As someone involved in both electronic engineering and information technology, I can only marvel at the level that ChatGPT-3.5 operated at. Its human-like responses to most questions or tasks put to it were astounding and downright scary at the same time. Not scary because it produced results and texts that were correct – after all, a Google search also produces correct results. It was scary in the way that it conversed. Scary in the same way as if you asked a child a complicated question and they produced a flawless reply using adult vernacular. Most people would not believe it.
Well, that is how I am feeling at the moment with the newly announced and now released into the wild GPT-4. This new AI model is going to blow the ‘old’ version out of the water. With around 100 trillion parameters compared to ChatGPT’s paltry 175 billion, the new engine will be roughly hundreds of times more powerful. In fact, at 100 million parameters, the AI is now reaching the neural connections and processing capability of the human brain!
GPT-4 has also been trained on much more information than its predecessor. This means that it will be able to make more ‘connections’ between the bits of data that it has access to, and this will ultimately be seen as answers that have been better researched and results that are more well-rounded.
Not only will it be able to respond to text-based queries like ChatGPT could, the new engine is also capable of producing images based on a description, websites based on a drawing or idea, video based on a storyline. The capabilities are mind-boggling.
It is a well-known story that a few months back ChatGPT managed to pass Google’s coding interview for a level 3 software engineer. So, imagine what GPT-4 can accomplish.
Well, OpenAI has said that GPT-4 can place in the top 10% of test-takers for the Bar Exam, the standard certification test for lawyers. It is now capable of both accepting and producing texts up to 25 000 words long (ChatGPT’s limit was 3000 words). This ability allows it to search and analyse longer texts.
GPT-4 is also quite at home solving complicated mathematical problems: it seems to be quite comfortable with questions dealing with calculus, linear algebra and statistics.
Developers will have access to the GPT-4 engine through an API and I am sure that it will not be long before we see applications developed which use it. Duolingo, the language learning app, has already incorporated GPT-4 to be able to provide an AI conversation partner to practice with.
The technology has not come cheap though. Microsoft has spent hundreds of millions of dollars on the hardware needed to allow the training of the GPT-4 model. Tens of thousands of Nvidia’s A100 graphics cards, the workhorse for training various AI models, have been used, with experts predicting that around 30 000 cards will be used for the current architecture build. This will be a pretty payday for Nvidia, with AI becoming one of the company’s biggest income generators.
For me, this will be the most significant technological breakthrough in modern times, which I believe is sure to usher in a new industrial revolution. Skynet from Terminator, anyone?
| Tel: | +27 11 543 5800 |
| Email: | malckey@technews.co.za |
| www: | www.technews.co.za |
| Articles: | More information and articles about Technews Publishing |
© Technews Publishing (Pty) Ltd | All Rights Reserved
printer friendly version