DatawhaleLearning
Sharer: Mr. Zhou Ming, special guest of Datawhale
This Datawhale open source study specially invited Mr. Zhou Ming to share:
Zhou Ming, founder and CEO of Lanzhou Technology, vice chairman of CCF of China Computer Federation, top expert in the field of NLP and large-scale models.
In 1991, he received a Ph.D. from Harbin Institute of Technology, an important NLP town, and then joined Tsinghua University to teach. In 1999, he was poached by Li Kaifu, the founding dean of Microsoft Asia Research Institute, and soon became the head of the NLP research group. Until leaving in 2020, Zhou Ming has been the vice president of MSRA for many years.
Zhou Ming is also one of the most accomplished Chinese in the field of NLP today. He is not only the developer of China's first Chinese-English translation system and the most famous Chinese-Japanese translation product developer in Japan, but also became the chairman of ACL (International Association for Computing Languages), the highest-level conference in the language field in 2019.
From an academic perspective, Zhou Ming's influence in the field of NLP is also among the top in the world. He is one of the scientists with the most published NLP papers in the past decade. According to the latest data from Google Scholar, the total number of citations of his articles exceeds 34,000.
In December 2020, he resigned from Microsoft, and then joined Sinovation Ventures as the chief scientist. In June 2021, Lanzhou Technology was incubated. In July, he launched a lightweight model - "Mencius", which ranked first in the CLUE list of Chinese language comprehension evaluation. Ranked first in many indicators.
Mr. Zhou has made very high achievements in personnel training, academics, and industry. This sharing is in the form of dialogue, mainly discussing and answering questions based on the interests of learners participating in open source learning.
In this study, Mr. Zhou Ming personally read the questions raised by each learner and provided feedback one by one . In the evening, some questions will be selected for discussion.
1. Live broadcast appointment: April 26 at 20:00 p.m.
Reply " Zhou Ming " in the background to get recording and sharing
The following are the questions asked by the learners. Mr. Zhou Ming personally categorized them (easy to meet with teachers, hard to meet with teachers, thanks to Mr. Zhou Ming), to see what everyone cares about:
Education
Hello, Mr. Zhou, can you talk about the possible future development of NLP, and the impact of this wave of enthusiasm for large models on NLP practitioners, researchers and students?
Hello Mr. Zhou, what do you think of the impact of AIGC on education? What do you think is the development direction of smart education under AIGC technology?
Teacher Zhou, hello, the current chatGPT may be smarter than anyone in the world, do you think the future education is still important, is it necessary to take the postgraduate entrance examination, because the society is changing so fast, the school’s set is getting worse and worse adapt to social development
Hello, Mr. Zhou, how should LLM learn for students who are just getting started?
Hello, Mr. Zhou, do you think chatgpt4 can help postgraduate students greatly improve their learning efficiency?
Professor Zhou, the era of big models has arrived. In your opinion, how can we make innovations under this background?
Hello teacher, I would like to ask, do you think under the background of the current large-scale model, whether the research and employment of nlp will be greatly affected and enter a recession zone, because the large-scale model actually leads to the ablation of many classic nlp tasks.
Hello, Mr. Zhou, how can a liberal arts student learn AI, especially the mathematical foundation of AI. Please plan a learning path for liberal arts students. Thank you Mr. Zhou.
Good evening, Mr. Zhou, I would like to ask about how to apply ChatGPT to do research in the field of social science research, how to take advantage of its LLM model, and what research points can be better cut in the field of social science. Thank you, teacher!
Hello, Mr. Zhou, I would like to ask the incumbents who are engaged in other industries and are interested in NLP. What kind of knowledge can they learn at this stage to make better use of these large models? Will they be replaced by others? It is convenient. Follow-up to seek better career development (such as financial auditing field, financial industry)
Mr. Zhou, what does this wave of large-scale models mean to individuals, and what are the important influences and opportunities? Can it achieve great empowerment for individuals? What are the main aspects? Thank you, teacher
Hello, Mr. Zhou! As an employee who has just graduated and joined the company, I am transforming into computer vision. The unit has requirements for innovative scientific research, but I feel that my innovation ability is limited. On the one hand, I feel that I have a lot of basic knowledge to learn. I feel that AI is developing so fast, I feel that I can't keep up, and I feel powerless. May I ask the teacher what should I do to catch up with the wave of AI and not be eliminated, thank you teacher~
Hello, Mr. Zhou, in today’s rapid development of ChatGPT, as a grassroots developer, how can we adjust our development to maintain our own value?
Hello, Mr. Zhou, do beginners still need to learn traditional machine learning theories and methods when large models are in power?
Hello teacher, I would like to ask how to find the direction of demand in the context of the emergence of AI, and how to find the direction of AI?
Technology category
Mr. Zhou, is it necessary to continue to develop knowledge graphs under the condition of rapid development of large models? If the development is llm+kg, how do you think the two should be combined?
Hello, Mr. Zhou, what do you think of the intelligence of LLM? How far is it from the real intelligence with self-thinking and creation? And the future development direction of NLP?
Hello, Mr. Zhou, I would like to ask: In traditional NLP tasks (such as extraction, generation, classification, question answering, etc.), what effect can a large language model achieve now? Do they have significant advantages over smaller models? If large models can surpass small models, does it mean that in the future everyone can complete NLP tasks without requiring too many NLP engineers?
Hello, Mr. Zhou, among the current gpt models, the relatively small 4b and 8b models seem to be able to achieve the effect close to GPT4 when trained based on the training materials generated by GPT4. If you only need a model for a specific scenario, such as the scenario of writing sql code instead of humans, can fine-tuning based on the 4b, 8b scale model also get good results? Will the future research trend be bigger and bigger along the route of GPT, or make the large model smaller while ensuring the accuracy, so that everyone can deploy it offline?
Hello, Mr. Zhou, I have two questions: Now more and more people mention that traditional NLP technology no longer exists, does this mean that application-oriented NLP beginners should spend less time intersecting in the past? In the study of traditional nlp theory?
May I ask to what extent the potential of the current large model has been tapped? If more data and computing power are added, how much will the model performance increase?
Hello teacher, I want to know whether artificial intelligence has really reached a bottleneck period, and it is necessary to start to improve the algorithm from the scale instead of from a more optimized model like chatgpt?
I would like to ask the teacher next week how to see the impact of the large model on CV. At present, it is generally done in a multi-modal way to complete the task on CV. I would like to ask what do you think of the purely visual large model. And for scientific researchers, the large model feels that many scientific researchers are in a situation where there is no research to do. What do you think of this?
Hello, teacher, the current big language models have a very serious common problem, that is, there will be serious hallucinations, and they will make up content that does not exist at all. What technical routes can be used to solve this disease?
Now that LLM is applied in the field of text generation, how to control and evaluate the quality of the generated data?
It seems that all kinds of AGI applications now use heresy methods to convert all kinds of required information into natural language texts, and then hand them over to LLM for API calls and self-iteration. This is obviously a stupid way. Is there any What promising research directions can replace the "glue" model of LLM?
Hello, Mr. Zhou. Compared with GPT-3.5, is there any change in the training phase of GPT-4? Is there any difference between the subsequent instruction fine-tuning process and GPT-3.5?
I am a business AI algorithm engineer, but when faced with the situation that a large AIGC model can unify all tasks, I feel particularly anxious and worried about being eliminated. How do you think algorithm engineers should deal with such changes?
Hello Mr. Zhou, what is your current view on the domestic big data model? What do you think are the exploration fields of domestic universities or companies that deserve attention? What do you think about the safety of AI?
Hello, Mr. Zhou, I would like to ask you how to use the open source big language model to build a personalized localized big language model for your own company?
Hello, Mr. Zhou, what are the research and development directions of LLM in the field of human-computer dialogue in the future? Which prospects are worth investigating?
Hello, Mr. Zhou, what is your opinion on the technical path of LLM's self-refine and self-improve?
What is the fundamental reason why ChatGPT and GPT4.0 are far superior to other similar LLMs?
What is the underlying reason or mathematical principle behind the emergent ability of LLM? Is ChatGPT ahead of similar products in terms of emerging capabilities? Are there specific indicators to measure?
After reducing the parameters to a small model by distillation and branch reduction from a large model, can the emergent ability also be produced through higher-quality data training and tuning?
application class
Hello, Mr. Zhou, in addition to natural language models, large models will also have an essential impact on the models of which industries. Will everyone be willing to share their own private domain data? If so, what is driving this change?
Hello, Mr. Zhou. What is your understanding of the finetune of LLM? In the future, may every company, even a small start-up company, be able to train LLM in its own business field cheaply? How much data does finetune need?
Hello, Mr. Zhou, currently OpenAI is not open to China, and there are compliance issues. The effect of domestic commercial LLM is still relatively weak. Many open source LLMs are not allowed for commercial use, and some Chinese that can be used for commercial use are not effective. May I ask for domestic individuals? Or small and medium-sized enterprises, how not to be left behind in this wave of changes
Hello, now that major companies and institutions are scrambling to study large-scale models as basic models, do we really need so many similar large-scale models, and will this cause a waste of computing power and energy?
How can the technology of natural language processing be applied to basic science? Such as materials, physics, biology and other subjects.
Hello, Mr. Zhou. At present, large models perform well on general unstructured data such as text and images. Do you think this advantage can be applied to structured data, such as time series data such as transportation and finance, to improve these vertical fields? Is the reasoning ability of the model?
Hello teacher, how to label data for vertical fields?
Hello, Mr. Zhou, what is the impact of AI on traditional software? Will it be embedded on a large scale in the future?
Hello Mr. Zhou! Can you talk about the development of the combination of multi-modal large models and robotics, and how to achieve AGI general artificial intelligence robots? thanks teacher
May I ask what breakthroughs have NLP made in sentiment analysis and psychological counseling? Is it the general trend to realize AI psychological counseling in the future?
Hello, Mr. Zhou, I feel that the open source ecology has played a big role in this wave of AIGC, such as github, huggingface, colab, LLaMa, etc. We not only rely on these foreign ecological learning, but even pay more and more. I would like to ask Mr. Zhou, is there any recommended domestic alternatives for such open source ecological products? And is this kind of domestic substitution worthy of starting a business?
Hello Mr. Zhou! Is there any successful case in the application of chatGPT and graph technology in the field of risk control?
Hello, Mr. Zhou, how to use chatGPT to better empower coding?
What are the possible difficulties in landing a dedicated large-scale model in the financial field?
What are the inflection points in the development of NLP, and what will be the latest or next inflection point?
Reference:
1. Dialogue with AI expert Zhou Ming: How many hurdles must be passed from scientist to entrepreneur
Participate in learning: AIGC team learning is here!
Like " Share " is a kind of support↓