Kye Gomez, OpenMythos developer
Kye Gomez
When Anthropic introduced its powerful new model, Claude Mythos, this spring, companies and countries freaked out. The general-purpose model, its creators claimed, could discover software vulnerabilities that no one knew existed.
Rather than release Mythos to the world, the company gave it to cybersecurity experts at major companies that build or maintain critical software infrastructure and told them to use it to find and fix bugs before Anthropic unleashed it on the world.
But less than two weeks later, 22-year-old developer Kye Gomez published OpenMythos, a theoretical implementation based on his hypothesis about Mythos’s architecture. Drawing on published research about recurrent-depth transformers, Gomez created code that he believes may approximate Anthropic’s approach—though the company has not confirmed the underlying design.
If true, this story has several startling implications: if a self-taught developer can rapidly implement architectural concepts from academic literature and credibly hypothesize about a major lab’s approach, then proprietary moats may rest more on training infrastructure and data than on secret architectures.
While Gomez has only trained a 770-million parameter version and acknowledges that scaling to frontier-level performance would require tens of millions of dollars in compute, the project demonstrates how architectural innovations can spread rapidly through the research community.
Finally, and most importantly, governments can no longer count on AI secrets staying in the hands of a few. If thousands of individuals and small teams around the world can reverse engineer the work of the biggest AI companies, global safety enforcement will be increasingly complex, if not impossible, as advanced capabilities spread.
Gomez’s background does not follow the standard route into machine learning. He grew up near Miami and began coding as a kid. He left high school early and never went to college. Instead, he built a checkered reputation through code, including troubling allegations of scams – linking himself to other people’s work, and overpromising on projects that reportedly did not function as advertised.
But he also created Swarms, a framework for coordinating large numbers of AI agents, and launched Agora AI Labs, an open-source research collective. Together, these projects form a small ecosystem built outside traditional institutions.
OpenMythos centers on Gomez’s hypothesis that Claude Mythos uses a new model paradigm that has been developing since 2022 and became reliably trainable at scale earlier this year. Rather than stacking neural network layers to give a model depth, researchers proposed looping data repeatedly through a smaller stack – in effect, giving the model depth through time.
Another way to describe it is giving the model time to think, with each loop improving its output.
As Gomez, himself, wrote on GitHub, the code repository favored by developers, “Claude Mythos is suspected to be a Recurrent-Depth Transformer (RDT) — also called a Looped Transformer (LT). Rather than stacking hundreds of unique layers, a subset of layers is recycled and run through multiple times per forward pass. Same weights. More loops. Deeper thinking.”
This architecture, documented in academic papers since 2022, has only recently become reliably trainable at scale. Gomez’s innovation was assembling these published concepts into a working implementation and hypothesizing their use in Mythos.
Gomez’s project does not yet include trained weights or full evaluations and has only trained a relatively small model. He has written instructions to scale to 1 trillion parameters, but that would take tens of millions of dollars to actually do.
Still, the impact has been immediate. Developers are experimenting with the architecture, writing tutorials and debating whether this kind of design could shape the next generation of models.
Whether OpenMythos matches Anthropic’s actual design remains unconfirmed—Anthropic has not disclosed Mythos’s architecture. But the project shows how quickly architectural concepts from academic literature can be implemented and tested once researchers believe they may be relevant to frontier systems.
Even if Gomez’s “theoretical implementation” is not what Anthropic is doing, it gives researchers something concrete to test. That alone can shape how the field evolves.
OpenMythos is another example of how fast cutting‑edge AI techniques propagate and how badly our safety and governance mechanisms lag.
In January 2025, DeepSeek, a relatively small Chinese lab, produced a reasoning model that rivals the performance of the biggest Western systems at a fraction of the cost. It has since been copied into downstream systems around the world.
In November 2025, a single developer created what is now known as OpenClaw, an open-source autonomous AI agent with none of the institutional safety machinery that usually wraps that kind of power.
The result is big‑tech autonomy without big‑tech safety, part of an emerging world of fast‑moving, bottom‑up hackers that clone and release frontier capabilities.
Gomez has said he plans to train larger versions of the model and connect them with his agent framework. Whether that leads to a serious competitor or a series of experiments will depend on resources and execution.
For now, OpenMythos has demonstrated that individual developers can rapidly implement and experiment with architectural concepts that may underlie frontier systems, even if training them to a competitive scale still requires institutional resources.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。