Mira Murati's First AI Model Inkling: Powerful but Underwhelming?

Mira Murati, who once helped run OpenAI (the company behind ChatGPT), has finally released her own AI model after leaving in September 2024. Her new company, Thinking Machines Lab, launched the model — called Inkling — on July 15.

Inkling is what's known as an open-source model (meaning anyone can download and use it for free, rather than it being locked behind a company's private servers). It's now considered the best open-source AI ever built "from scratch" by a Western lab, at a time when Chinese labs have been dominating this space.

Here are the key facts:

• 975 billion parameters (the internal "dials" the AI uses to make decisions) — but only 41 billion are active at once
• Trained on 45 trillion tokens (tokens are small chunks of text, like words or word-pieces)
• Can read text, images, and audio
• Available online at $1 per million input tokens and $4.05 per million output tokens (a token-based pricing system — you pay for how much text goes in and comes out)

Where Inkling shines is "agentic" tasks — meaning acting like a smart assistant that completes real jobs on its own. It scored 74.1% on one test of task completion and 77.6% on fixing software bugs, beating rival Nvidia's model.

But in hands-on testing, the results were mixed. When asked to build a simple zombie shooter game, Inkling produced only basic shapes and a rough result. Surprisingly, a much smaller AI model called Bonsai — small enough to run on an iPhone — created a better game.

This raises a big question: if a phone-sized model can outperform Inkling in some tasks, where are all those 975 billion parameters actually going?

For now, Inkling is a promising but imperfect debut — strong for professional automation tasks, but not always better than cheaper rivals for everyday coding.

This is an AI-generated summary. Read the original article at: https://decrypt.co/373884/review-inkling-mira-murati-first-open-source-ai