The rage over artificial intelligence or AI for short has been all over for the past 4–5 years and the up and coming market has been rocked the very foundations of how every system, company or nation is going to respond to it. The most famous of these AI tools is the Prompt software i.e. you give information or a specific phrase or sentence and the software acts on it based on the specific requirements in the prompts themselves. This can be a prompt interpreter and a prompt generator. Some of the tools that have been widely used are MidJourney and Dall-E which generates images, PromptPerfect which makes prompts better and understandable, Neural Writer which makes content and writing slightly easier and many more by smaller companies.
But the biggest of them all is ChatGPT which made searching for any reservations, planning trips, correcting coding mistakes, answering any questions and so on much easier and simpler. This has shaken the way searching on the internet far more sophisticated than search engines and Google knows this. Hence, they had in 2020 and 2021 started experimenting on their own AI and now has another on the ground AI as of February 2024. These new AI “tools” are PaLM, LaMDA and Gemini (previously called Bard).
LaMDA or Language Model for Dialogue Applications is a neural network powered bot unveiled in 2020. It was trained for human dialogue and natural language translation and it has power that can rival ChatGPT. This model has been trained on 1.5 trillion token words from text and documents. PaLM or Pathways Language Model is a transformer based model which was able to do a lot of tasks like common reasoning, joke explanations, code generation and arithmetic reasoning released privately among developers. This AI has been trained on social media conversations, open repositories, wikipedia articles, webpages and even LaMDA’s dataset. The Gemini AI is a prompt generator and receiver software which has been built on the LaMDA model with more than 4.5 billion token words and perfected enough to be out as a product.
These AI models if you were to call them are dervied from the basis of BERT or Bidirectional Encoder Representations from Transformers also by Google. Every output component that comes from the model are controlled by every input and weights in the deep learning model are dynamically updated but with a key difference. With embedding it converts one-hot encoded tokens into a vector and unembedding which reconverts them back to tokens. This is an important to perform natural language processing and works well in the English language. Since then, it has been adopted to over 70 languages in the search queries.
However, this brings up a big question on AI in general, about whether the training that AI does get is of quality. Yes, it does have the tools for logical reasoning, but it has no real tools to deal with abstract questions. For example: The video below talks about what AI can and cannot answer. In that, they had found that AI cannot really learn from learning experience and also struggles in reasoning on its own. Experts have also admitted that they use training data to test the AI as well. This does bring in questions of what the strengths of AI really means and it is important to go through the video to see the full explanation.
However, this will not stop companies to improve its capabilities. Google has incorporated image generation and interpretability in Gemini. However, it has received backlash over the images that have been created. Nevertheless, as more and more companies are developing or deploying their own AI to get ahead of the competition or catch the trend in the market. Google will definitely no longer have complete market control over searching but they are making changes in their company. This, hopefully leads to a brighter future ahead as the world is slowly heading to a recession.
Rejeti Kartik
Forum for Aspiring Computer Engineers (FACE),
Amrita School of Computing, Bengaluru
Amrita Vishwa Vidyapeetham, India
References:
https://blog.google/technology/ai/lamda/
https://blog.google/technology/ai/google-gemini-ai/
https://ai.google/discover/palm2/
https://blog.research.google/2018/11/open-sourcing-bert-state-of-art-pre.html
