AI Pioneer warns that funding hype is overshadowing progress
Demis Hassabis, co-founder of DeepMind, has expressed concern that the massive influx of capital into artificial intelligence is creating a hype similar to that of cryptocurrencies, which could overshadow the remarkable advancements in the field. As the head of Google’s AI research division, he said that the enormous investments being made in generative AI start-ups and products are accompanied by excessive hype and potentially deceptive practices, similar to other overhyped sectors like cryptocurrencies.
“In a way, AI’s not hyped enough but in some senses it’s too hyped. We’re talking about all sorts of things that are just not real”
Demis Hassabis
The introduction of OpenAI’s ChatGPT chatbot in November 2022 ignited a rush among start-ups to create and implement generative AI, attracting significant venture capital investment. Last year, venture capital groups poured $42.5 billion into 2,500 equity rounds for AI start-ups.
Investors in public markets have also been drawn to the leading technology companies, known as the Magnificent Seven, which include Microsoft, Alphabet, and Nvidia. These companies are at the forefront of the AI revolution, and their success has driven global stock markets to their best first-quarter performance in five years.
However, regulators have started to scrutinize companies for making misleading AI-related claims.
Pursuit for factual accuracy
Despite the hype and occasional misinformation surrounding AI, Hassabis, who was recently knighted for his contributions to science, believes that AI is one of the most transformative inventions in human history. He suggests that we are just beginning to realize its potential and predicts that the next decade will usher in a new golden age of scientific discovery, akin to a new Renaissance.
DeepMind’s AlphaFold model, released in 2021, stands as a testament to how AI can expedite scientific research. This model has facilitated the prediction of the structures of 200 million proteins and is currently employed by over a million biologists worldwide. DeepMind is leveraging AI to delve into various biological domains and hasten research in drug discovery, delivery, material science, mathematics, weather forecasting, and nuclear fusion technology.
DeepMind was established in London in 2010 with the objective of developing “artificial general intelligence” (AGI) that could emulate all human cognitive abilities. Despite some researchers suggesting that AGI might still be decades away or even unattainable, Hassabis believes that a couple more pivotal breakthroughs could lead us there. Given AGI’s potential power, Hassabis advocates for a scientific approach to its development, as opposed to the hacker approach favoured by Silicon Valley, underscoring the significance of AGI.
Recently, DeepMind researchers introduced a new methodology named SAFE, aimed at minimizing factual inaccuracies, referred to as hallucinations, produced by large language models like OpenAI’s GPT and Google’s Gemini. The inconsistency of these models has resulted in lawyers submitting fictitious citations and has discouraged many businesses from utilizing them.
Hassabis revealed that DeepMind is investigating various fact-checking methods for its models, such as cross-verifying responses with Google Search or Google Scholar. He likened this strategy to how the AlphaGo model mastered the game of Go by verifying its output. A large language model could also check if a response is logical and make necessary adjustments. Hassabis compared this to AlphaGo’s decision-making process, where it doesn’t immediately act on the first move it considers but takes time to plan.
When tasked with verifying 16,000 individual facts, SAFE concurred with human annotators sourced from the crowd 72 percent of the time, offering a solution that was 20 times more cost-effective.
