HomeAI News
AI points of the week: Countries are planning AI standards one after another, GPT open source rookies emerge in endlessly

AI points of the week: Countries are planning AI standards one after another, GPT open source rookies emerge in endlessly

Hayo News
Hayo News
April 14th, 2023
View OriginalTranslated by Google

With the popularity of ChatGPT, "AI" has become a hot topic in technology around the world. No rules, no standards. In the face of the new technological revolution, whether it is technology giants or governments of various countries, the speed of discussion on AI standards and norms is also unprecedentedly active. On the other hand, at the moment when ChatGPT has become a daily topic, many open source "AI dark horses" have sprung up on the fertile AI soil.

Countries are focusing on the standardized development of AI

Cyberspace Administration of China Solicits Opinions on AI Service Management

Recently, the governments of the United States, the United Kingdom, Italy, Canada, France, Spain and other countries have taken actions to regulate the development of AI. According to people familiar with the matter, on April 12 (Wednesday), executives and policy representatives from OpenAI, Microsoft, Google, Apple, Nvidia, Stability AI, Hugging Face, Anthropic and other companies will attend a high-level meeting of AI companies in San Francisco. Discuss how to develop AI responsibly, share best practice examples, and discuss related public policy frameworks and standards.

In China, the "Administrative Measures for Generative Artificial Intelligence Services (Draft for Comment)" issued by the Cyberspace Administration of China was released to the public on April 11 (Tuesday), which mentions that the provision of generative AI products or services should comply with The content of laws and regulations, public order and good customs should reflect the core values of socialism. At the same time, service providers must not illegally retain input information that can infer user identities, must not provide user input information to others, and must not use AI-generated content to damage the image and reputation of others and other legitimate rights and interests, and require the establishment of a mechanism for receiving and handling user complaints.

The draft for comments proposes management norms from various aspects, and is currently open to the public for comments.

Create "Western World" with ChatGPT

Watch "The Sims" with 25 AI agents

If you have ever watched "Western World" or played "The Sims", most of you will be fascinated by the daily sandbox simulation of robots/NPCs. Now with ChatGPT, everything is more possible.

On April 8 (last Saturday), a research team designed an experiment on the concept of "Generative Agents" (generating intelligent agents), using generative models to simulate human behavior agents and proving that they can produce believable The simulation of individual and sudden group behaviors, the researchers also put the demo on the Internet, allowing people to observe the "life" of 25 AI agents from the standpoint of "gods".

Throughout the experiment, there were surprising performances—these AI agents showed people’s “sentimentality” and even the “changeability” of human nature in communication, and there were also defects such as expectations—the lack of memory/simulation Creation, overly AI-based responses, etc...

But in any case, this research did arouse discussions this week, allowing people to see the rapid development of AI, and also achieved unexpected topical effects outside the experiment.

AutoGPT that can work autonomously

It's really "nothing to do with humans"

This week, the most "dark horse" AI project on Github should be AutoGPT, and it represents the "autonomous artificial intelligence" driven by GPT-4.

Some netizens tried it out for the first time. For example, they used React to build a basic website. As a result, AutoGPT built it in less than three minutes. Some media tried to make it "search online" and make a report on competing products. As a result, it can also search for information, identify authenticity, and analyze and summarize within 8 minutes.

However, such an AutoGPT has a lower limit that is unexpectedly friendly to the people. Friends who are accustomed to "one-click installation" can try the AgentGPT browser extension. Friends who are interested in studying can also deploy locally in only 30 minutes. The greater difficulty lies in In order to have an API preferably ChatGPT-4 and a deeper tuning. The omnipotence of AutoGPT also made some people exclaim: "In contrast, ChatGPT seems a bit outdated", which also shows that this AI dark horse is worth looking forward to.

👉Click to go to the "AutoGPT" data page

ChatGPT-like large model Dolly 2.0 released

True open source, free to modify for commercial use

On April 12 (Wednesday), Databricks released Dolly 2.0 and announced that this is the industry's first open-source, large-scale language model that follows instructions. It is fine-tuned on transparent and freely available open-source datasets. Model and data Sets are available for commercial purposes.

Compared with other free models before this, the advantage of Dolly 2.0 is: Compared with many existing open source models, the data set of Dolly 2.0 comes from 15,000 high-quality content "written in retreat" by more than 5,000 Databricks employees, without reference The components of ChatGPT avoid the prohibition of competing products in the OpenAI terms of service, and realize the possibility of commercialization. Compared with other commercially available large models on the market, Dolly 2.0 is proud of its intelligence and excellent communication performance.

I believe that in the open source community environment, we will see more and more powerful Dolly 2.0 in the future, and we also look forward to the better development of the open source CHatGPT large model.

👉Click to go to the "Dolly 2.0" information page

Microsoft Releases DeepSpeed Chat

Make ChatGPT-like model training more economical and efficient

In recent days, ChatGPT and similar models have sparked a wave in the AI field. In order to make the training and deployment of such models easier, the AI open source community has made various attempts. Microsoft experts have brought a A scalable system that supports end-to-end Reinforcement Learning with Human Feedback (RLHF).

Sound complicated? In layman's terms, if a research team wanted to study a ChatGPT-like model with hundreds of billions of parameters, it was very expensive and even unimaginable. It is not easy, fast, or economical to use less than 5% of the computing power of the machine. The DeepSpeed-RLHF system released by Microsoft has unparalleled efficiency in large-scale training, making complex RLHF training fast, economical and easy to promote on a large scale.

Of course, this still sounds very "far away" from us. After all, not everyone is doing super-large experiments, so on April 11 (Tuesday), a model suitable for CHatGPT-like training, which can meet various training time and resources And the DeepSpeed Chat for quality needs is here. According to the official example, just 1-2 hours of coffee or lunch break, you can try to use DeepSpeed-Chat on a consumer-grade GPU (A6000-48G) to train a training example of a 1.3 billion parameter model. When you return from your lunch break, the model will be ready.

Although this sounds a bit "far away" from us, the release of such a leap-forward and efficient model makes everyone more confident: one day, the era of one ChatGPT will come.

👉Click to go to the "DeepSpeed Chat" data page

Amazon officially launches AI programming service

Free for individual users, benchmarked against GitHub Copilot

On April 14th (Friday), Amazon officially launched CodeWhisperer, a real-time AI coding assistant, and also provides a free personal version for all developers to use.

CodeWhisperer helps developers write code quickly and safely, increasing productivity and streamlining workflows without much disruption. it

It can help developers filter out those controversial code suggestions, reduce the risk of using open source code, and at the same time scan code vulnerabilities and give suggestions for solutions.

At present, this AI coding assistant has provided services to all developers-whether you use Amazon's cloud services or not, originally regardless of whether you use Python, Java, JavaScript, TypeScript, C#, Go, Rust, PHP, Ruby, Kotlin, Write your code in C, C++, Shell Scripting, SQL or Scala.

Now, in addition to Microsoft's paid GitHub Copilot service, individual developers can also try CodeWhisperer. After all, it's free, and it may be very fragrant.

👉Click to go to the "CodeWhisperer" data page

This week's AI application recommendation

Forever Voices: Have voice conversations with celebrities such as Jobs and Musku

👉 https://www.hayo.com/entry/1993

BanterAI: Another AI voice software that can talk to celebrities

👉 https://www.hayo.com/entry/2043

Openai Bridge: OpenAI plugin designed for Blender

👉 https://www.hayo.com/entry/2020

AmyMind: AI mind map

👉 https://www.hayo.com/entry/2044

Duomo Intelligence: Generate documents, flow charts, and mind maps with one click of AI

👉 https://www.hayo.com/entry/2045

AutoGPT: A language model that can run autonomously

👉 https://www.hayo.com/entry/2134

TeamSmart AI: Build an AI think tank for your Google browser

👉 https://www.hayo.com/entry/2137

This week, the positive response of governments around the world has raised the topic of "AI" to a new dimension. It is no longer a cutting-edge tool for technologists, but a new thing related to people's lives. Just like computers that changed from expensive home appliances to production tools, AI is also integrating into our lives, and one day we will no longer be unfamiliar with its existence.

Because it's there.


no dataCoffee time! Feel free to comment