Blog
Research essays on what intelligence requires — and what it takes to build it.
The Intelligence Gap
Research
AI Can Talk — But It Can't Think
Large language models are genuinely capable — but capability isn't competence. The gap between producing likely outputs and knowing what's correct won't close with scale alone.
G6Solver Research Team
Research
The Hallucination Problem Is a Design Problem
Hallucination isn't a bug to be patched — it's an inevitable consequence of optimising for token probability rather than truth. The fix is architectural: separate generation from verification.
G6Solver Research Team
Research
Why Benchmarks Lie
When the benchmark becomes the target, it ceases to be a good measure. Goodhart's Law applies to AI evaluation with predictable consequences.
G6Solver Research Team
Research
The Context Window Is Not Memory
A context window is a buffer, not memory. It has no mechanism for prioritisation, consolidation, or retrieval by relevance. Real memory requires architecture.
G6Solver Research Team
Research
Scaling Laws Have a Ceiling
The ceiling isn't a wall — it's a signal that the next gains come from architecture, not scale. Post-training cognitive infrastructure changes the equation.
G6Solver Research Team
The Cognition Argument
Cognitive Science
Metacognition: The Skill AI Doesn't Know It's Missing
Knowing is not the same as knowing that you know. Without metacognition, AI systems cannot distinguish between high-confidence correct answers and high-confidence confabulations.
G6Solver Research Team
Philosophy
What Epistemology Teaches Us About AI Knowledge
Epistemology distinguishes true belief from justified true belief. LLMs hold statistical associations, not justified beliefs. Knowledge without justification is just lucky guessing at scale.
G6Solver Research Team
Cognitive Science
The Two Types of Thinking and Why AI Only Has One
Current AI is all System 1: fast, associative, pattern-matching. It has no System 2: slow, deliberate, logical. When hard problems need careful reasoning, it produces fast answers dressed in slow language.
G6Solver Research Team
Research
Learning Without Forgetting: The Continual Learning Problem
A doctor who couldn't learn from new cases, or an engineer who forgot old principles when learning new ones, would be useless. So why do we accept this from AI?
G6Solver Research Team
The Safety Imperative
AI Safety
Self-Improvement Without Self-Destruction
Safe self-modification is possible when the modification happens in an explicit, inspectable, formally verifiable layer — not in opaque neural weights.
G6Solver Research Team
AI Safety
The Alignment Problem Is an Engineering Problem
We have decades of safety engineering practice from aviation, nuclear, and medical devices. The problem isn't that we don't know how — it's that we haven't applied those disciplines to AI.
G6Solver Research Team
AI Safety
Why “Move Fast and Break Things” Doesn't Work for AI
A broken feature in a social media feed annoys users. A broken AI in healthcare causes real harm. The blast radius of AI failure is categorically different from software failure.
G6Solver Research Team
The Architecture of Intelligence
Architecture
The Operating System Analogy: Why AI Needs Middleware
LLMs are in the “bare metal” phase. Every application solves the same problems from scratch. What's needed is the cognitive equivalent of an operating system.
G6Solver Research Team
Architecture
Algorithms Are the Missing Intelligence
Neural networks learn parameters, not procedures. The frontier isn't bigger networks — it's systems that can select, compose, and synthesise algorithms at runtime.
G6Solver Research Team
Architecture
Formal Verification: The Safety Net AI Doesn't Use
Formal methods work in aviation, chip design, and nuclear systems. AI development has almost entirely ignored them. The question isn't whether to use them, but why we haven't been.
G6Solver Research Team
Economics & Power
Economics
The Economics of Intelligence: Why AI Gets Cheaper When It Thinks
A human employee gets faster over time; an LLM costs the same for its millionth query as its first. Systems that convert inference into reusable algorithms invert the economics entirely.
G6Solver Research Team
Economics
Democratising AGI: Who Gets to Build Intelligence?
Concentration of intelligence-building power is concentration of power, full stop. The alternative is systems where users build their own intelligence on top of any model.
G6Solver Research Team
Economics
The Human-AI Team: Augmentation, Not Replacement
The most productive arrangement is structured collaboration — AI handles computation and retrieval, humans handle judgment and creativity. The interface is the hard engineering problem.
G6Solver Research Team
Philosophy & Foundations
Philosophy
The Map Is Not the Territory
Statistical co-occurrence captures correlations, not causation. Models can describe gravity without understanding it. The map is getting more detailed, but it remains a map.
G6Solver Research Team
Philosophy
Antifragility in AI: Systems That Improve Under Stress
Current AI is fragile or at best robust. It is never antifragile. The difference is between a system that degrades over time and one that compounds value.
G6Solver Research Team
Philosophy
The Periodic Table of Intelligence
Intelligence isn't a single substance but a composition of discrete cognitive primitives. General not because it does everything at once, but because it can do anything by composition.
G6Solver Research Team