Beginnings
OpenAI has made the headlines again and this time for a valid reason. They have created a language model that is probably going to have a big societal impact going forward.
Language models are machine learning systems that were trained on a vast amount of textual data to then gain understanding from that data so that the machine system can be able to reason just like how any functioning human would.
These language models have come a far way and the use cases have evolved over time. In the early stages, scientists and engineers were developing rudimentary models that could do basic probabilistic predictions that would predict what word would come after based on the other previous words.
Other use cases would allow the system to do basic sentiment analysis by analyzing all the words in a corpus (collection of text) and do frequency analysis on the most common words which will then give an idea of what a particular corpus is mainly about.
These models were mainly based on statistical formulas you could find probably in your high school textbooks in statistical theory. They also were very specific and got the job done where ever needed. Those types of statistical models weren’t enough though to solve real-world challenging problems.
Thus, scientists and engineers worked tirelessly to improve on those foundations. Over the years these models have become more complex far beyond the comprehension of the average scientist due to the ingenuity that was involved and the basic fact that machine learning or deep learning models are hard to explain. Some scientists still see them as a black box.
Current State
Today these language models permeate through almost every sphere of our society, if you have a smartphone and you are texting, that autosuggestion engine is a language model whose lineage came from those primitive statistical algorithms.
If you are talking to Siri, Alexa, Cortana, etc. All of these systems are built on language models. But the problem is some of these systems’ how-to secrets are wrapped around non-disclosures, intellectual property, and all the corporate jargon we can think of.
This poses a problem, one of the problems is that if large multi-trillion-dollar corporations can be able to innovate quickly they will have a monopoly on certain sectors of the artificial intelligence market.
This monopolistic dream for some of these companies is already here but you do have companies who are allowing these technologies to be open source and one of them is OpenAI.
OpenAI was created to basically help balance the scale within the artificial intelligence market which is a very challenging task. Over the years their flagship product has been GPT-3 (Generative Pretrained Transformers), these generative models can be to generate textual data and can replace the writing ability of the average human being in certain cases.
It’s probably even there already for most, there have been cases where the text that these models can generate is pretty superior to individuals who have even graduated with an English degree, albeit these cases are localized within a smaller context.
Regardless it’s progress nonetheless and ChatGPT proves that we are in a new era of language models. ChatGPT feels as if you are talking to a sidekick who can ask certain questions and have that system answer them for you in a very eloquent way. It’s quite an extraordinary and thrilling experience.
Above you can notice that the user which has the red underlines and the ChatGPT system is basically answering the user and solving problems. If you notice that the user is asking the ChatGPT system to solve a software programming problem by reading the code.
In this other problem, you can see that the user is asking the system to write a basic introduction to introduce herself to the neighbor.
From looking at these examples, you can see that these systems can write stories and probably even dissertations for us if we know how to communicate with them. It will be very interesting to see how the developers evolve these systems over time.
These machines over time can be able to replace lots of digital skills that the common can do and that is very exciting but frightening at the same time.
Despite the amazing results of this new model, there are still drawbacks and it is discussed on the OpenAI website. Some of them can be summarized below:
- Ideally, the model would ask clarifying questions when the user provided an ambiguous query. Instead, our current models usually guess what the user intended.
- ChatGPT is sensitive to tweaks to the input phrasing or attempting the same prompt multiple times. For example, given one phrasing of a question, the model can claim to not know the answer, but given a slight rephrase, can answer correctly.
- ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers. Fixing this issue is challenging, as: (1) during RL training, there’s currently no source of truth; (2) training the model to be more cautious causes it to decline questions that it can answer correctly; and (3) supervised training misleads the model because the ideal answer depends on what the model knows, rather than what the human demonstrator knows.
Despite the challenges, these systems are moving in the right direction. It’s impossible to predict where they might take it but for now, it’s an interesting peek into a future that is already here.
