In the heart of San Sebastián, a peaceful Spanish city, a group of engineers is tackling a pressing issue: the energy-intensive nature of AI. Amidst the pine-covered hills, Multiverse Computing is developing a revolutionary approach to artificial intelligence, one that challenges the status quo and aims to make AI more accessible and environmentally friendly.
The Battle for Efficient AI
Multiverse's mission is to create smaller, more efficient AI models, believing that most users don't need the power of giant language models like ChatGPT. Their focus is on compression, shrinking large language platforms into manageable, affordable packages. This approach has gained traction, with clients ranging from manufacturers to defense contractors, including big names like Bosch and Moody's.
But here's where it gets controversial: some experts question the performance of these compressed models. However, Multiverse's clients swear by their effectiveness, with one recent project drastically reducing the cost of Telefonica's customer service system. The company envisions a future where 'smart' appliances, like fridges, use these small language models (SLMs) to provide instant updates on food replacements.
A Controversial Interpretation
Multiverse's CEO, Enrique Lizaso Olmos, argues that the AI arms race is unnecessary. He believes that most AI users have limited needs that can be met with smaller, more energy-efficient models. This perspective challenges the conventional wisdom that bigger is better in the AI world.
The Quantum Connection
The founders' journey to AI is an intriguing one. Lizaso, originally a doctor, found his passion in tech entrepreneurship after joining a WhatsApp group discussing quantum computing's potential in finance. This led to an academic paper in 2018, which caught the attention of Goldman Sachs and sparked the idea for Multiverse.
The company's products are now described as 'quantum inspired,' using quantum algorithms to train regular computers, resulting in faster, more efficient operations. This approach allows Multiverse to create SLMs with a fraction of the parameters found in large language models.
The Energy Crisis and AI's Future
Multiverse's work is timely, as the environmental impact of AI data centers is becoming increasingly apparent. U.S. data centers already consume a significant portion of the country's electricity, and this is set to rise dramatically by 2030. Multiverse's solution could stem this trend, but some experts remain skeptical. Théo Alves Da Costa, AI sustainability head at Ekimetrics, questions the performance of small language models, especially in languages like French.
Despite these doubts, Multiverse is pushing forward, determined to shrink open-source models to save energy and money. Their next project is a compressed version of DeepSeek, the Chinese generative AI model that made waves last year. Multiverse's version promises to be even cheaper and, true to their ethos, they've removed government censorship filters, allowing users access to politically sensitive information.
The future of AI is an exciting and controversial topic. Multiverse's approach challenges the norm and invites discussion: Is smaller AI better AI? What are your thoughts on this innovative startup's mission?
