For Thinking Machines, from 🤫 hussh
The work we admire.
Who Thinking Machines is, what it has shipped, and the science it shares, told from its own writing.
Who they are
AI that works for each person's needs.
“We're building a future where everyone has access to the knowledge and tools to make AI work for their unique needs and goals.”
- Founded
- February 2025, San FranciscoWikipedia
- Founder & CEO
- Mira Murati, former OpenAI CTOWikipedia
- Chief Scientist
- John Schulman, OpenAI co-founderConnectionism
- Funding
- About $2B at about a $12B valuation, led by Andreessen Horowitz in mid-2025, with Nvidia, AMD, Cisco and Jane Street among investors, as reportedTechCrunch
- First product
- Tinker, a fine-tuning API, launched Oct 1, 2025Announcing Tinker
- Research blog
- Connectionism: "shared science from the team"Connectionism
The product
Tinker: teach a model your own expertise.
A managed API for fine-tuning open-weight language models. It runs distributed training on their clusters and keeps the low-level knobs in the user's hands, using LoRA to share compute efficiently.
Researchers and developers who want full control over algorithms and data without managing GPU infrastructure.
Launched Oct 1, 2025 in private beta, free to start, with usage-based pricing to follow.
forward_backward · optim_step · sample · save_state
The science
Connectionism, shared in the open.
Distributed, customizable AI shaped by human knowledge, will and judgment, rather than one centralized model.
Real-time multimodal collaboration: audio, video and text in time-aligned micro-turns instead of turn by turn.
Train a student on its own outputs, graded by a teacher, matching RL results at a fraction of the compute.
LoRA can match full fine-tuning when applied across all layers within capacity. The science behind Tinker.
Constraining weight matrices to manifolds, with principled per-layer learning rates.
Batch-size variation is the true cause of nondeterministic LLM outputs. Batch-invariant kernels give bitwise-identical results.
The through-line
What their work keeps returning to.
Reproducible inference
Batch-invariant kernels so an LLM returns bitwise-identical results. Reliability as a first-class research goal.
Fine-tuning anyone can do
LoRA done right, so a small team can teach a model its own expertise without full-scale infrastructure.
Cheaper, stronger post-training
On-policy distillation that matches reinforcement learning at a fraction of the compute.
Real-time collaboration
Interaction models that listen while they talk: audio, video and text in time-aligned micro-turns.
AI shaped by people
AI that is distributed and adapted by the people who use it, guided by open science.
Open source
They publish the code, not just the paper.
batch_invariant_ops (MIT)
Drop-in batch-invariant kernels for bitwise-identical inference.
tinker-cookbook (Apache-2.0)
Open post-training recipes for the Tinker API.
manifolds (MIT)
Supporting code for the "Modular Manifolds" research.
Open science, owned compute