Just now, OpenAI engineers admitted that they can no longer understand the code written by AI.
SemiAnalysis reported that, while reviewing the low-level code of their in-house developed chip, OpenAI engineers reluctantly admitted: humans can no longer understand AI-generated code.
Even more astonishingly, at almost the same time, the Architect Labs team from Silicon Valley released a paper that could devastate the entire semiconductor industry:
Two human engineers wrote only a high-level specification in natural language, and within two weeks, the AI generated a real silicon chip capable of running large models—entirely without human intervention.
NVIDIA's CUDA moat has been directly breached.
Title: Redwood: An AI-Designed, Verified, and Deployed-from-Scratch Frontier AI Accelerator Built in 2 Weeks
Preprint: https://arxiv.org/abs/2608.26418
Revise the architecture specification once, and AI will complete code refactoring, revalidation, and deployment back to hardware in just 48 hours.
Moreover, this chip has successfully run open-source large models on an FPGA, with real-world performance demonstrating a 3.4-times higher energy efficiency compared to NVIDIA Jetson!
What shocked the entire internet most: the AI running on this first-generation chip has already begun designing its next-generation chip.
OpenAI engineers admit: "We no longer understand the AI code."
During the interview, a SemiAnalysis expert pointed to a section of code and asked: “How exactly do these lines of assembly logic activate the hardware units?”
The several senior OpenAI engineers present looked at each other, then shrugged and readily admitted:
To be honest, we have no idea what each line is doing—but it doesn’t matter. The AI understands it, tested it, and it runs incredibly fast with outstanding performance.
According to a major disclosure by SemiAnalysis expert Jordan Nanos, OpenAI has fully handed over the assembly-level code for developing the most critical underlying acceleration operators for its custom chips to AI.
On top of Triton, the OpenAI team built a low-level kernel programming language called Gluon.
The complex hardware instructions in Gluon are all automatically generated by AI in a short time.
The code doesn't need to be understood by humans; as long as the AI understands and verifies it, it is correct.
This is the most dramatic paradigm shift since the birth of software engineering over half a century ago.
Why hand over the underlying code to AI?
Only by doing so can we break through Huang Renxun's fortress of computing power.
According to SemiAnalysis, OpenAI’s secretly developed in-house AI chip outperforms NVIDIA in both generation speed and operational cost.
More critically, it is directly undermining the "CUDA moat" that has enabled NVIDIA to dominate the industry for two decades.
NVIDIA is indispensable because millions of developers and hundreds of thousands of companies worldwide are locked into the CUDA ecosystem. But OpenAI has demonstrated through action that:
When the underlying software-hardware co-optimization is fully handled by AI, humans no longer need to adapt to complex CUDA libraries. AI can directly generate optimal low-level assembly code tailored to the hardware architecture.
The chip cycle has been broken! Two people plus AI built real hardware in two weeks.
Meanwhile, the groundbreaking paper recently published by Architect Labs has completely transformed the semiconductor hardware design industry.
Paper URL: https://arxiv.org/abs/2608.26418
In the traditional semiconductor industry, what does it mean to design a specialized acceleration chip?
Team size: Top-tier chip architects, verification engineers, and backend teams numbering in the hundreds;
Development cycle: From project initiation, architecture design, RTL coding to tape-out and verification, typically takes 18 to 24 months;
Capital cost: tens of millions to hundreds of millions of dollars, with extremely low tolerance for error—a single misstep can lead directly to bankruptcy.
What Architect Labs has done has left the entire industry stunned:
Written solely by two human engineers using natural language, the high-level functional specifications generated a complete hardware design in two weeks.
Under these specifications, the entire process requires zero human intervention!
No commercially available accelerator IPs were used; the AI system fully autonomously generated register transfer level code, a hardware verification suite, and accompanying low-level firmware within just two weeks.
Traditional chip design fears requirement changes, as modifying one module may require months of re-running simulations and verification.
In Architect Labs' AI system, after modifying the specifications, the AI regenerates, validates, and deploys back to real hardware—all within 48 hours.
During the peak development phase, the system automatically merged 115 hardware changes per day, achieving 95% module coverage and zero bugs on the first release!
This chip, named Redwood, is no mere theoretical concept.
It is directly deployed on AMD Xilinx Versal FPGA hardware and runs open-source large models natively.
Calculations show that if Redwood were implemented as an ASIC chip using the same process node, its energy efficiency in edge-side Physical AI and low-power scenarios would directly reach 3.4 times that of NVIDIA’s flagship edge computing device, Jetson.
The semiconductor industry’s half-century-old “hardware R&D cycle law” has been completely shattered by AI.
AI is designing the next generation of chips on its own.
If you think this is just another case of AI reducing costs and improving efficiency, you’ve completely underestimated the massive storm behind this.
In the corner of this paper lies a description that sends chills down the spines of all computer scientists:
The domestic open-source large model running on the first-generation Redwood chip has fully engaged in the architectural design and code generation of the next-generation chip.
Please pause and carefully savor this statement.
This is the ultimate concept that the computer science community has theorized for decades: hardware self-recursive evolution.
In human evolutionary history, it took millions of years to progress from stone tools to iron tools, and two hundred years to advance from steam engines to integrated circuits. Because humans are limited by their physical bodies, the brain cannot directly iterate the material carriers of survival.
But the silicon-based world has no such limitation.
When smarter AI designs more efficient chips, and those chips, with three times the computing power and a 48-hour turnaround, run even larger AI models to design the next generation of chips…
Once the flywheel starts spinning on its own, its rate of iteration will no longer be linear but will surge exponentially.
Humans are being marginalized.
Programmers and engineers are becoming the caretakers of "creators"?
Today, the complexity of technology has for the first time surpassed the bandwidth of its creators' minds.
At the software level, traditional "programming" is dying. Future software development will involve constraining AI intentions and evaluating the resulting outputs.
At the hardware level, the billion-dollar barrier to chip design is collapsing. The semiconductor empire once dominated by giants may be defeated by countless "AI + two-person teams."
The underlying hardware and core operators supporting the future global computing network are becoming black boxes beyond human comprehension.
We can only stand outside the black box, watching AI build a kingdom beyond human understanding.
Perhaps a completely new silicon-based civilization, one that humans will ultimately be unable to understand or control, has already arrived?
Reference materials:
https://x.com/firesidealpha/status/2092421779008766021
https://x.com/firesidealpha/status/2093860831595499892
https://architectlabs.com/blog/redwood
https://arxiv.org/pdf/2608.26418
This article is from the WeChat public account "New Intelligence Yuan" (ID: AI_era), authored by Aeneas David.