Key Moments
Agents @ Work: Dust.tt — with Stanislas Polu
Latent Space PodcastKey Moments
Dust.tt's Stanislas Polu discusses AI agents, OpenAI, and building horizontal platforms.
Key Insights
Stanislas Polu transitioned from Stripe and OpenAI to found Dust.tt with a focus on building a horizontal AI agent platform.
His work at OpenAI focused on math and AI, exploring large language models' reasoning capabilities in formal mathematics.
Dust.tt aims to provide infrastructure for companies to deploy AI agents for their teams, emphasizing a product-driven approach over research.
The platform prioritizes enabling non-developers to create agents through natural language instructions, contrasting with vertical agent solutions.
Polu believes in the power of horizontal platforms to unlock emergent use cases and emphasizes API integrations over UI automation for internal company systems.
Dust.tt leverages tools like Temporal for orchestration and focuses on making AI accessible and non-intimidating to end-users.
JOURNEY FROM STRIPE AND OPENAI TO FOUNDING DUST.TT
Stanislas Polu shared his career path, which included roles at Oracle, Stripe, and OpenAI before he co-founded Dust.tt. His experience at Stripe, known for its engineering culture, and his time at OpenAI, particularly working on math and AI research, provided a strong foundation. Polu was part of the early wave of "OpenAI Alumnus Founders," leaving in September 2022, before the release of ChatGPT. This background shaped his vision for Dust.tt, moving from advanced research to building practical AI tools.
EXPLORING AI RESEARCH AT OPENAI
At OpenAI, Polu was dedicated to the "math and AI" research area, focusing on large language models' (LLMs) ability to perform formal mathematical reasoning. The goal was to combine the creativity of LLMs with the verification capabilities of formal systems, a challenging but promising endeavor. He noted that while Transformers are creative, they make mistakes, and formal math systems provide mechanical verification. This work was aligned with OpenAI's mission and received the necessary compute resources for progress.
THE BIRTH OF DUST.TT: A HORIZONTAL AGENTS PLATFORM
Polu's motivation for starting Dust.tt stemmed from the realization that immense value could be unlocked by productizing LLMs, particularly GPT-4, which he saw internally before its public release. He observed a gap between the models' capabilities and their deployment. Dust.tt was conceived as a horizontal agents platform, aiming to build infrastructure that enables companies to deploy AI agents within their teams. The core idea is to empower non-developers to create agents that solve operational tasks.
PRODUCT-DRIVEN DEVELOPMENT AND ACCESSIBILITY
Dust.tt's strategy is deliberately product-focused, eschewing its own research and model training to leverage existing frontier models. The platform prioritizes user experience, especially for non-technical individuals. Polu emphasized the importance of making the technology feel accessible and non-intimidating, using user-friendly language like "instructions" instead of "system prompts." This approach contrasts with deep research efforts and focuses on practical deployment and value generation.
INFRASTRUCTURE, CONNECTORS, AND ORCHESTRATION
A significant part of Dust.tt's work involves building robust infrastructure, including connections to various data sources and tools. This "boring infrastructure work" is crucial for enabling agents to access data and take actions. The platform uses tools like Temporal for asynchronous work and orchestration, ensuring that agents can handle complex workflows. Polu also discussed the challenge of integrating with diverse company systems, prioritizing API-based interactions over UI automation where possible.
THE HORIZONTAL VS. VERTICAL AGENTS DEBATE
Polu strongly advocates for a horizontal approach to AI agents, believing it unlocks emergent use cases across different teams and functions within a company. While vertical solutions might have easier go-to-market strategies, they often have limited potential. Dust.tt's horizontal platform aims to provide a general-purpose tool that can be adapted to various needs, allowing for more widespread adoption and value creation within an organization by empowering users to build their own solutions.
FUNCTION CALLING AND MODEL AGNOSTICISM
Dust.tt places a high emphasis on function calling, recognizing its critical role in agent performance. The platform supports multiple LLMs, allowing users to choose the model that best suits their needs, though a default is provided for ease of use. Polu noted that while models are improving, the core value lies in precise instructions and the ability for agents to correctly interpret and execute function calls, enabling more complex workflows and meta-agents.
THE FUTURE OF SAS AND ENTERPRISE UTILIZATION
Polu speculated on the future of Software as a Service (SaaS) in the age of AI, suggesting that highly capable AI agents might reduce the need for some traditional SaaS products, especially within tech companies that can build their own solutions. He also discussed the concept of "maximum enterprise utilization," where AI agents help companies tackle work that was previously impossible due to a lack of human resources, potentially leading to smaller, more efficient teams achieving significant output.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●People Referenced
Common Questions
Dust.tt aims to build the infrastructure for companies to deploy AI agents within their teams. It focuses on providing tools for non-developers to create agents for operational tasks, bridging the gap between complex AI capabilities and everyday business needs.
Topics
Mentioned in this video
Mentioned as the host of the podcast.
A research organization where Stanislas Polu worked before founding Dust.tt. It was a significant part of his career, particularly in AI research.
Mentioned for an early deep learning project involving image recognition.
A SaaS company whose customers might build AI directly on top of it, potentially reducing the need for the SaaS layer itself.
A company where Stanislas Polu worked for nearly 5 years, influencing OpenAI's culture.
An orchestrator used by Dust.tt for managing asynchronous workflows and agent execution.
A company where Stanislas Polu worked.
Mentioned in the context of the OpenAI split and as a competitor.
Achieved a significant result in AI and mathematics, scoring silver on the IMO, which impressed Stanislas Polu.
Mentioned as a platform for code and potential integration for AI agents.
A major enterprise software that requires API integration for AI agents, posing unique challenges.
Used as an example of a platform company with a successful product spin-off (AWS).
A company focused on data integration, discussed in the context of connecting to various SaaS platforms.
A company founded by Stanislas Polu, focused on building infrastructure for AI agents.
A formal proof system mentioned in the context of AI and mathematics research.
A key product from OpenAI that received significant compute resources.
A large action model from Salesforce mentioned on a leader board for function calling.
Mentioned as a previous generation of models in the research path at OpenAI.
A major model from OpenAI, its capabilities were recognized by Stanislas Polu as creating significant value.
An early success from OpenAI that used models to traverse and summarize the web.
A competitor to Dust.tt in the developer tooling space for AI.
An example of a highly autonomous AI agent approach, contrasted with Dust's more controlled method.
An example of a system with good API integration for AI agents.
A model family from Anthropic known for its function calling capabilities and chain-of-thought steps.
Mentioned as a product of OpenAI that influenced the talent flow to the company.
A browser extension developed as an early product from Dust.tt.
A specific model variant discussed for its function calling capabilities.
A model ranked on a function calling leader board.
A popular JavaScript framework used for building web applications, mentioned in the context of Dust.tt's development stack.
A productivity tool for which Dust.tt ensures structured data capture for AI models.
Mentioned as a company providing integrations.
A web framework that ChatGPT reportedly rewrote from Next.js into.
A multimedia framework that could be wrapped with GPT for a video editing agent.
Mentioned as an example of a company already selling AI products to enterprises.
A company building AI agents, discussed in comparison to Dust's approach.
A specific model within the Claude 3 family, noted for its function calling innovation.
The successful product spin-off of Amazon, highlighting a platform strategy.
Mentioned as a company providing integrations.
A company that spun off its internal search technology into a separate product, illustrating a product-over-platform strategy.
Mentioned as having a high position (COO) at OpenAI.
Taught an early AI class at Stanford that Stanislas Polu took.
An OpenAI employee involved in Stanislas Polu's hiring process.
Mentioned in the context of companies providing integrations.
Co-founder of Dust.tt, previously worked at Oracle, Thomsons, Stripe, and OpenAI.
Key figure at OpenAI who recruited Stanislas Polu and was deeply involved in the reasoning team.
Mentioned humorously in relation to a quote about single-person billion-dollar companies.
Worked with Stanislas Polu at Stripe and later at OpenAI. Influential in Polu's move to OpenAI.
Mentioned as a counter-example to a single-person billionaire creator, due to his Spotify deal and minimal team.
More from Latent Space
View all 145 summaries
86 minNVIDIA's AI Engineers: Brev, Dynamo and Agent Inference at Planetary Scale and "Speed of Light"
72 minCursor's Third Era: Cloud Agents — ft. Sam Whitmore, Jonas Nelle, Cursor
77 minWhy Every Agent Needs a Box — Aaron Levie, Box
42 min⚡️ Polsia: Solo Founder Tiny Team from 0 to 1m ARR in 1 month & the future of Self-Running Companies
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Try Summify free