‘Optimizing Language Models for Dialogue’

 

ChatGPT is and has been in the spotlight for the past few months, and the viral nature of this seemingly like-the-past ‘cliché’ AI algorithm has posed to us many new questions: How is ChatGPT different from other algorithms? What effect does it have on the immediate, and seemingly far away future? How should the world adapt, or ChatGPT itself adapt, to the imminent Fourth Revolution?

 

What is ChatGPT?

 

Created in November last year, OpenAI produces its prototype with the basis of GPT-3, a complex language model. This was fine tuned to using Transfer Learning, where over the basis of Machine Learning when AI learns how to solve a problem through masses of data sets. It also learns to solve similar problems in distinct yet similar ways without using too much excess data for training. It has been trained through supervised and reinforcement learning, where it uses labeled datasets to train algorithms until it feeds the correct output that has been marked in the datasets given. With this, reinforcement learning with human feedback (RLHF) awards the desirable outputs of the algorithm and punishes it when undesirable or inefficient solutions are given. In general, a reinforcement learning algorithm can perceive and interpret its environment, take actions, and learn through trial and error. OpenAI continues to gather data from ChatGPT users that could be used to further train and fine-tune ChatGPT. Users can upvote or downvote responses they receive from ChatGPT and fill out feedback for it to learn on.

 

In terms of marketing/accessibility, ChatGPT opens free basic service to the public, which had gone utterly viral over the course of December last year, reaching from 1 million users from the beginning of the month to 100 million in January. However, service and access to ChatGPT has often gone down in peak hours, and thus it opened a charged premium service for 42$/month of priority access and unrestricted features. However, ChatGPT remains un-open-sourced, which means that no one would have complete knowledge of how the algorithm works.

 

Its main services are, in contrary to the seemingly limited language model it uses, quite varied and wide ranged. From programming to theological discussions, it can provide answers that are summarized and gathered on the internet output this accordingly to the wording of the prompt. Its abilities are quite strong in school and work level, where from software developers and pupils in introduction courses to coding and all learn from the outputs given from ChatGPT. With enough prompts, it can output information accordingly to the desires of the user, say in the voice of a primary student who does not have specialized knowledge of subjects to university students and even, though not very coherently, a professor or someone who is very knowledgeable in a field.

 

How is ChatGPT different from other AI algorithms?

 

To put it simply, ChatGPT beat most general search algorithms by quite a bit. Compared to daily use Google Translate or, say, Twitter algorithms, it outcompetes by understanding the question being stated. As a chatbot, ChatGPT tends to give answers like a Q&A rather than listing possible sources to use. Even if there is no direct answer to an inputted question preexistent on the internet, it will digest all kinds of other useful information to come up with a relatively reliable solution. When asked ChatGPT how it can surpass previous AI algorithms, it answered in 4 main reasons.

 

1. Scale: Transformer-based models like ChatGPT are much larger than previous AI algorithms, which means they have more parameters and can be trained on more data. This results in a better ability to understand and generate language.
2. Contextual understanding: ChatGPT uses a technique called self-attention, which allows it to better understand the context in which words are being used. This means it can generate more coherent and relevant responses.
3. Pre-training: ChatGPT was pre-trained on a massive amount of data, which allowed it to learn a lot about how language works before being fine-tuned on specific tasks. This pre-training helps ChatGPT to be more accurate and effective at generating language.
4. Open-source community: The development of transformer-based models like ChatGPT has been driven by an open-source community of researchers and engineers who are committed to making advancements in natural language processing. This collaborative effort has led to significant improvements in AI language models in recent years.

 

To put it in context, ChatGPT firstly succeeds in RLHF, which current algorithms do not implement as much, which helps its tonality and style of response. Secondly it is based on a language model, which provides more scope to understanding problems rather than search engines and other algorithms which do not often use this. Most other corporations were originally disincentivized from using language models because it was quite an immature model that would not show as much development even when masses of parameters (at least that was what they thought) were inputted. OpenAI, however, was the first to realize and find out that when the number or parameters increased to over 10^9, the capability, studying efficiency and reliability of the algorithm increases dramatically. Currently, OpenAI’s model has over 175 billion parameters, whereas the language model proposed and created by Google merely has 13 billion. The advantage gained after the bottleneck zone of development and the head start made by ChatGPT is also a deciding factor for it to stand out so much in the AI world.

 

The wide usage is also quite appealing to the public, which is one of the main reasons for it to be so popular among students, workers and people of all age groups, backgrounds and needs. Rather than specialized algorithms that serve single purposes, say FSD (Full Self Driving) or Google Maps or social media, it can engage with all types of inputs due to a strong language model for it to interpret connotations and specific requirements for its output. This looks like it being able to write essays and plans for homework and prepare exam revisions by summarizing and condensing vast amounts of information online. Banks use ChatGPT to process lots of information and identify potential risk factors that would have needed lots of human staff to analyze and identify. Professional programmers or software designers can use ChatGPT to assist with the coding part of product creation, and so on and so forth. This was not sufficiently achievable by previous algorithms to the extent of usage and simplicity, thus leading to its massive attention on the general internet.

 

What are ChatGPT’s effects on the future?

 

As an algorithm trying to touch the lines of AGI, ChatGPT has brought about lots of social change since its release, some good and some bad. It depends on how society will adapt soon to the undeniable incoming of the fourth technological revolution, how well the future world is going to be with the AGI and tech that is expanding at record rates. This looks like, the redevelopment and restructuring of jobs and business markets that could eliminate advisory jobs, factory line workers, or secretaries and create new jobs such as AI trainers or prompt advisors that help maximize output by fine-tuning inputs.

 

One of the most prominent and controversial aspects of ChatGPT’s influence on the current system is on education. Enormous amounts of students around the world now have access to a free, advanced, and semi-intelligent essay writer which can help them quickly finish unwanted schoolwork and cheat on online exams. The extensiveness of its usages and accessibility leads to issues with efficacy and ethics. New York’s department of education has already imposed a total ban on ChatGPT in public schools, claiming “it does not build students’ critical thinking and problem-solving skills, which are essential to educational and lifelong success.” However, still students who use personal data can access the site and use them to gain ledges over short-term exam revision, which are still quite impactful in the current education system because immediate test scores can decide on future options for further education. Some who work in both systems try to argue that ChatGPT should be implemented or incorporated into curriculum which are still able to promote individualism and students critical thinking, which also reduce the amount of unfair competition that will exist even under the ban of the algorithm - this is similar to reasons to not banning alcohol or cigarettes - the inevitable existence of the underground market would lead to more misunderstandings and harder education to students what is actually correct. Many also argue that since it has been already used by experts and programmers and other workers in real workplaces, it should also be rational for students to learn to correctly use ChatGPT as an efficient tool through education because it would also be quite unlikely that in the near future the new generation would not have ChatGPT, if not way more developed AI that empowers the development of the planet.

 

From its influence on education to the workplace, further impacts on society can be noticed. Because of its nature as a developed language model, it can find user desires through connotations and wording of specific prompts, which allows it to provide personal feedback and output than other algorithms which provide general answers. The succession of knowledge, especially in deeper fields such as law or medicine was especially hard to achieve only in short periods of universal education. For this, the status quo used to offer, say, stock managers, lawyers, or doctors. The general knowledge and capability of ChatGPT to condense human knowledge and re-express or tailor advice to specific questions might mean that now increasingly ordinary people are able to access professional opinion without having to pay for a lawyer or hedge fund manager to explain what is happening. Does that mean the knowledge of professions becomes obsolete? Of course, no. The market will always exist, and it is a long-term process until that happens - ChatGPT is not yet as developed to be able to replace the vital roles lawyers and doctors play in society.

 

This leads us to the point on vulnerability of ChatGPT and possible new future changes because of this. ChatGPT isn’t going to be the best AI algorithm accessible in, say, 5 years because new models and theories are developed and created at a speedy rate. However, with its influence now I think it would be reasonable to find ways that new professions and jobs could be created soon. ChatGPT does provide some very outstanding responses to prompts. However, these prompts are hard to make precise or effective because most people do not completely understand how the algorithm works, thus asking less efficient problems (in terms of useful output it can give). Industries have already begun to develop in places like China, where professional advice is sold for college students or professors to enhance or outline their essays. Thus, even in the long-term future, it is likely that even in the presence of an AGI that can solve most if not all prompts effectively, humans are still the deciding factor on its efficiency and creation of creative, interesting, and appealing work. But still, mechanical, or repetitive work would be replaced. AI trainers, editors and programmers would still exist and flourish to optimize code which would be quite important in the future. Even the example of vets would thrive slightly more than the current situation because since productivity is increasing, people might have more time and effort to invest in pets, thus supporting another industry. The future is quite unclear, but a closed door can always lead to an open window.

 

What does the future look like and how should we adapt?

 

It is not yet clear whether ChatGPT will ever be officially accepted into society. Top academic journals such as Science do not allow the use of ChatGPT-produced text in their papers. Some schools, including the Institut d'Etudes Politiques de Paris, have also banned the use of ChatGPT in written texts or PPTs by students and staff. Internet corporations also do not allow their own engineers yet to write code in ChatGPT for fear of leakages.

 

ChatGPT has some issues of its own that also need to be addressed: Is the data used to train models copyrighted? Even the ultimate question of how to prevent technology from setting humanity to its grave will be on the agenda. Its accuracy of information and calculation are often horrendously bad, and lots more needs to be improved as it is still developing in an RLHF state.

 

Just like Berners-Lee, the father of the World Wide Web, never dreamt back in the day that the internet could change the world so drastically, no one really knows where ChatGPT is going to take us to, down the rabbit hole perhaps?