The Most Complex Artificial Intelligence

Super AI GPT-3 is destined to be one of the most important breakthroughs in the field of artificial intelligence in 2020. It is currently the most complex artificial intelligence language model in the world, developed by the renowned artificial intelligence open platform OpenAI.

How complex is GPT-3? Technically, the most important criterion for measuring the complexity of deep neural networks is the number of parameters in the network, and GPT-3 has a staggering 175 billion parameters, which is 10 times that of Microsoft’s super AI “Turing NLG”.

GPT-3 is a pure language model, where the input and output are text, and the training data used for GPT-3 is also pure text. The training of GPT-3 utilized a massive 45TB of plain text, which is roughly equivalent to reading all the books, articles, documents, and programming manuals available on the Internet. It is said that a complete training cycle on this 45TB of text costs about $12 million.

GPT-3 has already grasped most basic textual logic and can engage in smooth conversations with humans, even performing accurate logical reasoning. Its important applications include automatically generating complex programming code and conducting UI design through simple verbal instructions; it can also write articles, summarize classic literature, perform physical reasoning, and translate legal language, among other text-related functions that will be realized in the future.

How Far Is GPT-3 from the Singularity?

American futurist Raymond Kurzweil once predicted that with the accelerated development of artificial intelligence, a technological singularity would eventually emerge. The singularity is a state where artificial intelligence reaches its peak, completely surpassing all existing human wisdom and computational abilities.

So how far is GPT-3 from the singularity? There may only be three gaps.

(1) The ability to perceive more. GPT-3 is merely a pure language model, which is clearly insufficient for understanding the world. It needs to input visual, tactile, auditory, olfactory, and gustatory perceptions to be able to identify abstract content such as people, animals, objects, and events from images. With multiple senses, artificial intelligence can truly grasp the weight of words, thereby having the chance to pass the Turing test.

(2) The ability to learn continuously. GPT-3 also needs the capability to learn and train at any time, allowing knowledge to continuously increase. In other words, it needs to continuously input various texts and sensations for ongoing learning and training. Current artificial intelligence can only be in either training mode or usage mode at any given time; it can only learn from limited input content, not learn continuously and at any time.

(3) The ability of fundamental awareness. The singularity is a conscious entity, and an important manifestation of consciousness is the ability to ask oneself questions and provide answers, as well as to judge the answers to those questions. Only in this way can a complete individual consciousness be formed, and consciousness is one of the most important capabilities of the singularity.

Currently, although the above three conditions still seem distant, we can already see the possibilities. For example, artificial intelligence can already recognize people, objects, etc., in videos or images; moreover, it is not impossible for artificial intelligence to be continuously fed and trained at any time; and the idea of artificial intelligence asking and answering questions itself is not entirely out of reach. Perhaps the technological singularity will arrive faster than we imagine.

References:

01. “GPT-3: An Astonishing AI Model” – Wan Wei Gang

02. “50 Ways to Use GPT-3: It’s Cool, But It Hasn’t Passed the Turing Test” – AI Technology Review

03. “GPT-3 Starts Charging! OpenAI Announces: We Are Just Covering the Necessary Costs of a Great Mission” New Intelligence Source