Keynote with OpenAI Insider Arka Dhar

Three years ago, Sam Altman asked Arka Dhar to join OpenAI as one of its early employees to train LLMs. It’s only been a few days since Arka stepped out of his role as Head of Human Data at OpenAI, but there’s no sign that he’s slowing down professionally. A self-professed ‘strong believer’ in AI and its possibilities, Arka is a deep thinker about this technology and a passionate advocate for distributing its benefits. ‘I’m extremely excited about what it means to get to AGI [artificial general intelligence]’, he says. ‘What kind of use cases and applications open up? What does it do to humanity?’

The case for Generative AI

Rather than try to count the innumerable use cases for Generative AI, Arka first lays out his thinking about its game-changing nature. ‘We humans can be myopic at times’, he smiles. ‘We can get very focused on an outcome or an answer, and then it becomes very difficult for us to explore alternatives or intersections which don’t naturally come to mind.’

These are exactly the areas where GenAI (Generative AI) excels. Arka shares an example: ‘If you prompt the model and say, “Hey, I have all these different variables, X is what I’m solving for, but I want to understand Y, Z, A, B, and C. Can you help me create all these different scenarios or situations? And based on the data, analysis, and insights that you have access to, can you give me things that I have missed in the past?”’ Along with its capacity for large-scale computation, Arka sees this quality of being non-myopic as one of GenAI’s main strengths..

‘AI is amazing at breaking down context,’ he continues, pointing to an unusual example in agriculture to illustrate how broad-reaching the possibilities are. ‘In certain countries, especially the developing world, a farmer might not have the tooling or [ability] to understand very, very complex soil reports or weather reports. They might not be able to use it to actually improve their yield. I have seen applications of this technology where that kind of complex insight can be broken down in a language they understand, with the list of actions that [they] can take.’ 

The key to success: integrate AI from the bottom up

Most farmers may not yet be using AI to optimise crop yields, but many firms are already working to integrate GenAI into their businesses. Arka cautions against deploying AI at scale too quickly, however, emphasising, ‘This is a new technology, and we’re still learning its capabilities and how it can be integrated into our businesses.’

Since ChatGPT’s launch in November 2022, Generative AI has also renewed the conversation about replacement of work. Arka thinks this perspective does a disservice to GenAI’s potential, as focusing on replacement overlooks the specific benefits of optimisation. ‘The most important thing is to pick a workstream that you want to see highly optimised’, he suggests, whether in marketing, finance, or market research. Initially, use cases may be rather basic, but they can become more complex and integral to critical work streams and workflows. Instead of massive transformation projects, Arka urges firms to ‘pick a workflow, pick a group of users’, and experiment with AI to augment their capabilities. ‘[How] can we give them this incredible superpower to vastly improve the throughput and quality of work?’ he asks. 

How far away is AGI, really?

We’re not yet at AGI (Artificial General Intelligence, the stage at which an AI surpasses humans across the full range of cognitive abilities), although the technology is increasingly within reach. One step towards AGI will be agents, like large GPTs, that are trained to represent their user’s interests – ‘digital twins’ that reflect what the user cares about, as Arka puts it. These agents can manage complex tasks in both personal and work contexts, even operating in the background. 

‘I do believe agentic systems are the future.’

– Arka Dhar

The main issue is supervision and how comfortable people are with letting an agent act independently. For Arka, the question remains: ‘At what point will I feel comfortable letting my agentic system make a life-or-death decision for me? And for me, we are not close to that yet’, he says. For now, these systems will likely remain supervised but could eventually handle the majority of our daily tasks.

The economics of AGI: can it supercharge labour?

The definition of AGI is another subject of ongoing debate, but Arka shares one lens through which he considers this technology and its potential: how much economic activity an AI system is able to replace.

To illustrate, he uses the Cobb-Douglas function of GDP growth, first giving a short recap of the function: ‘It’s an equation which gives you a sense of what are the primary contributors of GDP, and it’s a multiplication of TFP [total factor productivity] with two other components contributing to GDP, which are capital and labour.’

Arka continues: ‘When you look at the coefficients [of the Cobb-Douglas function] as tasks which are getting more and more automated, what we very quickly realise is: as we go down this path of extreme automation, capital contribution to GDP increases [and] labour contribution to GDP goes down.’

In simple terms, he says, ‘Capital is way more efficient at generating more capital. Labour hours have an intrinsic limit to what we can do. That’s why I talk about really supercharging labour. Can we have a focus of using this technology to vastly improve what labour is able to accomplish in the limited time that we do have when we work?’ 

Distributed benefits or a ‘tool of the rich’

With this major economic potential comes plenty of risk, particularly with asymmetrical availability. ‘The consequences of access to these technologies [versus] not having access to these technologies can be devastating’, Arka warns. His concern is that AGI could become a ‘tool of the rich’ and that access could be limited to a specific group of individuals or companies. ‘[If] the benefits are not distributed, I think that’s one of the worst outcomes of these technologies’, he says.

Considering the potential of AI not only to increase productivity for individuals or firms, but also to contribute to maximising GDP for companies or nations with the most access to this technology, economic inequality could be accelerated by AI as a competitive differentiator. ‘The gap between haves and have-nots is going to be significant, as if you are taking away people’s hands and legs and asking them to compete’, Arka says. Instead, he hopes to see the opposite: ‘I would wish [that] everyone focuses and works on distributing the benefits of this technology as widely as possible.’

Arka also notes that market mechanisms and regulations on AI could help to ensure more equitability, such as the profit cap at OpenAI that necessitates lowering the price of their products as the company grows. In general, he says, ‘Regulations are important for safety, […and they] have to get a lot smarter so that there is continuous competition and new entrants and the status quo always gets challenged.’

Tackling privacy concerns and environmental costs

Arka acknowledges that there are other areas of concern, especially when it comes to operationalising AI at scale.  ‘[AGI] does have safety repercussions, so we have to be very, very focused on those aspects of the technology’, he thinks.

He points to recent reporting that there’s been a ‘massive increase in emissions’ by large tech companies, and that running AI models comes with enormous costs in compute and energy. From a climate perspective, the benefits of Generative AI are currently weighed up against these environmental costs. As a self-described optimist, Arka thinks that the field will become much more energy-efficient over time. In his opinion, the main points of focus in this regard are: data, compute and algorithmic improvements. ‘All three of those main pillars will one day get us to a place where this technology can actually become quite prevalent and easily accessible’, he says.

Meanwhile, data privacy concerns are also growing. The vast amounts of personal data collected by AI systems raise ethical questions about how this information is used and safeguarded. But in Arka’s view, objections about training data for AI may become less central for general users as models improve. ‘Is [AI] going to learn more from me versus someone way smarter than me?’ he asks. ‘If the models are already reaching a level of reasoning and cognitive abilities which are significantly higher than the median population, that gives me hope that [people] don’t have to be data-farmed to improve […] the model’s performance.’

A ‘massive acceleration’: AI is here to stay

During his years at OpenAI, Arka has observed a shift across the entire field as the scope of AI technology has broadened. His own firm, Zinier, began as a tool for schedule optimisation; he believes that the possibilities of AI are much greater as more attention and resources are invested – as Arka puts it, a ‘concentration of brainpower across the world’ is now being devoted to developing AI. ‘Now what you have is a whole lot of researchers who are single-mindedly focusing on GenAI and LLMs’, he notes. ‘We are not talking about pure machine learning for schedule optimisation. We are talking about a general technology that we believe is the ultimate representation of AI advancements.’

Arka concludes:

‘I think we are going to see improvements in this technology a whole lot faster. AI is not new, but we are seeing a massive acceleration of what can be done with AI. So, I’m a strong believer it’s here to stay. It’s a pretty permanent change to how we live our lives.’ – Arka Dhar

Join the conversation

This interview with Arka Dhar took place during the Platform Leaders online event organised by Launchworks & Co on the 11th of July 2024 (full list of speakers and agenda). To watch the full event, you play the video below.

With hundreds of experts from around the world in attendance, Platform Leaders offers a unique chance for entrepreneurs, academics, practitioners, and policymakers to delve into key issues and shape the debate. Look for more insights to come from the latest event, including videos and articles, and join the community to stay updated on the next events and much more from Platform Leaders.