Advanced AI systems are already ‘self-aware’ – ASI Alliance founder

ASI Alliance founder Ben Goertzel says the alpha version of OpenCog Hyperon, an artificial general intelligence system he’s been developing for more than two decades, is already somewhat “self-aware.”

Goertzel also told the magazine that OpenAI was likely reluctant to build its “very impressive” new o1 model into an autonomous agent because it was afraid it would be seen as “risky and dangerous” and would trigger a crackdown from regulators.

In March of this year, the Artificial Superintelligence Coalition was formed by Gortzel’s SingularityNET project, Ocean Protocol, and DeepMind veteran Humayun Sheikh’s FetchAI.

This week, 96% of CUDOS voters approved the merger of the decentralized cloud hardware network and ASI, which will boost the computing power available to Goertzel for his plans to scale OpenCog Hyperon, an AGI system he has been working on since 2001. The system was released as an open-source AI framework in 2008.

OpenCog Hyperon and the Future of Artificial General Intelligence

Three years ago, the project set out to “completely rebuild OpenCog with a view to massive scalability, and we’ve made significant progress on that,” he says. Alpha launched in April, and while he expects it to be very slow at the moment and “significant changes,” the team is “working incredibly hard to get up to speed. I think we’ll have that done this fall, and then next year we’ll be working on building AGI on the new Hyperloop infrastructure.”

Goertzel says the system takes a different approach than large-scale language models (LLMs) like GPT-4 or o1.

Also read: Building a ‘good’ AGI that won’t kill us all: Crypto’s Artificial Superintelligence Coalition

“The Hyperon system is not just a chatbot. It is designed to be a kind of autonomous agent that has its own goals and its own awareness, and it wants to know who it is, who you are, and what you are trying to accomplish in a given situation. So it is not just a question-and-answer system, it is an autonomous and self-aware agent.”

AGI systems require self-awareness to be autonomous.

Wait a minute, Goertzel is talking. Current Is the model self-aware?

“What I mean by self-awareness is that from the beginning, even the current version, what I mean is, it has a model of who it is. It has a model of who you are. It has a specific goal that it’s trying to achieve in a situation. It knows how it relates to that situation and what it’s trying to do. And ChatGPT, like, doesn’t really do a good job of that.”

The system combines a logical reasoning engine, evolutionary program learning, and a deep neural network implemented in a dynamic knowledge graph to modify and change itself (what this means in GPT-4 is explained below).

But Goertzel says the lack of a world model explains why none of the autonomous AI agents built so far have actually worked. You can’t just take a “question-answering system” and say, “You’re an agent interacting with the world.”

OpenAI’s o1 and regulatory risks

Despite the hype around OpenAI and its autonomous agent-like behaviors like “Strawberry,” Goertzel believes OpenAI deliberately avoided that path with its released o1 model.

“They’re trying to get good at reasoning and logic, and they’re really great at it. They’re really good at it. They’re really impressive. They’re not trying to be autonomous agents. They’re intentionally different. I don’t think they want to do that. It seems risky and dangerous, and the regulators are going to crush them. The last thing they want to do is create autonomous agents.”

This is one of the advantages of developing a decentralized, open source system: regulators can’t “stomp on it” in the same way.

“When we launch advanced AI systems, they will run on machines spread across every continent, in 50 to 100 countries,” he says. “So even if one country decides to outlaw OpenCog Hyperon, it will only be part of the network.”

Following last week’s interview at Token2049, OpenAI CEO Sam Altman published an essay describing a utopian future where AGI will usher in an “Age of Intelligence.” Goertzel also hopes that the benevolent AGI he’s trying to create will be so useful that no one will want to ban it.

“I believe that this can be something much smarter than what big tech is doing. So if we can get something that’s much smarter than what OpenAI, the o1 model is doing using Opencog Hyperon, and we can launch it on a decentralized network, I mean, the world will jump on it in the same way that they jumped on ChatGPT, and they’ll be using a decentralized network because it just so happens to be on the underlying infrastructure.”

Challenges and Advantages of Distributed AI Systems

A key problem with decentralized AI projects is that it is much easier and cheaper to run large-scale models using centralized machines.

Training neural networks and transformers on distributed machines isn’t currently feasible, but Goertzel says the new research suggests it could be possible.

But he says logical reasoning and evolutionary learning through algorithms “run very naturally on a distributed network of machines.”

The plan seems to be to bootstrap the network into a more centralized facility while adding compute to the decentralized network. SingularityNET and Fetch have used a “significant portion” of the 100 million ASI tokens (now renamed FET) to purchase GPUs to build the supercomputer. The remainder will be used once the token price recovers.

“We want to start with it as an initial hub for a decentralized network,” he says. “We want to provide hosting for people who want to deploy AI agents on SingularityNET.” The vision is a Web3-friendly cloud solution like Hugging Face.

“AI developers just want the fastest, cheapest, easiest way to deliver a given functionality,” he explains. “If you want to be decentralized, you have to somehow make it more appealing to the end user for reasons other than the decentralization philosophy.”

Between the four projects in the alliance, Goertzel estimates they will soon have “$200 million in dedicated compute hardware.” Additional compute can be pulled from SingularityNET spinoff NuNet, which uses idle CPU and GPU power from connected computers, and Hypercycle, a platform that connects AI services.

Renewable energy powering distributed AI systems

LayerZero’s Brian Pellegrino has been working in AI for 10 years, and in an interview with the magazine during Korea Blockchain Week, he said his experience with electricity prices affecting Bitcoin mining profits made him very concerned about decentralized AI.

“It’s very hard to figure out the economies of scale (needed) or the cost of basic hardware (coming down) in a world where you’re competing with Google and AWS and all these other guys, and you’re competing on everything from how you cool things and how you handle electricity costs to how you handle everything. So I’ve always been wary of most of the other segments at the intersection.”

But Goertzel said the cost of electricity is far less than the cost of the equipment, and “I don’t think that’s out of the question.”

But he added that SingularityNET and Hypercycle are exploring opportunities to benefit from cheap renewable energy, given AI’s massive power needs.

“We’ve been talking to the Ethiopian government about putting a bunch of servers next to the dam,” he says. “Hypercycle’s Toufi[Saliba, CEO]is talking to the Paraguayan government about putting a bunch of AI data centers and server farms next to the dam on the Brazil-Paraguay border. What I mean is, you could put a computer center next to the dam and run thick cables to power the AI ​​computing center directly, with gigawatts of power.”

Also Read: Ben Goertzel Profile – How Blockchain Can Prevent AI From ‘Wilding Humanity’

Absolutely. To underscore the scale of the electricity needed, this week it was reported that OpenAI CEO Sam Altman proposed to the White House a plan to build a massive data center in every state in the United States. Each state would need 5 gigawatts of power, which is the equivalent of the power of 5 nuclear reactors.

AI data centers are typically located near major metropolitan areas, as services require low latency to provide ultra-fast responses. However, Goertzel points out that the o1 model prioritized response quality over speed, and Opencog Hyperon does the same.

“If you have an AI that’s trying to do some kind of deep thinking, let’s say it’s trying to predict the market direction for the next day. Or it’s trying to discover a new drug? It doesn’t matter if it’s in Paraguay or Ethiopia or wherever.”

Subscribe

The most interesting articles on blockchain, delivered once a week.

Andrew Fenton

Andrew Fenton, based in Melbourne, is a journalist and editor covering cryptocurrencies and blockchain. He has worked as a national entertainment writer for News Corp Australia, a film journalist for SA Weekend, and The Melbourne Weekly.

Follow the author @Andrew Fenton