Small Is Better

Small Is Better

What has tech industry watchers and experts most concerned is the possible advent of AGI. At the same time, there’s no clear consensus on what exactly AGI means. Nvidia’s Jensen Huang said in March 2024 that AGI could soon become reality.53 Vested interests in furthering the already massive craze for AI (which has resulted in never-seen-before demand for the company’s semiconductor chips) aside, he clarified that he defined AGI as a machine being able to do something 8 per cent better than most people. This could be acing an exam or a logic-based game. However, there’s no clear consensus on what qualifies as AGI. In broad terms, it would be when machines reach a stage where they can do something with as much or more competence than a human being. Mustafa Suleyman, head of Microsoft AI and founder of DeepMind, said that artificial capable intelligence would be when a machine is given US$100,000anditcanturnthatintoUS$1 million without any assistance.54 The company believes—as of May 2024—that current AI models are equivalent to unskilled human beings if it were to map them on a six-level scale of AGI.

Meta chief AI scientist Yann LeCun said in an interview with Time magazine in 2024 that the current LLMs that power AI chatbots are not on a path towards human-level intelligence. He said that these LLMs work only based on how they have been trained and are incapable of doing anything else beyond that.55 While there are reports of Big Tech saying that AGI could be a reality within the next decade, it is important to keep in mind exactly where we are at currently in the AI development cycle. GenAI, while a large step forward in what AI systems can comprehend and achieve, is still predictive at best. It can provide responses only based on the information it has available and cannot think for itself.

Even while experts disagree over the exact definition of AGI, there is some consensus that it would need to meet some basic criteria, namely being able to outperform or match humans at a range of tasks and not just a single specific one. OpenAI goes on to add that to be considered an AGI, an AI system must also add economic value of some sort.56 At present, we do not have systems that are capable of general intelligence or intelligence across multiple fields. While DeepMind may be able to beat the world’s best players at a game of chess, it will fail miserably if asked to navigate a self-driving car—or even if asked to play another board game. ChatGPT or other LLMs can provide responses only within the boundaries of the information they have access to and the specific questions asked. There’s no adaptability or the ability to adjust their responses on their own unless they receive explicit instructions to do so. These instructions, or prompts, are still being provided by humans. Given this, leaders need to spend more time focusing on how they can harness the existing AI systems to benefit their business and customers instead of getting caught up in concerns about whether it will replace their entire workforce overnight.

One area that is showing scope for significant and more widespread impact is the development of small language models (SLMs). These SLMs are a more streamlined version of LLMs like ChatGPT or Gemini. These are focused on a smaller dataset or application, requiring less infrastructure and architecture, making them cheaper and quicker to run. SLMs currently comprise fewer parameters—a few million at best, compared to LLMs where it runs into billions and even trillions of parameters. This makes them quicker and more efficient to operate and makes them adaptable to applications like personalised chatbots or assistants. These can also be deployed relatively quickly and are more agile than their larger counterparts. At present, enterprises are experimenting with SLMs using their proprietary data, customising them for their specific needs. However, the small size also brings certain constraints, particularly the risk of producing inaccurate or less nuanced responses as compared to an LLM.

While the size of the LLMs enables them to cover a wider range of prompts, it also makes them unwieldy. Technology companies are increasingly working on creating effective SLMs that are easier to implement and monetise. These models are aimed at carrying out simpler tasks and are easier to fine-tune to meet specific customer needs. Big Tech is currently focusing a large part of its energies on shrinking these models to make them quicker and easier to scale.57 This will make GenAI much more accessible to a larger audience and will also make it easier for companies to adhere to ethical AI guidelines.

Another reason for the rapid speed with which GenAI is scaling up is the primarily open-source model that several companies like Meta and Elon Musk’s xAI have adopted. The rapid rise of automation and AI makes it critical for individuals and enterprises alike to focus on upskilling and adapting themselves to the new reality. One advantage that existing employees have over any possible external hire is the latent knowledge they possess about the industry, specific company and role. Because AI is still an evolving field, companies are struggling to find highly skilled specialised talent. They are realising that if they can upskill and retrain their employees, especially the high performers, then they would be in a far better position to take advantage of the automation and AI boom. This is something leaders will have to drive top down and make it a C-suite imperative to train not only themselves but also their top employees. This is a critical time in the transition of the organisational landscape, and it would help leaders to take a more long-term view of how humans and machines can work together, rather than thinking of it as one over another. It’s important to start thinking of how machines can be used to further augment human capability and find a way where this can work to the benefit of the enterprise and even society at large. Companies who effectively manage this transition will be well placed to take advantage of the changes and benefits this will drive in the next few years.

Re-imagining the workplace for a world in which humans and machines work together will require a lot more than just thinking about what parts of a task/job can be automated. It will require coming up with an entirely new approach to thinking of how the workplace of the future will look and redesigning it to enable optimal human–machine interaction. Given how widespread the adoption of GenAI is expected to be in the next few years, companies who are creating these AI models and those who are adopting them will have to work hard to mitigate the possible risks and security concerns. The possibility of a widespread cyber-attack driven through a chatbot is no longer the stuff of science fiction. The ability to create malicious code using natural language has made it extremely easy for a person to launch a cyber-attack. This could impact individuals, enterprises and nations, all in varying degrees of severity.