Last Edited : February-1, 2024
AI is evolving at an unprecedented pace, transforming how we interact with technology and redefining the boundaries of what machines can accomplish. Breakthroughs that once seemed like science fiction have rapidly become everyday tools. But where exactly do we stand in this journey? What comes next, and what opportunities does this new era present? This essay explores three broad themes: where we stand, what’s next, and the opportunities ahead.
Where We Stand
The AI landscape is shifting faster than ever, with innovations emerging at a breathtaking pace. One of the latest milestones in this revolution is DeepSeek, an advanced AI model that has captured attention for its sophisticated reasoning and generation capabilities. Its arrival underscores a broader technological pattern—progress thrives on commoditization and miniaturization. From the shrinking of transistors to the rise of cloud computing, history has shown that as technology becomes more accessible and efficient, its impact grows exponentially. AI is now following the same trajectory. I have previously explored this topic in a blog post, where I provided more historical examples illustrating this trend. For a deeper dive into these patterns and their implications, I encourage readers to refer to that blog.
To the un-initiated, DeepSeek, a Chinese startup, recently introduced DeepSeek-R1, an advanced large language model (LLM) that is disrupting the AI sector. Unlike many high-cost models from OpenAI, Meta, Anthropic etc, DeepSeek-R1 is open-source and highly efficient, operating at a fraction of the cost. Even under U.S. export restrictions on high-end AI chips, DeepSeek has leveraged software-driven optimizations to develop a cost-effective model. This development not only challenges the status quo but also demonstrates how AI might pivot toward accessible, lower-cost alternatives. The rapid adoption of DeepSeek-R1 suggests a shifting AI landscape, where users can transition between platforms more easily, increasing competition and disrupting the dominance of established AI companies.
To fully grasp how we got here, we must look back at a defining moment: November 2022, when ChatGPT was introduced. A useful mental model for understanding ChatGPT is that of a chess grandmaster. Just as a grandmaster has studied thousands of games, recognizing patterns and making optimal moves, ChatGPT has been trained on vast amounts of data. It does not "think" as humans do but generates responses with remarkable accuracy based on learned patterns.
However, recognizing patterns is not the same as true reasoning. A grandmaster can recall famous games instantly, but when faced with a novel challenge, they must strategize carefully. This distinction parallels the Two Systems Theory proposed by Daniel Kahneman and Amos Tversky. According to this theory, human cognition operates through two modes: System 1, which is fast and instinctive, and System 2, which is slow and deliberate. Earlier AI models resembled System 1, relying heavily on pattern recognition. However, the latest generation of AI is evolving toward System 2-like capabilities—engaging in structured reasoning rather than merely predicting plausible responses. This capability, known as chain-of-thought reasoning, allows AI to break down complex problems into smaller, verifiable steps, making it significantly more powerful in structured tasks.
This advancement is built on deep reinforcement learning, a method that enables AI to learn through trial and error, much like how humans learn from experience. A simple analogy is training a dog. When teaching a dog a trick, you reward it with treats when it performs correctly and withhold rewards when it does not. Over time, the dog learns which behaviors lead to positive outcomes. Similarly, AI models using reinforcement learning receive feedback, improving their decisions over time to achieve a goal. This technique has powered groundbreaking AI achievements such as AlphaGo, AlphaFold, AlphaZero, and ChatGPT. However, it is crucial to set realistic expectations. Despite being called “reasoning models,” these AIs do not possess human-like creativity or abstract philosophical thought. Instead, they excel in tasks where logical steps can be verified, such as mathematics and coding. The core technology behind these reasoning models is inspired by AlphaGo and AlphaZero, which successfully solved structured problems for games like chess and Go. The new insight researchers have gained is that structured problem-solving in general follows similar principles—where every move or decision must be evaluated for its long-term impact. This is why AI is poised to revolutionize autonomous coding tasks far beyond what traditional LLMs could achieve.
What’s Next?
Looking ahead, AI’s future can be framed through the lens of scaling laws—principles that have historically guided technological advancements. Consider Moore’s Law, which predicted that the number of transistors on a chip would double approximately every two years, leading to exponential increases in computing power. Similarly, Dennard Scaling once allowed power consumption to decrease as chips became smaller, enabling sustained performance improvements over time.
In AI, scaling laws dictate how increasing computational resources—such as model size, data, and training time—lead to performance improvements. The recent boom in LLMs has been driven by this principle: the bigger the model and the more data it consumes, the better it performs. However, we are now approaching a major bottleneck—the data wall. High-quality training data is finite, and as models continue to grow, they require exponentially more data to maintain their trajectory of improvement. This suggests that the era of simply scaling up LLMs may be reaching its limits.
To put this into context, it’s worth appreciating the pace of AI innovation. Dennard Scaling lasted for decades, and Moore’s Law has shaped computing since Gordon Moore first articulated it in 1965. In comparison, the rise of LLMs has been astonishingly fast, but we may be nearing the end of this particular scaling law.
However, this does not mean AI progress is slowing down. Instead, we are entering a new phase—one driven by the scaling of reasoning models. Their capabilities are expected to improve rapidly, especially in tasks requiring structured problem-solving and autonomous execution. However, an exciting technical challenge remains: where will reasoning datasets come from? For LLMs, the internet acted as a vast gold mine of language data, but reasoning tasks require a structured and curated dataset, which is far less abundant. Addressing this challenge will be crucial for the next phase of AI development. One of the most immediate and transformative applications will be in coding, where AI-generated software development will become even more seamless and powerful.
This shift marks a profound evolution—one that will redefine industries, reshape workflows, and push the boundaries of what AI can achieve.
The Opportunities Ahead
With a clearer picture of where we stand and the trajectory of AI’s future, we now turn to the exciting opportunities that lie ahead. This new frontier is being shaped by two core technologies: i) Large Language Models (LLMs) and ii) Reinforcement Learning (RL).
The potential for impact in these fields can be examined through four fundamental dimensions:
Applications – where AI-driven disruption is already reshaping industries.
Algorithms – the backbone of innovation, driving advancements in structured problem-solving.
Data – the essential fuel for AI models, with new challenges and opportunities emerging in its collection and curation.
Compute – the infrastructure powering AI’s rapid growth and scalability.
Applications
This is where the most immediate transformations are taking place, setting the stage for massive disruption across industries. LLMs are being used as black-box technologies to build AI-native software. One way to conceptualize this shift is to view LLMs as a new class of computers and AI-native applications as the software built on top of them.
This evolution is similar to the development of the software economy in enterprise, personal, and mobile computing:
Enterprise and personal computers led to the rise of operating systems like Windows and Linux, followed by the internet and enterprise/consumer applications.
Mobile computing followed a similar trajectory with mobile hardware manufacturers, OS platforms like Android and iOS, and application ecosystems such as the Apple App Store and Google Play Store.
An interesting pattern has emerged in both cases—one closed-source platform and one open-source platform—indicating that open-source LLMs could play a major role in the future.
While entering the AI-native application space is relatively easy, establishing a strong competitive edge is far more challenging. In this landscape, active user engagement becomes the true differentiator. Today, numerous companies are developing AI-powered applications, often referred to as vertical agents, each targeting specific market niches. However, a significant opportunity exists for a broader approach—creating an aggregator platform that functions as an operating system for the LLM-driven computing environment. Such a platform could seamlessly integrate various AI-native applications, offering users a more cohesive and efficient experience while driving competition and innovation in the AI ecosystem.
For the Indian market, IT services companies have a unique advantage. Since they are deeply embedded in customers' processes and businesses, they are well-positioned to build AI-native applications and help clients reimagine their businesses. However, to fully seize this opportunity, they will need to disrupt their own traditional business models.
Algorithms
The core idea here is to leverage advances in deep and reinforcement learning to build models tailored to specific applications. For example, reinforcement learning and AI can be used in drug and material discovery to predict molecular structures, optimize synthesis pathways, and accelerate the development of new pharmaceuticals and advanced materials. By simulating chemical interactions and testing thousands of potential compounds in a virtual environment, AI can significantly reduce the time and cost associated with traditional laboratory experiments. Companies like DeepMind's AlphaFold have already revolutionized protein structure prediction, demonstrating the immense potential of AI-driven discovery. Similarly, AI-driven material discovery is being explored in fields such as battery technology and semiconductor development, where optimizing molecular structures can lead to breakthroughs in energy storage and computing efficiency. DeepMind's GNoME project has discovered over two million new materials using AI, with hundreds already synthesized in labs, accelerating the search for advanced materials. These examples demonstrate how AI is reshaping material science by reducing discovery time, minimizing costs, and unlocking new functional materials for a range of industries.
While AI-native applications focus on reimagining existing digital experiences, application-specific AI algorithms have the potential to create entirely new markets. This requires significant R&D, with inherent uncertainty in outcomes. However, successful algorithmic breakthroughs would provide a sustainable competitive advantage that is difficult to replicate.
Data
Modern LLMs require vast amounts of training data, a factor often overlooked. For example, OpenAI collaborated a data annotation company, to employ human labelers in Kenya who were tasked with filtering toxic content and improving ChatGPT’s responses. These workers played a crucial role in reinforcing safety mechanisms by reviewing and categorizing harmful content, ensuring the model aligns with ethical standards and social norms. The initiative was integral to OpenAI's reinforcement learning from human feedback (RLHF) process, refining ChatGPT’s behavior through curated human input. Scaling laws in AI inherently demand increasing amounts of high-quality data, making data labeling an ongoing business opportunity.
The BPO industry is well-positioned to capitalize on this shift by integrating technology into human-centric work. Several high-impact opportunities exist:
LLMs today are trained by companies primarily based in California, meaning their responses are aligned with Western sensibilities. A significant opportunity lies in curating training data to align these models with regional perspectives, such as the Indian market. Some suggest training LLMs exclusively for specific regional languages, but most regional language data is already available on the internet in some form and has likely been included in the training data of frontier models. The real value, therefore, lies in aligning existing large-scale models to better reflect local sensibilities and behaviors, ensuring AI systems remain adaptable to diverse cultural and linguistic contexts.
Training AI for robotics requires substantial real-world data, as these models need extensive exposure to diverse environments to improve their decision-making capabilities. Tesla, for instance, collects vast amounts of driving data through cameras on its vehicles, capturing real-world scenarios, traffic patterns, and edge cases. This data is then used to refine its self-driving AI, enhancing its ability to predict and respond to complex situations on the road. However, collecting raw data alone is not enough. A significant portion of this data must be labeled and structured to effectively train AI models. This process typically involves a combination of automated tools and human annotators working together to classify objects, define behaviors, and verify model predictions. Beyond self-driving cars, robotics applications in warehouses, manufacturing, and healthcare require similarly vast datasets, where labeled information is essential for ensuring precise and adaptable automation across industries.
Compute
This is the most obvious dimension. As the saying goes, “When everyone digs for gold, the real winner is the one selling shovels.” In AI, compute resources serve as the foundation for all advancements. However, this space is already dominated by established players like NVIDIA, which played a crucial role in kickstarting the AI revolution. The opportunities for new entrants in this sector are limited, but optimizing compute usage through innovations in AI model efficiency remains an area worth exploring.
Conclusion
The rapid evolution of AI is transforming industries, challenging existing paradigms, and unlocking unprecedented possibilities. From the commoditization of AI models to the scaling of reasoning capabilities, we are witnessing a profound technological shift. The journey so far has been marked by breakthroughs in large language models, reinforcement learning, and material discovery, demonstrating AI's potential to revolutionize science, business, and everyday life.
Looking ahead, the future of AI will be shaped by its ability to overcome critical challenges such as data curation, algorithmic improvements, and compute efficiency. While scaling laws have driven AI advancements thus far, new paradigms in reasoning models and AI-native applications will define the next frontier. The opportunities ahead span multiple dimensions—applications, algorithms, data, and compute—offering avenues for innovation across industries.
Ultimately, AI is not just a tool but a paradigm shift that is reshaping how we solve problems, interact with technology, and expand human potential. As we navigate this brave new world, those who innovate, adapt, and explore will lead the way in shaping the future of intelligence and automation.