AI Can Replace CEOs When It Starts Thinking Like Us

Whether its training an AI to take over as CEO, or help you book tickets, the requirements are the same. Every single AI model worldwide needs to be trained on data.

The way training an AI works is that you provide it with a dataset and a specific use-case in mind. Then you refine the results it produces by constantly aligning the parameters to provide the answers required. But in the case of training an AI CEO, what sort of datasets does one need? Corover’s Sabharwal says that common behaviours across around 4,000 CEOs may be enough for a robust dataset.

For context, that’s over half of public companies listed on the National Stock Exchange (2,266 as of Dec. 31, 2023) and roughly 75% of the ones listed on the Bombay Stock Exchange (5,309 as of Jan. 24, 2024). There’s the size of companies to consider. The roles and responsibilities of a CEO at a mid-sized corporation would be vastly different as compared to one of larger company. But overlaps are likely.

“Every role has its own requirement and specific assessment indicators. The AIs we create will also have to work accordingly,” said Sanjeev Menon, co-founder of E42.ai. His company has been building AI co-workers that automate complex, procedure intensive processes, that have normally been performed by people.

But even to build an AI that can handle the role that a CEO plays, there still needs to be training on generalised capabilities. In that regard, the dataset requirement that Sabharwal mentions is important. It’s only once that is taken care of can one potentially think of narrowing training down to cater to a specific company.

“It’s never going to be a ‘one size fits all’ model. An AI will have to be trained on an organisation’s data to be really effective and it will have to go through reinforcement learning through human feedback for a while, for it to at least start being able to provide some level of decision-making support,” Menon said.

A core part of the human experience, CEO or otherwise, is the understanding of “concepts” and their application across different situations. The reason humans are able to apply concepts widely is because we’re able to draw on our own past experiences and knowledge to understand when and where to use them. This isn’t the case for LLMs and genAI models.

At the core, all generative AI models and LLMs work based off pattern recognition.

Technically, via machine learning, an AI model can be taught a specific “concept.” For example, to teach an AI what a cat is, researchers end up providing it with a lot of data—in this case, picture of cats.

For the AI to learn, it uses a set of rules already fed into the system to analyse the pictures and find patterns. Based on these rules, it is able to create its own image of what a cat looks like and then identify images of cats based on the patterns it has identified. This is broadly how most AI models are built, implemented and constantly tweaked.

“AI learning using conceptual knowledge in a completely new environment for a completely different problem altogether is the ‘aha!’ moment,” according to Menon.

Conceptual learning in AI has been a longstanding debate. Since at least the 1980s, philosophers and cognitive scientists have proposed that AI isn’t capable of interconnecting concepts and apply them in new settings, called “conceptual generalisations.” But recent work suggests that we’re a step closer.

A researcher duo from New York University and Spain’s Pompeu Fabra University published a paper last year, showing off a new approach to teaching AI, called multi-learning for compositionality. The technique works by “training neural networks—the engines driving ChatGPT and related technologies for speech recognition and natural language processing—to become better at compositional generalisation through practice”, according to a press release.

The research has found that this method outperforms existing approaches and is on-par and in some cases, better than human performance. While models like ChatGPT-3 and GPT-4 have previously showcased such behaviour in some instances, it’s not on the same scale as the researchers have found. AI models have previously struggled in emulating the kind of thinking that is seen in humans.

Multi-learning for compositionality can be used to teach AI models novel concepts, where we’ve previously struggled and let it apply concepts in cases where it couldn’t before.