Transformer Embeddings and Tokenization
Evolution or Revolution: The Ascent of Multimodal Large Language Models
Esteemed historian and author Yuval Noah Harari, celebrated for his seminal works "Sapiens: A Brief History of Humankind" and "Homo Deus," discussed an intriguing theme at the 2023 Frontiers Forum on AI - the burgeoning prowess of Artificial Intelligence (AI) in the realm of language. The conversational eloquence and user-friendly interface of OpenAI's ChatGPT has brought millions of everyday users into the fold of Generative AI, fanning the flames for advancements in the field, and bringing sizeable investment dollars to the table. However, Harari also gave a grave warning regarding the potential hazards AI poses if we develop Artificial General Intelligence, that it could wield language in ways that are detrimental to humankind. This leaves us contemplating: Should we fear this tightening weave of AI and language?
The Dance of Thought and Word: The Human Perspective
"The limits of my language mean the limits of my world," wrote Ludwig Wittgenstein, the great 20th-century analytical philosopher whose words and ideas echo through the ages. Indeed, language is integral to human thought, shaping our abstract concepts, reasoning abilities, and capacity to convey complex ideas. Furthermore, language is crucial in organizing, storing, and recalling information, contributing to our cognitive processes. From syntax to vocabulary, written scripts to cultural narratives, language has co-evolved with our social and technological progress. It has colored the tapestry of human experience and has been an essential element in shaping the societies we live in. As AI and Large Language Models (LLMs) become more entrenched in our daily lives, we face an intriguing question: How do we adapt?
Deciphering AI's Linguistic Abilities: Mimicry or Understanding?
In his book "The Stuff of Thought," Canadian-American cognitive psychologist and psycholinguist Steven Pinker presents language as an adaptation to the cognitive niche. Pinker argues that humans constantly innovate language within the bounds of their environment. We live in an age where the use of AI-powered tools is growing in their importance in our day-to-day lives. How will the increasing use and reliance on the synthetic imitation of human language (e.g., the text outputs of the LLMs) shape communication, our thoughts, and our ideas?
It is important to note that LLMs mimic and mirror human language structures, presenting a simulacrum of human communication. They utilize probabilistic models to predict the next word in a sentence, employing advanced algorithms to decipher word sequences, contextual relevance, sentiment, and implied meanings. As a result, LLMs produce increasingly human-like responses, which has led some to posit they possess a form of comprehension. However, while these AI models may demonstrate a facsimile of understanding, they lack the grounding to fully comprehend semantic depth, a central feature of ‘language in meaning'.
A New Era: Promises and Perils of AI
LLMs herald significant possibilities, from idea generation to translation, educational support, and beyond. These models are poised to revolutionize sectors, from education to business, augmenting productivity and quality and hastening the 'time to value’. They hold the promise of making communication more seamless and efficient. They will become our sophisticated writing assistants, a clever copilot to improve the quality of our work. Yet, as we stand at this promising frontier, we must not overlook the potential societal concerns associated with LLMs.
The specter of AI misuse for language manipulation, evident in cases like the Facebook scandal during the 2016 US Presidential Election and Brexit in the UK, raises legitimate concerns and we need to keep a keen eye on the training data and the outputs of the models - we know there are biases and limitations. Some argue that interacting frequently with AI might change the way we interact with other humans, potentially leading to changes in social norms.
There are also ethical implications, such as the potential for AI to manipulate people's decisions or emotions, or the question of whether AI should ever be made to appear indistinguishable from a human. As Bill Gates noted in his recent blog, The risks of AI are real but manageable, "The idea that technology can be used to spread lies and untruths is not new. People have been doing it with books and leaflets for centuries. It became much easier with the advent of word processors, laser printers, email, and social networks. AI takes this problem of fake text and extends it." Added to this, we are on the verge of the large multimodal Gen AI models taking center stage; we must remember that language is more than text, and the impact will be felt through audio, images and videos (as an example, the sophistication of deepfakes on the internet is rapidly advancing).
High Stakes Game
The stakes are undeniably high, and our responsibility lies in managing these risks, fostering transparency in AI models, and advocating for regulation governing AI use. Initiatives like AI Godfather Geoff Hinton's calls for clear labeling of AI-created data and content are vital to safeguarding the public and maintaining information integrity.
Over-reliance on AI tools could potentially impact the skill and art of human communication and language learning. As we weave our way through the linguistic labyrinth of AI and thought, we must ensure that the boundaries of our language do not become the boundaries of our world. By harnessing the potential benefits of LLMs and Gen AI models and diligently working to mitigate their risks, we can ensure a future where AI enhances rather than confines our collective narrative.
