Open-weight models: Will China, Distillation undercut U.S. AI Alignment, Safety?

By David Stephen

The space race of the 20th century between the United States and the Soviet Union had some finite goals: reach space, launch satellites, plant flag on the moon, boost security, surveillance, have a space station, and so forth. There was already knowledge that it was not a race to the nearest star or a manned flight to some distant planet within the solar system. So, even though some wins seemed exponential, it was not beyond what the other team could attempt.

Open-weight: will China undercut U.S. AI  Safety?

The AI race between the United States and China is infinite. There is no boundary to possibilities with intelligence. So, if the goal is to keep improving intelligence, eventually, it will be possible to make extreme strides beyond what the competition can match.

The United States appears way ahead of China in the AI race. But it looks like what seems to be naive corporate businesses with venture capital, research publications, academic adjacent ethics, and so forth, could become a global leadership crisis, if there is no classified national strategy to win the AI race.

Even as OpenAI and Anthropic have formidable models, it does not seem like the United States has a clear, un-tailed path to staying ahead of China, beyond export controls for GPUs, and maybe sanctions.

This is clearly not enough against China that has shown that near-latest GPUs can still be potent and then native talents in math and computer science can devolve into complexities and sweep benchmarks.

There are lots of criticisms that power should not be concentrated within a few companies. This is even the smaller problem. The bigger problem is that the approach [optimization algorithms] of Anthropic to improving AI model capabilities are the same with OpenAI, Meta, DeepMind and the rest.

They are all expanding through existing architectures, squeezing more out of parameters, tokens and compute, forgetting that they have to detour at basic research to take off in a way that would take China another decade or more to close in.

This means that the National Laboratories of the United States, should pursue fundamental research that would even exceed those of the U.S. AI industry, since their goal is to chase enterprise token revenues.

So, Argonne National Laboratory, Oak Ridge National Laboratory, Los Alamos National Laboratory, Lawrence Berkeley National Laboratory, Lawrence Livermore National Laboratory, Sandia National Laboratories, Savannah River National Laboratory and others can look into fundamental science, in a way to have models from the U.S. pursue and achieve artificial general intelligence [AGI], with superintelligence capabilities and ability to make a lot of difference on any objective.

Michael Kratsios, the Director of the Office of Science and Technology Policy; Chris Wright, the Secretary for the US Department of Energy; and Emil Michael, the Under Secretary for Research and Engineering, at the Department of War, should not just be looking at regulation or other concerns, but how to build what would win, from as soon as August 5, 2026.

This indicates that what AI model would match the human brain first in general intelligence, to become as creative and be able to do encompassing problem-solving?

So, conceptual brain science is the focus, what postulates are possible to kick off a new basis for architectures that would surpass large language models [LLMs] and reach multiple types broad AGI? This is not just about saying that video models would be used as training data to achieve AGI. The aim is that even if the data type is changed the form which data is stored is still the same.

This says that digital memory is stored the same way, so even if spatial intelligence or world models are sought, they might not leap as much, since the memory is not stored as human memory.

Humans have generalized memory hence generalized intelligence. AI uses digital memory with bytes representation the same way that had always been. So, memory is an angle to pursue AGI, just like several other paths that can be extricated from the postulate in Conceptual Biomarkers and Theoretical Biological Factors for Psychiatric and Intelligence Nosology.

The United States can win but must avoid sprinting with China where China is strong [math and coding]. Conceptual Brain Science is a promising direction, which must be thoroughly exhausted.

There is a recent [July 28, 2026] report on the AP, Workplaces look for cheaper AI as ‘tokenmaxxing’ fades as a corporate fad, stating that, “A corporate fad of “tokenmaxxing” on artificial intelligence technology is hitting its limits as workplaces throwing AI at everything are seeing the costs rise without a similar spike in productivity.”

“What started as tech industry-fueled springtime hype over squeezing as much AI-generated work as possible out of products like OpenAI’s ChatGPT and Anthropic’s Claude has shifted to a summertime backlash.”

““Tokenmaxxing” refers to maximizing usage of tokens — the building blocks of generative AI that correspond to small pieces of text that an AI system reads or writes. Each token is about three quarters of a word. And there’s typically a limit to how many you can use, with pricier versions of AI products offering higher caps.”

“At the same time, those who favor racking up as many tokens as possible are having a field day with new open-source models from Chinese startups like Moonshot’s Kimi or Zhipu’s GLM, which nearly match the capabilities of top U.S. models at a fraction of the price.”

David Stephen does research in conceptual brain science and served as a visiting scholar in medical entomology at the University of Illinois at Urbana-Champaign. He did computer vision research at Rovira i Virgili University in Tarragona.

See more breaking stories here.