Thinking Machines Labs Lifts the Curtain: Multimodal AI for “Collaborative General Intelligence”

Just weeks after shattering records with a $2 billion seed round — the largest ever by a wide margin — Thinking Machines Labs, the secretive San Francisco startup founded by ex-OpenAI CTO Mira Murati, is finally starting to reveal what it’s building and looks like our assessment that they're building their own ChatGPT was largely on point although their ambitions seem greater than that.

In a post on X (formerly Twitter), Murati laid out the vision:

“We’re building multimodal AI that works with how you naturally interact with the world — through conversation, through sight, through the messy way we collaborate.” (full tweet below)

The first product, expected in the next couple of months (that sounds really fast), will include a significant open-source component designed to support researchers and startups building custom models. Murati also promised to share the company’s best science to help the broader AI community understand frontier systems. It's unclear how Multimodal their first product will actually be.

What is “multimodal AI”?

Simply put, multimodal AI systems can process and combine multiple types of input — text, images, maybe video or even audio — to understand or generate responses. For example, OpenAI’s GPT-4V (vision) can look at an image and answer questions about it, while Google DeepMind’s Gemini is designed to handle both language and image reasoning.

Feature / Focus	Thinking Machines Labs (planned)	ChatGPT-4V (OpenAI)	Google Gemini
Core approach to multimodality	Building native multimodal systems that integrate conversation, sight, and collaborative tasks in a unified architecture	Primarily a text model with image understanding added on top (vision is an “add-on”)	Native multimodal model, but largely still outputs text — tightly controlled pipelines
Openness & customization	Planning significant open sourcecomponents, with tools to help researchers and startups build custom models	Fully closed source; no user-level customization of underlying models	Closed model; limited external customization, mostly via API tuning
Collaboration / agentic interactivity	Aims to advance “collaborative general intelligence” — systems that work alongside humans in messy, natural workflows	Designed as a conversational assistant with limited autonomous collaborative planning	More integrated multi-turn reasoning, but still focused on user-as-prompt-master, not equal collaborator

Thinking Machines Labs is less than a year old, yet it’s already making waves, largely thanks to its pedigree: the startup was founded by ex-OpenAI CTO Mira Murati and is stacked with former OpenAI engineers and researchers. Their own team chart (see below) reads like an OpenAI alumni roster, underscoring just how much this ambitious new player draws from the original builders of ChatGPT.

TheMarketAI.com Take:It’s unclear whether Thinking Machines’ “first product” will truly be a fully multimodal system—building that typically takes longer than a few months to come out with such a breakthrough. More likely, we’ll see them unveil their own foundational LLM, laying the groundwork for the richer multimodal collaboration they promise.

Meanwhile, on prediction markets like Polymarket, Google still leads the pack for “Best AI model by end of 2025,” with OpenAI trailing at 26% and xAI at 15%. We’ll be watching closely to see if Thinking Machines earns a spot on the scoreboard in the coming months.

Thinking Machines founding team (source: https://thinkingmachines.ai):

Name	Bio
Alex Gartrell	Former Leader of Server Operating Systems at Meta, expert in Linux kernel, networking, and containerization.
Alexander Kirillov	Co-creator of Advanced Voice Mode at OpenAI and Segment Anything Model (SAM) at Meta AI, previously multimodal post-training lead at OpenAI.
Andrew Gu	Previously working on PyTorch and Llama pretraining efficiency.
Andrew Tulloch (Chief Architect)	ML systems research and engineering, previously at OpenAI and Meta.
Barret Zoph (CTO)	Formerly VP of Research (post-training) at OpenAI. Co-creator of ChatGPT.
Brydon Eastman	Formerly post-training research at OpenAI, specializing in human and synthetic data, model alignment and RL.
Chih-Kuan Yeh	Previously Building data for Google Gemini and Mistral AI.
Christian Gibson	Formerly infrastructure engineer at OpenAI, focused on supercomputers used in training frontier models.
Devendra Chaplot	Founding team member & Head of Multimodal Research at Mistral AI, co-creator of Mixtral and Pixtral. Expert in VLMs, RL, & Robotics.
Horace He	Interested in making both researchers and GPUs happy, formerly working on PyTorch Compilers at Meta, co-creator of FlexAttention/gpt-fast/torch.compile
Ian O'Connell	Infrastructure engineering, previously OpenAI, Netflix, Stripe.
Jacob Menick	ML researcher, led GPT-4o-mini at OpenAI, previously contributed to the creation of ChatGPT and deep generative models at DeepMind.
Joel Parish	Security generalist, helped ship and scale the first versions of ChatGPT at OpenAI.
John Schulman (Chief Scientist)	Pioneer of deep reinforcement learning and creator of PPO, cofounder of OpenAI, co-led ChatGPT and OpenAI post-training team.
Jonathan Lachman	Operations executive, former head of special projects at OpenAI and White House national security budget director.
Joshua Gross	Built product and research infrastructure at OpenAI, shaping ChatGPT's learning systems and GPU fleet; previously on product infra at Meta.
Kevin Button	Security engineer focused on infrastructure and data security, formerly at OpenAI.
Kurt Shuster	Reasoning at Google DeepMind, full-stack pre-training and inference at Character.AI, and fundamental dialogue research at Meta AI.
Kyle Luther	ML researcher, previously at OpenAI.
Lia Guy	Previously at OpenAI and DeepMind, working on model architecture research.
Lilian Weng	Formerly VP of Research (safety) at OpenAI. Author of Lil'Log.
Luke Carlson	Former ML Engineer in Apple's Machine Learning Research group, designed ML frameworks for model orchestration, speech generation, private federated learning, and image diffusion.
Luke Metz	Research scientist and engineer, previously at OpenAI and Google Brain. Co-creator of ChatGPT.
Mario Saltarelli	Former IT and Security leader at OpenAI.
Mark Jen	Generalist, most recently infra @ Meta.
Mianna Chen	Previously at OpenAI and Google DeepMind. Led advanced voice mode, 4o, 4o-mini, o1-preview, and o1-mini launches.
Mira Murati (CEO)	Former CTO of OpenAI, led OpenAI's research, product and safety.
Myle Ott	AI researcher, founding team at Character.AI, early LLM lead at Meta, creator of FSDP and fairseq.
Naman Goyal	Previously distributed training and scaling at FAIR and GenAI @Meta, most recently LLAMA pretraining.
Nikki Sommer	Formerly VP HRBP at OpenAI and Director, HRBP at Twitter.
Noah Shpak	ML Engineer, loves making data go vroom while GPUs go Brrr.
Pia Santos	Executive Operations Leader, previously at OpenAI.
Randall Lin	Previously babysitting ChatGPT at OpenAI and co-tech leading 'the Twitter algorithm' at X.
Rowan Zellers	Formerly at OpenAI, working on realtime multimodal posttraining.
Ruiqi Zhong	Passionate about human+AI collaboration, previously PhD at UC Berkeley, working on scalable oversight and explainability.
Sam Schoenholz	Led the reliable scaling team and GPT-4o optimization at OpenAI. Previously worked at the intersection between Statistical Physics & ML at Google Brain.
Sam Shleifer	Research engineer specializing in inference, previously at Character.AI, Google DeepMind, FAIR, HuggingFace.
Saurabh Garg	Researcher, formerly working on all things multimodal at Mistral AI. Deep into the magic of pretraining data and loving every byte of it!
Shaojie Bai	Avid ML researcher to make audio-visual models better and faster, previously at Meta.
Stephen Roller	Previously full-stack pre-training at DeepMind, CharacterAI, and MetaAI.
Yifu Wang	Passionate about novel ways of overlapping/fusing GPU compute and communication. Formerly PyTorch @ Meta.
Yinghai Lu	ML system engineer, formerly led various inference efforts at OpenAI and Meta.