Chief AI Officer Blog - Generative AI, A New Paradigm in Computing: Part 1

Generative AI (GenAI) has emerged as a groundbreaking shift in the evolution of computing and marks a transformative new paradigm. By automating complex tasks and democratising knowledge generation, Generative AI is redefining the boundaries of what machines can achieve.  

One of the most remarkable aspects of this paradigm shift is the reduction in the marginal costs of action. In Part 1 of this blog, we will explore how GenAI decreases the marginal cost of content production leading to significant business model innovations along the whole value chain and enabling companies to create and capture value in new ways. Key developments include the declining cost and increasing efficiency of information processing, automation and intelligent workflows. These advancements are not just reducing operational barriers; they are enabling entirely new possibilities.

Democratising the power of computing

By enabling the creation of novel content, solutions and automation, GenAI is poised to reshape industries, redefine workflows and democratise access to advanced problem-solving capabilities. This innovation represents the latest chapter in the historical journey of information technology, where successive waves of advancement have reduced the costs and barriers to processing, sharing and now generating information. To understand its significance, we need to explore how GenAI will disrupt the status quo and further evolve information technology.

The history of technology can be viewed as a sequence of advancements aimed at reducing the effort associated with information processing, input and distribution. Over the past few decades, two significant waves of democratisation have transformed the way we access knowledge.ⁱ Historically, the sharing of knowledge was a costly and inefficient endeavour, reliant on physical methods like pen and paper or word of mouth. The introduction of computers in the mid-20th century marked the first major shift. By automating and democratising data processing, computers drastically reduced the marginal cost of crunching data and enabled businesses and individuals to handle information at an unprecedented scale.

The advent of the internet marked the second wave of democratisation, as it made information accessible across the globe. By minimising the marginal cost of sharing and accessing data, the internet catalysed extraordinary levels of distribution and communication across geographical barriers.

Today, we stand at the precipice of a third wave. GenAI goes beyond processing and sharing information—it generates new knowledge, automates workflows, and enables intelligent systems to act independently. This reduces the marginal cost of information generation, content creation and action, and opens the door to a future where innovation is accessible at scale and fundamentally alters how we interact with knowledge. ⁱⁱ

GenAI takes on the once human-dominated task of ‘inputting’, freeing us from this bottleneck and unlocking new realms of progress. No longer limited by the speed at which we can type or the information we can process, AI's ability to analyse vast datasets and generate novel outputs allows us to transcend these previous limitations. AI lowers the cost of capital for marginal tasks, improves the efficiency of existing machinery or algorithms performing these tasks, and fosters new complementarities between tasks, enhancing labour productivity in areas where it is utilised. ⁱⁱⁱ

Lower costs, greater automation

Parallel to this, GenAI has become increasingly economically viable due to significant cost reductions and improvements in performance. Training and inference costs for GenAI models have decreased, driven by advancements in GPU price-performance, model quantisation, and software optimisation. Smaller models, trained on larger datasets and enhanced through techniques like Reinforcement Learning with Human Feedback (RLHF), further lower costs while maintaining or even improving performance. The venture capital firm Andreessen Horowitz has termed this rapid increase in tokens at a certain price as "LLMflation,"  noting that this phenomenon has outpaced the historical cost reductions in other technologies, such as compute during the PC revolution or bandwidth in the dot-com boom. ^iv  Notably, the cost of an LLM of equivalent performance is decreasing by 10x every year.

These advancements have not only reduced costs but also increased the value of GenAI by enhancing the quality of generated content and automating knowledge work.^v Consequently, new use cases are becoming commercially viable at lower price points, paving the way for broader adoption and innovation.

This shift towards AI-driven input and the declining cost of LLMs has profound implications for automation. These implications can be categorised:

1. Automation of digital workflows

Generative AI introduces a new economic model for information technology, as it drives the cost of content generation, learning, and problem-solving toward zero. For businesses, this means that automating digital workflows becomes increasingly affordable and scalable, as GenAI systems reduce the reliance on explicit programming by enabling intelligent agents to orchestrate and execute tasks. Recent developments in technology, such as Anthropic's Claude, with its new computer use, has been trained to navigated user interfaces, with the ability to take control of a computer, navigate various applications and websites, and execute actions.^vi This marks a critical shift, where the marginal cost of digital action - processing and deploying solutions - continues to fall, bringing down costs not only for data processing, but for decision-making and creative work as well.

2. Reducing the cost of action

As GenAI drives down the cost of creation and decision-making, it enables businesses to scale operations without proportional increases and allows them to explore and experiment with new business models, products and services.^vii For example, companies might use GenAI to expedite research and development, creating prototypes or even testing new products virtually before investing in physical production. GenAI also fosters new operating models by facilitating co-piloted workflows. In areas like customer service, content generation and software development, AI agents work alongside human professionals by augmenting their capabilities and expanding the scale of their output. This model of co-piloted work will allow organisations to move quickly, explore uncharted knowledge and experiment with innovative service delivery without incurring traditional costs.

3. Enabling intelligent automation in robotics

GenAI’s integration into robotics is paving the way for smarter physical automation. From manufacturing to logistics, AI-powered robotics can perform complex tasks at a lower marginal cost, enhancing efficiency across industries. One of the most exciting areas where GenAI’s cost-lowering effects are evident is in robotics. As AI models improve, their integration into physical robotics becomes more feasible and economical. By embedding GenAI into robotics, companies can automate physical workflows at a lower marginal cost, making processes like inventory management, manufacturing and logistics more efficient and less reliant on human labour. The potential for AI-powered robotics could be transformative across industries, enabling smarter, self-optimising systems capable of executing complex tasks.

GenAI is more than an advanced tool—it is a new paradigm that redefines how we think about creation, automation and collaboration. By driving down the costs of content generation and process automation, GenAI democratises the power of computing, making it accessible to more people than ever before. As this technology evolves, it will continue to unlock opportunities cutting-edge innovation, transform industries and foster new business models.

GenAI research trends shaping the market

The democratisation of computing power, achieved through reduced costs and increased accessibility of technology like Generative AI, has paved the way for groundbreaking advancements. This evolution in computing is not merely about enhancing existing processes but also pushing the boundaries of what's possible. As GenAI continues to evolve, cutting-edge research is already transitioning into commercially viable trends, poised to reshape the future of various sectors.

Key trends making an impact include but are not limited to:

• The proliferation of Small Language Models (SLMs) – while the current technical landscape is dominated by LLMs, SLMs are gaining traction due to their cost effectiveness and efficiency. Leveraging advancements in parameter optimisation and fine-tuning, SLMs can deliver high performance in specific applications with the same high resource demands of larger models
• AI on the edge – deploying LLMs on edge enables AI systems to run natively on smartphones, laptops and other devices.  Localised training reduces latency, enhances privacy and lowers the cost of LLM training.^viii
• Neuro-symbolic AI – combines the data-driven learning abilities of neural networks with the knowledge-based logical rigour of symbolic reasoning. Neuro-symbolic AI can help improve interpretability and decision-making, addressing limitations in common-sense reasoning and causality which are critical for complex tasks
• Infrastructure as a competitive edge – Companies are joining the race to optimise AI infrastructure, as improvements in GPUs, TPUs and software stacks drive efficiency and reduce costs
• AI security and red teaming – it is paramount to address vulnerabilities in GenAI systems. Red teaming, which interactively simulates adversarial attacks, can increase trust in AI applications as it ensures robust defence against misuse, bias, misinformation and security breaches.

These new trends will make AI part of our operating fabric and AI driven workflows. At the same time, researchers face significant challenges as they strive to push the boundaries of what AI systems can achieve. Current AI systems excel at pattern recognition but still struggle with abstract reasoning and adapting to novel scenarios. The lack of common-sense understanding and the ability to infer causation rather than correlations hinder the decision-making reliability of AI. The ability for AI systems to be able to simulate intricate systems, plan multi-step tasks and sustain performance over extended interactions are also ongoing challenges in research.

Open-ended AI and the future of research

In terms of future research directions, an exciting new phase of research focuses on the emergence of open-ended AI. Open-ended systems continuously learn and are self-improving, refining their own capabilities through feedback and exploration. An influential paper from a group of Google Deepmind researchers defined a system as open-ended if, from the perspective of an observer, ”the sequence of artifacts it produces is both novel and learnable.” ^ix

Open-ended AI has the potential to tackle increasingly complex problems and drive groundbreaking innovation, as it mimics the process of human discovery and creativity, allowing AI systems to evolve dynamically over time. By pushing the limits of what machines can achieve, these open-ended systems have the potential to transform fields like robotics, automation, and knowledge. For example, the POET system demonstrates open-ended AI by training a population of agents in dynamic, evolving environments.^x Each agent is paired with an environment that changes over time, fostering the development of new skills and behaviours. Through billions of training steps, POET generates a diverse set of specialist agents that can navigate novel and intricate environments. 

The foundation for open-ended AI is being laid by a combination of techniques and advancements. Early advancements in Reinforcement Learning (RL) are enabling AI to operate without human intervention, learning autonomously in dynamic and unpredictable situations. Evolutionary algorithms and adaptive task frameworks ensure that challenges remain forever engaging, fostering continuous growth. Similarly, task generation techniques adjust the difficulty level of tasks to an agent’s capability so that they remain perpetually challenging yet learnable. The path towards open-endedness can also be augmented by self-improvement loops to facilitate the generation of new knowledge, as agents engage in tasks that expand the limits of their abilities. These innovations represent a stepping stone towards artificial superhuman intelligence (ASI), where AI systems surpass human performance on an unprecedented scale. ^xi

The road to open-ended AI and ASI needs to be tread carefully and responsibly.  A crucial hurdle lies in developing systems capable of self-improvement without compromising safety, ethics, or transparency. Open-ended AI must be guided towards generating knowledge and artifacts that are understandable and beneficial to humanity. Incorporating human oversight, explainability, and robust validation mechanisms will be critical to ensuring these systems enhance society responsibly.

Despite these challenges, the possibilities are immense. Open-ended AI, with its ability to evolve and adapt beyond predefined limits, represents an unprecedented leap toward artificial superhuman intelligence that could unlock a multitude of new discoveries, advancing scientific understanding and reshaping the way humans interact with technology. As we continue to refine the foundations of AI and explore the boundaries of open-endedness, we stand on the brink of innovations that could redefine modern society.

Conclusion

In closing, the democratisation of AI, driven by reduced costs and increased accessibility, is revolutionising knowledge creation and automation. However, this is merely the tip of the iceberg. As these cost reductions converge with the commercialisation of cutting-edge GenAI research, a new era of robotics is dawning, one where the digital and physical realms seamlessly intertwine.

In Part 2 of this blog, we will explore the exciting opportunities presented when we bridge this gap and integrate the power of GenAI into physical robotics. We will delve into the profound implications for businesses, society, and the future of work in a world increasingly shaped by intelligent machines.

__________________________________________________________________________

References

i The Third Wave of Mass Democratisation of Knowledge, Medium, 2023.

ii GenerativeAI — The Third Computing Era, Medium, 2023.  

iii The Simple Macroeconomics of AI, Daron Acemoglu, 2024.  

iv Welcome to LLMflation, Andreessen Horowitz, 2024.  

v Welcome to LLMflation, Andreessen Horowitz, 2024. 

vi Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku, Anthropic, 2024.  

vii The GenAI is out of the bottle, Dominik K. Kanbach, Louisa Heiduk, Georg Blueher, Maximilian Schreiter & Alexander Lahmann, 2024.  

viii How LLMs on the Edge Could Help Solve the AI Data Center Problem, AI Business, 2024.  

ix Open-Endedness is Essential for Artificial Superhuman Intelligence, Edward Hughes, Michael Dennis, Jack Parker-Holder, Feryal Behbahani, Aditi Mavalankar, Yuge Shi, Tom Schaul, Tim Rocktäschel, 2024. 

x Open-Endedness is Essential for Artificial Superhuman Intelligence, Edward Hughes, Michael Dennis, Jack Parker-Holder, Feryal Behbahani, Aditi Mavalankar, Yuge Shi, Tom Schaul, Tim Rocktäschel, 2024. 

xi Open-Endedness is Essential for Artificial Superhuman Intelligence, Edward Hughes, Michael Dennis, Jack Parker-Holder, Feryal Behbahani, Aditi Mavalankar, Yuge Shi, Tom Schaul, Tim Rocktäschel, 2024.