Scientists hope new supercomputer network could lead to AGI, with first node set to launch within weeks

Scientists plan to accelerate the development of artificial general intelligence (AGI) using a global network of extremely powerful computers — starting with a new supercomputer that will go online in September.

Artificial intelligence (AI) encompasses technologies including machine learning and generative AI systems such as GPT-4. The latter offer predictive reasoning based on training from a large set of data—and can often outperform human performance in one specific area, based on their training data. However, they are below average in cognitive or reasoning tasks and cannot be applied across disciplines.

Artificial general intelligence (AGI) is a hypothetical future system that surpasses human intelligence in many disciplines and is also capable of learning from itself and improving decision-making based on access to more data.

The supercomputers built by SingularityNET will form a “multi-tiered cognitive computing network” that will be used to host and train architectures necessary for artificial general intelligence (AGI), company officials said in a statement.

These include elements of advanced AI systems, such as deep neural networks, which mimic human brain functions; large language models (LLMs), which are large data sets on which AI systems are trained; and multimodal systems, which combine human behaviors, such as speech and motion input, with multimedia output. This is similar to what you might see in AI movies.

Building a New AI Supercomputer Network

As company representatives told LiveScience, the first supercomputers will be launched in September, and work will be completed in late 2024 or early 2025, depending on delivery dates from suppliers.

Related: 22 Jobs Artificial General Intelligence (AGI) Could Replace — and 10 Jobs It Could Create

The modular supercomputer will feature advanced hardware components and infrastructure, including Nvidia L40S graphics processing units (GPUs), AMD Instinct and Genoa processors, Tenstorrent Wormhole server racks with Nvidia H200 GPUs, and Nvidia GB200 Blackwell systems. Together, they represent some of the most powerful AI hardware available.

“This supercomputer will be a game-changer in its own right in the transition to AGI. While the new approaches to neural symbolic AI developed by the SingularityNET AI team reduce the data, processing, and energy requirements somewhat compared to standard deep neural networks, we still need significant supercomputers,” said SingularityNET CEO Ben Goertzel LiveScience reported in a written statement.

“The mission of the computing machine we are creating is to provide a phase transition from learning on large data sets and subsequent reproduction of contexts from the semantic memory of a neural network to non-imitative machine thinking based on multi-step reasoning algorithms and dynamic modeling of the world based on cross-domain pattern matching and iterative distillation of knowledge. A paradigmatic shift towards continuous learning, smooth generalization and reflective self-modification of AI is taking place before our eyes.”

The Path to AI “Superintelligence”

The goal of SingularityNET is to provide access to data for the development of artificial intelligence, AGI and the future artificial superintelligence — a hypothetical future system that is cognitively far more advanced than any human. To do this, Goertzel and his team also needed unique software to manage a federated (distributed) computing cluster.

Federated computing clusters enable the abstraction of user data and the sharing of aggregated data necessary to perform large-scale but protected computations on data sets that contain high-security elements such as PII.

“OpenCog Hyperon is an open-source software framework designed specifically for AI systems architecture,” Goertzel added. “This new hardware infrastructure is specifically designed to implement OpenCog Hyperon and its AGI ecosystem framework.”

To grant users access to the supercomputer, Goertzel and his team use a tokenized system that is common in AI. Users gain access to the supercomputer and, through their tokens, can use and add data to existing sets that other users rely on to test and implement AGI concepts.

In their simplest form, AI tokens mimic the tokens found in independent video games in arcades. Players had to buy tokens to then insert into a video game to gain a certain number of chances to play. In this case, the data collected during play is available to all other players, not just in one arcade, but wherever that instance of the game is located in other arcades around the world.

“GPT-3 was trained on 300 billion tokens (usually words, parts of words, or punctuation marks), and GPT-4 on 13 trillion” he wrote Mercatus scholar and software engineer Nabeel S. Qureshi. “Self-driving cars are trained on thousands of hours of video footage; OpenAI Copilot, for programming, is trained on millions of lines of human code from the Github website.”

Leaders in AI, notably DeepMind co-founder Shane Legg, have said the systems can meet or exceed human intelligence by 2028. Goertzel previously estimated that the systems reach this point by 2027.while Mark Zuckerberg is actively pursuing AGI with $10 billion invested In January, construction of the infrastructure for training advanced artificial intelligence models will begin.

SingularityNET, which is part of the Artificial Super Intelligence Alliance (ASI) — a collective of companies dedicated to open-source AI research and development — plans to expand the network in the future and increase the available computing power. Other ASI members include Fetch.ai, which recently invested $100 million in a decentralized computing platform for developers.