This $12 billion startup finally shipped something...

Overview

This episode dives into the story of Thinking Machines, a $12 billion startup founded by Mira Murati after her departure from OpenAI. The focus is on their newly released AI model, Inkling, which is deliberately designed to be mid-tier in performance but excels in efficiency, customization, and practical applications. The episode explores its unique features, the philosophy behind its development, and its potential impact on the AI landscape.

Notable Quotes

- Inkling was rewarded for admitting what it doesn’t know instead of guessing with confidence. – On the model's epistemic training.

- Why waste time saying lot word when few word do trick? – On Inkling's token-saving inner monologue during training.

- The real play is to hand you a mid model for free, then charge you to fine-tune it on Tinker into a specialist that destroys your one specific problem. – On Thinking Machines' business strategy.

🚀 The Birth of Thinking Machines

- Mira Murati left OpenAI with a team of top researchers, including co-founder John Schulman and VP of Research Barrett Zoff, forming a dream team in Silicon Valley.

- The company initially launched Tinker, an API for fine-tuning open-weight models, but it received lukewarm reception from the ML community.

🧠 Inkling: A Mid-Tier AI Model with a Purpose

- Inkling is a 970-billion-parameter mixture-of-experts model, but only 41 billion parameters are activated per token, making it highly efficient.

- It processes raw audio and pixels directly, bypassing traditional encoders for images and audio.

- Despite being mid-table in benchmarks, its thinking effort dial allows users to balance speed and computational cost.

🔍 Unique Features and Innovations

- Epistemics Training: Inkling is designed to admit uncertainty, making it highly reliable for forecasting and decision-making.

- Self-Modification Demo: The model demonstrated the ability to rewrite its own training script and adapt its weights, showcasing advanced autonomy.

- Token Efficiency: During training, the model developed a minimalist caveman speak inner monologue to save tokens.

💡 Business Strategy and Practical Applications

- Thinking Machines offers Inkling as an open-weight model under an Apache license, encouraging customization through their Tinker platform.

- The goal is to provide a free, general-purpose model that can be fine-tuned into specialized solutions for specific problems.

- This approach targets industries needing cost-effective, scalable AI solutions without competing directly with top-tier models like GPT-5.

📊 Competitive Landscape

- Inkling launched just as Moonshot released Kimmy K3, a 2.8-trillion-parameter model that outperformed it in raw benchmarks.

- However, Inkling's efficiency and customization options position it as a practical alternative for businesses prioritizing cost and adaptability over raw power.