Conference Presentation, Panel
Architects of Transformation: Foundational Models in Generative AI | RAISE Summit 2024 | Paris
- Business customers anticipate the need for legal certainty in generative AI applications, specifically avoiding copyright risks from training datasets like the Lion set while demanding content that maintains strict brand consistency in style and objects, despite current limitations in balancing creativity with this determinism.
- Enterprises expect the industry to develop deterministic metrics, evaluation standards, and a single accountability number for AI output quality to counter the stochastic nature of current models, moving away from the current lack of solutions for testing and quality control.
- Market participants predict that AI will be underestimated in the long term while intermediate effects are currently overestimated, advising patience for the next 16 months while exercising caution during the immediate three-month period.
- Providers are expected to facilitate specialization through instructions and guardrails while delivering day-to-day evaluation systems, and data connectivity to agents or models is projected to become increasingly seamless over the next 16 months.
- The industry forecasts a shift where businesses target niche markets or small applications where generative AI excels within this year or the next, though the cost-to-ROI balance for Retrieval-Augmented Generation (RAG) remains unresolved.
- Future developments in image generation will focus on controllability and on-brand personalization, with video generation anticipated to become the next frontier either later this year or next year, contingent on the availability of increased compute resources.
- Financial services adoption is expected to require a 99.999 error rate Service Level Agreement (SLA) before transitioning from human-in-the-loop verification to fully automated AI resolution.
- The open-source community is predicted to maintain its role in driving transparency and forcing big tech to release models similar to Google's Gamma and Meta's Llama, though commercial providers like Adobe Firefly are expected to gain trust over open-source options due to guaranteed data respect and IP protection.
- The panelist notes a larger pool of top AI researchers and engineers in the US compared to Europe, alongside a paradox of choice for AI scientists navigating the decision between large companies and smaller firms, while talent acquisition trends favor systems that respect creators in an "iTunes way."
- The adoption of generative AI in the media industry was previously stalled by the Hollywood strikes and copyright mistrust, and businesses are currently expected to use giant AI for testing without constant traffic only until a clear ROI is identified in production.