HomeAI News
Autonomous AI that everyone can have: can use tools with memory, and can learn by itself! The underlying self-developed framework is about to be open source

Autonomous AI that everyone can have: can use tools with memory, and can learn by itself! The underlying self-developed framework is about to be open source

Hayo News
Hayo News
April 18th, 2023
View OriginalTranslated by Google

Why did AutoGPT become popular so quickly?

It's not that the human imagination of AI has been opened again.

There is no need to guide the AI in a sentence, let alone study the prompt project.

It is not a dream for everyone to have a super AI Jarvis like Iron Man.

And just now, this Auto AI Copilot has been mass-produced.

Probably the style of painting be like:

If you are browsing a travel app and intend to book a hotel, click on the AI assistant, and it will send a thoughtful inquiry based on the information it has:

I remember you said that you plan to travel to Singapore, is there anything I can do for you?

Then tell your situation and needs to the phone, it immediately finds out 13 solutions , and gives you the most recommended choice.

Not only that, it will also query the local weather to remind you what clothes to prepare, and can plan routes according to the weather conditions. The whole process is exactly the feeling of talking with a real assistant.

In addition to travel apps, such an AI can be built into various software such as shopping and home improvement. And the official also revealed that the underlying platform will be open source in the near future .

This makes people curious, where do these AI come from?

Produce AI Copilot with memory and self-learning

The AI Copilot seen above comes from an AGI application assistant platform called MindOS .

Its main function is to create a super AI for all kinds of software, which can not only answer questions raised by human users, but also have memory , self-learning ability , and personality at the same time.

For example, when you open the home improvement software and plan to buy some decorations, but you have no idea what to buy, AI Copilot can recommend items based on what it knows about you.

In addition to recommending good things, AI Copilot can also guess intentions based on human questions.

For example, let it inquire whether the purchased table and chairs have been shipped. AI first answered the progress of the logistics, and then felt that humans might hope that the tables and chairs could be delivered together, so they immediately put forward their own guesses.

Human: Where is my order?
AI Copilot: Hi, your order has been shipped by DHL Express and is expected to arrive at 7pm today.
Human: Are the chairs and tables delivered together?
AI Copilt: They are sold separately. You have 4 chairs that will be delivered today. You want to get the tables today too so you can have a full set on Saturday?
Human: Yes, can I cancel the order and reorder the whole set?

And it will intelligently give recommended solutions according to the situation.

For example, tell humans that the chairs that have already been shipped cannot be canceled. Instead of returning the chairs and ordering a new set of tables and chairs, it is better to buy another table to make up for it.

It is understood that the MindOS platform now supports 30 languages including Chinese and English, provides more than 1,000 preset AI characters with personalities and functions , and the accuracy rate of accurately inferring intent reaches 97% .

It is actually not difficult to customize these AI Copilots. You don’t need to understand algorithms and programming, and you don’t need to mark the data. It can be done in a few minutes.

There are many types of AI assistants to choose from on the platform, including shopping, finance, website guidance, etc.

After selecting the type, you can start personalization.

The platform supports access to APIs, data and documents, etc., and provides some skills for AI assistants, which can be selected and used as needed.

Then only use natural language description to complete the initial setting of the AI assistant.

The content that can be controlled includes the image of AI, self-introduction, dialogue style, which tasks to complete and which things cannot be done, etc.

The whole process is the same as recruiting a new employee, just tell it the work rules.

After the setup is complete, insert the code generated by the platform into the website or software to complete the access to AI Copilot.

It is understood that the principle behind MindOS is to use a large model as the underlying foundation. On top of this, the development team behind it has built a set of self-developed framework UMM (Unified Mind Model) that imitates the macroscopic structure of the human brain, adding complex thinking, memory, Services and stronger self-learning capabilities make large model content more controllable, capabilities more autonomous, deployment more flexible, and integration more in-depth .

Among them, the large model is like a general-purpose computing platform, similar to the frontal lobe of the human brain, with strong comprehension and logical reasoning abilities. But if people want to complete various tasks in the real world, they must also need the cooperation of vision, execution, motivation, memory and other areas.

Therefore, the UMM framework connects these capabilities to the large model, such as the ability to perceive and process multimodal information, remember professional knowledge, historical information and data, and self-learning. In this way, AI can complete a long chain of tasks , which is very similar to AutoGPT.

And compared to AutoGPT, the framework behind MindOS is more autonomous.

For example, AutoGPT autonomously plans an appropriate execution plan given publicly available skills (such as search).

The AI in MindOS, when it finds that its own skills are not enough, will take the initiative to search for new skills on the Internet, test and connect independently, and use the newly learned skills to further complete complex tasks.

For example, when a user asks "I want to know which Seattle customers the company has?"

MindOS's AI found that its existing capabilities could not solve this problem, so it found a CRM interface (customer information system) in the company's network, and then learned how to use it independently with the user's permission, and finally gave the company's Seattle The head customer and the specific introduction of each customer.

At the beginning of November last year, MindOS released version 1.0, a few weeks earlier than ChatGPT.

After several months of iterations, the 2.0 version was recently launched, and a round of beta testing will be opened, at which time you can experience it first.

And the behind-the-scenes team revealed that its underlying framework UMM will also be open sourced in the future.

This is definitely a big news. After all, the popularity of ChatGPT and AutoGPT is high enough, but there are still some concerns about the use and autonomous controllability. A truly commercial and mature AI framework, coupled with a complete configuration platform, I believe many people will be eager to try it.

This inevitably makes people more curious, who is the development team behind MindOS?

Former Facebook senior research scientist led the team to build

MindOS comes from Mind Universe , was founded in January 2022, and is positioned as an AGI company .

The company's team members come from Meta, Google, TikTok, Alibaba, etc. Since its establishment, it has received investment from top funds such as Sequoia China Seed Fund, Linear Capital, and Ginkgo Valley Capital.

Founder and CEO Tao Fangbo is a familiar face in the field of data mining and AI.

He graduated from Tsinghua University with a bachelor's degree, and then went to the University of Illinois at Urbana-Champaign to obtain a doctorate in computer science, under the tutelage of Professor Han Jiawei, the originator of the field of data mining.

Dr. Tao Fangbo has successively engaged in research and development work in Microsoft Research, Facebook Research, NASA, etc., and is the founder of the Neural Symbol Laboratory of Alibaba Dharma Academy. Its research and development results have been applied to NASA, Boeing, etc.; Facebook's large-scale content understanding platform has been built, with an average daily service of more than 200 million users.

In addition, he has also served as a review expert for top conferences in EMNLP, CIKM, ACL, SIGMOD, WWW and other fields.

There are two main reasons for the creation of the mind universe.

On the one hand, it is driven by the underlying technology, and on the other hand, the demand really exists.

Among them, the thinking and preparation from the technical dimension should start from 3 years ago.

The shocking release of GPT-3 in 2020 made Tao Fangbo see the unlimited potential of LLM.

This is the first model that can complete many complex tasks with one model, carrying almost all human knowledge.

At that time, discussions on AGI in the industry had been endless, and Tao Fangbo also agreed that AGI would bring changes to the future world.

How to get here? The path of the large model has gradually become clear. Several years ago, the then-research lab team started building a prototype of an AGI framework.

The technical understanding they put forward is: the large model is a brand-new computing revolution, which provides a new general-purpose computing resource, that is, a digital brain resource that can be accessed through natural language.

Compared with the traditional computing power that provides "connectivity", this computing power can now directly provide "understanding and reasoning" capabilities.

But the problem is that at present, this kind of computing power is still indiscriminate, and there is no deep integration (Grounding) with scenarios and personal needs. And this kind of integration requires very complex scene understanding and personal demand understanding.

Therefore, it is necessary to build a valuable scheduling framework on top of the large model, similar to the operating system of the new era, so that this kind of computing power not only has the ability to understand, but also deeply integrates with the scene, and can grow independently and self-sufficiently driven by the goal. Construct.

And technological innovation is only the underlying support. To go one step further, there needs to be a real demand in the industry.

In fact, in recent years, many software and applications have tended to build an AI assistant for themselves. Generally, it is some low-frequency just-needed apps, such as banks and governments; and some apps with information integration and long user decision-making paths, such as e-commerce and tourism.

For low-frequency just-needed apps, users are not familiar with the operation interface because of the low frequency of opening. Once there is a need, it often takes a while to explore the functional interface, and sometimes you even have to search for tutorials on the Internet, which is really cumbersome.

Therefore, the APP will tend to add an AI assistant to allow users to interact through natural language, just like asking a guide to complete the operation quickly.

For apps with long decision-making paths, there are generally two situations.

First, users do not yet know what they specifically need .

For example, if you are going to attend a wedding next week, what kind of gift should you prepare? Many people have no idea when they open the shopping software, so they need to search for inspiration by themselves, and it will be a relatively long process to finally confirm the product and place an order.

In the second case, the user needs to make complex shopping comparisons before completing the consumption .

For example, when buying cars, real estate, or sneakers and skin care, different people will give priority to comparing different specifications when purchasing; or when booking a hotel, you must know the price, environment, location and other information of many hotels in advance. Such decision-making process will also be more complex.

Then, if there is AI that can quickly integrate comparative information or give clear recommendations, it will improve the user experience and increase the conversion rate of platform merchants.

From this, it is not difficult to understand why Lin Songqi, co-founder and COO of Mind Universe, repeatedly expressed in the early days of entrepreneurship:

All distribution can be done again with AI, all UI can be done again with natural language, and all software (Application) is worth doing again with virtual characters.

It is undeniable that the AI-based software was initially questioned by many from the market. But with a series of combined punches from OpenAI and Microsoft, this trend has gradually been seen and believed by everyone.

After ChatGPT ignited the trend, Microsoft took the lead in integrating GPT-4 into New bing, allowing AI to penetrate into the entire process of people's daily use of search engines. This wave of operations directly made New bing’s daily active users exceed 100 million, and one-third of the millions of active users are new users.

Immediately afterwards, the Office Family Bucket also announced access to GPT-4, and launched a new feature, Microsoft 365 Copilot. Microsoft CEO Nadella said directly at the press conference: Today, we have entered a new era of human-computer interaction and reinvented productivity .

Afterwards, major domestic manufacturers also followed suit, and Ali announced that it would integrate Tongyi Qianwen into all products.

But the problem is that those who have the ability to launch large-scale models and complete AI upgrades by themselves are, after all, the "privilege" of very few technology giants. Even direct access to the API requires the enterprise itself to have a certain development team. Therefore, under the new trend, many software and application service providers are urgently looking for an AI-based software application solution with good effect and low threshold.

As an enterprise-level partner of Microsoft and OpenAI, Mind Universe, which has leading commercialization results, will naturally be "broken through the threshold" in the past few months.

According to Mind Universe, since the release of MindOS, they have received trial applications from hundreds of companies, especially after the trend exploded in January this year, the number of applications has increased unprecedentedly.

At present, they have reached cooperation with one of the top 3 e-commerce platforms in the world, a leading financial brokerage in the Asia-Pacific region, a top cross-border e-commerce platform in the world, a leading virtual human company in China, and a leading Metaverse platform.

A new starting point for the revolution of human-computer interaction

In short, the emergence of the universe of consciousness has sent a very exciting signal:

A more autonomous AI Copilot is making great strides into our lives.

Coupled with the recent explosion of autonomous artificial intelligence tools such as AutoGPT and AgentGPT, it has once again been verified that more flexible, reliable, and easy-to-use AI is the direction of the new trend.

As the founder of the AGI field who saw the trend ahead of time and took the lead in launching technology research and development and commercialization, Tao Fangbo put forward his own cognition and judgment, that is, the development of the future world will be divided into three stages:

  • AI-based applications and services : Any process that requires users to make decisions should be assisted by AI Copilot to better provide in-depth demand understanding and scenario-based reasoning.
  • The emergence of individual-centered AI : Everyone can gradually cultivate an AI Copilot that is deeply bound to themselves and can help connect everything in the process of interaction.
  • AI Copilot is connected in series to form a network : the AI Copilot of applications and services and each individual AI Copilot are connected to form a brand new network, and the distribution of services will take place on this AI Network. Instead of traditional connection-based distribution, it is distribution based on understanding and reasoning. These AI Copilots will cooperate and collaborate with each other to help humans complete complex tasks.

At present, the universe of consciousness is advancing rapidly, which is the first stage.

The second phase is also close at hand. In the second half of this year, they will release MindOS for consumers. At that time, everyone will have the opportunity to create their own AI Copilot.

Under these trends, a more shocking change is also quietly advancing, that is, the way of human-computer interaction is being rewritten .

In the past 100 years, the way of human-computer interaction has developed from the initial button-based to the command line, and then made a further leap to complete the interaction through the graphical interface.

The current computers, smart phones, etc., and the information exchange between people and software are all based on the GUI, which also allows the computer-computer interaction to develop from the initial limited to the professional crowd to the scope of ordinary people.

But anyone who has a little understanding of human-computer interaction knows that the most natural way for humans to interact is actually language.

For example, if you want to check the weather, if you use language, you only need to say "what's the weather in Beijing tomorrow?" If you use GUI, you need the user to find the software, select a city, and then check the specific date.

In the past, due to the limited capabilities of AI, this kind of language-based interaction can only be achieved at a glance.

AI voice assistants often make mechanical replies by identifying fixed sentence patterns and keywords, which brings limited experience upgrades.

Today, in the context of the soaring capabilities of large models, new possibilities have emerged.

When AI can fully understand human language and link various capabilities, human-computer interaction through language alone will become the most convenient way of interaction. The traditional way of defining the UI interface by the product manager may be rewritten.

Lin Songqi said:

In the future, when users explain the workflow through natural language, AI can connect to the APIs of various software, independently build workflows and user-specific UIs, so that all screens can become AI canvases, and each AI assistant is the user's "Ma Liang". Magic Pen".

In other words, AGI is to software, just as Made to order is to manufacturing. Before that, it was supply to demand, and then demand to supply. This is a brand new production method.

The resulting impact may not be limited to the software field, and even the hardware ecology may be rewritten.

In the future, the carrier of AI may be a mobile phone, a computer, a headset, or a robot like Iron Man.

Perhaps in the past, this kind of statement would be defined as "beautiful infinite imagination", but under the trend of ChatGPT, innovation is happening "in units of hours".

Two months ago, students used ChatGPT to write homework and take exams, which was already staggering; now, people are trying to let ChatGPT find a way to develop websites by themselves, and let GPT-4 control the robots in the laboratory to do chemical experiments...

It seems that nothing is impossible.

And people's infinite vision and expectation may also be one of the necessary factors for AI to create infinite possibilities.

So, do you think an AI Copilot for everyone will be the future of mankind?

Reprinted from 量子位 明敏View Original


no dataCoffee time! Feel free to comment