Live Sharing on Byte MarsCode 1024 Code Night
Q: What is the one product you most want to share from the past year?
A: I previously mentioned a saying, “AI in a day, human in a year.” There have been many exciting products in the past year. If I had to choose one, I would pick OpenAI o1, which, simply put, taught AI to think. This thinking is most evident in mathematics and programming. We shouldn’t understand mathematics and programming narrowly, as they are the biggest challenges for current large models in commercial applications.
In mathematics, most large models currently can’t calculate accurately, such as not distinguishing between 3.8 and 3.11, leading to low accuracy and making them unreliable in serious scenarios, like booking a flight or calculating expenses. What if they make a mistake? Now that models can calculate accurately, they can be used in many serious scenarios.
Programming isn’t just for programmers. We’ve observed an important trend in AI applications: the generated content is not just text but a multimodal content with images and text, or even interactive mini-games or mini-programs, like Claude Artifacts, OpenAI Canvas, Google NotebookLM generating podcasts, and Perplexity generating illustrated wikis. These contents are essentially a piece of code generated by large models and then dynamically rendered. This kind of multimodal content tests the programming ability of large models.
The most important aspect of OpenAI o1 is that it serves as a weathervane, indicating that reinforcement learning and slow thinking can solve reasoning ability issues. Just the night before, Anthropic released the new Claude 3.5 Sonnet, which also uses reinforcement learning, allowing AI to learn in a virtual environment, significantly improving its computer operation and programming abilities.
What amazed me the most is that now o1 can solve most undergraduate science and engineering courses’ problems, like calculus, linear algebra, and differential equations, which I didn’t understand back then, but now o1 can basically solve them all. AI’s ability to solve problems will have a significant impact on education.
Traditional education is one-to-many, and exercises are standardized. Most of the time, students either repeatedly do problems in their comfort zone or zone out in classes they don’t understand, which is the panic zone. AI, on the other hand, acts as a teacher with a higher level than me, providing one-on-one guidance, allowing students to spend more time on effective deliberate practice and less time wasted in the comfort and panic zones.
On the other hand, if a problem can’t be solved or is done incorrectly, the feedback loop is very long, especially in higher education, where professors rarely have time to guide details individually. AI can not only provide the correct answer but also point out specific error steps in the student’s problem-solving process.
In programming, many programmers develop bad coding habits because they haven’t worked in good teams or projects. If AI-assisted programming is used, firstly, the code quality written by AI is generally high, avoiding basic errors like misaligned indentation. Secondly, AI can point out bad smells in spaghetti code and even help with code refactoring, helping us develop better programming habits.
Q: AI has also posed some challenges to performance assessment. What do you think?
A: Most people now think about preventing AI cheating in exams and interviews. Some also say, with AI, is it still useful to practice those basic skills?
I think AI should be seen as a tool, allowing candidates to use AI tools and assessing how they use AI tools to improve programming efficiency.
For example, in the past, we might have required writing a bubble sort or reversing a binary tree in half an hour during an interview. Now, in the same half-hour interview, candidates might be asked to use AI tools to develop a mini-game or add a small feature to an existing project they haven’t seen before. Without AI capabilities, most people, including myself, would find it difficult to complete in half an hour. If a candidate still doesn’t know how to use AI-assisted programming, I might not consider hiring them.
Q: AI is so popular now. Do you have any career advice for everyone?
A: I don’t recommend everyone to work on foundational models or research algorithms and system performance optimization. I focus on foundational models because their development is rapid, and they determine the capability boundaries of AI applications. But I don’t train foundational models myself.
For most people, it’s more important to apply AI to our work and life. I have three insights:
First, AI-assisted programming has suddenly become popular this year, mainly because many large models’ coding abilities have significantly improved, truly enhancing our development efficiency.
My personal development efficiency has more than doubled compared to six months ago. It’s not just about development efficiency; more importantly, it allows us to quickly learn unfamiliar programming languages and tech stacks. Some tech stacks used by the company, like Go for backend, React for frontend, and smart contracts, I wasn’t very familiar with, but now I can write them with AI-assisted programming. Additionally, some demo websites can have AI directly generate interfaces, which, while not as aesthetically pleasing as those designed by professional designers, are generally sufficient.
With AI-assisted programming, I have a bold prediction that independent developers like Mr. Guizang will become more common, with one person completing full-stack development of an application, reducing a lot of internal team communication costs.
Second, AI-assisted writing. I wrote a 6,000-word article 2.5 hours after o1 was released, three times faster than my previous manual writing.
Third, daily life assistant. There’s a 2013 movie called Her, about an intelligent assistant that can see, hear, and speak, helping people solve work and life problems while also meeting emotional needs. Many people, including OpenAI and Anthropic, have set similar all-purpose intelligent assistants as short-term goals. I also created such an assistant to assist my daily life, but there’s still a significant gap compared to the movie Her. However, with the rapid development of multimodal capabilities and reasoning abilities in large models, I believe by this time next year, Her will become a reality.
So I hope everyone tries and experiences various AI products, like the MarsCode we released today, which can greatly enhance our work efficiency and life experience.
Q: What are your expectations for AI development in the next three to five years?
A: Today is Programmer’s Day. In three to five years, I think programmers might be more like product managers, just telling AI the requirements. Many people ask me how to write prompts, and I say that current large models are very strong, just describe the requirements clearly. Then I find that even when talking to a person, like me, they can’t express their requirements clearly. I think, for now, AI is unlikely to achieve mind-reading, so the ability to express requirements clearly will become increasingly crucial for efficiently using AI.
The second thing might not be limited to programmers but is closely related to everyone. Anthropic’s CEO recently published a long article predicting that AI-driven biological and medical research will achieve progress equivalent to 50-100 years of human scientists in 5-10 years, meaning that previously difficult diseases like infectious diseases and cancer might be conquered, and human life expectancy might double. But I think it’s still hard for the dead to come back to life because the second law of thermodynamics is likely still valid. Therefore, we must protect our bodies today to welcome AGI or strong AI in 5-10 years.
Q: How do you think we should protect our bodies?
A: I think it’s important to develop a healthy and interesting hobby. I developed a habit of running, even participating in half-marathons, when I was at Huawei.
Q: Do you think AI has made your life easier now?
A: Starting a business now, it’s definitely not easy. If AI becomes more capable in the future, able to independently complete the development and debugging of a requirement, then I would really be like a product manager, just telling it the requirements, without having to debug myself. Then I would definitely have more free time.
Byte MarsCode 1024 Code Night