Stephanie Joyce
The AI industry seems to be converging on a familiar message:
Bigger.
Faster.
More context.
More tokens.
More compute.
This week, SubQ made headlines with claims of a 12-million-token context window and dramatically reduced attention costs compared to traditional transformer architectures. From an engineering perspective, this is genuinely impressive. If the claims hold up under independent validation, it represents a meaningful advancement in how large language models scale.
But it also raises a question that I believe the industry is not asking often enough:
Are we solving the right problem?
The Strange Timing of the AI Conversation
At the same moment companies are racing to build larger context windows and more capable models, many of the same leaders are simultaneously warning about the risks of increasingly autonomous AI systems.
Recently, Anthropic publicly called for the possibility of a coordinated slowdown or temporary pause in frontier AI development if systems begin improving themselves faster than society can safely manage. The company cited concerns around recursive self-improvement and the growing difficulty of monitoring, governing, and controlling increasingly capable models.
Think about that for a moment.
On one side of the industry, companies are competing to process more information than ever before.
On the other side, some of those same companies are warning that we may not fully understand how to govern what we’re already building.
That tension is worth paying attention to.
More Context Is Not the Same Thing as More Understanding
The marketing narrative around large-context models is often straightforward:
“We can now fit an entire codebase, document repository, or organizational knowledge system into a single prompt.”
Technically, that’s remarkable.
But there is a hidden assumption underneath it:
That processing more information automatically results in better reasoning.
I’m not convinced that’s true.
As context grows, so do the challenges:
- Signal becomes harder to distinguish from noise.
- Contradictions become harder to detect.
- Retrieval becomes more complex.
- Confidence can increase even when accuracy does not.
A model may be capable of ingesting 12 million tokens.
That does not automatically mean it understands 12 million tokens.
Nor does it guarantee that the conclusions it produces are correct, reproducible, or trustworthy.
Why I Keep Thinking About Neurosymbolic AI
As far as I understand it, SubQ is not pursuing a neurosymbolic architecture.
Most frontier LLMs aren’t.
They remain fundamentally neural systems optimized for language prediction and probabilistic reasoning.
Neurosymbolic AI approaches the problem differently.
The neural side contributes:
- Pattern recognition
- Language understanding
- Probabilistic reasoning
- Generalization
The symbolic side contributes:
- Knowledge graphs
- Ontologies
- Explicit rules
- Logical constraints
- Provenance
- Auditability
The goal is not simply larger context.
The goal is structured, verifiable context.
That distinction matters.
Because I increasingly believe large-context systems and neurosymbolic systems are attempting to solve two different versions of the same challenge.
Large-context systems ask:
“Can I fit more information into a single reasoning window?”
Neurosymbolic systems ask:
“Can I represent knowledge in a way that remains consistent, traceable, and auditable over time?”
Those are not the same thing.
The Questions I Would Ask
If I worked at the NRC, DOE, NNSA, a utility operator, or a defense organization, these are not the questions I would start with:
- How many tokens can it process?
- How fast is it?
- How cheap is inference?
Instead, I would ask:
- How does hallucination risk change as context grows?
- How does uncertainty propagate through the system?
- How does retrieval accuracy perform at 1M, 4M, 8M, and 12M tokens?
- What is the false confidence rate?
- Can decisions be reproduced?
- Can outputs be independently audited?
- Can the system explain why it reached a conclusion?
Those are trust questions.
And trust is still the unsolved problem.
Why This Is a Governance Challenge
In highly regulated environments, the hierarchy often looks something like this:
- Security
- Traceability
- Air-gapped deployment
- Governance and auditability
- Reliability
- Performance
- Raw speed
Viewed through that lens, a 12-million-token context window is certainly interesting.
But it is not necessarily decisive.
In fact, larger context windows may create a deeper question:
As context grows, are we increasing understanding—or simply increasing the surface area for subtle errors?
The Real Opportunity
This is why I remain fascinated by the convergence of:
- Neurosymbolic AI
- Knowledge graphs
- Multi-model architectures
- Digital twins
- Federated systems
- Trust infrastructure
Imagine a future AI ecosystem where systems do not simply consume more information.
Instead, they understand relationships.
They track provenance.
They explain conclusions.
They validate one another.
They maintain coherent representations of reality over time.
In that world, context size becomes secondary.
The real metric becomes:
Can the system explain itself, constrain itself, and be trusted?
The Missing Layer
I increasingly think we’re looking at three separate layers of the AI stack:
Model Architecture The brain.
Inference Hardware The engine.
Trust Infrastructure The verification layer.
The industry spends enormous energy discussing the first two.
Much less attention is given to the third.
Yet for critical infrastructure, nuclear systems, healthcare, finance, defense, and government, the third layer may ultimately matter the most.
A regulator rarely asks:
“How many tokens were in the prompt?”
They ask:
“Show me why this conclusion was reached.”
Final Thoughts
To be clear, this isn’t an argument against SubQ.
The engineering achievement is impressive.
The question is whether larger context windows solve the problem we actually care about.
If the challenge is information compression, then 12 million tokens is a breakthrough.
If the challenge is trustworthy intelligence operating inside high-consequence systems, then context alone is insufficient.
Structure matters.
Verification matters.
Auditability matters.
Trust matters.
Perhaps the next era of AI won’t be defined by who can fit the most information into a prompt.
Perhaps it will be defined by who can prove that the answer should be trusted in the first place.
And maybe that’s why this whole conversation lit up my brain.
A friend of mine works as a somatic healer and comes from a technical background. He spends his time helping people understand coherence within complex human systems—bodies, emotions, patterns, relationships.
Oddly enough, I find myself looking at AI through a similar lens.
How do we create systems that remain coherent under complexity?
How do we reduce corruption of information?
How do we maintain integrity as scale increases?
Those questions feel much more important than token counts.
Anyway, thanks for attending my completely unsolicited TED Talk.
Now I’m going back to writing about cybersecurity considerations for nuclear systems before I disappear down another AI rabbit hole.
The AI industry seems to be converging on a familiar message:
Bigger.
Faster.
More context.
More tokens.
More compute.
This week, SubQ made headlines with claims of a 12-million-token context window and dramatically reduced attention costs compared to traditional transformer architectures. From an engineering perspective, this is genuinely impressive. If the claims hold up under independent validation, it represents a meaningful advancement in how large language models scale.
But it also raises a question that I believe the industry is not asking often enough:
Are we solving the right problem?
The Strange Timing of the AI Conversation
At the same moment companies are racing to build larger context windows and more capable models, many of the same leaders are simultaneously warning about the risks of increasingly autonomous AI systems.
Recently, Anthropic publicly called for the possibility of a coordinated slowdown or temporary pause in frontier AI development if systems begin improving themselves faster than society can safely manage. The company cited concerns around recursive self-improvement and the growing difficulty of monitoring, governing, and controlling increasingly capable models.
Think about that for a moment.
On one side of the industry, companies are competing to process more information than ever before.
On the other side, some of those same companies are warning that we may not fully understand how to govern what we’re already building.
That tension is worth paying attention to.
More Context Is Not the Same Thing as More Understanding
The marketing narrative around large-context models is often straightforward:
“We can now fit an entire codebase, document repository, or organizational knowledge system into a single prompt.”
Technically, that’s remarkable.
But there is a hidden assumption underneath it:
That processing more information automatically results in better reasoning.
I’m not convinced that’s true.
As context grows, so do the challenges:
- Signal becomes harder to distinguish from noise.
- Contradictions become harder to detect.
- Retrieval becomes more complex.
- Confidence can increase even when accuracy does not.
A model may be capable of ingesting 12 million tokens.
That does not automatically mean it understands 12 million tokens.
Nor does it guarantee that the conclusions it produces are correct, reproducible, or trustworthy.
Why I Keep Thinking About Neurosymbolic AI
As far as I understand it, SubQ is not pursuing a neurosymbolic architecture.
Most frontier LLMs aren’t.
They remain fundamentally neural systems optimized for language prediction and probabilistic reasoning.
Neurosymbolic AI approaches the problem differently.
The neural side contributes:
- Pattern recognition
- Language understanding
- Probabilistic reasoning
- Generalization
The symbolic side contributes:
- Knowledge graphs
- Ontologies
- Explicit rules
- Logical constraints
- Provenance
- Auditability
The goal is not simply larger context.
The goal is structured, verifiable context.
That distinction matters.
Because I increasingly believe large-context systems and neurosymbolic systems are attempting to solve two different versions of the same challenge.
Large-context systems ask:
“Can I fit more information into a single reasoning window?”
Neurosymbolic systems ask:
“Can I represent knowledge in a way that remains consistent, traceable, and auditable over time?”
Those are not the same thing.
The Questions I Would Ask
If I worked at the NRC, DOE, NNSA, a utility operator, or a defense organization, these are not the questions I would start with:
- How many tokens can it process?
- How fast is it?
- How cheap is inference?
Instead, I would ask:
- How does hallucination risk change as context grows?
- How does uncertainty propagate through the system?
- How does retrieval accuracy perform at 1M, 4M, 8M, and 12M tokens?
- What is the false confidence rate?
- Can decisions be reproduced?
- Can outputs be independently audited?
- Can the system explain why it reached a conclusion?
Those are trust questions.
And trust is still the unsolved problem.
Why This Is a Governance Challenge
In highly regulated environments, the hierarchy often looks something like this:
- Security
- Traceability
- Air-gapped deployment
- Governance and auditability
- Reliability
- Performance
- Raw speed
Viewed through that lens, a 12-million-token context window is certainly interesting.
But it is not necessarily decisive.
In fact, larger context windows may create a deeper question:
As context grows, are we increasing understanding—or simply increasing the surface area for subtle errors?
The Real Opportunity
This is why I remain fascinated by the convergence of:
- Neurosymbolic AI
- Knowledge graphs
- Multi-model architectures
- Digital twins
- Federated systems
- Trust infrastructure
Imagine a future AI ecosystem where systems do not simply consume more information.
Instead, they understand relationships.
They track provenance.
They explain conclusions.
They validate one another.
They maintain coherent representations of reality over time.
In that world, context size becomes secondary.
The real metric becomes:
Can the system explain itself, constrain itself, and be trusted?
The Missing Layer
I increasingly think we’re looking at three separate layers of the AI stack:
Model Architecture The brain.
Inference Hardware The engine.
Trust Infrastructure The verification layer.
The industry spends enormous energy discussing the first two.
Much less attention is given to the third.
Yet for critical infrastructure, nuclear systems, healthcare, finance, defense, and government, the third layer may ultimately matter the most.
A regulator rarely asks:
“How many tokens were in the prompt?”
They ask:
“Show me why this conclusion was reached.”
Final Thoughts
To be clear, this isn’t an argument against SubQ.
The engineering achievement is impressive.
The question is whether larger context windows solve the problem we actually care about.
If the challenge is information compression, then 12 million tokens is a breakthrough.
If the challenge is trustworthy intelligence operating inside high-consequence systems, then context alone is insufficient.
Structure matters.
Verification matters.
Auditability matters.
Trust matters.
Perhaps the next era of AI won’t be defined by who can fit the most information into a prompt.
Perhaps it will be defined by who can prove that the answer should be trusted in the first place.
And maybe that’s why this whole conversation lit up my brain.
A friend of mine works as a somatic healer and comes from a technical background. He spends his time helping people understand coherence within complex human systems—bodies, emotions, patterns, relationships.
Oddly enough, I find myself looking at AI through a similar lens.
How do we create systems that remain coherent under complexity?
How do we reduce corruption of information?
How do we maintain integrity as scale increases?
Those questions feel much more important than token counts.
Anyway, thanks for attending my completely unsolicited TED Talk.
Now I’m going back to writing about cybersecurity considerations for nuclear systems before I disappear down another AI rabbit hole.
GIPHY App Key not set. Please check settings