AI tools are arguably changing the nature of our work. Large language models (LLMs) can already produce reports on my own research papers that can rival reports made by humans. Unlike humans – who are always short on time – an LLM has access to a lot more information at a moment’s notice and generally has fewer biases. The AI also points out my analytical weaknesses, proof checks and makes suggestions for improvements. Reports created by humans are rarely better than this. Nevertheless, the excessive enthusiasm being shown in the market regarding AI has now become a cause for concern. So it makes sense to consider where things could go wrong in the AI supply chain. The AI supply chain starts with infrastructure producers and designers: companies like TSMC and Samsung make chips; Nvidia designs them and Cisco provides the connectivity. After this come hyperscalers like Amazon, Google and Microsoft. They are building data centers for use with their own AI models and to sell computing capacity to others. Apart from hyperscalers there are also companies like Equinix (data center) and developers of arguably the seminal LLM – Anthropic and OpenAI. Last but not least, there are the end-users of AI services. The use of AI by individuals is growing rapidly and corporate use in some areas (such as software development and customer support) has also grown at an explosive pace. But most large companies have not yet implemented end-to-end uses. Many still need to organize their historical data and reorganize traditional operations so AI can be deployed to drive further improvements. Additionally, many firms are concerned about data security, errors and misinformation that could destroy their brand image. Due to all this, AI rollout can be disrupted in many ways, which will create risks. For example, if graphics processing units, CPUs, and memory chips become faster and more energy-efficient, the value of equipment in existing data centers could decline rapidly. Currently, AI labs are investing huge amounts of money to train new and bigger models. Their idea is that the first model to reach a magical point – where it becomes self-improving – will rule the AI world and reap huge profits. But so far no model has gained any lasting advantage. Unless Gemini (Google), Cloud (Anthropic) and ChatGPT (OpenAI) differentiate themselves by attracting specific segments of users (or by merging or partnering with each other), it is difficult to see where the returns on their huge investments in training will come from. Politicians may be watching from the sidelines for now, but they will sooner or later make policy interventions to address AI risks and concerns. As data centers consume huge amounts of electricity – increasing the cost of electricity for everyone – there will be increasing political pressure on governments to limit their construction. For example, several counties in Indiana have recently announced moratoriums on data-center construction. Projections already made for next year suggest that hardware manufacturers and data centers will be unable to supply enough US computing power. While widespread use of AI may take longer than many expect, its malicious use by hackers and deepfakers, and its uncontrolled use by children, is growing rapidly. In such a situation, the voices of those demanding regulation and more accountability for AI models will become louder. The risks posed by AI could also spur dialogue among the world’s major powers, perhaps leading to some kind of AI Geneva Convention. Given all these uncertainties, it is not clear how widely and how quickly AI will be rolled out, and who will benefit from it. So far, no AI model has gained any lasting edge over the other. It is difficult to see how the huge investment in training by Gemini, Cloud or ChatGPT will be remunerated. (@Project Syndicate)
Source link
Raghuram Rajan’s column: ‘AI’ is proving to be a loss-making deal
