The world, 1 year later

OpenAI released its first public Large Language Model (LLM) GPT-3 one year ago. It unleashed an AI frenzy comparable to the Gold Rush era with new interest in computers. Yours truly wrote no less than 21 articles on that topic, trying to separate the reality from the hype.

So where are we one year later with this technology? What can we expect going forward?

Since the launch of GPT3, OpenAI has released a ChatGPT - a chatbot, which is using GPT3.5 as a backend. Later, it released GPT-4 with ChatGPT Plus. OpenAI also released an API for developers and organizations to connect their systems and take advantage of LLM’s functionality.

The functionality of GPT/ChatGPT products started conversations about the future of humanity. Those conversations have oscillated between “We are all going to die, AI is taking over” and an alternative perspective, “AI will do everything for us and nobody will have to work.” (A few conversations still boil down to “AI is useless.”)

More practically, OpenAI’s innovation already created a new category of self-declared jobs with titles like 'Prompt engineer'. Importantly, it also started a bigger conversation about what this technology can do, what it can not or should not do and how to make it useful. Excited politicians are already drafting laws around how to tame this wild, uncontrollable, undefined animal.

Let's start with what ChatGPT doesn't do. ChatGPT can't distinguish between correct and incorrect answers. It does provide an answer, but when told by humans that it is not correct, it has no problem changing it. It is a Language Model not a Knowledge Model.

Another problem is that it was trained on so much content, where the accuracy of the information has not (could not) been verified. The fact that ChatGPT was used to pass various tests, including simulated school bar exams, only means that it was able to take the training content and (correctly) answer the questions. However, it didn't answer all the questions correctly. An argument that it ranked in the top 10% alongside human test-takers doesn't hold, since we should expect 100% accuracy from machines. That, despite various polished demos, disqualifies it from writing computer code or providing a medical diagnosis.

Where it does provide value is in the language analysis department. Creating an abstract or summary from large sets of documents on its own is a major achievement and has many practical use cases. Being able to ask questions about a content with different context is another powerful example.

The real advancement will come when the LLMs will be combined with other technologies. A good example is Graph Databases, where properties and relationships are explicitly defined. When combined with an LLM, you get an accurate answer with rich context. Thinking that LLMs are the only way forward is parallel to the hammer and nail analogy.

The concern about a never ending stream of machine generated content flooding the Internet is legitimate and it will negatively affect the training of future models. Using the AI to generate more and more content will only lead to an elevated level of noise. That will render the systems useless and the hunt for quality content will only emphasize the real value of humanity - original thought. That thought brought us here and will keep carrying us to the future. The recurrent pattern you can bet on.