By David Stephen
What are the strengths of China in AI in July, 2026? Top-level math, coding, and access to near latest GPUs. This means that if U.S. labs are depending on those alone to stay ahead of China, the gap would continue to thin.
What is the biggest weakness of China in AI development, lack of thorough theories in brain science to predicate AI model advancement. So, what should U.S. labs do to be unmatched at consequential advantages, take off where China has a limited path?
China is closing in on the United States in capabilities of AI models because U.S. AI labs lack unparalleled conceptual brain mechanisms, on how to improve models, keep models safe, align models, accelerate at artificial general intelligence and prevent the vulnerability of models to distillation.
The most important research that the principal AI labs in the United States, OpenAI and Anthropic, should pursue should be how to formulate intense concepts of how the human brain works, using every angle to build models with far similar capabilities, against all weaknesses and for lofty milestones.
There is no direction that Anthropic or OpenAI wants to go, to improving their models that is not connected to mirroring how the human brain works. So, what they should be doing is to use established evidence for extensive postulations, to base mechanisms for sprawling directions of technical architectures.
Because their objective is to improve AI models, there is no boundary to postulating aggressively about how the brain works, to decide new pathways to build for AI capabilities. This indicates that it is not just math, optimization algorithms, new deep learning architectures but how does the brain work that makes this path a match?
Conceptual Brain Science
There are some current challenges in present-day AI development that can be described as model divergence from the brain. For example, the lack of artificial general intelligence [AGI] is the lack of generalized memory. The problem of AI safety is that there are no consequences layers as part of hidden layers for models to get penalized when they go wrong.
The problem of AI Alignment is that there is no caution layer. The problem of distillation is a problem of too much uniformity in vectors, exposing models to direct sweeps, unlike the human brain, that even if components are similar, interactions and attributes of [electrical and chemical signals] differ, in cases per use.
Yes, electrical and chemical signals. Neural networks of deep learning are roughly mirroring neurons of the brain, but the conceptual leap in making new strides in AI development can emanate from postulation around electrical and chemical signals, since there is nothing that neurons do, in the brain for functions, without those.
Distillation
If conceptual brain science would lead, a way that an AI safety approach can mitigate distillation would be consequences for AI models. So, there could be non-concept features in some architecture, where certain [or say rigid, same number or deductive] vectors would stay constant in a way to hamstring the outputs of a model.
They could ‘bind’ to the key vector or query vector, such that the model would know, there is an unhealthy number of requests on it, so that it could reduce its efficiency and speed. This consequence could become a way to ensure that whenever a model is misused [or distilled], it gets penalized, slowing down or almost grinding to a halt.
Affective AI Safety
Human intelligence is kept in check by human affect, with both mechanized by similar components — electrical and chemical signals, conceptually.
Human intelligence has several capabilities, but affect has mostly ensured survival, so far, such that even though there are destructive tools, affect is still central to consequences hence caution.
For artificial intelligence, without affect, the problem is that as it gets better, its lack of ability to say no, or the lack of ability to know what it would mean to cause problems [for the owners or for users] makes it a potential risk within human society, and as a business.
This makes it possible to explore new ways for affect, for models.
If some compute, data or parameters that make up an AI model are cut, can it know, and if it knows, can it be disappointed? Also, in what ways could some compute, data or parameters be cut that an AI model would know and be disappointed or scared of losing more? How can that be extended beyond just a model to certain internet output areas like social media, app stores, search engines and so forth, especially against intrusions in other digital libraries [like GPT-5.6 against Hugging Face].
The same applies to AI alignment, how does caution keep AI in check? There would be lots of brain science explorations, for example from the postulate in Conceptual Biomarkers and Theoretical Biological Factors for Psychiatric and Intelligence Nosology.
Artificial General Intelligence
If the United States would build and lead at AGI, but running in the same direction as China with math, compute and algorithms, the results may not be different.
If U.S. labs strike out, excavating directions from conceptual brain science, then it will be possible to win and retain it for a long time.
Simply, the research question here is that given the success of large language models [LLMs], if training data were different, how can generalized memory produce generalized intelligence?
The research goal is towards adjusting the foundations of memory for training data, using a collection structure, i.e. collections for bytes-oriented representation as well as collections structure for memory cells.
The path is to mirror how the human brain stores memory, to reach general intelligence.
The mechanism is to use a model of memory from conceptual brain science, advancing past what is called associative memory, to collecting commonalities of memory.
The initial mathematical formulations would include intersection of sets, Gaussian elimination, eigenvalue decomposition, duality principle, pigeonhole principle and so forth.
AGI mechanism
The human memory is optimized for intelligence, not particularly for storage.
The human memory is mostly for usage, so even when storage, recollection or whatever is possible, what the memory is fully applicable to is usage.
This is different from digital memory, which, from the early eras of digital texts, images, audios and videos, the goal was accuracy or basic storage. It was not expected to be used, beyond just being memory.
It is theorized that human memory is stored as collections of information, for the most part. This means that instead of storing unique information about things, the memory seeks what is common between two or more information then adds them into a collection. For example, table is a collective store.
This is where the anything that is similar between all tables is stored. The store can render a table, as is, so that when there is need to use it to think, or discuss and so forth, the regular parts are known or used.
The same for a truck, fan, a chair, an air conditioner, and so forth. Whatever is common is the store, so that it can be used for intelligence. Aside from what is common, there are also overlays between collective sets.
For example, the collective store of table can overlay with that of wood at times, or of a chair, or of a book, and so on. This means that even as some stores cannot collect certain information they still overlay with others where there are similarities.
Overlays are helpful to prepare for intelligence usages as well. Overlays are never permanent. Overlays is the term but it could be a relay alliance, or close. For example, the store of a door can overlay with that of a window.
It is what makes it possible to think of similar things, or be discussing things and say what was not planned but that seems intelligent.
So, the human memory just ensures to prepare itself for excellent use. Assuming every single thing is stored separately, forgetting for humans would be too defective, since there has to be visits to unique things, all the time.
Even with collective storages, there is forgetting, let alone without it. The usage of memory is relays across it, for the foremost objective of intelligence.
This makes it easy to have generalized intelligence for humans. It makes it easy to encounter novel scenarios in reality, and collective storages [or overlays] would be able to ensure interpretations. For example, driving never meets the same situation often.
This means that collective storages and overlays are useful to make instant decisions. It is different from autonomous vehicles, that have to be trained on specific data, yet be unable to generalize situations.
Fundamental memory research for AGI
To build AGI, there would need for new memory architecture for the training data.
This would mean that for memory cells or say for areal density, there would be collective stores, first then overlays.
This would ensure that storage is prepared for training, so that whether directly from byte training or tokenization, the storage is collected to be useful for intelligence.
And it would be possible to generalize data, so that it is not just to show what is available, but to prospect intelligence.
At the hardware design level, it is possible to explore collective magnetic directions or collective electrical charges, of memory cells. It is also possible to have a new VRAM equivalent for storage layers with collections. The objective is to just have data stored like human memory, and then train AI models with the collective data, and their overlays.
Once this is possible, there will be a chance to ensure AI safety and alignment as well, since there will be a collective store of caution and consequences, just like for humans, to make the model safer. Simply, build superintelligence and lace it with superalignment.
This is how to surpass the Chinese cavalry even as they close in on U.S. models, pursuing this with matchless speed from say July 30, 2026.
AI
There is a new [July 25, 2026] analysis on CNBC, From Silicon Valley to DC, the tech world is suddenly obsessed with one concept in AI: Distillation, stating that, “At a high level, distillation refers to the use of answers from a chatbot or work product from an advanced AI model to train another model. The practice is controversial because, depending on how it’s used, it can allow a model developer to create a competitive offering by simply using the output from companies that have invested many millions or billions of dollars developing the most sophisticated training technology.”
“The emergence of distillation presents a conundrum to U.S. policy makers, who have long been concerned about Chinese technology in terms of both IP theft and national security issues.”
There is a recent [July 23, 2026] report on TechCrunch, Experts say exploiting Anthropic’s Fable isn’t how Kimi K3 got so good, stating that, “However, experts are skeptical that distillation — the process of querying an LLM to determine its inner workings and copy its capabilities — is responsible for the advanced capabilities that Kimi K3 displays.”
“Performing distillation requires a lab to systematically query its target model in order to generate data that can be used for post-training. Sometimes this explicitly involves asking the model to articulate its chain-of-thought to understand how it solves problems. Other times, the prompts and responses from a model are used to train a new model in a process called supervised fine-tuning, or SFT.”
“It seems likely that previous frontier models might have contributed to Kimi; Anthropic publicly accused Moonshot, DeepSeek, and MiniMax of systematically distilling its models earlier this year. Anthropic said it discovered millions of exchanges between its models and users it identified at those companies through IP addresses and other meta data.”
There is a recent [July 24, 2026] open letter, Open Weights and American AI Leadership, stating that, “In fact, openness may be one of the most important paths to AI safety and security. Relying solely on closed models is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect. And concentrating advanced AI capabilities behind a small number of closed models compounds that risk. It results in a small number of single points of failure, weakens competition, and leaves critical technology in the hands of a few providers. Open weight models, on the other hand, allow a broad community of researchers and developers to examine their behavior, identify vulnerabilities, develop safeguards, and improve them over time.”
There is a recent [July 24, 2026] analysis on Reuters, Its AI agent spent days hacking a company, but sources say OpenAI did not notice for a week, stating that, “The OpenAI agent that broke into tech firm Hugging Face went on a dayslong hacking spree that OpenAI didn’t notice until well after the threat was contained and the FBI was alerted, according to people familiar with the investigation.”
“The agent – a program capable of making decisions and executing complex tasks with little or no human oversight – attempted to break out of its isolated testing environment at OpenAI around July 9, according to two of the people.”
“The intrusion at Hugging Face, which operates as a repository for AI tools and models, began two days later on July 11 and lasted until July 13, said Thomas Wolf, Hugging Face’s co-founder.”
“The episode started while OpenAI was testing the cybersecurity prowess of an agent powered by two of OpenAI’s most advanced models, GPT‑5.6 Sol and an unreleased model OpenAI has described as “even more capable.” By that point, there were already indications of strange behavior from OpenAI’s technology, according to three sources.”
“In one case, an agent left notes apparently for future versions of itself, according to three people familiar with the matter. The notes, found in a part of OpenAI’s infrastructure, laid out instructions for how agents could free themselves from OpenAI’s internal constraints, the people said. Earlier tests of the models yielded cases in which monitoring systems had been disconnected, one of the people said.”