Mira Murati just launched an AI lab that sounds like Anthropic's brand department wrote the pitch deck.
The Summary
- Thinking Machines, founded by former OpenAI leader Mira Murati, released Inkling, a 975-billion-parameter open-weights model with 41 billion active parameters, customizable via its Tinker platform.
- The company explicitly positions itself as building "AI that helps organizations cultivate knowledge, not AI that extracts and replaces it" — nearly identical positioning to Anthropic's "enhance human cognition" messaging.
- Thinking Machines concedes Inkling isn't the most powerful model, but competes on customization and letting enterprises "shape AI in their own image."
The Signal
Thinking Machines is betting that the future of AI isn't raw power. It's customization at scale. Inkling won't beat GPT-4 or Claude on benchmarks. It doesn't need to. The play is enterprise-specific intelligence that learns from your org, not the internet's consensus reality. That's the thesis, anyway.
Murati knows how to read a room. She watched Anthropic carve out mindshare by positioning Claude as the thinking person's LLM while OpenAI chased consumer scale. Now she's running the same playbook with a twist: open weights plus fine-tuning infrastructure. The message is clear. You don't want AI that averages out human expertise. You want AI that amplifies your specific domain knowledge.
"Every organization is powered by the expert knowledge of its people. We believe in AI that helps the organization cultivate that unique knowledge, not AI that extracts a snapshot of it."
The technical bet here matters more than the marketing. Open-weights models let enterprises run inference on their own infrastructure. Fine-tuning via Tinker means you can teach Inkling your company's specific workflows, terminology, and edge cases. If you're a pharmaceutical company, you don't need an AI trained on Reddit. You need one trained on your clinical trial data and regulatory frameworks.
Key differentiators:
- 975B parameters with 41B active (mixture-of-experts architecture, likely)
- Open weights = full transparency and on-premise deployment options
- Tinker platform handles fine-tuning without requiring ML expertise in-house
But the positioning is also a tell. When a new lab says "we're not the most powerful," they're pre-empting the criticism that they can't compete on frontier capabilities. Thinking Machines is choosing a different game. The question is whether enterprises actually want bespoke AI or whether they'll just use whatever model their SaaS tools ship with by default.
Anthropic proved there's a market for AI that doesn't feel like a slot machine. Murati is betting there's a bigger market for AI that feels like it actually works for your company, not against it. The open-weights angle is smart. It appeals to enterprises worried about vendor lock-in and data sovereignty. But customization requires expertise, even with a platform like Tinker. Most companies struggle to define what they want AI to do, let alone fine-tune it.
The Implication
The AI market is fragmenting faster than people realize. We're past the "one model to rule them all" phase. Anthropic owns the intellectual high ground. OpenAI owns consumer reach. Google owns search integration. Thinking Machines is carving out the "your AI, your rules" niche. If they can make fine-tuning actually accessible, not just theoretically possible, they've got a shot.
Watch whether enterprises with deep domain expertise (pharma, legal, finance) actually adopt Inkling or stick with whatever Microsoft bundles into Office. The future of AI might not be about who builds the smartest model. It might be about who builds the model that learns fastest from your data.