GPT-4 is Here 🤖

OpenAI has unveiled GPT-4, the new model that exhibits "human-level performance" on various academic and professional benchmarks.

GPT-4 is multimodal, meaning it can accept input in both text and image form. The company has embedded its new software into multiple apps and third-party programs.

Open AI put GPT-4 through stress tests to test out the harms of the technology, revealing the potential for harm, such as hate speech and advice on planning attacks. It concluded that GPT-4 was not yet capable of autonomous actions without human input.

📝 Cloudly PRO: It is clear that GPT-4 surpasses ChatGPT in its advanced reasoning capabilities, featuring broader general knowledge and problem-solving abilities. GPT-4 was trained on Microsoft Azure AI supercomputers and is available as an API. Microsoft has also revealed that Bing is already running on GPT-4.

The new model is much more powerful than its predecessors, GPT-4 can manage 64,000 words of text, enabling applications such as lengthy content production, extended conversations, and document exploration and analysis to put into context ChatGPT has a 3,000-character maximum. OpenAI also states that there is an 82% less risk of responding to disallowed content and a 40% higher likelihood of factual responses than GPT-3.5 on internal evaluations. GPT-4 also scored much higher on bar tests. Down below is an illustration of the increased power of GPT-4.

The improvement in GPT-4 is massive and will unlock new use cases that was not possible before. What is astonishing is the rapid development of these models. ARK put out very interesting data on the cost of training models, decreasing by about 70% a year. According to Wright's Law the unit cost of production and software costs should decline by 57% and 47%, respectively. This means that training these types of models will cost almost zero by 2030.

Exponential growth is hard for the human brain to understand. However, that is what we are experiencing and will experience regarding AI. New use cases will emerge, and others will die out. The new GPT-4 release might have made hundreds of apps/software obsolete, as their unique selling point now is something that is embedded into the model. This is the natural way when there is a new and emerging technology. What we deem very interesting is to look for "bets" with less "feature" risk. We have said before that Nvidia is an interesting play on the AI boom. We also believe even more data vendors that sell unique datasets that the models will train on will emerge. Data will truly be the oil of the AI age.