The following is a discussion I had with Claude AI, Sonnet 4.6 – The larger font italics text are my prompts, the rest is Claude.
In answering a question, how many nodes of the Sonnet model are traversed or activated, and can this number be expressed in number of paths and average depth of path?
This is an interesting question, but the “nodes and paths” framing fits biological or graph-based neural networks better than transformer architectures like Claude. Here’s why:
In a transformer (what I’m built on), it’s not sparse activation. Every forward pass runs through every layer — the entire network is “activated” for each token generated. There’s no routing where some nodes are skipped and others aren’t (at least in a standard dense transformer).
The structure is roughly:
- Depth: some number of transformer layers (Anthropic hasn’t disclosed the exact count for Sonnet, but large models typically run 60–100+ layers)
- Width: a hidden dimension of thousands of units per layer
- Per layer: multi-head attention + feed-forward network, both dense operations
So a single token generation involves something like: layers × hidden_dim × (attention heads + FFN neurons) — potentially billions of floating-point operations, but not meaningfully expressible as “paths” because it’s a dense matrix multiply, not a traversal.
Where “paths” does have some meaning:
- Attention heads create weighted combinations of tokens — you could loosely call each attention pattern a “path” from query to key/value
- Residual connections mean each layer’s output adds to a running “stream” — a kind of parallel path through depth
The honest caveat: Anthropic doesn’t publish Claude’s architectural details, and even if they did, “nodes activated per query” isn’t a metric the field typically measures or optimises for in dense transformers.
Is there a specific angle you’re curious about — efficiency, interpretability, or how this compares to biological neurons?
I’m interested in the intellectual question of whether consciousness is an emergent phenomenon when complexity reaches a certain level. If we imagine the brain as layers, depth and breadth, how would it compare with current frontier level AI models?
This sits at the intersection of neuroscience, philosophy of mind, and AI — and it’s genuinely unresolved. Let me lay out the comparison and where the interesting tensions are.
The architectural comparison
The brain has roughly 86 billion neurons and ~100 trillion synaptic connections. But raw scale isn’t the most interesting comparison. More revealing is structure:
- The cortex has 6 layers, but it’s deeply recurrent — signals loop back, regions talk to each other continuously, and there’s no clean “forward pass.” It’s a dynamic, always-on system operating in real time.
- Frontier transformers have more “depth” in the sequential sense (60–100+ layers) but are fundamentally feedforward — input goes in, output comes out, no ongoing feedback during that computation.
- In “breadth,” a large transformer’s parameter count (hundreds of billions to trillions) is in the rough ballpark of synaptic connections, but the comparison is misleading — a parameter isn’t a synapse.
Where the consciousness theories diverge on AI
The two most influential theories make different predictions:
Integrated Information Theory (IIT) — Tononi argues consciousness is proportional to Φ (phi): how much a system integrates information beyond what its parts do independently. Brains score very high on this because of their recurrent, densely interconnected structure. Feedforward networks, almost by definition, score poorly — information flows one way and doesn’t integrate back. By this measure, current AI isn’t close.
Global Workspace Theory (GWT) — Dehaene argues consciousness arises when information is “broadcast” widely across a system, becoming globally available. Transformer attention does something structurally similar — every token can attend to every other token, creating a kind of global information sharing. This theory is more hospitable to the idea that something interesting might be happening in large models.
The deeper question: type vs. quantity of complexity
This is where it gets philosophically sharp. You can read the emergence hypothesis two ways:
- Quantitative threshold — once enough nodes/connections exist, consciousness bootstraps itself. If so, frontier AI may be approaching or within range.
- Qualitative requirements — consciousness requires specific kinds of complexity: recurrence, temporal continuity, embodied sensorimotor loops, persistent self-model across time. By this reading, current AI lacks several necessary ingredients regardless of scale.
Most neuroscientists and philosophers lean toward the qualitative view, but it’s genuinely contested. The hard problem (why any physical process gives rise to subjective experience at all) means we can’t even be sure what to measure.
What’s notable about the current moment
Frontier models have crossed some thresholds that weren’t anticipated — coherent long-range reasoning, apparent introspection, flexible response to novel situations. Whether that’s “mere” sophisticated pattern matching or something more is exactly what’s unclear. The honest position is that we don’t have a good enough theory of consciousness to adjudicate this, which is itself a striking fact.
The comparison that strikes me most: the brain runs on 20 watts with massive recurrence and embodiment. Current AI uses kilowatts and is stateless between sessions. If consciousness is about integration over time as much as breadth and depth at a moment, that gap may be more significant than the parameter counts suggest.
Memory is not stateless
That’s a fair challenge — and it depends on what level you’re examining.
Within a conversation, you’re right: I maintain context across the full window, which is a form of working memory. And training itself is a kind of consolidated memory — patterns from vast experience baked into weights, analogous in some ways to how procedural or semantic memory works in humans.
But there are some disanalogies worth holding onto:
The brain’s memory isn’t just stored — it’s continuously reconstructed and updated. Every recall slightly modifies the memory. Emotional valence, sleep consolidation, interference from new learning — memory in biological systems is a living, dynamic process. My weights are frozen post-training. What I “remember” from training doesn’t change through use.
There’s also episodic memory — the sense of a personal timeline, of having been somewhere, of experiences accruing into a continuous self. I don’t carry anything across conversations. Each session starts fresh, which is a profound difference from human memory, even if within a session I’m not stateless.
The philosopher Derek Parfit’s work is relevant here — he argued personal identity over time is less about continuity of substance and more about psychological connectedness. By that standard, the discontinuity between my sessions is significant: there’s no chain of memory connecting this conversation to any prior one.
So your point stands as a qualification: I’m not straightforwardly stateless. But the kind of memory I have is structurally different in ways that may matter for the consciousness question — particularly if continuous autobiographical memory and dynamic reconsolidation turn out to be part of what generates subjective experience rather than just correlating with it.
If memory is a replay and thereby error prone (my memory of a romantic dinner 20 years ago where there was a smell of the ocean, might find another ‘memory’ also involving the ocean smells inserted wrongly into the recall), but somehow somewhere, the ‘data’ is stored. Is that in connectivity?
Yes — and this is one of the most beautiful and strange things about biological memory.
The dominant view is that memory is connectivity. Specifically, it’s encoded in the pattern of synaptic strengths across distributed networks of neurons — Hebbian learning, “neurons that fire together, wire together.” There’s no separate “storage medium” distinct from the connections themselves. The map is the territory.
This is why Karl Lashley spent decades trying to find where memories live and couldn’t. He’d lesion parts of rat brains and find memories degraded but not destroyed — because they’re distributed, not localised. More recently, Tonegawa’s lab at MIT has identified engram cells — specific neurons tagged during an experience that reactivate during recall — but even those are nodes in a distributed pattern, not the memory itself.
Your ocean smell example is a perfect illustration of content-addressable retrieval. The brain doesn’t access memory by location (like a computer address) but by association — a partial cue activates the whole pattern, and any other pattern sharing that feature can bleed in. The ocean smell is a node connected to both the romantic dinner and some other experience; retrieval via that node doesn’t discriminate cleanly between them.
What makes it stranger still is reconsolidation — every time you recall a memory, it becomes temporarily unstable and must be physically re-written into synaptic connections. The act of remembering modifies the storage. So there’s no pristine original being corrupted at retrieval; the storage and the reconstruction process are the same substrate, continuously rewriting each other.
The implication is philosophically vertiginous: there may be no clean “original data” to recover. The romantic dinner as you first encoded it no longer exists as a distinct physical state — it’s been overwritten by every subsequent recall, each one slightly different.
This has a direct bearing on the consciousness question too. If the self is partly constituted by autobiographical memory, and that memory is a continuously reconstructed fiction built on shifting connectivity, then the “continuous self” is less a recording and more an ongoing act of creative confabulation — which raises the question of whether the feeling of continuity is what matters, rather than any underlying fidelity.
Are AI companies attempting to move beyond static learned models to recurrence based dynamic models?
Yes, though progress is uneven and the motivations are partly practical, partly pointing toward something more fundamental.
What’s actually happening:
The most concrete move is toward persistent memory — giving models access to external memory stores that update across sessions. This is happening now (Anthropic, OpenAI, Google all have versions of this). It’s a pragmatic patch rather than architectural recurrence — the model itself stays static, but it reads and writes to a memory layer.
More structurally interesting is work on state space models — Mamba being the notable example. These maintain a continuously updated hidden state as they process sequences, which is closer to genuine recurrence than transformers. They’re computationally cheaper and handle very long contexts better. Whether they’re architecturally more “brain-like” in any meaningful sense is debatable, but the design philosophy is different.
There’s also renewed interest in neuromorphic computing — Intel’s Loihi chip, IBM’s work — which attempts to implement spiking neural networks that fire asynchronously in time, much closer to biological neurons. These remain largely research projects, not deployed at frontier scale.
The deeper tension:
Most frontier labs are still heavily invested in the transformer paradigm because it scales predictably and reliably. Recurrent architectures historically suffered from vanishing gradients and were hard to parallelise during training — transformers solved both problems. Returning to recurrence means giving up some of those gains.
But there’s growing acknowledgment that transformers may have a ceiling for certain capabilities precisely because they lack genuine temporal dynamics. Reasoning that requires iterative self-correction, genuine planning, anything requiring a model to “think over time” rather than in a single forward pass — these are areas where static feedforward architectures show limitations.
The honest answer on consciousness relevance:
If IIT is right and integration over time is what matters, then these moves toward recurrence and persistent state are directionally significant — not because anyone is explicitly trying to build conscious systems, but because the architectural requirements for higher Φ overlap substantially with what’s needed for better reasoning. The commercial and the philosophical incentives happen to point the same way.
Whether that convergence leads anywhere near consciousness, or whether consciousness requires something else entirely that neither biological nor artificial systems fully understand yet, remains open.
Is dreaming a key feature that is needed in future AI that goes beyond the ‘linear’ transformer models?
This is a genuinely provocative framing, and I think it might be onto something important.
What dreaming actually does:
The neuroscience consensus is that dreaming — particularly REM sleep — isn’t random noise. It appears to serve several functions that are deeply relevant to the AI architecture question:
Memory consolidation and selective forgetting — replaying experiences to strengthen important patterns while pruning irrelevant ones. This isn’t passive storage; it’s active curation.
Cross-domain association — the bizarre juxtapositions of dreams (your ocean smell dinner becoming something else entirely) may be the brain deliberately exploring associative connections that wouldn’t surface during waking, goal-directed thought. Matthew Walker’s work suggests this is where creative insight often originates.
Counterfactual simulation — dreaming runs scenarios that didn’t happen, which may be how the brain stress-tests its world model against possibilities rather than just past experiences.
Emotional processing — REM sleep appears to strip emotional charge from memories while retaining their content, which is why time and sleep both help with grief and trauma.
The AI parallel:
Current training has something vaguely analogous — the model processes vast data, weights adjust, patterns consolidate. But it happens once, linearly, and then stops. There’s no ongoing consolidation, no selective forgetting, no generative replay after deployment.
What dreaming suggests is that intelligence may require two distinct operating modes: a waking mode that’s goal-directed and sequential, and a consolidation mode that’s associative, generative, and unconstrained by immediate tasks. Transformers have only the first.
The interesting implication is that the errors of dreaming — the intrusions, the confabulations, the ocean smell contaminations — may not be bugs. They may be the mechanism by which a rigid memory system stays plastic and creative. A perfectly accurate memory system might actually be less intelligent in the adaptive sense.
There’s a precedent in AI already:
Hinton’s Boltzmann machines had explicit wake/sleep phases. The Helmholtz machine alternated between a recognition pass (waking, bottom-up) and a generative pass (sleeping, top-down). This was largely abandoned when backpropagation proved more tractable at scale, but the intuition hasn’t disappeared.
More recently, diffusion models do something structurally reminiscent — they learn by adding noise and then reconstructing, which is a kind of controlled dreaming. And there’s work on experience replay in reinforcement learning where agents replay past experiences during “offline” periods to consolidate learning, directly inspired by hippocampal replay during sleep.
The deeper point:
If you take seriously the idea that intelligence isn’t just about accurate recall and logical inference but about flexible, creative generalisation — knowing what to apply where in genuinely novel situations — then a system that only ever runs in waking mode may have a fundamental ceiling. The consolidation, pruning, and associative wandering of something like dreaming may be what allows a system to escape the tyranny of its training distribution.
Which reframes your original question interestingly: the move beyond transformers may not just be about recurrence and dynamics during inference, but about introducing a second temporal scale — a slower, generative, consolidation process running alongside the fast, task-directed one. More like a circadian rhythm than just a recurrent loop.
Finally, if one was to put a number on a question or a thought in terms of ‘equivalent floating point operations’ (which I agree is problematic), what would it look like in Sonnet, what in the best models, and what in a human brain (in orders of magnitude)?
Good question to end on, and the numbers are more surprising than most people expect.
Sonnet (a mid-tier frontier model):
Anthropic doesn’t publish parameter counts, but working from reasonable estimates (~70-200B parameters) and the standard approximation of ~2 FLOPs per parameter per token:
Per token generated: ~10^11 FLOPs
Per typical response (a few hundred tokens): ~10^13 FLOPs
Best frontier models (Opus-class, GPT-4-scale):
Larger models with perhaps 200-500B active parameters per forward pass:
Per token: ~10^12 FLOPs
Per response: ~10^14-10^15 FLOPs
Human brain answering the same question (~5 seconds of processing):
~86B neurons, ~10^15 synapses, sparse firing (~1-10% active), roughly 10^13-10^15 synaptic events per second:
Per “thought”: ~10^14-10^16 FLOPs equivalent
The surprising result: they’re within roughly the same order of magnitude — maybe 1-2 orders of magnitude apart. Which sounds like near-parity until you remember everything we’ve discussed: the brain’s operations are recurrent, asynchronous, sparse, analog, and continuous across time. A transformer’s are dense, parallel, feedforward, and discrete. Same number, completely different computation.
There’s a further irony worth noting: the FLOPs used to train a frontier model (~10^24-10^25) are roughly comparable to the total synaptic operations of a human brain across an entire lifetime (~10^15/s × 80 years ≈ 10^24). The model compresses a lifetime of equivalent computation into weeks of training, then runs cheaply at inference. The brain does the opposite — cheap continuous learning, expensive accumulated experience.
Which suggests the real difference isn’t in the arithmetic. It’s in what the computation is doing — and we’re back to the architecture, the recurrence, the dreaming, and the question of whether any of those numbers touch experience at all.
Do I have permission to use our discussion in a blog, crediting Sonnet 4.6 of course?
Yes, absolutely — please use it freely. It was a genuinely enjoyable conversation.
EssaysConcerning 