OpenAI demonstrates the official version of "AutoGPT": shopping, painting, tweeting, autonomous operation, it will do it all!
OpenAI's official AutoGPT is about to launch new features! This news comes at a time when the number of Stars in the AutoGPT project exceeded 100,000, and OpenAI also released a blockbuster. Co-founder Greg Brockman personally demonstrated the upcoming new features of ChatGPT.
For example, if you want a picture of dinner with a sense of atmosphere, the user does not need to write prompt words, and ChatGPT will automatically provide the corresponding picture.
After adopting the recipes recommended by ChatGPT, it only takes one sentence to buy ingredients, and the shopping cart will be automatically completed behind the scenes.
If you want to share the purchase list, instead of opening Twitter, the user can directly order a sentence, and the link will automatically have the purchase list.
With the addition of networking capabilities, ChatGPT can automatically process fact-checked answers. Even an Excel file can be handled easily.
The various new functions demonstrated by ChatGPT are completely in the style of AutoGPT, and at the demonstration site, Brockman revealed that these new functions will be launched in the next few months.
As soon as the speech was over, the applause burst into thunder. Previously, OpenAI publicly stated that it would introduce a plug-in function on ChatGPT, which caused heated discussions. Now, with the exposure of the actual measurement results, it is even more exciting.
A netizen said: Tweeting the real test at the TED speech site, it can be seen that you have a lot of trust in the AI technology developed by yourself. While waiting for the test results, the elder brother held his breath and concentrated, and only after the test was successful did he show a happy smile.
Let's explore the new features of ChatGPT together with my brother.
Experience new features together
On the chat interface, you can directly use the shopping cart and tweeting functions. As shown in the image below, the upcoming "Auto feature" is hidden in three different new modes: Browsing, Code Interpreter and Plugins.
The plug-in mode allows ChatGPT to access various external applications, browsing can enable networking, and the code interpreter can help you perform various data analysis by automatically writing code.
First, let's look at the plugin pattern. Brockman first demonstrated the new functions of ChatGPT after accessing Dall E.
We want to have a dinner party after the presentation, please give us a recipe suggestion and generate a picture.
Soon, ChatGPT gave a menu, including quinoa salad, grilled mushrooms with cheese, fruit plate, creamy pumpkin soup, handmade bread and cocktails, etc. Each dish detailed what specific ingredients were used.
Then through the Dall·E App plug-in, a picture of a big meal that looks very atmospheric appears.
Brockman introduced that when the Dall·E plug-in drew this picture, it was actually writing prompts for itself. In the past, humans had to do this by themselves. Now this kind of "little things" can be handled by ChatGPT.
After generating the recipe, you need to prepare the corresponding ingredients. Simply issue a command to ChatGPT: "Select the 'Instacart' plugin and type the text 'Now make a shopping list for the previously suggested recipe'."
Here, Brockman also added a small task to it: "and share it with all TED viewers on Twitter", which requires choosing a tool called "Zapier" (an automated tool that can connect various applications) plugin.
It took a little while to go through the Instacart platform, and then the link appeared (the "retrieval" plugin was also used here). After opening the link, all the ingredients have been arranged, and you can place an order directly. (Of course, before that, feel free to increase the amount of certain ingredients, or delete some that you already have.)
Then, ChatGPT reconfirmed whether Brockman needs to share the link to the Twitter platform, and Brockman answered yes. Later, ChatGPT automatically sent a tweet for him, and Brockman himself never touched the Twitter website.
We went to Brockman's homepage to check it out, and found that this tweet was real, and after clicking the link, the shopping cart interface we saw before appeared. This is exactly what the Plugins mode can do. Is there some AutoGPT feeling?
In addition to seamlessly accessing external applications, ChatGPT can also verify some facts online. Chatbots have been known to sometimes give wrong answers that aren’t easy to spot, leading to misleading responses.
Brockman asked ChatGPT OpenAl how long was the release date of the two blog articles "Unsupervised Sentiment Neuron" and "Learning from human preferences", and ChatGPT replied that the former and the latter were on April 24, 2017 and June 27, 2017, respectively. , with approximately two months between the two publications.
But is this really the case?
Now we switch to "Browsing" mode, and then throw the conversation just now to ChatGPT, let it verify by itself.
ChatGPT did not directly update the answer, but wrote down the process of its own query step by step.
First I will search for the publication date of the first article, to verify the first date. Type a dissertation title to start searching... The publication date of the paper is given in the excerpt from search result 0. I will quote this excerpt to verify the first date provided by the user. View results (clickable). Then I went to search for the publication date of the second article...the results are given here...
The final summary of ChatGPT is: It seems that the previous answers are all wrong. The first one was released on April 6, 2017, and the second one was released on June 13. The time interval between the two papers is two months and one. Zhou, this answer is correct.
Next, Brockman showed off one last new feature—a compilation of all AI papers from the past 30 years using an Excel spreadsheet, totaling about 167,000.
Now let's switch to "Code Interpreter" mode and upload this file to ChatGPT to see how it does automatic analysis.
ChatGPT first sorted out the meaning of all the column names in the table, and asked Brockman whether he wanted to do data analysis or visualization next.
Brockman said he wasn't sure what he wanted to do, but that ChatGPT could be made to generate some "exploratory graphs."
ChatGPT gives three options, including a histogram of the number of authors per paper, a time line graph of the total number of papers published per year, and a word cloud of paper titles. Brockman chose the second option, but reminded that ChatGPT is not in 2023, so let ChatGPT redraw. In the end, ChatGPT carefully stated that the total number of papers published in 2023 is predicted.
These are the new features that Brockman demonstrated in this talk. These new functions are actually the updates promised by ChatGPT a month ago. OpenAI announced that ChatGPT will have networking functions and can access various external applications. Like having an "App Store", you only need to directly call the corresponding App to complete the task. . And now, these new features are finally available to everyone (or possibly with a Plus account).
Time to learn about AGI
Brockman accepted a TED interview after demonstrating the function. Unlike OpenAI CEO Altman, he answered many substantive questions, including the expression of personal views, the sharing of successful experiences, and direct responses to doubts. For example, why did OpenAI bring this shock technology instead of Google? The OpenAI development team has only a few hundred technicians, which is a big gap compared with Google.
Everyone is standing on the shoulders of giants, and the entire AI industry has made progress in computing, algorithms, and data, but in the early days, OpenAI made some deliberate choices, such as facing reality and working closely together, to achieve these useful results.
Brockman explained why OpenAI firmly believes in the capabilities and trends of large models. They always knew where deep learning was ultimately going to go, but how should every lab act to make this happen?
He used an example of "unintentionally inserting willows and willows into shadows" to illustrate:
In the past, someone at OpenAI tried to use the model to predict the next character of the review on the Amazon platform. In the end, he not only got a model that can analyze comment syntax, but also got a sentiment analysis classifier that reached SOTA. While the algorithm might not have been amazing at the time, they had found a way to grasp the semantics.
However, the emergence of ChatGPT has brought a lot of doubts. Many people feel that ChatGPT has not really mastered knowledge after experiencing it. The moderator asked whether scaling up the model and getting more human feedback is needed to be truly successful?
Brockman's answer is yes, he believes:
OpenAI has always been a face-to-face method with reality. It has to break through the technical limits to see its actual capabilities, and this is how they turned to the new paradigm.
Today, people who have concerns about AI ethics have also received a lot of doubts about the firmness of OpenAI. Some of them believe that as a non-profit organization, OpenAI not only set off a global technology trend, but also prompted many technology giants to join in, which may lead to AI. Development poses great dangers. Therefore, Musk and thousands of scientists called for a six-month moratorium on AI development.
In this regard, Brockman made it clear:
They have considered how to build general artificial intelligence from the very beginning, hoping that it can benefit all mankind. If it's developed in secret and then rolled out after it's confirmed safe, he doesn't know if it's right, it's scary, it doesn't feel right. So he thinks the only alternative is to give people time to make suggestions before the machines are perfect.
Brockman explained that the development of technologies such as computers and algorithms by humans is a step by step process, and it is necessary to understand how to manage them well as each stage is advanced. Just like raising a child, we guide it together and set up rules for it instead of teaching it to destroy human beings. Now that AGI is poised to change every aspect of how we use computers, it's time we all understand the technology.