Beyond the AI Scaling Wall: Trident G and the Future of General Intelligence

From Crystallised Knowledge to Fluid Reasoning: AI’s New Frontier?

Dec 27, 2024

Is AI stalling out, or are we simply shifting gears? Over the past year, that question has fuelled a surprisingly intense debate among AI insiders. Some major voices - like OpenAI co-founder Ilya Sutskever - hint that blindly scaling Large Language Models (LLMs) may no longer yield the transformative jumps we enjoyed when GPT-4 eclipsed GPT-3.5. Meanwhile, others in leading labs insist we’ve still only scratched the surface of what advanced data generation, multimodality, and deeper “thinking loops” (where the model to iterates on its own output) can achieve.

So, who’s right? Or is the AI landscape just more nuanced? Today, I’m drawing on two fascinating takes by AI guru Rohit Krishnan to frame how we might balance these competing perspectives. Along the way, I’ll introduce Trident G, my own cognitive-neuroscience–inspired theory of intelligence. The punchline: scaling knowledge alone (sometimes called crystallised intelligence, or Gc) won’t solve deeper challenges of memory, iterative reasoning, and metacognitive oversight - critical components of fluid intelligence (Gf). If we want truly adaptive AI, we need an architecture that mixes both raw knowledge and dynamic problem-solving acumen.

1. A Quick Snapshot of Trident G: Gc Meets Gf

Before we dive into Krishnan’s articles, let’s set the stage. Trident G is a theory that unifies two classical ideas:

Crystallised Intelligence (Gc): The store of facts, domain knowledge, and patterns absorbed over time.
Fluid Intelligence (Gf): The flexible ability to solve novel tasks, reason abstractly, and manage multi-step planning.

1.1 The 3D ‘Fork’ of Fluid Intelligence

In Trident G, Gf is not just a vague label. We model it as a 3D space, with each axis capturing a different problem-solving dimension:

Representational Axis:
- Dimensionalisation (extracting relevant concepts or components of a problem) vs. Configuration (fixing on a way of initially representing the problem - such as a spatial or verbal).
Computational Axis:
- Relational Reasoning (finding logical relationships) vs. Evidence Accumulation (deciding when to commit to a conclusion based on tradeoffs and evidence).
Working Memory Axis:
- WM1 (in-the-moment scratch-pad manipulations) vs. WM2 (standing constraints, rules and subgoals).

These three axes together let you picture how a system handles novel tasks: does it reframe the problem at will, does it carefully gather evidence, does it keep track of constraints over multiple steps? Humans excel precisely because we can juggle each axis adaptively. This 3D space of fluid intelligence (Gf) is the central fork of the trident.

The Trident G theory of general intelligence

1.2 Creative vs. Executive Modes

Layered on top of Gf are two overarching modes:

Creative Mode (right fork): Wide attention, leaps of insight, willingness to explore half-baked ideas.
Executive Mode (left fork): Narrow focus, logic-driven filtering, high thresholds for final answers.

In practice, we toggle between these stances as tasks demand. Think: initially brainstorming a puzzle solution (Creative), then methodically zeroing in once we spot a promising path (Executive).

1.3 Metacognition and Near-Criticality

At the core is what I call the φ near-critical region, metacognition or mindfulness. This is the system that monitors our performance, nudges us to shift from Creative to Executive (or vice versa), and keeps us from drifting into chaos or locking up in rigid patterns. In neural network terms, this kind of minfulness underpins the brain adaptively hovering near a ‘critical point,’ balancing adaptability and structure. This Quanta Magazine video introduces the ‘brain criticality’ hypothesis for anyone who wants a deep dive into this physics-based advance in our understanding of intelligence.

2. Where Does Gc Fit In?

Crystallised intelligence (Gc) - the ‘shaft’ of the trident in our model - is all about knowledge: facts, domain-specific patterns, vocabulary, code snippets - exactly what current LLMs devour by the terabyte. If you think about models like ChatGPT or Claude, they’re essentially Gc powerhouses. But as we’ll see, stockpiling knowledge is only half the battle when it comes to multi-step tasks requiring a stable sense of constraints and real-time adaptation - where Gf truly shines.

3. Two Perspectives from Rohit Krishnan: The Scaling Dilemma and Goal Drift

3.1 “Is AI Hitting a Wall?”

In one article, Is AI hitting a wall?, Krishnan asks if we’ve squeezed most of the juice from ‘just scale it’ approaches. Yes, we can keep throwing more data and compute at LLMs, but the flashy leaps from GPT-3.5 to GPT-4 might not repeat quite so dramatically. On the bright side, he points to new paradigms that could keep momentum going:

Synthetic Data: Re-shaping the same data in multiple ways, or generating entirely new tasks so the model ‘sees’ material from different angles.
Multi-Modal Integration: Feeds from video, robotics, or symbolic systems can drastically expand the Representational Axis (dimensionalisation).
‘Think for Longer’ Strategies: Chain-of-thought prompts, reflection loops, or RL-based reasoning cycles. These directly boost the Computational Axis (and approximate WM2 if they store constraints across steps).

From a Trident G lens: simply maxing out Gc doesn’t do the trick. We need scaffolds - like synthetic data or iterative prompting - that begin to replicate elements of fluid intelligence. The old playbook of ‘let’s scale bigger’ might be hitting diminishing returns, but that doesn’t mean we’re at a dead end. More like switching from brute force knowledge to better structural design.

3.2 “What Can LLMs Never Do?”

Krishnan’s other essay, What can LLMs never do?, highlights how LLMs stumble on tasks that demand stable constraints over multiple steps - like Wordle, Sudoku, or Conway’s Game of Life. These puzzles are child’s play for a moderately bright human, yet LLMs often lose track of partial solutions or fail to notice obvious contradictions.

The culprit is goal drift. Because many LLMs generate tokens in a purely forward pass, they don’t spontaneously re-check their own logic or revert to earlier steps. They’re missing a robust WM2 (a ‘map’ of where you are in the problem) and that all-important metacognition that might say, “Hey, wait—this is inconsistent with what we said two steps ago!” Without those, the model’s single-stream approach can unravel quickly.

In Trident G terms, we see:

Weak WM2: No stable store for subgoals or constraints.
Missing φ (metacognition): No built-in mechanism to pivot modes or revise mistakes mid-sequence.
Huge Gc alone can’t patch these holes: LLMs ‘know’ loads but still fail at iterative reasoning.

4. Bridging the Gap: Toward Human-Like G

4.1 Scaffolding Gf Processes

Researchers are starting to tack on chain-of-thought prompts, external memory modules, or reinforcemebnt learning with human feedback. In Trident G lingo, these are partial attempts to replicate WM2 or meta-level oversight. They help, but as Krishnan notes, the fundamental single-pass architecture might remain a bottleneck.

4.2 Creative vs. Executive Toggling

Humans naturally decide when to think outside the box (Creative) and when to clamp down on detail (Executive), harnessing the critical dynamics of a neural network. AI systems still lack a robust ‘mode-switch’ framework. They often produce creative flourishes or systematic reasoning, but rarely pivot seamlessly between them.

4.3 Gc + Gf = The Real Power Move

Long-term, Trident G suggests synergy: combine the vast knowledge base (Gc) of an LLM with a fluid, near-critical system that can maintain stable constraints, re-check partial solutions, and shift modes on the fly. That’s not a trivial engineering fix. It’s more like an architectural overhaul -one that might borrow from multi-agent hierarchies, neural criticality theory, or dynamic memory frameworks. But it’s a path beyond the shortfalls we see today.

5. Looking Ahead

So, is AI hitting a scaling wall? Sort of, if you define ‘wall’ as the point where pure expansions of knowledge alone (Gc) yield diminishing returns. But from another angle, we’re just turning a corner: the next wave may hinge on evolving our models from Gc-rich ‘fuzzy processors’ into well-rounded cognitive engines with fluid memory, contextual awareness, and mode-switching capabilities.

For me, this is where Trident G feels like a practical roadmap. It breaks down intelligence into:

Raw knowledge (Gc): necessary, but insufficient.
Multi-axis fluid reasoning (Gf): bridging representational expansions, iterative computation, and stable working memory.
Creative–executive toggling: so we don’t get stuck in one style of thinking.
Near-critical metacognition (φ): ensuring we catch mistakes, shift strategy, and coordinate the entire show.

If that sounds ambitious - well, so was building GPT-4 in the first place. The takeaway is that scaling isn’t done; it’s just no longer the only game in town. Researchers are deep into exploring new “S curves,” as Krishnan puts it, from advanced synthetic data pipelines to more sophisticated “reasoners.” And, ironically, that might be exactly what AI needs to step past the old plateau and reach a new horizon of true general (or at least more general) intelligence.

Final Thought

In short, the challenge of the alleged ‘AI wall’ hasn’t yet won the day. There is an opportunity to embrace the chance to address deeper structural gaps - particularly those around working memory, iterative reasoning, and metacognition. In a world where ChatGPT can confidently chat all day but can’t reliably solve a Sudoku puzzle, there’s clearly room for growth. The Trident G framework spells out how we might get there, weaving creative leaps with executive rigour and a whole lot of stable context in between.

And isn’t that the essence of human intelligence? Balancing knowledge with flexible reasoning, anchored in good old-fashioned “wait, does that make sense?” If our AI systems can inch closer to that sweet spot, we may see fewer references to ‘walls’ and more to ‘new frontiers.’

(As always, this story is still unfolding—if you spot a novel approach or a puzzle that baffles the best model out there, I’d love to hear about it. After all, the best inspiration often comes from the strangest corners of AI’s self-proclaimed boundaries.)

References

Krishnan, R. (2024, April 23). What can LLMs never do? On goal drift and lower reliability. Or, why can't LLMs play Conway's Game Of Life? Strange Loop Canon.