Stop crying about Mira Murati copying China.
The breathless tech headlines are practically dripping with panic. They want you to believe that Thinking Machines—the highly anticipated startup from OpenAI’s former chief technology officer—is suffering an early identity crisis because its debut model draws heavily from Chinese architectures like DeepSeek. Building on this idea, you can also read: Why Europe Cannot Easily Escape American and Chinese Tech.
They are calling it a loss of American dominance. They are calling it a shortcut.
They are completely wrong. Experts at Gizmodo have shared their thoughts on this situation.
This is not a failure of Western innovation. It is the long-overdue collapse of Silicon Valley’s favorite marketing lie: the illusion of the proprietary algorithmic moat.
For five years, venture capitalists have dumped billions into AI startups under the delusion that these companies were building unique, unreplicable digital brains. The reality is far grittier. The elite AI labs have been quietly copying each other's homework since day one. Murati is simply the first founder with the guts to do it in broad daylight, abandoning the theater of "pure organic research" to build something that actually makes economic sense.
The Myth of the Proprietary Moat
I have watched founders set fire to $100 million of venture funding trying to train large language models completely from scratch. Why? Because their pitch decks promised investors a "fully proprietary, ground-up foundational model."
It is a vanity metric. It is also a commercial death sentence.
In the real world, AI development has become an engineering discipline, not a mystical science. The fundamental math behind transformers is public. The optimization techniques are shared in open-access papers. The idea that any single company holds a secret, magical formula for intelligence is a fantasy sold to LPs to justify outrageous valuations.
When Thinking Machines adopts architectural choices pioneered by Chinese labs, they are not admitting defeat. They are practicing basic engineering hygiene.
The Illusion of "First-Principles" Innovation
| The Venture Capital Lie | The Industrial Reality |
|---|---|
| Every elite startup must discover its own proprietary architecture to win. | Architectural designs are commodities; execution and data pipelines are the real differentiators. |
| Copying a competitor's structural design is a sign of technical weakness. | Reusing proven, hyper-efficient open architectures saves hundreds of millions in compute costs. |
| The US and China operate in completely isolated technological silos. | AI research is a global, open-source feedback loop where everyone borrows from everyone. |
If you are building a car company today, you do not spend three years reinventing the internal combustion engine just to prove you are smart. You take the best existing engine design, optimize it, and focus your capital on the suspension, the software, and the manufacturing scale.
Yet, the tech media expects AI startups to reinvent the wheel with every single funding round.
What Thinking Machines is Actually Borrowing (And Why It Is Smart)
To understand why Murati’s strategy is correct, we have to look at the specific mechanics of modern Chinese models like DeepSeek-V3 and its reasoning counterpart, R1.
These models did not achieve near-parity with Western frontier models by utilizing brute-force compute. They did it through radical architectural efficiency.
Multi-head Latent Attention (MLA)
Standard attention mechanisms in transformers are compute-heavy and memory-throttled. When a model generates text, it has to store a massive amount of data in its memory (the Key-Value cache). This makes running large models incredibly expensive.
Chinese researchers optimized this by introducing Multi-head Latent Attention. MLA compresses the Key-Value cache into a smaller latent vector during generation, drastically reducing the memory footprint without sacrificing performance.
If Thinking Machines is integrating MLA-style structures:
- They are cutting their inference hardware requirements by up to 50%.
- They are enabling longer context windows without exponential cost increases.
- They are choosing practical unit economics over academic pride.
Mixture-of-Experts (MoE) and Fine-Grained Routing
Instead of activating all 100+ billion parameters for every single token, modern MoE architectures route specific queries to specialized "expert" sub-networks.
Historically, Western models used coarse routing—sending a query to one or two massive experts. Chinese architectures refined this by using highly granular, smaller experts and a dedicated "shared expert" that captures general knowledge. This prevents specialized experts from redundant learning.
[Input Token]
│
▼
[Router] ───► [Shared Expert] (Always Active for Base Logic)
│
├─────► [Active Expert A] (Specialized in Math)
└─────► [Active Expert B] (Specialized in Code)
By borrowing these exact routing mechanisms, Murati’s team can bypass the ruinous research-and-development cycle of trial-and-error training run failures. They are inheriting a highly optimized structural framework and focusing their compute budget on what actually matters: custom post-training and high-quality data curation.
Dismantling the Geopolitical AI Panic
"Is Western AI losing its lead to China?"
This is the question burning through Washington and Wall Street right now. It is entirely the wrong question.
The premise of the question assumes that AI progress is a nationalistic sprint where the winner takes all and the loser gets nothing. It ignores how open science actually functions.
The gap between the absolute frontier of closed-source Western models and the best open-weight architectures (many of which now originate from Chinese research teams or global collaborative communities) is virtually zero.
Why the Nationalistic AI Narrative is Broken
- The compute bottleneck is physical, but the software is liquid. While export controls restrict physical hardware from entering certain jurisdictions, they cannot stop the global flow of algorithmic insights. A paper published on arXiv is readable in San Francisco, Beijing, and London simultaneously.
- Post-training is the real battleground. The base pre-trained model is just raw clay. The actual utility of the AI—its safety, its reasoning capabilities, its system integration—is determined during post-training (Reinforcement Learning from Human Feedback, Direct Preference Optimization, and targeted fine-tuning).
- Execution beats origin. It does not matter who wrote the initial code for an attention mechanism. What matters is who can deploy it to enterprise customers with the lowest latency, the highest reliability, and the most intuitive integration.
Thinking Machines is not a puppet of Chinese technology. It is an American business utilizing the global state-of-the-art to build an enterprise-ready product. Expecting them to ignore Chinese research is like expecting Boeing to ignore aerodynamics research conducted in Europe.
The Hidden Danger of the Copycat Playbook
While Murati's strategy is economically rational, it is not without severe risks. Copying structural designs from competitor models introduces architectural debt that is incredibly difficult to pay off.
When you adopt an existing architecture, you inherit its inherent structural limitations:
- Blind Spots in Alignment: If you build on top of or closely mirror the architecture of models trained on foreign datasets, you can inherit subtle biases in data representation and routing priorities that are hard to debug.
- Loss of Deep Intuition: When you do not build the engine yourself from scratch, your engineering team lacks the deep, intuitive understanding of why certain structural failures occur under extreme scale. When a customized model begins to hallucinate or collapse at 10-trillion-token scales, fixing it requires retroactively reverse-engineering your own borrowed codebase.
- The Fast-Follower Trap: If your entire strategy relies on adopting the architectural breakthroughs of others, you are permanently relegated to playing catch-up. You will never be the one who discovers the next transformer-level leap.
This is the trade-off Thinking Machines has made. They have traded the remote possibility of a foundational breakthrough for the immediate certainty of a highly efficient, cost-effective product. In today's tight venture climate, that is a trade most rational founders would make in a heartbeat.
Stop Romanticizing Originality
The tech industry has a pathological obsession with the myth of the lone genius inventing the future in a vacuum. We see it in the hagiography of Steve Jobs, Elon Musk, and Sam Altman.
But the history of technology is actually a history of aggressive refinement.
Apple did not invent the smartphone; they refined the capacitive touchscreen and the user interface. Google did not invent search; they refined the ranking algorithm.
Mira Murati’s move is a signal that the first phase of the AI boom—the era of wild, speculative, unoptimized research spending—is officially over. We have entered the industrial phase. In this new phase, efficiency is king, compute budgets are finite, and pride is a luxury no startup can afford.
If borrowing architectures from Chinese competitors makes Thinking Machines’ models cheaper, faster, and more accessible to the market, then Murati isn’t failing.
She is winning.
