Building the Impossible: Technical Frontiers in GenAI for Enterprises | RAISE Summit 2024 | Paris

Adaptive (Baptiste Pannier, Co-founder & CTO)

• Builds a platform enabling companies to test, select, and improve Generative AI (GAI) models via user interactions.
• Maintains a strong bias toward building all infrastructure in-house, citing the team's history with the open-source Falcon LLM series.
• Proposes "Adaptive" as a drop-in replacement for inference engines (e.g., vLLM, TGI) featuring A/B testing tools.
• Utilizes RLHF (Reinforcement Learning from Human Feedback) and RLAIF (Reinforcement Learning from AI Feedback) to adapt custom LLMs using proprietary user data without exposing that data to the platform.
• Identifies hiring senior talent and non-technical co-founders managing product/sales functions as primary challenges.

PhotoRoom (Elias, Co-founder & CTO)

• An app with ~25 million users that leverages GenAI for background generation on product images.
• Initially adopted Stable Diffusion post-DALL-E 2 release, but shifted to in-house model development after validating high user demand for background generation.
• Switched to in-house architecture to overcome limitations of open-source models (high training time, speed constraints, inability to modify architecture, and safety issues like nudity/violence).
• Employs a validation-first strategy: builds initial features with third-party APIs (e.g., ChatGPT) to validate product-market fit before internalizing infrastructure.
• Cites the EU AI Act as a catalyst for proactive safety measures, training models specifically to prohibit violence and nudity prior to regulatory mandates.

Google Cloud (Mathieu Valland, AI/ML Specialist)

• Observes non-AI-native end-users currently prefer "off-the-shelf" infrastructure due to the prohibitive Total Cost of Ownership (TCO) of building and maintaining custom GPU infrastructures.
• Notes that bleeding-edge applications or state-of-the-art models still require self-hosting.
• Acknowledges the difficulty of tracking rapid market innovation, describing it as a "full-time job" with inevitable information gaps.
• Describes the "catch-up" challenge as the primary organizational hurdle following a late start compared to competitors in late 2022/early 2023.
• Proactively engages with regulators and adheres to internal AI principles, a practice deemed resource-intensive for smaller startups.

Crisp (David Begassarian, CEO & Co-founder)

• A 7-year-old company focusing on real-time, on-device AI for noise cancellation and accent localization with latencies of 20–100ms.
• Keeps real-time, latency-critical speech tech in-house while utilizing external APIs for LLM-based tasks like meeting summarization and action item extraction.
• Developed its own on-device speech-to-text technology by fine-tuning open-source models over two years.
• Leverages a unique talent pipeline by recruiting from Armenia (where the founders originated) and the Paris ecosystem, avoiding global competition for senior talent.
• Argues that defensibility in AI exists in niche, horizontal-unfriendly use cases where deep technical focus (e.g., real-time voice) allows startups to outcompete generalist giants.

Infrastructure Strategy & Build vs. Buy Decisions

• Building in-house is preferred when AI is core to the product, requiring deep control for evolution; external APIs are used for validation or non-core functions.
• The transition to in-house models occurs when existing solutions hit technical roadblocks regarding speed, safety, or architectural flexibility.
• A common industry pattern involves using multiple API providers for LLM tasks and building internal benchmarks to switch between models as needed.

Data Moats & Defensibility

• Building a moat with data is feasible in niche industries but difficult in generalist markets where "no one has a moat."
• High-quality, unbiased proprietary data remains the primary constraint for creating defensible custom models.
• RLHF and RLAIF act as "multipliers" for data efficiency, allowing models to improve rapidly with fewer samples.
• Network effects, unique technology, and specific use-case focus remain the core frameworks for defensibility alongside data.

Talent Acquisition

• High talent concentration exists in Paris, facilitated by the DeepMind club and local academic output.
• PhotoRoom hires across Europe to access researchers in academic labs, leveraging public research papers to identify and recruit top talent.
• Crisp capitalized on a non-existent AI sector in Armenia to build a dominant local talent pool.
• Junior developer roles offer higher gender diversity potential compared to senior profiles or classic engineering roles; research scientist roles already show more mixed gender ratios.
• Top talent is attracted by ownership stakes, exciting projects, and competitive compensation.

Fundraising & Valuations

• GenAI features increase cloud inference costs by 10x to 20x compared to traditional operations, necessitating significant capital for growth.
• High valuations are justified by "scaling laws," where deterministic model performance increases with compute investment.
• Investment flow is heavily skewed toward training (capital intensive) rather than inference (value creation), creating a risk of imbalance.
• The industry risks a bubble if training investments outpace actual market exchange and value generation (e.g., Stability AI reportedly ran out of runway).
• PhotoRoom recently closed a $43 million Series B led by Balderton.

Regulation (EU AI Act & US Dynamics)

• Regulation is viewed as a source of uncertainty; the "spirit of the law" and enforcement mechanisms often lag behind written text.
• PhotoRoom adopted proactive safety filters (blocking violence/nudity) before the AI Act, driven by internal ethical principles.
• Large enterprises (e.g., Google) can afford proactive regulatory engagement; startups face higher resource barriers.
• Concerns exist that regulations (like compute limits) could stifle innovation while benefiting larger incumbents who can navigate compliance.
• Consensus suggests regulating specific AI use cases is acceptable, but regulating the underlying science or compute limits is detrimental.

Major Challenges in Building AI Companies

• Hiring: Securing senior talent and diverse teams remains the primary hurdle.
• Product Focus: Avoiding the trap of building features based on "shiny" new papers that do not deliver user value.
• Research-to-Product: Integrating deep technical research into viable products, marketing, and sales simultaneously is uniquely complex.
• Operational Scale: Managing the high cost of GPU inference and maintaining competitiveness against rapidly evolving global competitors.