Subvolts take Specialized AI models can outperform larger ones by focusing training on specific tasks. May 26, 2026 - Developers
- Learn about building specialized, high-performance coding models using efficient training techniques.
- Business
- Understand how a top-down approach can accelerate AI product development and improve inference efficiency.
This episode features a firsthand account of a novel approach to developing high-performance AI models for a specific task.
How Cursor Trained Composer on Fireworks: Distributed Infrastructure for High-Performance RL
Audio streams from Sequoia Capital. Your place is remembered on this device.
From Subvolts
Summary and analysis by Subvolts editorial, AI-assisted.
Sequoia Capital's Training Data podcast features Federico Cassano of Cursor and Dmytro Dzhulgakov of Fireworks discussing their collaboration on Composer. They built Composer, a specialized foundation model for software engineering, using an unconventional top-down approach. This involved mid-training and reinforcement learning on an open-source base, then refining it with real user data. Fireworks provided distributed infrastructure, enabling Composer to achieve frontier-class coding performance with high efficiency.
This episode explains how two companies built a smart AI assistant for writing code. Instead of starting from scratch, they adapted an existing AI and trained it specifically for coding tasks. This makes the AI very good at coding and fast to use, like a specialized tool.
This episode offers insight into a practical, accelerated approach to building specialized AI models. Listeners learn how Cursor and Fireworks combined a top-down strategy with distributed infrastructure to create a high-performance coding model quickly. It's for AI engineers and product leaders interested in efficient model development and deployment. This work parallels frontier labs' efforts to create more capable and efficient models, particularly in coding assistance, a key area for many AI companies.
- HostSequoia Capital
- GuestFederico CassanoCursor
- GuestDmytro DzhulgakovFireworks
- Guests: Federico Cassano (Cursor), Dmytro Dzhulgakov (Fireworks)
- Show: Training Data (Sequoia Capital)
- Length: 46 minutes
- Model Focus: Composer, a specialized foundation model for software engineering.
- Training Insight: Models have finite capacity in weights; allocating bits to a singular task frees the model for better performance and efficiency.
- Training Approach: Unconventional top-down: mid-training and RL on an open-source base, then specializing around real usage.
Cursor's Federico Cassano and Fireworks' Dmytro Dzhulgakov explain how they collaborated to build Composer as a specialized foundation model. The core insight: models have finite capacity in their weights, and allocating all those bits to the singular task of software engineering in Cursor frees the model to be both better at the task and far more efficient at inference. Rather than start from pre-training and work up, they took an unconventional top-down approach — mid-training and RL on top of an open-source base to get a useful model into users' hands fast, then specializing the model around real Cursor usage. With Fireworks providing distributed infrastructure, Composer delivers frontier-class coding performance with the speed of a much smaller model.
Hosted by Sonya Huang, Sequoia Capital
- foundation model
- Dmytro Dzhulgakov
- A large, general-purpose AI model that can be adapted for many specific tasks.
- weights
- Federico Cassano
- The numerical parameters within an AI model that store its learned knowledge.
- inference
- Federico Cassano
- The process of using a trained AI model to make predictions or generate outputs.
- RL
- Dmytro Dzhulgakov
- Reinforcement Learning, a type of machine learning where an AI learns by trial and error, receiving rewards for good actions.
1 h 08David Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience
WithDavid Silver, Lasse Espeholt
1 h 041 h 0552 minMaking Cities Awesome: Peregrine’s Nick Noone & Ben Rudolph
WithNick Noone, Ben Rudolph
1 h 44Opus 4.8 Beats GPT 5.5, the $220B OpenAI Foundation, and Hassabis’s 2029 AGI Prediction | EP #260
WithSalim Ismail, Dave Blundin, Dr. Alexander Wissner-Gross
1 h 12
53 minHow AI Is Reinventing Elder Care | Chia-Lin Simmons of LogicMark
WithChia-Lin Simmons
56 minQuestions people ask
- What is Composer?
- Composer is a specialized foundation model built by Cursor and Fireworks for software engineering tasks.
- What was the training approach for Composer?
- They used an unconventional top-down approach, starting mid-training and applying reinforcement learning on an open-source base, then specializing it with real user data.
Page generated Oct 19, 2026. Summaries are Subvolts' own; the audio, artwork and notes belong to Sequoia Capital.