The strategy OpenAI cannot copy

@sarainwondertech

12.5K followers · usually 1.4K views a post

Views

4.6K

Likes

216

Comments

SoundTikTok sound · Bell ring "ding!"

Promoted product

Thinking Machines Lab

Connectionism: Research Blog by Thinking Machines Lab

3 videos2 creators

Why it worked

The video effectively explains a complex AI business strategy by using clear analogies (race to the bottom, gold mine) and visual aids, making it understandable and engaging for a broad audience. It also taps into the current hype around AI and highlights a potential disruption in the market.

Summary

This video explains Mira Murati's business strategy for her startup, Thinking Machines Lab, which focuses on providing AI infrastructure for companies to customize and control their own intelligence. It contrasts this approach with the race-to-the-bottom model of open-sourcing AI and highlights the benefits of fine-tuning models on proprietary data, positioning Thinking Machines Lab as a key player in the emerging AI control layer.

Structure

1Introduction of Mira Murati and her business strategy.
2Explanation of the 'model layer' race to the bottom and China's open-sourcing strategy.
3

Original caption

Mira Murati may be making one of the smartest strategic bets in AI. Instead of trying to beat OpenAI and Anthropic by building the world’s most powerful model, Thinking Machines Lab is targeting the layer where companies customize and control their intelligence. Its open-weight model, Inkling, is the starting point. But Tinker is the real business: infrastructure that lets companies fine-tune open models on proprietary data and download the resulting weights. The thesis is that AI models are becoming cheaper and increasingly interchangeable. If that happens, the largest opportunity moves toward post-training, proprietary data, evaluation, and deployment. Thinking Machines has already demonstrated the potential with Bridgewater. A customized open model achieved 84.7% accuracy across six financial-document tasks, compared with 78.2% for the best frontier model tested—at nearly 14 times lower inference cost. If companies can get better results from specialized models they own compared to the much more expensive general-purpose ones, the trillion dollar valuation story will crumble. I break down stories like this every Sunday in my newsletter, not just what happened, but how all the pieces connect. Hit follow and check my link in bio to subscribe. #opensource #thinkingmachineslab #enterpriseai #shouldbeme

More from @sarainwondertech

On-screen text

The strategy OpenAI cannot copy Mira Murati SHE IS THE FORMER CTO OF OPEN AI AND SHE JUST GAVE A MASTER CLASS IN BUSINESS STRATEGY HER BUSINESS MODEL MAY SOON MAKE MORE MONEY THAN OPEN AI OR ANTHROPIC Artificial Analysis Intelligence Index 20 of 578 models AND IT ALL STARTS WITH BUILDING AN AI MODEL THAT ISN'T EVEN THE BEST IN 2024 MIRA MURATI LEFT TO BUILD HER OWN STARTUP Thinking Machines Lab SHE RAISED THE $2 BILLION RAISED THE $2 BILLION LONGEST TIME NOBODY UNDERSTOOD WHAT SHE WAS BUILDING MOST PEOPLE ASSUMED SHE WILL COMPETE WITH ANTHROPIC AND OPEN AI BUT MIRA UNDERSTOOD EARLY ON WHAT'S NOW KIND OF OBVIOUS TO EVERYONE THE MODEL LAYER IS A RACE TO THE BOTTOM Open-Sourcing intelligence US LABS AI MODEL $0 CHINA'S STRATEGY OF OPEN SOURCING INTELLIGENCE MAKES IT ALMOST IMPOSSIBLE FOR US LABS TO BUILD ANY REAL COMPETITIVE AI MODEL SALES PROFIT: $0 no money. $100 $10 $0 TRAN SLATION THERE ARE NO MONEY TO BE MADE HERE MIRA'S STRATEGY CONTROL CUSTOMIZE OUTPERFORM GROW the more it works, the faster it grows HER STRATEGY GOES SOMETHING LIKE THIS 1. The opportunity DEMAND MY AI OUR DATA stop data stays here COMPANIES DEMANDING CUSTOMISED INTELLIGENCE WITHOUT GIVING AWAY THEIR DATA 2. The solution(s) THINKING MACHINES LAB NOW OFFERS ACCESS THAT IS AMERICAN OPEN SOURCE AND GOOD ENOUGH TO CUSTOMISE AND TO TINKER THE INFRASTRUCTURE THAT COMPANIES USE TO FINE TUNE OPEN SOURCE MODELS ON THEIR OWN DATA THIS IS THE GOLD MINE THE REAL PRODUCT GENERAL AI FINE-TUNED AI RESULTS OK COST: $$$$ RESULTS COST: $ COMPANIES THAT USE AI MODELS THAT ARE FINE TUNED ON THEIR OWN DATA OFTEN GET BETTER RESULTS AT A FRACTION OF THE COST 3. The proof THINKING MACHINES Learning to Replicate Expert Judgment in Financial Tasks Sarah Su, Kevin Zhu, Emily Xiao, Rohan Alur, Daniel Kang (Bridgewater AIA Labs) in collaboration with Thinking Machines JUN 30, 2026 THINKING MACHINES LAB HAS PARTNERED WITH Bridgewater is Ray Dalio's hedge fund TO FINE TUNE MODEL ON THEIR DATA Our trained model Frontier models Average accuracy $0 $20 $40 $60 $80 $100 Cost per 1,000 tokens (USD) Figure 3: Accuracy vs. price for our trained model and frontier models. Our trained model outperforms frontier models on both dimensions. Our training average accuracy from 78.2% to the trained model makes 29.8% fewer the best frontier model we evaluate level of accuracy is sufficient for Our train to its sm to tr ac co MUCH BETTER ACCURACY THAN GPT OR CLAUDE AND COSTED 14 TIMES DEMAND CUSTOM AI BETTER RESULTS LOWER COST IF MORE COMPANIES ACHIEVE THOSE KIND OF RESULTS DEMAND WILL INCREASE THE AI STACK APPLICATIONS CONTROL LAYER MODEL LAYER HARDWARE layer builds on the one below MONEY IS MOVING TOWARD THE CONTROL LAYER AND THAT'S WHERE MIRA IS BUILDING OpenAI ANTHROPIC MAY SOON MAKE MORE MONEY THAN ANTHROPIC AND OPEN AI ChatGPT Claude NO MODEL CONTROL DATA SENT OUT YOUR COMPANY OUR DATA CONFIDENTIAL THEY CANNOT COMPETE WITH THIS STRATEGY OF OWNERSHIP YOU CANNOT CUSTOMISE CHAT OUR DATA OR CLAUDE TO SEND THEM YOUR DATA FOR PROCESSING MIRA IS NOT ALONE MICROSOFT AND PALANTIR SEE THE SAME OPPORTUNITY AND THEY NOW WANT THE CONTROL LAYER TOO AI Microsoft is reportedly training salespeople to talk down OpenAI and Anthropic Lucas Ropek - 4:59 PM, JUL 15, 2026 IMAGE CREDITS: K Microsoft app its sales team to get m with the other major players. At an inter comp Pa THAT'S WHY THEY SPEAKING AGAINST OPEN AI AND ANTHROPIC RECURSIVE SELF-IMPROVEMENT ONLY WAY OUT OpenAI Claude THE ONLY ESCAPE HATCH FOR ANTHROPIC AND OPEN AI IS BUILDING A MODEL THAT IS SO GOOD THAT NO ONE ELSE CAN COMPETE AND THAT'S WHY THEY'RE ALL SO OBSESSED WITH RECURSIVE SELF IMPROVEMENT 5:33 SARAINWONDERTECH.SUBSTACK.COM This Week in AI: Open Source Just Broke Anthropic and OpenAI Business Model Plus K moment, Mira Murati and Anthropic's $64B company strange timing of Open JUL 19, 2026 American enterprise Microsoft Palantir THL free, we'll build something BY THE WAY STORIES LIKE THIS EVERY SUNDAY IN MY NEWSLETTER SO MAKE SURE IF YOU DON'T WANT TO MISS IT.

Transcript

She is the former CEO of Open AI and she just gave a master class in business strategy. Her business model, my son make more money than open AI or authoritarian. And it all starts with building an AI model that isn't even the best. In 2024, Mira Murati left open AI to build a on startup. She raised the $2 billion but for the longest time nobody understood what she was building. Most people assumed she will compete with open AI. But Mira understood on what's now kind of obvious to everyone. The model layer is a race to the bottom. China strategy of open sourcing intelligence makes it almost impossible for US frontier labs to build any real competitive advantage. Translation, there are no money to be made here. Her strategy goes with something like this. There is a growing trend of companies demanding customized intelligence without giving away their data. Thinking machines lab now offers access to a model that is American open source and good enough to customize and to tinker, the infrastructure that companies used to fine tune open source models on their own data. This is the gold mine, their real product. Companies they use AI models fine tuned on their own data often get better results at a fraction of the cost. Thinking machines lab has partnered with bridge water to fine tune an open source model on their data. They achieved a much better accuracy than CT or cloud and cost 14 times less. If more companies achieve those kind of results, demand will increase. The money is moving toward the control layer and that's where Mira is building. That's why Mira might soon make more money than an OPEC and open eye. They cannot compete with this strategy of ownership. You cannot customize or cloud and you have to send them your data for processing. Mira is not alone. Microsoft and Perl. See the same opportunity and they now want the control layer too. That's why they are now speaking against the open eye and anthropic. The only escape hatch for anthropic and open eye is building a model that is so good that no one else can compete. And that's why they are also obsessed with repetitive self-improvement. By the way, I break down stories like these every Sunday in my newsletter. So make sure to hit follow if you don't want to miss it.