Key Takeaways
- Thinking Machines has released its first open AI model, Inkling.
- Inkling features 975 billion parameters but uses only a fraction for specific tasks.
- The model is designed for customization by users through the Tinker platform.
- Thinking Machines emphasizes adaptability over a one-size-fits-all approach.
Introduction of Inkling
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has launched its first in-house AI model named Inkling. This model stands apart from those offered by major players like OpenAI, Anthropic, and Google due to its open-weight nature, allowing developers and companies to download and modify it directly.
Technical Specifications
Inkling operates as a mixture-of-experts system with a total of 975 billion parameters. However, it utilizes only about 41 billion parameters for individual tasks, a design choice that enhances efficiency and reduces operational costs. The model was trained on an extensive dataset of 45 trillion tokens, encompassing text, images, audio, and video, enabling it to reason across these formats. Currently, its outputs are limited to text, which includes code and structured data.
Market Positioning
This release marks a significant milestone for Thinking Machines after a year and a half of development largely conducted behind closed doors. The company aims to challenge the prevailing notion that centralized AI models are superior, arguing that AI tailored by organizations will yield better results. Inkling is positioned as a starting point for enterprises to refine and adapt through Tinker, the company’s customization platform.
Performance and Customization
Inkling is designed to provide calibrated responses, indicating uncertainty when necessary. Users can adjust the model’s “thinking effort” to balance speed and accuracy. According to the company, Inkling requires significantly fewer tokens than Nvidia’s latest model to achieve comparable coding performance.
Strategic Insights
Thinking Machines does not claim that Inkling is the strongest model available. Instead, it focuses on delivering well-rounded performance. The company’s strategy emphasizes that organizations can benefit from customizing AI to meet their specific needs. This approach contrasts sharply with the general-purpose models developed by competitors.
Industry Perspectives
Concerns about proprietary AI models have been voiced by industry leaders. Microsoft CEO Satya Nadella noted that enterprises effectively pay twice when using closed models, once for subscriptions and again by sharing valuable business insights. Similarly, Hugging Face CEO Clem Delangue suggested that the future of AI will see a shift towards open-source alternatives for most production tasks.
Case Study with Bridgewater Associates
A notable example of Thinking Machines’ approach is its collaboration with Bridgewater Associates, the largest hedge fund globally. Researchers from both organizations enhanced an existing open-source model with Bridgewater’s financial expertise, achieving impressive results in financial reasoning tests at a fraction of the cost of proprietary models.
Development Timeline and Future Plans
Thinking Machines claims to have developed Inkling in about nine months, a significantly shorter timeframe compared to competitors. While there are questions about the model’s training methods and cost management, the company has focused on efficiency rather than competing directly with larger firms in terms of spending.
Company Culture and Staffing
Thinking Machines has stabilized its workforce, employing around 200 people despite earlier departures. The company promotes a culture of continuity, minimizing reliance on individual personalities, which helps maintain stability as team members transition between roles.