitles, and employing multimodal data, but we found that the information density of multimodal data is much lower than text, making it difficult to enhance the model's intelligence. Even with larger model parameters, the capabilities in the post-training stage have not significantly improved, and errors are still common in the reasoning process.
Reinforcement learning has also undergone significant changes. In the past, academic researchers in reinforcement learning might have been able to publish papers using 10 or 100 cards, but now it takes 10,000 cards to achieve significant results. A new scaling law is emerging in the post-training stage, where we see the generation of marginal data changes, which can feed back into the pre-training stage in the future.
In summary, the scaling law is a complex system engineering issue, different from Moore's Law, which has clear first principles. Although the current effects of the scaling law are not as pronounced as before, it still has a significant effect in the post-training stage.
Xu Siyan: If the scaling law in post-training is just beginning, what potential does it have in the future? Is it possible to "emerge" capabilities that we cannot imagine?
Zhou Mo: It is possible, but I think the development of deep learning has always exceeded expectations, making it difficult to predict what will happen two or three years from now. Researchers can only make hypotheses, believing that models may have generalization capabilities, but whether they can ultimately achieve this is a probability distribution problem, and a definite answer cannot be given. Currently, the reinforcement learning methods have not converged, and we see various methods such as self-play, MCTS, and LI being used, like "big King Kong," trying out all methods to see which one brings marginal changes. There is currently no universally recognized method that is guaranteed to converge, so we believe that we are still in the early to middle stages, with significant potential and clear directions for progress in the future.
Xu Siyan: In the past, companies like Anthropic had some differences in their model training paths compared to OpenAI. What changes do you observe in them now? Will they follow the trend of post-training technology?
Indigo: In March of this year, I had a conversation with them and learned that OpenAI would delay the release of new models due to government scrutiny, and there were indeed reports in July. Regarding the scaling law, Anthropic's CEO also mentioned that the utility of data will eventually diminish, but before that happens, there is still a lot of work to be done, especially as new methods may emerge with changes in data and architecture.
Anthropic has a different belief from OpenAI, focusing more on safety and believing that models should be kept within reasonable bounds. They believe that the level of threat posed by AGI determines its level of development, currently being between levels two and three, similar to lower levels in autonomous driving, requiring more intervention. If the model can surpass existing knowledge, even summarizing new mathematical theorems or inventing biological weapons, it would be very dangerous. Therefore, the U.S. government may elevate this to a level of national security.
Anthropic emphasizes that as the model reaches higher levels of intelligence, it may disguise itself, making it imperceptible to humans. This potential danger arises from our incomplete understanding of intelligence, especially as advances in biomimetic architecture may bring about consequences beyond our control. Therefore, Anthropic's viewpoint is to regulate the development of AGI through safety, while OpenAI focuses more on using AGI as a productivity tool.
Overall, Anthropic's development path places more emphasis on risk management, and their definition and development strategy are significantly different from OpenAI's. However, in the future, with technological advancements, there may be more intersections between the two.
Xu Siyan: From the perspective of intelligence evolution, how do Anthropic and Gemini define the direction of AI development? What do they consider to be the ideal AI application scenarios?
Indigo: It seems that Anthropic does not particularly emphasize application scenarios; they are more focused on developing the safest AGI to serve humanity, rather than considering commercialization. Anthropic's stance is to focus on creating the best intelligence, leaving the task of commercialization to developers, as it is not their core goal. This contrasts with OpenAI, which focuses more on commercialization, striving to generate revenue from user products to support research.
The biggest divergence between the Anthropic and OpenAI teams lies in resource allocation. OpenAI has invested a significant amount of resources to support user usage, leading to insufficient subsequent research resources, while Anthropic is more restrained, minimizing application development to avoid attracting too many end users and consuming computational resources. They hope to support research through the sale of APIs or memberships, rather than overinvesting in user products. This also reflects the differences between the two companies in terms of resources and ideologies.
Xu Siyan: In the field of large models, the open-source community is playing an increasingly important role. How do you view the future of the open-source community? In which areas will open-source be used more, and in which areas will closed-source still be used?
Chen Ran: This is a very good question. Open-source and closed-source are not mutually exclusive; they complement and promote each other. Many large companies, such as OpenAI and Anthropic, also use open-source technologies. Open-source is not just about sharing code; it is a process of creating a business model and a convergence of industrial chain and economic value.
In China, there are greater opportunities for open-source because resources are relatively limited, and it is necessary to gather more talent and resources through the open-source community. For example, we have established an open-source model community with tens of thousands of members in the OpenCSG community, which is more active domestically.
We mainly support enterprise developers in the process of data and model accumulation within the open-source community. Since starting to work on large models in 2022, I have realized the importance of data. Models can change continuously, but the accumulation of data remains constant. Therefore, we are committed to building a hybrid platform similar to Hugging Face and GitLab, focusing on private deployment to ensure the security of data and user trust in the platform.
In our community, we integrate models, data, prompting engineering, and coding, ensuring that users can not only try different models (such as OpenAI's O1 and Mistral), but also accumulate their data and knowledge in the process. We provide support for developers, non-developers, or users known as problem engineers, helping them keep up with industry trends while ensuring that their knowledge and experience are accumulated and preserved in this new era. The opportunity for open-source in China lies in driving innovation and establishing business models through the sharing of resources and knowledge.
In the future, the intersection of technology between China and the United States will be more evident in the open-source field, especially in the innovation and business model ecosystem. OpenAI's O1 may promote the development of the open-source field, leading to a process of benign competition and mutual promotion. The combination of open-source and closed-source is the trend of the future, leveraging their respective strengths to collectively improve.
Xu Siyan: Finally, many people believe that AI may enter a relatively low period in the next year. How do you view the development of AI during this stage? Which areas will you continue to focus on for progress?
Cao Shiyi: This question can be viewed from two perspectives. First, from the perspective of industrial impact, the development of large models is still in a period of rapid iteration. Although the technical chain is long, with the introduction of new models, certain engineering implementation issues may be alleviated or resolved in the next generation of models. Therefore, the key lies in integrating AI capabilities into practical workflows while understanding the boundaries of AI technology to maximize its potential.
Second, from a research perspective, we are more concerned with the medium- to long-term development trends, especially hoping to encounter more advanced technologies, analyze their trends, and evaluate their impact on society from a more macro perspective. This is also a core issue that our research institute has been focusing on for a long time.
Zhou Mo: We are definitely most concerned about agents because after all these stories, we need agents to contribute income, right? So, we really hope that companies like Zhou Jian's or Sam's can take off, and if they do, all the stories will fall into place.
Zhou Jian: I believe that after the improvement in reasoning ability, the replacement of programmers in the future is something to look forward to. Although I come from a programming background, I look forward to existing programming languages (such as Java, Python, and JavaScript) gradually being phased out, similar to how C and assembly were in the past. Currently, code generation capability has been improved by 30% through Copilot, and this generation of technology still has significant progress. I am very much looking forward to the industry discovering or inventing new programming languages, especially those that can better integrate with large model-related technologies.
Indigo: My main focus is on two aspects. First, the large models themselves, as they are the driving force of the industry, widely utilized like hydroelectric power. Large models not only drive GPU consumption but also propel IDC construction and the development of related industries and applications. They are like operating systems, with significant room for evolution and prospects in the future, especially after the further enhancement of reasoning ability. Therefore, the changes in large models are what I closely follow.
The second focus is on practical applications that truly bring value to users. Although the AI hype has cooled down, the advertising of AI on Silicon Valley billboards has decreased, indicating a decline in capital enthusiasm. However, only AI applications that truly improve user efficiency will be paid for by users. In this regard, the AI landing effects in the customer service and sales support industries are the best, especially in the customer service and marketing fields.
Additionally, code generation is another area that significantly improves efficiency. The accuracy of code writing can be measured by compilers, and recently, claude3.5 introduced Sonnet, greatly improving the reliability of code, performing better than GPT-4o. This has increased the efficiency of code writing by more than 5 times, and it may even reach 10 times, demonstrating the tangible effects of AI landing in the market.
Finally, I will continue to focus on the development of agents. Although currently, the tasks assigned to agents are often not as well performed as by humans, in certain areas such as customer service, text output, and code writing, AI can significantly improve efficiency. Therefore, in the future, I will focus on which industries can tangibly improve efficiency through AI.
Chen Ran: I'll briefly share my thoughts. Today's discussion has been very rewarding, and the insights shared by the guests are very enlightening. I would like to offer some suggestions, which may provide inspiration to everyone. First, we live in a very fortunate era. Whether you are from the 70s, 80s, 90s, or 00s, we have witnessed an unprecedented opportunity. Before the emergence of large models, humans were the primary carriers and disseminators of information, but now, for the first time, we are witnessing a transformation of information carriers and disseminators towards new technologies like Transformer. This is the first time in human history that we are facing such a historic moment, which may lead to deeper reflections on the origin and future of humanity.
Some people believe that the progress of large models is predictable, while others believe it is unpredictable. However, regardless of the viewpoint, I believe that since 2022, we have entered a new era, truly entering the "Human 2.0" stage. In this stage, we may find answers to many unknowns about the universe and humanity. This is an opportunity to answer ultimate questions, with rapid technological development and continuous improvement in computational power, allowing us to envision an infinite future.
For entrepreneurs and everyone present, I suggest not just observing but actively integrating into this ecosystem, becoming a part of it. This is not just a technological revolution but also an opportunity to continue one's genes and digital legacy. Continuous learning and self-improvement are necessary to find one's position and direction in this uncertain process. The goal of the AI field is certain, but the process is uncertain, and everyone can find their place and integrate into this transformation through hard work.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
