Panel
Building the Impossible: Technical Frontiers in GenAI for Enterprises | RAISE Summit 2024 | Paris
- Panelists anticipate discussions on talent acquisition, funding dynamics, product strategy, and regulatory impacts, with specific focus on the timeline of AI adoption.
- Product roadmaps include leveraging GenAI for features like background generation within approximately two years, with validation of external needs preceding internal development.
- Infrastructure strategies favor building custom solutions from scratch for core AI products to ensure deep control, though off-the-shelf infrastructure is preferred for non-AI native users due to high self-hosting costs.
- In-house development is planned for tasks requiring low latency (20 to 100 milliseconds) or where existing APIs face limitations in speed, safety, or architecture, while companies may aggregate different providers for varied tasks like summarization.
- Competitive moats are expected to be built through proprietary data in niche industries, reinforcement learning methods (RLHF/RLAIF), and user interaction loops, whereas generalist models lack defensibility without unique technology or network effects.
- Talent acquisition relies on local clusters in Paris and Armenia, with plans to hire globally to access state-of-the-art research and broaden the pool, acknowledging challenges in diversity for senior roles but optimism for future parity.
- Financial projections indicate GenAI integration will increase cloud operating costs by 10 to 20 times, creating a capital-intensive environment where high funding is necessary to avoid runway exhaustion and scale training.
- Investment trends show a current imbalance with a 1-to-10 ratio of training to inference spending, raising concerns of a bubble where much capital is allocated to training without immediate value generation.
- Regulatory expectations include uncertainty regarding the EU AI Act's enforcement and alignment with the law's spirit, with startups potentially facing heavier burdens than established entities, though the US market may offer a less regulated environment.
- Operational challenges involve managing the pace of innovation to avoid building non-useful features based on hype, as well as the difficulty of integrating research into product and sales simultaneously.
- Market consolidation and competition are anticipated where specialized startups can outperform large tech companies in specific verticals like real-time voice AI by focusing deeply on use cases rather than horizontal generalization.
- Long-term economic forecasts suggest a fraction of the economy will be automated, justifying the current high capital intensity required to scale models as the industry remains far from the end of scaling laws.
- Product evolution plans include offering drop-in replacements for inference engines like VLLM and using A/B testing to redirect traffic for model comparison and adaptation.