Conference Presentation, Panel
AI Unleashed : Navigating the Future with Next Gen Applications | RAISE Summit 2024 | Paris
- Panel Context: The "AI Unleashed" panel at the RAISE Summit, moderated by Nicolas Guillon, features representatives from Animage (Sixt), X-Plane/Explain (Arthur), and L'Oréal (Jean-Paul) to discuss the transition from first-generation GenAI (ChatGPT, Midjourney) to second-generation verticalized, agentic applications.
- First-Generation Limitations: Initial GenAI tools are viewed as "insane" but insufficient for niche, high-stakes industrial needs requiring proprietary data sets and specific human-centric constraints, such as the unique body proportions and movement logic of cartoon characters.
- Verticalization Strategy: Animage requires proprietary data sets and verticalized applications to create specific media franchises, as generic models cannot replicate the "specific human body proportions" and behaviors needed for their content.
- Cost of Specificity: Sixt notes that creating niche AI applications often necessitates access to private, non-public data sets, which are critical for providing more viable information than generic models can offer.
- Second-Generation Focus: Arthur Levy-Explain argues that the second generation of GenAI must move beyond "where are the real use cases?" to solve complex, text-heavy corporate workflows, such as public sector tendering.
- X-Plane Application: X-Plane is building an agentic platform to automate the public tender response process for clients like Veolia, Vinci, and Engie, addressing the pain points of reading hundreds of pages of bureaucratic text and writing proposals.
- Agentic Architecture: The proposed second-generation apps utilize "agents" (chains of small AIs) rather than single solutions to fully automate or facilitate integrated workflows, moving from the "end of the answer" to the "last mile" of action.
- Microsoft vs. Vertical Debate: Jean-Paul highlights a strategic divergence where companies must choose between building vertical, domain-specific agents or integrating AI into all-in-one generalist platforms like Microsoft Copilot, Google Assistant, or SAP.
- Interface Criticality: Success depends heavily on interface design; Mustafa Suleiman's concept of "interaction and action" requires tools to be packaged into user-friendly interfaces that empower workers, not just theoretical cloud APIs.
- Demo-to-Production Gap: The primary challenge identified is bridging the gap from experimental demos (e.g., Sora, AutoGPT) to live, professional production workflows where reliability and quality are non-negotiable.
- Implementation Barriers: Animage notes that implementing tools like Midjourney or Sora requires overcoming cultural resistance and ensuring "controllability," as artists must adapt to new workflows where jobs are significantly changing.
- Orchestration Workflow: Animage has created a fully AI-assisted workflow by integrating off-the-shelf tools (Suno, 11 Labs) with proprietary text-to-motion models, allowing for manual adjustments in 3D environments like Unity or Unreal Engine to ensure Hollywood-level quality.
- Cost Constraints in Production: Arthur Kleinmann states that production-scale AI cannot rely on premium models like GPT-4 due to prohibitive costs (e.g., €10/user/day), necessitating the use of lower-quality models optimized for efficiency.
- Latency Requirements: Unlike demos, production applications must minimize latency; delays of even 40 seconds render a tool unusable for daily workflows, requiring architectural hacks that are impossible in experimental settings.
- Verifiability and Hallucinations: Arthur emphasizes that production apps must account for inevitable AI hallucinations by building interfaces that allow users to easily verify, source, and edit information, rather than assuming perfect accuracy.
- Determinism vs. Reality: While first-gen demos tolerated errors, second-gen business applications face higher stakes where hallucinations are unacceptable; the solution lies in robust systems that function despite errors, not in eliminating them entirely.
- L'Oréal's "Beauty Genius" Demo: L'Oréal is testing "Beauty Genius," a conversational AI beauty coach, to address the need for ultra-personalized diagnostics and product recommendations in a consumer-facing context.
- Predictability Challenges: Jean-Paul identifies predictability as a key hurdle for L'Oréal, noting that users expect consistent computer outputs, whereas generative AI often varies, requiring significant work to ensure consistency.
- Empathy and Tone: For the beauty sector, AI must avoid "robotic" language; the interface must be conversational and fluid to match the empathetic nature of personal beauty advice, a challenge not present in industrial applications.
- Liability and Accountability: A major unresolved issue across industries is determining liability for AI errors (user vs. software vendor), which complicates the adoption of tools in high-stakes environments like legal tendering or medical advice.
- Media Industry Disruption: Sixt predicts AI will reduce animation production costs by 90%, drastically shrinking the workforce required for traditional methods (e.g., 800-1,000 people for a Pixar movie) and shifting value from fabrication to creativity.
- Workforce Transition: Legacy studios like Disney face a "cultural challenge" in transitioning 80% of their traditional workforce to AI workflows, whereas native AI companies like Animage can build from scratch with AI-centric production methods.
- Historical Comparison: Sixt frames the current AI shift as the third major media revolution, following the transition from silent to sound films and 2D to CGI, with the first fully AI-generated movie expected in the coming months.
- Public Procurement Scale: Public tenders represent 12% of global GDP but are highly inefficient; clients currently spend 4-5% of their total budget just to respond to tenders without winning a contract.
- Efficiency Gains: X-Plane's AI tools have already enabled clients to consume documents 5x faster and extract 3x more information, with the long-term vision of reducing tender response times from one month to a few days.
- Quality over Quantity: The ultimate goal for public sector AI is not to spam governments with volume but to improve the quality of proposals and government documents, avoiding the "Kafkaesque" flooding of administration.
- Personalization Trend: Jean-Paul notes that L'Oréal's shift toward "beauty for each" aligns with AI's ability to deliver ultra-personalized visual and content engagement, moving away from the standardized beauty ideals of the 2000s.
- Competitive Differentiation: Companies cannot differentiate solely by plugging in generic AI; they must integrate unique solutions into their workflows to add value, as the volume of generated content may actually increase the need for human review and customization.
- Recommended Tools: Panelists recommended Suno.ai for music generation, Microsoft Copilot/Google equivalents (which were deemed disappointing for business), and Obsidian for personal knowledge management ("second brain").