How Can The Reader Use LLMs
Introduction
The term Large Language Model (LLM) does not (yet) have a generally agreed-upon definition. However, the general principle is that it is a
“Machine Language Model That Can Manipulate Text”
This brief article will be very useful to all readers who want to learn more about the tools and mechanisms behind Large Language Models. The first three chains will concisely discuss three general tasks that these programs do well. The fourth chain will discuss in no-nonsense terms the mechanism for how these models work.
Disclaimer: It should be noted that LLMs make mistakes and all users follow the advice in this article at their own risk. Chainspeech will not have responsibility over reader decisions or ChatGPT text.
Chain 1: Entertainment
Many LLMS (we will use ChatGPT as an example) have a large bank of entertaining riddles and knowledge. ChatGPT has access to large amounts (but not all) of the internet (more on that in the chain on mechanisms). This means that LLMs are capable of generating a broader scope of riddles, would-you-rathers, truth or dares etc than any human.
Chain 2: Therapy/Friends
There is an epidemic of poor mental health in the world, especially with the younger generation.
Above is an illustration of the Google searches against time (from Google Trends):
Red = Anxiety Help
Blue = Depression Help
Yellow = Mental Health Help
While it can be debated that this increase is due to the popularisation of the internet. When searches for anxiety are compared with the gold-standard search: “Cat Videos”:
Red = Anxiety Help
Blue = Cat Videos
(Also from Google Trends)
It is once again evident that Poor Mental Health is on the rise.
LLMs allow users to discuss their emotions without fear of judgement for stigmas surrounding their emotions. LLMs are adept at giving neutral advice having been trained on thousands of internet pages for psychology.
Chain 3: Education
It is possible (big surprise incoming) to use ChatGPT for help with schoolwork. This is not advisable in many scenarios, as there is an arms race between AI detection and AI non-detection and if somebody is detected using AI - their academic reputation may be permanently tarnished. Interestingly, many schools have decided to scrap the idea of certain types of homework altogether. St Alleyn’s School, a famous independent school in London, UK are “looking to move away from homework essays”. (Source: https://www.dailymail.co.uk/news/article-11691571/Teachers-concern-OpenAis-ChatGPT-powerful-end-homework.html#:~:text=A leading independent school has, was awarded an A* grade.)
However, LLMs can be used productively. For example, LLMs (unsurprisingly) are adept at language manipulation. This can be useful for students to learn complex syntax and vocabulary in their first language as well as learning new languages!
Chain 4: Mechanism
It is difficult to generalise the mechanism for all LLMs - as many have different specific algorithms and training. For example, GPT (standing for generative pre-trained transformer) went through multiple versions before becoming ChatGPT on GPT models 3.5 and 4. However: the general principles
Training
Pretraining
LLM is trained on lots of data (e.g. large amounts of the internet)
LLM learns to predict the next word in a sentence by the context of the sentence
LLM Transformer
A Complex step that can be summarised as making sure information is contextually relevant.
Fine Tuning
Fine Tuning
Training on more specific data sets
Further Fine Tuning
User responses help companies to further refine their AI
N.B
GPT is not the only type of LLM - another notable variety is BERT (Bidirectional Encoder Representations from Transformers) which is used by Google.
Conclusions
There you have it, a brief summary of useful implementations of LLMs as well as the basics of its mechanism. If you want to learn more about associated topics, read our other articles!
