news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

New Intelligence Report

Editor: Peach

【New Wisdom Introduction】How far are we from Skynet? Now, scientists hope to accelerate the development of human-level artificial intelligence through a powerful supercomputing network, which is expected to be fully operational by 2025.

The world's most powerful new supercomputer will be officially launched in September.

It is said that this supercomputer will bring about the AGI moment.

This was built by SingularityNET, forming a "multi-level cognitive computing network" that is mainly used to host and train the architecture required for AGI.

This includes deep neural networks that can mimic the human brain, LLMs that AI system training relies on, and multimodal systems for multimedia input and output.

After all, AGI is a hypothetical future system that can surpass human intelligence in multiple disciplines and has the ability to self-learn and self-improve.

NVIDIA's most powerful GPU supports the construction of a new AI supercomputing network

This modular supercomputer will be equipped with advanced components and hardware infrastructure.

These include NVIDIA L40S GPUs, AMD Instinct and Genoa processors, Tenstorrent Wormhole server racks equipped with H200 GPUs, and NVIDIA GB200 Blackwell systems.

This supercomputer itself will be a breakthrough in the transition to AGI.

Currently, their team is developing a new neural-symbolic AI approach that reduces resource requirements but still requires powerful computing power.

SingularityNET said that the mission of this supercomputer is to enable the transition to more advanced non-imitative machine thinking based on big data learning and simple contextual reproduction.

In other words, the new machine thinking is based on multi-step reasoning algorithms and dynamic world modeling, involving cross-domain pattern matching and iterative knowledge extraction.

Right now, before our eyes, a paradigm shift is taking place:

a. Continuous learning: AI can continuously learn and update knowledge.

b. Seamless generalization: AI is able to apply learned knowledge to new, unseen situations.

c. Reflexive AI self-modification: AI is able to autonomously improve and adjust its own algorithms and structures.

Insiders said the first supercomputer will begin to be put into use in September and the entire project will be completed by the end of 2024 or early 2025, depending on the supplier's delivery schedule.

The road to AI “superintelligence”

Maybe many people don’t know much about SingularityNET.

The company was founded with the goal of providing access to data for the development of AI, AGI, and future artificial superintelligence (future systems that hypothetically far exceed human cognitive abilities).

To do this, Goertzel and his team also needed unique software to manage federated (distributed) computing clusters.

Federated computing clusters allow abstraction of user data and expose necessary aggregated data for large-scale and protected computing.

These calculations involve data sets that contain highly secure elements such as personally identifiable information.

Goertzel added, “OpenCog Hyperon is an open source software framework designed specifically for AI system architecture.”

This new hardware infrastructure is specially built to implement OpenCog Hyperon and its AGI ecosystem environment.

Paper address: https://arxiv.org/pdf/2310.18318

To grant users access to the supercomputer, Goertzel and his team used a word segmentation system that is common in the field of AI.

Users gain access to the supercomputer via tokens, and they can use and add to existing datasets that other users rely on to test and deploy AGI concepts.

GPT-3 was trained on 300 billion tokens, while GPT-4 was trained on 13 trillion tokens.

Also, self-driving systems are trained on thousands of hours of video, and OpenAI Copilot, which is used for programming, is trained on the basis of millions of lines of human code on the Github website.

Shane Legg, co-founder of DeepMind, has said that AI systems could reach or surpass human intelligence by 2028.

However, Goertzel now estimates that this level is expected to be reached in 2027.

Some netizens directly shouted that Skynet is coming.

Another netizen said, "Let's imagine a super-intelligent toy computer."

It learns differently from other computers—not just by memorizing large amounts of information, but by trying to understand and think independently. This new way of learning may help us create computers that think more like humans.

The toy keeps learning and getting smarter, and may even figure out how to improve itself. This is a major change in the way we build intelligent computers.

References:

https://www.livescience.com/technology/artificial-intelligence/new-supercomputing-network-lead-to-agi-1st-node-coming-within-weeks

https://x.com/SingularityNET/status/1822264864057741759