Getting the Most From AI With Multiple Custom Agents ft Dust’s Gabriel Hubert and Stanislas Polu

Market Dynamics and Model Landscape

• Dust predicts a "bimodal" future for AI where multiple models will coexist rather than a single "one model rules all" paradigm.
• Scenario A (Continuous Scaling): Technology continues to improve rapidly, necessitating a switchable architecture where users select the best model for specific use cases at any given time.
• Scenario B (Technology Plateau): The field reaches a Gaussian distribution where a few standardized models emerge, allowing every company to train their own high-quality models locally (e.g., on consumer hardware like a MacBook M6) in hours.
• The platform is designed to remain viable regardless of which scenario prevails by enabling seamless switching between API-based frontier models and local/on-device inference.
• Security and Latency: Enterprises increasingly require the ability to route sensitive tasks to smaller, local models while reserving high-cost, low-latency API calls for non-sensitive, high-reasoning tasks.
• Current reasoning capabilities have plateaued since the end of GPT-4 training roughly two years ago, despite advancements in context length and media support.
• Technical Bottlenecks: The delay in reasoning breakthroughs is attributed to the immense complexity and instability of scaling GPU clusters (cluster failures and synchronous training challenges) rather than inherent algorithmic limits.
• Mathematics and formal verification remain a critical vector for unlocking advanced reasoning, providing a perfect environment for testing model capabilities without human verification bias.

Business Strategy and Product Philosophy

• No GPUs Before PMF: The founders adopted the internal mandate to avoid training proprietary models until Product-Market Fit (PMF) was achieved, prioritizing application-layer innovation over foundation model risks.
• Data Silos as the Key: Access to proprietary, internal data within silos is identified as the primary constraint and opportunity for unlocking full AI value, rather than raw model intelligence.
• Rejection of Fine-Tuning for Context: Dust argues that fine-tuning for company-specific data is economically and technically inefficient; instead, the solution lies in Retrieval-Augmented Generation (RAG) and controlling the data context presented to the model.
• Augmentation vs. Replacement: The core philosophy is "augmenting humans, not replacing them," focusing on providing "exoskeletons" for employees to increase impact rather than just optimizing for effort reduction.
• Interface Evolution: The current conversational interface is viewed as a transitional "Bash shell" phase; future interactions will evolve into multi-processing agents and co-editing interfaces that abstract data retrieval entirely.
• Fragmentation over Verticalization: Dust deliberately avoids building vertical-specific applications, opting for a horizontal "sandbox" platform that allows users to build specialized agents, mirroring the flexibility and ubiquity of spreadsheets.
• Adoption Curve: Usage typically follows a pattern of identifying a single pilot use case, plateauing, and then skyrocketing to 70% company-wide adoption once the "Lego blocks" of the platform enable unforeseen, organic use cases.

User Demographics and Adoption Trends

• Power User Profile: The most active and innovative users ("makers") are predominantly under 25 years old, who are less set in traditional workflows and more comfortable with the "stochastic mindset" of AI.
• Adoption Metric: Successful pilots are often initiated by surveying teams for personal ChatGPT usage; high personal users correlate strongly with organizational adoption.
• ROI Calculation: The ideal user adopts a risk-reward ratio that accepts occasional model errors in exchange for a 10x upside in time savings and draft generation.
• Predictive Patterns: High engagement correlates with individuals who have previously scripted out repetitive tasks and demonstrate a desire to focus on high-value, non-repetitive work.

Specific Use Cases and Performance

• Cross-Functional Translation: High-value use cases involve assistants translating technical code changes (e.g., Pull Requests) into plain English for non-engineers, or extracting technical insights from sales calls for engineers.
• Personal Coaching: Younger employees use assistants to analyze past interactions (Slack, Notion) for feedback on conciseness, tone, and goal alignment.
• Global Expansion: One case study showed a company shaving 8,000 hours annually by using assistants to review foreign incorporation documents and policy checkers, enabling expansion without a local team.
• Model Performance Trends: Users are showing increased stickiness with Anthropic's Claude Sonnet and interest in coding-specific models like CodeStrile, though switching behavior is often driven by curiosity rather than definitive superiority.
• Productivity Gains: Measured productivity gains range from 5% to 80%, heavily dependent on the specialization of the assistant to the specific workflow.

Founder Insights and Ecosystem Views

• Second-Founder Advantage: As second-time founders, they prioritize "exploration over exploitation," trusting their team more, and leveraging a pre-established "graph of trust" (inspired by their time at Stripe) to enable rapid decision-making.
• French Ecosystem: The French AI ecosystem is recognized for its deep talent pool and scale-up experience, though founders must maintain "ruthless ambition" and target US markets to succeed.
• Admired Figures: Stan Polu cites Ilya Sutskever (visionary leadership) and the "resistance" figures at OpenAI (Shimon and Yakov) for maintaining balance in the hype cycle; Gabriel Hubert admires Yann LeCun for his temperance and questioning of the status quo.
• Short-Term Pessimism: The founders predict a period of reduced excitement and "tough times" as the hype cycle corrects and the true value of AI diffuses slowly through society.
• Long-Term Optimism: Despite short-term stagnation in reasoning breakthroughs, the long-term trajectory remains one of massive value creation and a fundamental shift from deterministic to stochastic tool usage.