Interview, Fireside Chat
Getting the Most From AI With Multiple Custom Agents ft Dust’s Gabriel Hubert and Stanislas Polu
- The transition from deterministic tools to stochastic AI is viewed as the most significant shift in computing since the advent of the computer, though workforce adoption remains incomplete.
- Organizations are expected to adopt risk-reward ratios accepting imperfect results and a distribution of ROI in exchange for significant time savings, driven by users' willingness to trade accuracy for high upside scenarios.
- Multi-model integration is predicted to be essential as no single model will dominate, with access to proprietary data within silos identified as critical for unlocking AI power while maintaining privacy.
- Data security sensitivity will dictate model deployment strategies, potentially separating smaller on-device models for sensitive tasks from API calls to frontier models for less critical needs.
- Future trajectories depend on two potential paths: continuous rapid progression requiring dynamic model switching, or a plateau where every company can train large models on consumer hardware like a MacBook M6 within hours.
- If a technology plateau occurs, the need for model routing may disappear as token production becomes commoditized, potentially allowing open-source models to surpass closed-source alternatives if the latter offer no distinct value.
- Reasoning capabilities are currently viewed as having not advanced as quickly as optimistic predictions suggested, with concerns that the technology may be asymptoting despite ongoing research into formal mathematics.
- Infrastructure challenges, including massive cluster complexity and frequent GPU failures, make scaling to the next order of magnitude for generative models extremely difficult.
- AI adoption bottlenecks are identified as product and engineering layers rather than model quality, with a strategic preference for controlling data access over fine-tuning company data.
- Industry challenges involve applying established software guardrails and access controls to new AI interfaces like agents, while the user interface is expected to evolve from conversational models to co-editing and human-in-the-loop proposal systems.
- User adoption patterns typically show flat usage after initial use cases before reaching a critical mass of 70% company-wide, with a predicted trend toward specialized, distinct tools rather than unified agents.
- Long-term market value in five to ten years is expected to belong to companies providing human augmentation ("exoskeletons"), shifting the definition of productivity from speed to impact and better outcomes.
- Search functionality is anticipated to become abstracted as the focus shifts to task completion and proof generation, while over-automating customer support risks losing critical product development insights.
- A difficult period is forecasted where excitement will wane and the diffusion of AI's massive value through society will take a considerable amount of time.
- Founders are advised to accept market realities rather than fighting them, with the belief that success is achievable from France in the US market through ambition and strong conviction despite the lack of "magical capabilities" in the US ecosystem.