'GenAI at Scale Measuring Enterprise Outcomes in the AI-First World' | RAISE Summit 2024 | Paris

Morgan Stanley & Financial Services

• Deployment Status: Morgan Stanley became the first financial firm to partner with OpenAI, deploying dozens of production use cases.
  • Specific implementation includes an assistant for financial advisors to retrieve regulations and analyze institutional research.
  • Call summarization tools allow advisors to quickly generate client takeaways after proofreading recordings.
• Team Composition: The Machine Learning Research Group consists of 20 researchers, led by Dr. Yuri Nevmevaca.
  • Output Metrics: The team publishes approximately a dozen papers annually at top-tier conferences (NeurIPS, ICML, ICLR).
  • Talent Acquisition: Recruitment has shifted from high friction in years 1–3 to a "flywheel" effect where the team now attracts top French and global talent through demonstrated research credibility.
• Open Source Strategy: The firm utilizes open-source libraries (PyTorch, TensorFlow) to build upon existing foundations but maintains a clear boundary.
  • Proprietary code tuned for production and trade secrets remain internal.
  • Publicly released code includes research byproducts that do not reveal core production logic.
• Future Research Focus: Priority is shifting toward applying LLM and diffusion model research to time-series and tabular data, a "forgotten modality" in current AI trends.
• Regulatory Outlook: Regulation is expected to become a primary constraint, necessitating a trade-off between rapid experimentation and strict centralized governance for safety.

BCG & Strategic Consulting

• Leadership Change: Managing Director François Candelot announced he will leave BCG in May to join French private equity firm 72 (formerly Apex).
  • Investment Mandate: He will focus on helping portfolio companies implement AI and invest in AI startups, with a specific criterion for target startups: profitability (or pro forma profitability).
• Human-AI Collaboration Findings: Internal experiments at BCG revealed a divergence in value creation capabilities.
  • Creativity: AI excels at idea generation and obtaining a high volume of ideas.
  • Problem Solving: Humans currently outperform AI in complex problem-solving tasks.
  • Diversity: Human-AI collaboration results in lower solution diversity compared to human-only teams.
• Process Re-engineering: Performance prediction based on human-only skills is not predictive of performance with AI; recruiting processes must be revised to account for new skill mixes.
• Implementation Model: BCG utilizes a "10-20-70" effort model for AI adoption.
  • 10% of effort is dedicated to the technology itself.
  • 20% is dedicated to infrastructure and data.
  • 70% is dedicated to change management, human adoption, and workflow integration.
• Risk of Disillusionment: The primary risk to AI adoption is not technological failure but the inability of traditional companies to integrate humans and technology effectively within workflows.

Databricks & Data Infrastructure

• Product Launch: Databricks launched DBRX, its first open-source Large Language Model (LLM).
  • Performance: DBRX is reported to be twice as powerful as the Llama 2 model while being less expensive to run.
  • Data Sovereignty: The model allows companies to retain intellectual property and data privacy while utilizing open-source technology.
• Market Positioning: Databricks claims to be the only vendor currently leading Gartner's Magic Quadrants for both Data Management Systems and AI.
• Adoption Readiness: Despite a high volume of customer experiments, the panel noted that few companies have achieved GenAI at true enterprise scale due to data platform blockers.
  • Prerequisite: Successful GenAI strategy requires a mature data platform; the same companies that maximize value from predictive AI are likely to succeed with GenAI.
• Case Study (Michelin): Used LLM-driven supply chain optimization to align demand forecasting with product distribution efficiency.
• Case Study (Burberry): Implemented LLMs to classify hundreds of images for advertising campaigns.
  • ROI: Achieved a 70% reduction in classification time, equating to approximately 10 Full-Time Equivalents (FTEs) saved.

GenSyn & Decentralized Compute

• Business Model: GenSyn operates as a decentralized machine learning protocol, described as "Airbnb for GPUs."
  • Utility: Allows individuals and entities to lease idle GPU capacity (from gaming PCs, research labs, or data centers) to a global network.
  • Cost Advantage: The protocol is priced 75–80% cheaper than standard cloud providers (e.g., AWS).
  • Value Drivers: Elimination of vendor margins, deflationary scale effects, and utilization of underused hardware.
• Technical Capabilities: Supports distributed training for small foundation models using techniques like LoRA, PEFT, and quantization on consumer-grade hardware (e.g., RTX 3090).
• Security Posture: Current deployment focuses on open-source training with no privacy guarantees, as this is the primary market interest.
  • Future Roadmap: The team is exploring fully homomorphic encryption to enable secure, private training in future iterations.
• Ecosystem Impact: The protocol is expected to complement NVIDIA by increasing utilization of data center chips while stimulating demand for consumer-grade gaming GPUs.

Alaya & Enterprise AI Platforms

• Platform Function: Alaya serves as middleware for the AI stack, bridging cloud environments and applications.
  • Deployment Scope: Supports on-premise, traditional cloud, and European sovereign clouds (e.g., Outscale, Scaleway).
  • Focus: Secure AI and GenAI bundles for handling sensitive data.
• Strategic Philosophy: Successful GenAI requires an adversarial combination of predictive AI (for scoring/risk) and generative AI (for content creation) to build trust and utility.
• Adoption Drivers: Success depends on integrating corporate culture, local reinforcement learning, and continuous user feedback loops.
• Success Metrics:
  • Nuclear Industry: A project combining predictive and generative AI for specific industry language achieved a 30–35% business ROI.
  • Market Analysis: Generic use cases involving knowledge base creation, risk scoring, and insight generation are identified as highly deployable across industries.

Consensus Challenges & Forward-Looking Statements

• Adoption as the Critical Path: The panelists identified user adoption and the integration of AI into daily workflows as the single most significant hurdle, surpassing technological limitations.
• Open Source Preservation: Harry Grieve (GenSyn) emphasized that preserving the open-source nature of AI models is critical to preventing dystopian regulation of future brain-machine interfaces.
• Regulatory Landscape: Both Yuri Nevmevaca (Morgan Stanley) and Antoine Courret (Alaya) highlighted the necessity of "open regulation" and centralized governance to ensure models remain safe and interoperable.
• Cultural Integration: Leadership must understand the technological revolution to drive adoption; without leadership buy-in, the 70% "human effort" required for implementation fails.
• Data Sovereignty: European companies face a specific challenge in balancing scalability with data privacy, making sovereign cloud providers a key differentiator in the region.