Artificial intelligence: the hottest topic of the decade!

The most discussed and debated technology of the 21st century, resulting in artificial intelligence like Chat Gpt and Bard being household names. We are implementing AI in new and innovative ways every second, with whole industries predicted to be completely automated in the next few years. In this blog, I will discuss the secrets behind the 2-second responses, the dangers and adversity AI brings, and our AI-integrated future. ( In this blog I will mainly be discussing a specific type of AI: Large Language Models.)

What is a Large Language Model(LLM)?

Artificial intelligence like Chat Gpt 4, Bard, and Bert are known as large language models. Large language models can be compared to machines. These are structured just like the neural network in your brain. Just like your brain, large language models have more than half a trillion parameters( also called model parameters), which act just like the neurons in your brain: helping with decisions and generating responses. They are called “large language models” specifically because they have been trained on and consumed massive amounts of all types of English text.

How do Large Language Models work?

Large language models are in simple terms: next word predictors. A simple analogy could be used here when someone says “Twinkle twinkle little dash”, the next thing you think of is “star”. This is because since you were young, your brain has been trained and fed the rhyme “Twinkle twinkle little star”, not “Twinkle twinkle little moon”. Large language models function just like this when generating the next word. In reality, large language models are much more complicated and we are just scratching at the iceberg's peak.

Various LLMs and Comparisons:

The Dark Side of AI:

Hallucination is a side effect that occurs when you ask an LLM a particular question that it has not encountered from its massive amount of training data. What results is the next word predicted being an answer that is not factual or an answer that is not correct in terms of the context of the question you asked. Whenever the LLM does not have an answer, it answers with what is the most probable summary of the correct answer.

There is a rampant panic about the amount of misinformation being spread through the use of AI. They can be manipulated to give biased, inaccurate, or harmful information. AI can generate thousands of articles' worth of misinformation per second. This means that anybody can spread misinformation so widely and rapidly that people start believing the misinformation just due to the sheer amount of it. AI can also create specific misinformation such as deepfakes, false videos, and photos( I am sure a lot of you have seen those Instagram accounts with millions of followers that just post celebrity deepfakes). An example of the damage that can be caused by this is that someone generated a video of the Pentagon being bombed, and that video caused worldwide hysteria for a massive 30 minutes, causing economies worldwide unquantifiable damage.

Evolution of AI:

I am sure many of you have observed, if not I have already mentioned this at the start. AI is trained mainly in “English” text. This not only results in a linguistic block for people who do not know English and want to use AI but also a lack of perspective: Cultures around the world, hold valuable information that would benefit users of AI greatly. Some AIs like Bard are already trying to integrate multilingualism into their user interface, whether that integration has been successful or not is debatable. However, what I am trying to convey is the fact that approximately 80% of the world has yet to experience AI. This is why we can expect companies to focus on successfully breaking the language barrier of AIs in the foreseeable future.

Nobody can see the future but with the rate at which AI is developing, it is going to be involved in our lives in some way or the other, ( for a lot of us, it already is) whether we like it or not. That is why in my upcoming blogs I will keep you updated on the latest developments regarding AI alongside the other technology and topics I will discuss.

Useful References

1.      Attention is all we need: The original paper by Google Deep Mind on which all LLMs are based: https://arxiv.org/abs/1706.03762

2.      What are language models - https://www.analyticsvidhya.com/blog/2023/03/an-introduction-to-large-language-models-llms/

3.      Technical comparison of ChatGPT vs Bard - https://contentdetector.ai/articles/google-bard-vs-chatgpt-statistics

4.      Hallucination - https://en.wikipedia.org/wiki/Hallucination_(artificial_intelligence)

5.      Open Source LLMs - https://github.com/eugeneyan/open-llms

6.      5 Best open source LLMs - https://www.unite.ai/best-open-source-llms/

7.      CodeLlama model card - https://github.com/facebookresearch/codellama/blob/main/MODEL_CARD.md

8.      Hugging Face open LLM leaderboard - https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard

Terminologies 

• GPT - Generative Pretrained Transformers

• T - trillion, B - billion, M - million, K - thousands

• Tokens - 75 English words are equivalent to 100 tokens, approximately