Exclusive: Mira Murati’s Stealth AI Lab Launches Its First Product

Summary

Thinking Machines Lab, a new startup cofounded by Mira Murati and several former OpenAI researchers, has launched its first product: Tinker. The tool automates the creation and fine‑tuning of frontier AI models, lowering the technical barrier for researchers, businesses and hobbyists to adapt large language models for specialised tasks.

Tinker currently supports fine‑tuning Meta’s Llama and Alibaba’s Qwen using supervised learning and reinforcement learning techniques. Users can call a simple API, run experiments, then download and deploy the resulting customised models wherever they choose. The company is offering API access initially without charge and plans to monetise later.

Key Points

Tinker automates distributed training details while exposing control over data and training algorithms.
The product supports both supervised fine‑tuning and reinforcement learning from human feedback (RLHF).
Supported base models at launch include Meta’s Llama and Alibaba’s Qwen; fine‑tuned models can be downloaded and run off‑platform.
Thinking Machines Lab is backed by a substantial $2bn seed round and is staffed by high‑profile ex‑OpenAI researchers (including John Schulman and others).
The API is free for now, with plans to charge in future; early beta users report Tinker is more powerful and user‑friendly than many competitors.
The company vets access to mitigate misuse and intends to add automated safeguards, but concerns about malicious fine‑tuning persist.
Thinking Machines has published research on efficient fine‑tuning and network performance that underpins Tinker’s tooling.
Murati frames Tinker as a push to keep frontier AI more open at a time when many leading models are closed behind APIs.

Context and Relevance

This launch comes as the AI field polarises between closed commercial models and more open, modifiable open‑source alternatives. By simplifying fine‑tuning for frontier models, Tinker could accelerate specialised research and productisation of advanced capabilities outside a handful of closed platforms.

The team’s pedigree and the company’s deep pockets amplify the potential impact: a tool that makes reinforcement learning and other tuning methods widely accessible changes who can iterate on and discover new model behaviours — for better or worse. Policymakers, safety researchers and product teams should watch how access, controls and commercialisation evolve.

Why should I read this

Short version: it’s Mira Murati plus ex‑OpenAI brains, $2bn in seed cash, and a tool that could let lots more people tweak the most powerful models. If you care about who gets to build, test or misuse frontier AI, this matters — and you’ll want the specifics on Tinker’s capabilities and safeguards.

Author style

Punchy: this isn’t fluff. The story signals a potential shift in the AI ecosystem — from tightly controlled, closed models to more widely tunable frontier models. If you work in AI research, safety, product or policy, reading the details could save you from being blindsided by faster‑moving bespoke models.

Source

Source: https://www.wired.com/story/thinking-machines-lab-first-product-fine-tune/