Seeing Silicon | Concerns for society are high among AI researchers

At ICML 2024, scientists are divided between pushing the possibilities of machine learning and reflecting on how this technology affects society.

The first keynote at the 41st International Conference on Machine Learning (ICML) 2024, in Vienna, was about how the machine learning field is moving from being open source to closed, from collaborative to increasingly competitive and how resources and the compute power required for this research makes it exclusive and expensive.

Before 2022, AI technology was mostly limited to research projects inside labs in universities and companies. When OpenAI released ChatGPT for public in late 2022 – more as a test than anything else – its rapid adoption by users turned the mostly-academic field into a goldmine.(Representative Image/REUTERS)

“GPU compute and pretraining of data needs the deep pockets of billionaires, VCs and large corporates,” said Soumith Chintala, who leads AI projects at Meta in New York, as he addressed 500 attendees in one of the venue’s largest spaces. As compute and engineering become expensive, and issues in our society rise about data legality, safety and societal impact, companies have increasingly stopped open research. Though big companies are in a rush to be leaders in this research, they also care about data legality, safety, and societal impact, he added.

The excitement, the money, disruptive directions in AI and machine learning (ML) technology, whispers of new startups and job conversations were palpable in the hallways of Messe Wien Exhibition Congress Center, the glass fronted conference venue which was filled with thousands of students, researchers, scientists, professors, and industry professionals. As were conversations about visiting Vienna, attending jazz concerts, taking a swim in the Danube river in the last week of July.

“We started off as a close-knit worldwide community of about 500 researchers 20 years ago,” said Katherine Heller, research scientist in Responsible AI at Google Research and the program chair of ICML 2024, who has attended this AI conference regularly for two decades. For the past few months, she has been reviewing submissions, creating schedules, inviting speakers, deciding on workshops and tutorials for the week-long event. “This year we have reviewed more than 10,000 papers from 12,000 authors, and selected about 2,600 to showcase,” she said, describing the growth of the community involved in ML has been as “exponential”.

Before 2022, AI technology was mostly limited to research projects inside labs in universities and companies. When OpenAI released ChatGPT for public in late 2022 – more as a test than anything else – its rapid adoption by users turned the mostly-academic field into a goldmine. Overnight, ChatGPT became a household term and companies who had been working on generative AI and large-language models smelled opportunity – a general purpose technology that could not only be used for other scientific research, but perform code, search, and generate synthetic words, images, and videos for everyday use.

According to Bloomberg Intelligence, the market size of consumer generative AI is expected to grow from $40 billion in 2022 to $1.3 trillion by 2023. The potential of the market has since then rapidly infused talent, money, and more research into the field.

“AI research can dramatically impact issues of healthcare, education and climate and its important research which we shouldn’t miss out in India,” said computer scientist Sandeep Juneja who heads the Center for Data Learning and Decision Sciences at Ashoka University and was at the conference to present a paper, understand new directions of research in this fast moving hot scientific field.

Relentless pursuit of profit and talent

High corporate interest is pushing and highlighting certain type of machine learning, while pushing other kinds of ML research on the back burner. Thematically the most frequent keywords in the research papers were large language models (LLM), reinforcement learning, deep learning and graph neural network – all research that is focused on building generative AI and multimodal AI products for profit.

“The research community is the leading edge of what’s happening technically in the AI universe,” said Sheila Gulati, managing director and founder of Seattle-based Tola Capital, an enterprise AI venture capital firm, who was at the conference to find new directions in AI for investment.

One of the 10 best paper awards was given to GENIE (Generative Interactive Environments) a program developed by scientists at Google DeepMind that showcased a technology that created synthetic virtual environments based on text. Another award went to VideoPoet, a LLM which generated video from scratch created by scientists at LumaAI, Google DeepMind and Google Research.

“It takes money to do this research, to scale it up and impact more people,” said Heller who was an academician before moving into Google to work on healthcare systems. There’s a huge amount to be gained from having access to that kind of resources in the industry, but you have to align your goals with what the for-profit company is interested in most at the time, she explained. “Industry is capitalistic and researchers recognize that,” she said.

There’s a flipside to this breathless pace though. “Making sure that AI delivers consistently is driving research and stressing researchers out,” said Dr Sunita Sarawagi, professor and founding head, Center for Machine Intelligence and Data Science, Indian Institute of Technology Bombay, who has been in the field since 2003. “While it’s great that machine learning has so much impact, there are a lot of dangers in uninformed use of AI, and we need to be careful,” she said, adding that because of such a high demand for AI researchers in the industry, she’s finding it hard to attract students to her center to do academic research.

For AI/ML experts are in high demand not only for technology companies and startups but also for finance industry and academicians like Sarawagi need to compete with all. This showed in the recruiting channel for the conference. Everyday, a few dozen jobs were posted for Phd candidates, engineers, AI scientists and more, while recruiters actively interacted with students in the hallways. “I have seen some interesting candidates with good research and software skills,” said M Saquib Sarfraz, technical lead at Mercedes Benz Tech Innovation, based in Germany, whose one job posting for an engineer has already gotten more than 50 responses.

California-based Matthew Leavitt, co-founder and chief scientist at Datalogy AI, had sponsored the conference in hope of getting the brand name of his newly minted startup into the community. “The cost of sponsoring the conference and travel would be recovered if we hire even one candidate, for its less than what we pay a recruiter for an AI researcher,” he said, adding that there’s tough competition (their stall is right across from large Google installations) but they have been able to garner interest thanks to the fact that they’re working in data curation, “a frontier research problem”, which is an important research focus for a data-hungry AI community.

Reflection on how AI affects societies

As society understands the capability of the technology, the owners of AI models are facing governmental regulations, copyright violation suits and fears of potential security risks. This has caused an increased interest in ethical, social, copyright and safety implications of this technology. “There’s been an increased number of workshops, researchers, research topics, groups and labs that have been thinking about the societal impact of AI, how to mitigate the negative effect and accentuate the positive impact in a nuanced way,” said Weiwei Pan, assistant director for Graduate Studies in Data Science at Harvard University, who attends multiple AI conferences every year.

As the data used to train all these models is increasingly questions, another subfield of research has become data - its sources, legality, curation and ethical issues around its procurement. One of the best paper awards went to ‘Measure Dataset Diversity, Don’t Just Claim It’, a research paper which focused on diversity of data, and brought in social science principles into defining the diversity of datasets.

“We need clearer definitions of concepts like diversity, which can mean different to different people,” said Dora Zhao, a scholar at Stanford University and the primary author of the paper. Like Zhao, many researchers who are pursuing PhDs are increasingly focusing their work on the social implication of AI, which wasn’t the case a couple of years ago.

Another paper which also won the best paper award, made two AI models debate with each other to find the right answer for humans. “It makes us, the users, judges with two AI experts arguing,” said Akbir Khan, the primary author of the paper, who is doing a PhD at University College of London. Khan who ran a startup before getting into research feels that with multiple AI models available today, it should become easier for humans to spot the right answer between all of them. His research is the first step in that direction.

Is AI only for the rich?

Science is not equal, and that's especially true for machine learning. Of the 10000 attendees at this conference, half were from the US and Western Europe, followed by China which had about 1,000 people attending. The cost of the conference aside, there was minimal representation from countries in the Global South with only 100 from India. This is thanks to the high cost of doing AI research.

“The Global South lacks compute, resources and capability to compete in the AI space,” said Vukosi Marivate, professor at the University of Pretoria in South Africa, co-founder of Africa’s biggest research conference Deep Learning Indaba who gave an invited keynote at the conference. “There’s the US, China, a bit of Europe and then there’s India which like everybody else including Africa has the crumbs,” Marivate suggested that instead of pursuing profits, these countries should focus on the needs of their societies.

“I worry about whether the companies are overselling AI,” said Sarawagi, adding that computer science researchers need to figure out their community's needs when it comes to this technology. “AI might help developed countries reduce their dependence on humans but for us, it cannot solve societal issues like living conditions, hunger or climate change,” she said. Instead of following the developed world, we should build our own systems, encourage local and relevant R&D, and change governmental policies.

Shweta Taneja is an author and journalist based in the Bay Area. Her fortnightly column will reflect on how emerging tech and science are reshaping society in the Silicon Valley and beyond. Find her online with @shwetawrites. The views expressed are personal.