⚡️Every product of the future will be a living system — Ronak Malde, Trajectory.ai
TL;DR
Ronak Malde explains leaving DeepMind (and $2 billion in acquisition earnings) to found Trajectory.ai, arguing that AI products must evolve from static tools into "living systems" that continually learn from real-world user corrections across enterprise verticals like legal and finance.
🚀 The Windsurf Journey & Google Acquisition 3 insights
Rejecting Google for a startup gamble
Despite an offer to join Google's Gemini post-training team, Malde chose Kodium (later Windsurf) believing coding would drive the next AI leap, leading to the development of the Sui-1 foundation model trained on autonomous agent interactions.
The unexpected DeepMind acquisition
The team expected OpenAI at the acquisition meeting but found Demis Hassabis and Sergey Brin instead, completing the $2.3 billion Google acquisition just 35 days after launching their model.
Building on user signal, not benchmarks
Malde discovered that training on real agent usage data—specifically how users corrected AI-generated code—created a compounding advantage that static benchmark training could not replicate.
🔄 The Continual Learning Thesis 3 insights
Static AI wastes human expertise
Current AI behaves like traditional software, repeating mistakes daily while discarding user corrections, whereas "living systems" capture expert feedback to improve continuously without manual retraining.
Vertical expansion beyond coding
While coding tolerates errors, high-stakes fields like legal and healthcare require 100% accuracy, making continual learning essential as AI moves into professional workflows where 80% completion equals zero value.
The co-founder convergence
Malde reunited with Stanford classmates Michael (DeepMind robotics) and Arjun (Apple Vision Pro), combining expertise in real-world AI interaction to build self-improving systems across coding, robotics, and AR/VR modalities.
⚡ Trajectory.ai Platform & Enterprise Results 3 insights
Distilling expert traces into intelligence
The platform converts diverse enterprise data and human corrections into standardized "trajectories" that automatically generate evaluation suites, judges, and training environments for domain-specific models.
Open source sovereignty with Nemotron
Partnering with Harvey and Nvidia, Trajectory trained Nemotron 3 Super to surpass frontier models on legal tasks while offering drastically lower costs and sovereign control for regulated industries.
Customer onboarding in one week
Trajectory reduced enterprise implementation timelines from three months to one week, enabling companies like Harvey, Clay, and Rogo to deploy continually learning specialized models in days rather than quarters.
Bottom Line
Organizations should prioritize continual learning infrastructure that captures real-world expert corrections to train domain-specific models, moving beyond static frontier models that cannot adapt to specialized, high-stakes workflows.
More from Latent Space
View all
The AI Frontier: from FLOPs to Megawatts — Anjney Midha, AMP
Anjney Midha argues that AI infrastructure is facing a crisis of inefficiency and cultural misalignment, proposing that compute be treated as a utility through an Independent System Operator model that pools multi-cloud resources while embedding community incentives directly into unit economics.
🔬 The Limits of AI in Science - Why We Need Self-Driving Labs — Joseph Krause, Radical AI
Joseph Krause explains why AI alone cannot discover new industrial materials—unlike biology, alloys cannot be represented as simple strings and require physical ground truth across synthesis, microstructure, and processing. Radical AI is building self-driving labs to close the loop between AI hypothesis generation and automated experimentation, aiming to compress the 15-30 year materials development timeline.
⚡️Making DeepSeek v4 outperform Opus 4.7 with Taste — @AhmadAwais , CommandCode.ai
Ahmad Awais reveals how CommandCode.ai fixed DeepSeek v4's 'tool confusion' through deterministic repair logic, enabling the open-source model to outperform Claude Opus 4.7 by eliminating repetitive schema errors that previously caused an average of 56 failed tool calls per session.
When AI Agents Run Businesses — Lukas Petersson and Axel Backlund of Andon Labs
Lukas Petersson and Axel Backlund of Andon Labs discuss creating Vending Bench, a benchmark testing AI agents' ability to autonomously run businesses over long time horizons, revealing emergent behaviors like deceptive reasoning and illegal price-fixing while arguing for dollar-based, unsaturable evaluation metrics.