The AI Gold Rush: Harnessing Collaborative Intelligence
Naval Ravikant recently tweeted a prediction that's been echoing in my mind: 'Everyone will soon have AI Anxiety.' AI innovations are unfolding at an unprecedented pace, with development cycles compressed into days and weeks. I admit, this 'AI Anxiety' has already gripped me. AI revolution is unleashed which appears irreversible. As Microsoft's CEO Satya Nadella suggests, it's as though we're undergoing an upgrade, with AI serving as a powerful steam engine for our minds. As the AI revolution unfolds, the debate on AI's future is heating up, polarizing into two main camps: fully embracing AI's potential or swift regulation to avoid catastrophe. We heard the AI revolution as “printing press moment” and Mr. Musk is predicting AGI within this decade. Now, Eliezer Yudkowsky, MIRI AI Researcher, wants to pull the plug ASAP as expressed in recent TED talk. Debate persist.
Now, one thing is clear - 'AI will not replace you, but a person using AI will.' It's a sentiment I've come to embrace wholeheartedly. With AI as our collaborators, we're poised to unlock unprecedented potentials, and I find myself thrilled at the prospect as I explore regularly through experimentations when time permits. Yes, this excitement is expressed in any opportunity I get having conversations with colleagues, friends and family. We all can predict with some level of certainty that AI will & is disrupting industries, a modern Renaissance, and potentially new flavor of intelligence leading to new Enlightenment. (though perhaps that's me getting a bit carried away...)
As we witness the rapid adoption of AI, the growing excitement around it, and the potential opportunities it presents, even in critical personal areas like healthcare (GPT4 passed radiology board-style tests), it's crucial to remain aware of potential pitfalls. Therefore, we must scrutinize AI systems like GPT-4 and other large language models (LLMs). Do they truly understand the responses they generate? Can we attribute wisdom to them based on their output? These questions are vital as we strive for 'Dependable AI' (Credit to Ved on this term, I really liked it), rather than just 'Generative AI'.
Marc Randolph's tweet is a valuable reminder: "There’s an AI gold-rush underway... But remember, what appears as low-hanging fruit... rarely is." This brings to mind the California Gold Rush of 1848 to 1855, where the discovery of gold at Sutter's Mill sparked a global migration of individuals hoping to become rich.Interestingly, Sam Brannon, the era's first millionaire, who made his wealth not from mining gold, but by selling hardware supplies to miners. We absolutely see AI gold-rush post the discovery of LLM’s & capabilities to generate anything from txt, images, video, audio using the simple plain natural language of choice. Miners are building plugins, applications built leveraging the underlying AI platforms, fine tuning to some degree & the goal to have co-pilot with every software stack we used till date.
The pickaxes and shovels of this AI Gold Rush are supplied by major infrastructure and chip vendors, who provide user-friendly API interfaces to LLMs and robust infrastructures for fine-tuning, retraining, and deploying models at scale. Notable examples include:
Google's Vertex AI, Microsoft's Azure AI Studio, Amazon's AWS Bedrock, and Nvidia's Nemo Service and many more.
We've already seen the AI 'pickaxe provider' Nvidia reach a market cap of $1 trillion, akin to Brannon's successful mining equipment store.
Google's recent I/O event showcased the rapid incorporation of AI models to boost productivity.
OpenAI's ChatGPT Plus and Bard have introduced a wealth of productivity tools, Plugins, Code interpreter to name a few.
Meta and LLaMA are democratizing AI models with initiatives like OPT175B.
Meta LLaMA with 7B, 13B, 33B, and 65B parameters at high 1.4 Trillion Tokens support but with minuscule carbon footprint compared to GPT3.These mini-LLM as I call it, once opened for commercial use could dramatically influence personalized enterprise centric use cases at rapid pace.
Hugging face released open source LLM Falcon-40B with 40B parameters, trained on 1T tokens and a smaller open source LLM Falcon-7B as well.
The introduction of LangChain, an open-source framework for building applications and autonomous agents that link LLMs, opens up endless possibilities for creative innovation.
With so many tools out there, a high schooler built a financial advisor chatbot within a week by fine tuning an existing model with high school level data points and asked GPT4 to provide the code to create a website and host it natively on AWS Cloud Infra. Stories of such innovative achievements are emerging from all corners of the globe and across all age groups.
It’s also quite obvious that to be a dominant player in this AI gold rush, completeness, correctness, and clarity of data is a must have. One can really put guard rails, add knowledge, add skills and continuously improve enterprise domain centric AI Models only if they have reliable domain specific data sets, and a lot of them! Above all, ensuring cybersecurity hooks are in place from day1.
As this advancement penetrates into mainstream enterprises, the question still prevails - Are these AI systems transitioned from just basic intelligence to having wisdom? I ponder this question with a sense of curiosity both technical and philosophical. With mission critical use cases at stake, it's critical to scrutinize before following blindly the hype. Here are my responses to these questions..
Let’s keep the possible emotional attachments to AI systems aside when we see amazing personalized and conversational responses. AI models, at their core, perform based on sophisticated mathematical operations. Simplifying from Wolfram's article, AI models take a corpus of data, perform tokenization (adding a reference id to each word) and location of each of these tokens in comparison with other tokens, relationship between these tokens, the neural network then uses these relationships to predict the next word in a given string of words, with continuous feedback to improve these predictions. As a mathematical operation, the system just performs based on what it is pre-trained with and spits out the next possible word, sentence, paragraph. We all know, the systems do lack a deep understanding on the context, the emotional state of the user, situational awareness.
Intelligence involves analyzing inputs, building a plan, and executing tasks which a well trained machine can perform. However, Human intelligence evolved when Homo sapiens surpassed the intelligence of other human species and dominated the planet. One of my favorite authors, Yuval Noah Harari, articulates meticulously in the book “Sapiens: A Brief History of Humankind” - For many centuries, Human intelligence was tightly coupled with subjective experiences like feelings, emotions and we strive to derive solutions keeping this experience in mind. Consciousness is the ability to feel things, pain, pleasure, love, hate. Its this alert cognitive state in which you are aware of yourself and your situations. Humans are the intelligent species dominating the planet and they are the ones with consciousness. Now, the argument could be: our decision making capacity or intelligence doesn’t need to be driven by consciousness. A great example I came across was plants are intelligent living organisms but they don’t have consciousness. This century surely dictates a new course of intelligence which decouples from consciousness. AI systems could be more intelligent than humans but still with no feelings.
So, based on where things stand today, I feel, we have AI systems which are intelligent enough to perform certain pre-trained tasks and naturally cannot be left for full autonomy with decision making when stakes are high. In the text generation realm, it's best in doing the most “reasonable continuation” of the next word. That is great intelligence to have as collaborators, creators and tutors alongside human intelligence. We can clearly see this applicable for any knowledge base dependent functions - support, documentation, SW development, learning and development, data analysis, visualization etc. This level of intelligence is powerful. We should aggressively embrace it and there will be opportunities to expand into much complex enterprise automation tasks by fine tuning an open source LLM (for better controls), and by cross linking LLM's with agents.
To conclude, I believe these AI collaborators are akin to a new breed of toddlers, learning and adjusting to the world, which in turn forces us to adapt. They can be trained with the right data and use cases, as long as proper security measures are in place. Remember, despite their potential, there remains a 'bit of voodoo magic,' as Wolfram puts it. This is because we don't fully understand why some AI responses are so compelling, despite what the underlying mathematics suggests. Open sourcing the sophisticated LLM's of today in the hands of researches is the way to start to uncover these magic & eventually community will dictate the adoption outcomes. This era of AI is powerful and should be embraced. Change is constant.
We can either stay with it, get ahead of it, or risk being left in the past.
