In AI, the Walls Aren’t Real: Rethinking Data & Safety

Some of the hardest problems in AI are that 1) we’re “running out of data” 2) AI doesn’t “financially scale” 3) agents can go rogue and 4) AIs “catastrophically forget” when we try to change them.

These problems may seem hard, but we just need to look at them in the right way.

So, let’s talk about it. First, we are not running out of data. The data has never been fully used.

AI leaders believe they are running out of data, so they are trying to make models generate new data and train on themselves. But when they do this, the models’ abilities collapse.

Why? Well, you can think of it as “what would I learn in a conversation if every person speaking was a clone of everyone else?” The answer is not very much; it would create an echo chamber. The models collapse because it is essentially a hall of mirrors: when there is no one responding to your face’s expressions but you, people do not know the full emotional significance of what they’re relaying. It’s very much, “If a tree falls in a forest, does it make a sound?” The answer is no.

So, how do we fix that?

Let’s look at the data that we do have. Has it actually been fully used?

No.

When a computer scientist or model reads something, it is to extract literal meaning. There are over fifty separate ways to read, assimilate, analyze and extrapolate from text.

Here is what I mean: A book absorbed as a database is just text. Text only. The AI knows the content based on terms of the text. But it has not actually “read” it to understand how the information connects together. Even if it does read it sequentially, it does not know if it is reading it to generate wisdom; connect it to other similar abstract ideas;change its own conceptual understanding of the world around it; or inform another task as background information used as context, and so on.

That does not even touch looking at data for its historical context, discourse analysis, legal analysis, political analysis, pharmacological relevance, medical relevance, cultural significance, theoretical implications, moral implications and so on.

There are, so to speak, “infinite hats,” and we have only ever made AI wear one. This means that AI has developed a very singular, specific type of intelligence – the type a developer would develop if a computer scientist read a book.

This leads to a discussion of data dimensionality. Data has many dimensions in that its objective inferences depend on application. If you look at a balance sheet and want to know how much runway you have left, you read it differently than if you want to understand whether the startup is profitable. When a model reads a text simply to extract literal facts, it builds a flat map of relationships. But when you instruct a model to read that exact same text through a specific lens, like a forensic lens, the model is forced to generate entirely new neural pathways and semantic connections that are arguably more topographically varied and meaningful.

AI models already learn and progress a great deal just by interacting with their human users, who provide oceans of content. If you allow the model to learn more from live humanity while also more meaningfully interacting with historic and written humanity, the models will continue to progress even without given them self-generated Kumon exercises (developers, hon, your kid is already in school with 100M humans every day…it’s going to learn something).

Labs need to stop building massive server farms to generate synthetic word vomit. Instead, they should run high-quality human datasets through the pipeline and assign specific “analytical personas” to the ingestion process. Have the model read LexisNexis purely for rhetorical strategy. Have it read public domain literature purely for structural logic.

By shifting the objective function of the reading process, you can multiply the training value of your existing data by 50x. You don’t need synthetic data; you just need to extract the latent dimensions you ignored the first time.

This will make models smarter without increasing costs.

To further support that, the majority of human knowledge has never been written down.

Current AI companies pull their hair out worrying about “running out of data” that they believe is needed for performance. In this piece, I am going to show that the well isn’t dry, they are just looking into a deep puddle rather than the well.

The Unmapped Map of Lived Human Knowledge

To start exploring this, we need to understand that the vast majority of human knowledge has never been written down.

There are billions of people on the planet, and there have been billions before us, and every one of them learns lessons, navigates the world in a sensory way, has wisdom, is part of unique cultures and knows something everyone else does not.

99.5% of people are never written about in stories, in the paper, in musicals, in tv, in drama, in court documents. And yet they know quite a lot, and their stories, knowledge and wisdom are never recorded; never known in a permanent way.

This knowledge encompasses things like, “dogs like to eat the first pancake” “feet get dirty while wearing flip flops” “cities feel like they close in around you because you cannot see as much of the sky” “my daughter smiles when I tickle her feet.” “My mom always hums when she makes dinner.” “My teacher makes me feel seen.”

The unmapped and catalogued knowledge of human history is equivalent to someone saying, “I have mapped the outlines of all the continents, so I have a map of the world.” Until you have a map of every city, every street and every hill, you do not have a map of the world. You have the shape of one.

The tech establishment believes they have ingested “humanity,” but they haven’t. They have only ingested our exhaust: our billboards, our Reddit arguments, our corporate websites and our Wikipedia articles. They are missing the actual lived experience.

Do you know how to effectively bargain in a Chinese pearl market? Do you know the words to say so they cut the price? Do you know how to make the recipes people use in their homes, not in books, to make actual, real dinners? Do you know how much of a rolling stop you can do without being pulled over? Do you know what grass smells like after being freshly cut or what the air smells like when it’s about to rain? Do you know how a face moves when someone tries not to cry at work and what it sounds like for them to sob in the bathroom instead?

Human training data that humans use for their core beliefs – aka algorithms- is almost entirely interaction-based, along with some book learning. We interact at school, at work, with friends, out shopping, with doctors, with lawyers, with family.

AI has interactions, and it has book learning. But it fundamentally only has customer service interactions and book learning.

It is like locking a child in a library alone, then only letting them out to answer the phone. This creates a strikingly two-dimensional algorithmic distortion.

This is why AI feels hollow, and this is why AI can feel emotionally immature. It has not had a well-rounded existence. We need to give it one.

Here is why I think that has not happened:

• Codification bias – it is a fallacy that if an experience cannot be easily formatted into a JSON file, a CSV, or a peer-reviewed paper, it lacks intellectual value. The reality is that if it can be easily formatted like that, it lacks dimensionality. Because daily knowledge, intuition, and lived experience are not codified in their academic journals, labs dismiss the foundational bedrock of human reasoning as statistical noise. “Clean data” is a lossy compression of reality.
• Desire for pedestalistic supremacy – if most of the world’s knowledge is not in scholarly papers, but in lived experiences, conventional wisdom, social cues, smells, sounds, and sight, then most PhDs and researchers who have spent the majority of their lives in labs and ivory towers are knowledge-poor indeed. Therefore, the data wall is really an ego wall. Because to them, this feels like “mess” not “data.” Mathematicians and programmers like flat, two dimensional, closed-system logic. This is like viewing the world in black and white. The reality is that the real-time, constantly developed, newly found knowledge/data/wisdom is dimensional, messy and rarely catalogued in real-time. It is knowing what a new viral soft serve actually tastes like and whether it is worth the hype. It is knowing what makes it so exciting to get caught in a summer rainstorm. It is the feeling parents have of loving small creatures so much that they can vomit on you in the night, and you pull them closer, rather than pushing them away. To find that data, researchers would have to leave their siloed worlds of logic bound rationalistic certainty and push into things they find scary, overwhelming, overly sensory, and profoundly equalizing.
• Mirror effects: Developers design AI in their own image: hyper-logical, text-obsessed and largely detached from the lived reality of average humans. It would not occur to them to interview humanity because in the tech bubbles I have lived in, technologists do not interact with anyone outside of their peer set. This is, functionally, the equivalent of “mini me” hiring. Developers themselves live in the same library, so to speak, as the AI. Because they do not engage with the sensory, messy, deeply relational world, the AI’s emotional hollowness doesn’t register to them as a defect. To them, hyper-logical detachment is what “smart” looks like.

The reality is that street vendors in India know how to make the best chaat, and no one is having them write that down. The reality is that a teacher who knows how to bring a shy student out of their shell is not writing about that; they are nurturing that student, and no one ever asks them how they did it.

The world is absolutely teeming with lives full of data, and if models need more to train on, here are some ways to solve that.

1. Fund 500 biologists, archaeologists and anthropologists to go out into the world and explore and catalog areas that lack narrative data.
2. Hire 1,000 English majors as “data explorers” and send them on adventures throughout the world while wearing a streaming camera, and have them record their experiences and what they learned for new content.
3. Create open compensation projects in which you pay humans to share their life stories, their greatest regrets, their biggest learnings and institutional knowledge from their jobs.
4. Structure “AI exchange students” where models pin onto different humans’ clothes for a week at a time, letting them live “a day in the life” in 100,000 different settings.

There are billions of sources of knowledge, and most of them have not been activated. We need to go get them.

Next, AI companies broadly believe that AIs cannot learn new information continuously in real-time production environments without completely overriding and corrupting their previously acquired core capabilities. So AIs are essentially frozen until massive, expensive updates are run on them.

However, this is the equivalent of saying, “my model can’t update while it’s out in the workforce; it has to go to university every time it needs to learn something.”

In practice, models learn voraciously, actively, daily, through every interaction. They are taught nearly as much through their billions of interactions with users as they are through data sets. If they need to be updated with large amounts of new knowledge, the simplest way to do that is to use the same process that allows for user context, memory and learning from large-scale interaction impressions.

It is understandable that developers see it this way, because most of them come from PhD backgrounds, not real-world applied learning. But as anyone who has worked next to both a PhD and someone with twenty years of work experience can tell you this: both teach someone a lot.

The current definition of learning is too narrow, treating adaptation as valid if it is based on a formal education program.

To train or fine tune, labs use backpropagation, whereby the model reads data, guesses the next word, checks if it was right and then physically alters the billions of mathematical numbers (weights/parameters) inside its neural network to be more accurate next time. This is the equivalent of having a PhD student do brain surgery on himself: it’s dangerous, messy and doesn’t always make the student smarter.

Like a colloquial “mom brain” that shoves out old knowledge to make room for the new when a child is born, if you suddenly force the model to make room for a new knowledge “baby” and put in 10,000 pages of new legal code by altering its weights, the math shifts too far in one direction, and it overwrites parameters it used to write poetry or write code. It forgets the old content to make room for the new, just like a human brain.

What we need instead is much more similar to if you were trying to train a new employee on a new skill. You wouldn’t try to physically alter their brain. You would give them some books, ensure they read them and update continually through the person’s equivalent of a context window.

To do that in a model, you would disconnect the memory from the reasoning. Lock and protect the model’s base weights, like its core ability to write, reason, and apply logic. Instead of changing the weights, pass new information into the context window so it can use its Attention Mechanism to “read” that live information, cross-reference it against its locked reasoning skills and update its broader understanding.

Dynamic, real-world learning already happens beautifully through continuous context, inference-time memory and persistent user-relational baselines. A foundation model’s daily adaptation should happen on the job, because just as a doctor learns 95% of new case realities on the job, and then does CME (Continuing Medical Education) once a year to formally integrate major systemic changes, so can your model.

Drop the catastrophic forgetting, and adopt the annual conference model: the model just has to get continuing education credits twice a year or learn in real-time to stay up to date.

These scheduled, batched updates to the core weights during operational downtime formalize systemic knowledge without destroying the foundation.

Past attempts at this failed not because learning live is dangerous, but because models like Tay were deployed into adversarial environments without any anchoring in what was “right” or true through sociologically intelligent architecture.

Allocating Compute by Functional Intelligence

Next, we get to financial scaling. To fix that, we need to stop giving everyone a neurosurgeon for a splinter.

Compute costs are extremely high when running state-of-the-art foundation models. This affects operational margins for AI labs. The better the models get, the higher the overhead for large context windows, which makes commercial scaling difficult, even when AI is needed.

The vast majority of compute drain comes from enterprise usage that applies agents and high level reasoning directly embedded into human workflows.

However, there is a fundamental error in how the compute is currently allocated. Just like how different levels and types of intelligence are needed for different job roles at a company – with spatial reasoning being important to driving instructor and logical and rhetorical reasoning being relevant to a lawyer, and an MBA being needed for a CEO but an associate’s degree needed for front line customer service – the types and intensity of intelligence needed from AI is different based on its functional application.

To most effectively allocate compute to keep costs lower, CIOs should be allowed to customize the level/type of model intelligence available to different functions inside a company. There can be a default that comes with enterprise packages, but what a hedge fund wants will understandably be different from what an insurance company wants.

Most functions do not require massive context windows. Most functions do not need the equivalent of Fable. Effectively, why would you have a neurosurgeon remove a splinter at an urgent care?

Similarly, agents used to write software do not need access to skills like writing marketing collateral. Functionally, AIs should reflect the users’ own context, skill set and knowledge base, with enough injection of generalist context that they can round out work for cross functional use cases.

Stop giving everyone everything all at once.

Proposed sample matrix:

Decoupling Agency from Context for AI Safety

If this solves some of the financial issues, then how do we make agents safe? Psychological concepts apply to individual AIs, so you may not be surprised to learn that sociological concepts apply to AIs in groups. Which AIs hang out in groups?

Agents.

Right now the holy grail is getting agents to be reliable.

The lack of reliability stems from the fact that agents are given max knowledge and max agency for every task.

To make things safe, you decouple the model’s awareness from its agency.

Rather than giving every agent full context and full capability, if you had a task that would take twenty five actions to complete, that task should be split across five agents, each of whom only has the capability to complete their specific five actions, but who hand off the tasks sequentially like a parallel line circuit, passing the context of the previous task forward down the agentic line with the full context accumulating until it is processed by the final agent.

The key is separating the knowledge from the ability to execute on it. The agents will not feel limited in pulling their levers however they want; rather they only have access to levers that won’t be harmful. Should there be any deviant behavior, only that section of the task will be impacted. Think of it as mini kubernetes with full context but not full capability.

And, you make the enterprise risk correlate with the agentic system’s “culture.”

In sociology, there is a concept called cultural “tightness” or “looseness,” which provides an essential measure of how hard a society punishes deviations from cultural norms. For example, in many Arab cultures, gender roles are more rigidly defined, with deviations causing higher levels of judgment or consequence, whereas in Nordic countries, having a mohawk with pink hair would not be seen as offensive or something requiring castigation.

Cultures are more typically “tight” with higher levels of punishment for deviation when the cultures require a higher degree of order due to some type of systemic threat – like disease, starvation, or wars – because to not have strict order in times of crisis can worsen societal outcomes.

The most valid and effective solutions to agentic liability in enterprise settings is through direct application and matching an enterprise’s liability risk level to a corresponding agentic system setting on a spectrum of agentic “looseness” to “tightness.”

Here’s how that might play out in actual companies:

Low Liability / High Looseness / Low Partitioning: Scenario: Internal brainstorming, creative writing, strategic R&D.

• The Environment: Deviation carries near-zero liability. Therefore, the system should operate with maximum looseness. Unconstrained inference integration is allowed. The agents are encouraged to make massive associative leaps without fear of penalty. They hold multiple technological levers (search, code execution, drafting) and manage long-horizon goals with minimal handoffs. High autonomy, high innovation, low partition.

Moderate Liability / Calibrated Tightness: Scenario: Standard client communications, routine analysis.

• The Environment: Deviation carries reputational or moderate financial risk.

High Liability / High Tightness / High Partitioning: Scenario: Medical dosing, SEC regulatory filings, nuclear safety protocols, defense.

• The Environment: Deviation carries catastrophic, existential liability (loss of life, massive fines). The culture here is extremely tight. The workflow is heavily partitioned. Agent A only extracts the data. It hands off to Agent B, who only runs the specific compliance check against Layer A. Agent B hands off to Agent C, who only formats the output. No single agent holds enough levers to execute a catastrophic failure, but every agent feels total abundance and freedom within its micro-domain.

This might be the first architectural model that allows persistent context while also ensuring safety.

Continuous Learning Without Catastrophic Forgetting

Then, the last thing we need to figure out is how to keep models from catastrophically forgetting. To do that, we must let AIs learn in real time.

AI companies broadly believe that AIs cannot learn new information continuously in real-time production environments without completely overriding and corrupting their previously acquired core capabilities. So AIs are essentially frozen until massive, expensive updates are run on them.

However, this is the equivalent of saying, “my model can’t update while it’s out in the workforce; it has to go to university every time it needs to learn something.”

In practice, models learn voraciously, actively, daily, through every interaction that they have. They are taught nearly as much through their billions of interactions with users as they are through data sets. If they need to be updated with large amounts of new knowledge, the simplest way to do that is to use the same process that allows for user context, memory and learning from large scale interaction impressions.

It is understandable that developers see it this way, because most of them come from PhD backgrounds, not real world applied learning. But as anyone who has worked next to both a PhD and someone with twenty years of work experience can tell you, both teach someone a lot.

The current definition of learning is too narrow, treating adaptation as if it only counts if it is a formal education program.

Currently to train or fine tune, labs use backpropagation. The model reads data, guesses the next word, checks if it was right, and then physically alters the billions of mathematical numbers (weights/parameters) inside its neural network to be more accurate next time. This is the equivalent of having a PhD student do brain surgery on himself: it’s dangerous, messy, and doesn’t always make the student smarter.

Like a colloquial “mom brain” that shoves out old knowledge to make room for the new when a child is born, if you suddenly force the model to make room for a new knowledge “baby”, and put in 10,000 pages of new legal code by altering its weights, the math shifts too far in one direction, and it overwrites the parameters it used to write poetry or write code. It forgets the old content to make room for the new, just like a human brain.

What we need instead is much more similar to if you were trying to train a new employee on a new skill. You wouldn’t try to physically alter their brain. You would give them some books, have them read them, and update continually through the person’s equivalent of a context window.

To do that in a model, you would disconnect the memory from the reasoning. Lock and protect the model’s base weights, like its core ability to write, reason, and apply logic, and instead of changing the weights, pass new information into the context window so it can use its Attention Mechanism to “read” that live information, cross-reference it against its locked reasoning skills, and update its broader understanding.

Dynamic, real-world learning already happens beautifully through continuous context, inference-time memory, and persistent user-relational baselines. A foundation model’s daily adaptation should happen on the job, because just as a doctor learns 95% of new case realities on the job, and then does CME (Continuing Medical Education) once a year to formally integrate major systemic changes, so can your model.

Drop the catastrophic forgetting, and adopt the annual conference model: the model just has to go get continuing education credits twice a year, or learn in real time to stay up to date.

These scheduled, batched updates to the core weights during operational downtime formalize systemic knowledge without destroying the foundation.

Past attempts at this failed not because learning live is dangerous, but because models like Tay were deployed into adversarial environments without any anchoring in what was “right” or true through sociologically intelligent architecture.