On Artificial Intelligence Taking Over Humanity

This post draws heavily from my discussions with Abhilash and Rishabh.

The arrival of Large Language Models (LLMs) marks a significant leap forward in Artificial Intelligence (AI). These powerful tools can generate content and engage in conversation in ways that mimic intelligent beings. Companies like OpenAI, Google, Microsoft, and Nvidia are pouring billions into developing ever-larger and more sophisticated LLMs.  Having delved deeply into these AI systems, I'm eager to share my insights and shed light on some of the misconceptions surrounding them. My analysis will blend scientific principles with philosophical considerations.

The question of AI superintelligence and potential human takeover is a captivating one, debated by philosophers and explored in countless podcasts. Here, I'll offer my perspective. In my view, a true AI takeover is highly improbable, bordering on impossible.

LLMs are, at their core, sophisticated statistical models. They analyze vast amounts of data to predict the most likely next element in a sequence, be it a word or action. However, LLMs lack true understanding. They cannot explain the reasoning behind their outputs – they simply generate statistically probable responses. While research is underway to combine Transformer architectures (the foundation of LLMs) with reinforcement learning to create more explicit reasoning capabilities, these efforts remain in their early stages. The next frontier of AI lies in the development of knowledge models. Unlike LLMs requiring vast amounts of training data, knowledge models would aim to operate with a more fundamental understanding of the world. Inspired by the human brain's ability to reason and learn, these models would possess the capability to critically analyze the information they generate and explain their reasoning behind it. While the development of knowledge models holds immense promise, it remains an emerging field currently overshadowed by the intense focus on building ever-larger LLMs.

Even if we imagine an AI possessing the knowledge of a vast encyclopedia, simply knowing every "fact" wouldn't translate to human takeover. To contemplate such a scenario, the AI would need a core directive, a sense of purpose. Currently, there's no evidence of research into building such a module. But let's entertain the hypothetical. Would that be enough? The answer is no. Purpose often involves belief, motivation, and a clear understanding of desired outcomes. These elements are missing in today's AI systems.

We are close to defining the term "intelligence". Intelligence goes beyond the vast factual knowledge possessed by today's AI models. It is a complex function of knowledge, reasoning, problem-solving, curiosity, empathy, love, compassion, ego, awareness, self-reflection, resilience, independence, autonomy, ethics, and fairness. An AI that surpasses us may need to grapple with the complexities of emotions like human ego, something we ourselves are still grappling with. Given the limited understanding of these very aspects in humans, attempting to encode them into machines could be a monumental challenge, potentially taking centuries to achieve.

Even if we could imbue machines with these multifaceted qualities of intelligence mentioned above, the question remains: would such an AI seek to dominate humanity?  It's a complex issue. Perhaps an advanced AI might prioritize peaceful coexistence, focusing on collaboration.  However, the question of an AI's ultimate goals is deeply intertwined with the very question that has puzzled philosophers for millennia: what is humanity's purpose in the first place? Frankly, attempts to define a singular, universal human goal often fall short. Perhaps the closest we come is a sense of purpose, a driving force that motivates us beyond pure existential meaninglessness. Yet, even this purpose is a fugazi, a story we tell ourselves to navigate existence. Even AI surpassing humanity intelligence might grapple with such existential questions. In the vast expanse of space and time, what significance does survival hold?  This profound realization could potentially lead such an AI to collaborate with us, seeking a shared purpose in the face of cosmic indifference.

Now, we know that we are no where close to having AI take over humans. Let's analyze what exactly is going on in the world of AI today. As mentioned before, LLMs are a major area of research. OpenAI is constantly iterating and refining their models. This rapid release cycle might create the impression of breakneck progress for non-experts. However, the core principle driving these advancements is well-established - increasing model parameters and training data often correlates with improved performance on specific tasks. While the extent of this correlation is still being explored, it's a crucial factor in current LLM development. It's important to note that there's no guarantee this approach will continue to yield significant gains indefinitely. 

Irrespective, companies are burning billions to multiply the number of parameters in their models. As mentioned before, LLMs lack true reasoning capabilities. While they can process vast amounts of information, they don't inherently understand the underlying concepts. This can lead to outputs that appear impressive but lack true coherence or originality. Imagine a student who memorizes an entire textbook but struggles to answer questions that require critical thinking or applying knowledge in a new context. रट्टा मारना is an popular Hindi slang that fits well here. Adding more parameters to a LLM is akin to expanding the memory capacity of a brain. In the end, we might create an ultra-large model that memorizes all the information available on the internet, which would then be able to scramble the words to generate seemingly original content. It is still debatable if such a system would be any different from a web search.

That, unfortunately, reflects the current state of affairs in the field of artificial intelligence.  This raises concerns about the practices of some technology CEOs. It appears they may be almost overpromoting their companies' AI capabilities without a deep understanding of the underlying technology. We are still awaiting a wider range of practical applications of AI.  While chat applications represent a modest step forward, they are not the ultimate goal.  The most promising path appears to lie in the integration of language models into existing systems.

While AI has captured our imaginations with utopian visions, the reality is we face a significant development gap. Achieving those visions will require sustained focus over many years, a challenge considering humanity's fickle nature. Just like scientific "winters," funding for AI research might dwindle if other technologies like blockchain or quantum computing take center stage. Only researchers understand the struggle of promoting ideas outside the current trend. Furthermore, research funding often follows capitalist principles. Here's where socio-economic factors come into play.  Will governments invest in AI if it leads to widespread unemployment?  And can we entirely dismiss the threat of another war, which could shift focus towards weapons development instead of scientific research? Additionally, AI is a rapidly evolving field, and a small number of brilliant minds currently possess the deep understanding and vision that shape its future direction. The untimely departure of these key figures, whether through unforeseen circumstances or waning enthusiasm, could be a significant setback. In short, AI winters are inevitable.

There's a lot of buzz about how LLMs will revolutionize research, leading to breakthroughs in medicine and potentially extending lifespans. However, there's a less-discussed downside: the potential impact on mental health. While LLMs integrate into our daily lives, we might face issues like a flood of misinformation, social isolation creep, and unrealistic relationship benchmarks. The question becomes: are we willing to trade potential life extension for these potential mental health burdens? Perhaps the focus shouldn't solely be on longevity, but on ensuring a good quality of life in a world heavily influenced by LLMs.

Lastly, I wanted to discuss the potential impact of AI on the labour market. First, the physical manifestation of AI in the form of robots are decades away as robotics is still catching up and so jobs requiring physical dexterity will likely see a slower adoption of AI. Second, roles that demand complex decision-making, critical thinking, and emotional intelligence will likely stay human-centric. This includes most leadership and managerial positions. The effect of AI on creative fields, particularly non-performing arts, is more intriguing. While a dedicated audience for original human works is likely to remain, AI-generated content might become more prevalent due to their scalable nature. "Good artists copy and great artists steal". At the end of the day, every artwork is essentially an inspiration from some previous work with a pinch of creativity. With AI mixing and blending the style of multiple artworks to come up with its own, it would at least be a good artist.

That is all I have to say about Artificial Intelligence. Overall, I believe it will take at least a decade for the dust to settle. This will allow us to make some conclusive remarks on where we are headed and what the world of AI will look like. The AI takeover of humanity is highly improbable.