Conference Presentation, Panel
'International Perspective on GenAI Views from the Frontiers' | RAISE Summit 2024 | Paris
RAISE SummitRodrigo Liang, Patrick Perez, Michael Kratsios, Benjamin Cistecky, Chandra Khatri, Antoine Moyroud, Antoine Moreau
- AI is projected to deliver significant upside in defense while lacking risk-free technology, with anticipated adverse societal impacts requiring proactive management to address potential winners and losers.
- Near-term risks include large-scale disinformation and deepfakes exacerbated by upcoming election schedules in key jurisdictions.
- Governments and leaders are expected to prioritize workforce reskilling and upskilling to mitigate adverse effects on employment.
- Large language models are predicted to increasingly permeate citizen lives, healthcare, and mission-critical sectors including home security and defense within the next five to ten years.
- The trajectory of the European Union's hypothetical 27 large language models remains uncertain, though the region aims to anchor itself as a global AI trust hub and innovation center for safety assurance due to regulatory clarity from the EU AI Act.
- The US is expected to maintain its top-tier leadership in breakthroughs, while a global race is anticipated where nations develop sovereign LLMs to control their destiny, leveraging customized datasets that reflect local culture, tradition, and heritage.
- Middle East countries, particularly following the UAE's success with the Falcon model, show obsessive interest in deploying models, whereas Western Europe may see slower adoption rates compared to the US.
- Samba Nova plans to deploy privately and securely accessible trillion-parameter models via a full-stack solution to simplify access, while also releasing Samba Lingo with record-breaking performance in low-resource languages to facilitate fine-tuning for business use.
- Hardware innovations are expected to improve cross-language learning efficiency, enabling better accuracy in subsequent languages (e.g., the 10th, 11th, or 13th).
- Crewtrim anticipates the Indian government will increase oversight as the market grows, focusing on indigenous tokenizer development to reduce inference costs and address code-mixing challenges in Indian languages.
- Local relevance in India requires building custom evaluation benchmarks to replace the current lack thereof, ensuring queries return region-specific data and leveraging fast data acquisition for powerful local models.
- Open models and open science are viewed as crucial for fine-tuning, maintaining data quality, and addressing shortages in evaluation benchmarks and knowledge transfer across modalities, with a specific push to maintain this openness in Europe and France.
- Performance drop-offs and cultural or political biases are expected in low-resource languages, potentially slowing the transition from hundreds of millions to billions of users unless technologies successfully cross language and cultural borders.
- Future adoption will require significant B2B progress alongside B2C growth, driven by the industry's ability to technically address specific security, safety, and robustness requirements for enterprises and the public sector.
- The next five to ten years are characterized by high complexity and movement, where conviction based on current observations is held loosely given the rapid pace of change in the AI landscape.