Network Security Internet Technology Development Database Servers Mobile Phone Android Software Apple Software Computer Software News IT Information

In addition to Weibo, there is also WeChat

Please pay attention

WeChat public account

Shulou

The father of reinforcement learning joins AGI to start a business! Join hands with legendary programmer Carmack, saying that we don't rely on big models.

2025-01-30 Update From: SLTechnology News&Howtos shulou NAV: SLTechnology News&Howtos > IT Information >

Share

Shulou(Shulou.com)11/24 Report--

Legendary programmer John Carmack has teamed up with Richard Sutton, the father of reinforcement learning, All in AGI.

The goal of showing general artificial intelligence to the public in 2030 is feasible.

And different from the mainstream methods, it does not rely on the large model paradigm and pursues real-time online learning.

The pair made the announcement at a special event at the Institute of Machine Intelligence (Amii) at the University of Alberta where Sutton teaches.

Sutton will join Carmack's AI startup Keen Technologies while maintaining his teaching position in Alberta.

Both men admitted at the event that Keen Technologies's team was small compared with large companies with hundreds of employees.

At present, it is still in its infancy, and the entire technical team of the company is on the scene-only these four people are standing.

Its financing scale is 20 million US dollars, which is incomparable with billions of dollars such as OpenAI and Anthropic.

But they believe that in the end, the source code of AGI is of the order of magnitude that one person can write, perhaps only tens of thousands of lines.

And the current AI field is in the most leveraged special moment, small teams also have the opportunity to make a big contribution.

The legendary story of Carmack, the legendary programmer and father of reinforcement learning, is well known from the development of the world's first 3D game, to the transformation of rockets, to joining Oculus and becoming a key figure in Meta VR.

Later, he formed a bond with AI and was also related to OpenAI.

In another interview, he revealed that Sam Altman had invited him to join OpenAI and thought he could play an important role in system optimization.

But Carmack thought he didn't know anything about the modern AI of the machine learning paradigm, so he didn't agree.

This was an opportunity for him to get to know AI.

He asked Ilya Sutskever, the chief scientist of OpenAI, for an entry-must-read list, taught himself from scratch, and first got a basic understanding of traditional machine learning algorithms.

When he had free time and planned to continue to dabble in deep learning, he had a week-long programming challenge:

Print several classic papers of LeCun and practice in the case of network disconnection, starting with pushing the reverse propagation formula.

A week later, he ended up with a convolutional neural network rubbed with his C++ hands, without the help of the modern deep learning framework on Python.

I can only say that I admire the god.

At this time, his main business is still in Facebook (later renamed Meta) Oculus Research VR, leading the team to launch products such as Ouclus Go and Quest.

However, in this process, he and the management of the company also gradually developed contradictions and differences, that the company's internal inefficiency, but also publicly expressed dissatisfaction.

In 2019, he resigned as Oculus CTO to become a "consultant CTO" and began to turn more of his energies to AI.

In August 2022, he announced that Keen Technologies, a new AI startup, announced that it would raise US $20 million, with investors including Sequoia Capital and former GitHub CEO Nat Friedman.

Later, he also revealed that he could handle it with a mere US $20 million.

But taking money from others can give him a sense of crisis and urgency and a stronger determination to get things done.

At the end of 2022, he officially left Meta and saw VR as a stage of his life that had passed, and then turned entirely to AI.

In addition to this obvious main line, Carmack and AI also have some inexplicable fate.

At that time, his 3D games inspired the demand for graphic computing, and GPU began to grow from the game field.

To this day, it was GPU's math that supported AI's outbreak, and he is still proud of his contribution when he talks about it.

……

Sutton, another protagonist today, is also a legend.

He is known as the father of reinforcement learning, making important contributions to strengthening methods such as time difference learning and strategy gradient, and is also the co-author of reinforcement learning standard textbooks.

In 2017, he joined DeepMind as a distinguished scientist and participated in a series of AlphaGo studies, and his student David Silver was one of the principal directors of AlphaGo.

Sutton wrote a famous essay, The Bitter Lesson, arguing that it would not work to try to teach human experience to AI. So far, all breakthroughs have relied on the promotion of arithmetic, and the right way is to continue to make use of the scale effect of arithmetic.

Before the formal communication between the two men, Carmack had expressed his concern and approval for this article.

But it was Sutton who took the initiative to communicate directly between the two.

A few months ago, after Carmack announced financing for AGI startups, he received an email from Sutton.

Sutton wants to ask him whether he should take a purely academic, commercial or non-profit organization on his way to research.

However, in the follow-up email exchange, the two found amazing consistency in the direction and concept of AI research, and gradually established a cooperative relationship.

Specifically, the two reached four consensus:

They all believe that the current development of AGI is limited to a few narrow directions, relying too much on big data and big computing power to the neglect of innovation.

It is believed that commercialization too early will hinder the development of AGI.

It is believed that in the end, AGI will not be too complex, one person can master all the principles, and even one person can write the main code.

It is agreed that the emergence of an AGI prototype in 2030 is a feasible goal.

Not only rely on large models, small teams also have the opportunity to have bold goals, and the live audience also thinks so.

Faced with the question of "how can a small team achieve such an ambitious goal", Carmack believes that the amount of data and computing requirements required to achieve AGI may not be as large as expected.

Videos shot at 30 frames per second of what humans see throughout the year can be mounted on thumb-sized USB drives.

On the other hand, 1-year-old children only have so much empirical data that they have shown obvious intelligence.

If the algorithm is right, there is no need to use the entire Internet data for AGI to learn.

As for the demand for computing power, he also uses this intuitive thinking to consider: the computing power of the human brain is also limited, far less than the level of a large computing cluster.

Larger than a server node (node) and larger than a cabinet (rack), but the maximum is only an order of magnitude higher.

And with the passage of time, the algorithm will be more efficient, and the required computing power will continue to decline.

If Carmack has one thing in common in the seemingly disparate areas of 3D games, rockets and VR, it's the optimization of large real-time feedback systems.

That's what Sam Altman liked when he invited him to join OpenAI.

The AGI architecture he envisioned should be modular and distributed, not a huge centralized model.

Learning should also be continuous online learning, not that most of the parameters will not be updated after the current pre-training.

My bottom line is that if a system can't run at the frequency of 30hz, that is, update every 33 milliseconds or so during training, I won't use it.

He further said that as a low-level system programmer who can write raw Cuda code and manage network traffic on his own, he may do some work that no one else would even consider.

Even not limited to the existing deep learning framework, will try a more efficient network architecture and computing methods.

The overall goal is to simulate a virtual agent with intrinsic motivation and continuous learning ability to learn continuously in the virtual environment.

No robots, because the experience of making rockets made him think that the fewer physical objects he was dealing with, the better.

Sutton has spent decades on this issue than Carmack has just set foot in AGI, and he has more specific research plans.

Although not much has been said about this event, the main part has been written in an arXiv paper in the form of the Alberta Project.

The Alberta Plan proposes a unified agent framework that emphasizes universal experience rather than special training sets, focuses on time consistency, gives priority to methods that produce economies of scale with computing, and multi-agent interactions.

A 12-step road map was also proposed.

The first six steps focus on designing the continuous learning method of model-free, and the last six steps introduce environmental models and planning.

The last step is called intelligence enhancement (Intelligence Amplification). According to some general principles, one agent can use what it has learned to magnify and enhance the action, perception and cognition of another agent.

Sutton believes that this enhancement is an important part of realizing the full potential of artificial intelligence.

In this process, it is very important but difficult to identify indicators to assess AI progress, and the team is exploring different developments.

In addition, Carmack has always been an advocate of open source, but he said he would keep some openness on the issue of AGI, but would not disclose all the details of the algorithm.

As a small team, Carmack believes that it is necessary to maintain a pioneering spirit and focus on long-term development rather than short-term interests.

Commercialization is not considered prematurely, and there is no intermediate form that can be released publicly like ChatGPT.

As for what can be done in 2030, Carmack believes that "there is an AGI that can be shown to the public", Sutton said that "the AI prototype can show signs of life (signs of life)."

2030 critical node 2030 and AGI are not the first to appear at the same time.

The top AI teams all regard the period before and after 2030 as the key node to realize AGI.

OpenAI, for example, wrote in the announcement that 20% of its total computing power was set up to set up a super-smart alignment department. We believe that super-intelligence will come this decade.

Even similar views have emerged in the investment community. Son has just presented such a PPT at Softbank Corp. 's World Enterprise Congress.

Apart from OpenAI and Keen Technologies, there are not many organizations dedicated to developing AGI.

The CEO Dario Amodei of Anthropic, OpenAI's biggest competitor, which has just raised $4 billion, mentioned in a recent interview that AI can behave like a well-educated human within two or three years.

When Transformer authors Vaswani and Palmer left Google, they founded AdeptAI, with the goal of building universal intelligence.

But now the two suddenly left the company earlier this year, leaving only one David Luan (far right) among the co-founders.

The two Transformer authors founded another Essential AI, which has a less "starry-looking" vision and a more pragmatic commercialization of large models.

There are also not many people who clearly call out the goal of AGI in China, mainly the dark side of the month newly founded by MiniMax and Yang Zhilin.

Reference link:

[1] https://www.amii.ca/latest-from-amii/john-carmack-and-rich-sutton-agi/

[2] https://www.youtube.com/watch?v=uTMtGT1RjlY

[3] https://arxiv.org/abs/2208.11173

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.

Views: 0

*The comments in the above article only represent the author's personal views and do not represent the views and positions of this website. If you have more insights, please feel free to contribute and share.

Share To

IT Information

Wechat

© 2024 shulou.com SLNews company. All rights reserved.

12
Report