The Next Architecture of Intelligence
AI’s first industrial revolution was powered by scale. Its next one may be powered by the cortex.
The modern AI boom began with a humbling discovery: intelligence can emerge from arithmetic at terrifying scale.
Behind the poetry, code, images, and conversation produced by large language models sits a machine doing enormous amounts of linear algebra. Words become high-dimensional vectors. Images become tensors. Those vectors flow through layers of weight matrices. Training adjusts billions or trillions of numbers until the system learns a statistical geometry of language, images, and behavior.
That description strips AI of mysticism without making it less extraordinary. A large model is a gigantic learned transformation. It takes an input point in a high-dimensional space, repeatedly stretches, rotates, compresses, and reprojects it through matrices, then emits the most likely next token, pixel, action, or representation. The miracle is that when the matrices become large enough, the training data broad enough, and the optimization stable enough, this statistical machine begins to look like reasoning.
The question now is whether this architecture is the final form of machine intelligence or simply the first scalable platform that worked.
The answer matters far beyond academic computer science. Transformers turned AI into a compute market, a data-center market, a chip market, a cloud market, and a new operating layer for enterprise software. If the next architecture changes the cost, memory, and learning profile of AI, it could redraw the entire value chain again.
Before the Transformer, AI Remembered One Step at a Time
Before 2017, AI had already gone through several eras.
The early symbolic systems tried to encode intelligence as rules. They worked when the world was narrow and formal: chess positions, expert systems, hand-written decision trees. They broke when reality became messy. The world contains ambiguity, context, common sense, and exceptions. Hard-coding all of that turned out to be impossible.
Neural networks offered a different path. Instead of writing rules by hand, engineers would build a structure capable of learning from examples. A model would start with random weights, make predictions, compare them with correct answers, and use gradient descent to adjust its internal parameters. Intelligence became less like a library of rules and more like a landscape of weights.
Convolutional neural networks became the breakthrough architecture for images. Their key insight was local pattern recognition. A CNN could learn edges, textures, shapes, and eventually objects by applying small filters across an image. That structure matched vision well because nearby pixels are meaningfully related. Images have spatial locality.
Language created a harder problem. Words arrive in sequence. Meaning depends on order, memory, and context. The sentence “the animal didn’t cross the street because it was too tired” requires the model to connect “it” with “animal.” Change one phrase and the reference changes. Sequence models had to carry information across time.
Recurrent neural networks, and later LSTMs and GRUs, became the standard solution. They processed text token by token, passing a hidden state forward like a rolling memory. LSTMs improved the system by adding gates that decided what to keep, what to forget, and what to expose. This helped with longer dependencies and made machine translation, speech recognition, and text modeling more useful.
Yet the architecture had a bottleneck. The model still moved through the sentence sequentially. Token one influenced token two, token two influenced token three, and so on. Training could not fully exploit parallel hardware because the computation depended on time order. Long-range memory also remained fragile. Important information could fade as the sequence grew.
Attention existed before the Transformer as an add-on to encoder-decoder systems. It allowed a decoder to look back at different parts of the input sequence instead of compressing everything into a single hidden state. This was powerful. The model could translate a word by paying attention to the most relevant source words. But attention was still attached to recurrent or convolutional machinery.
Then came the 2017 paper that changed the trajectory of AI.
“Attention Is All You Need” Was an Architecture Shock
In June 2017, Ashish Vaswani and colleagues released Attention Is All You Need. The title sounded almost too simple. The claim was radical: sequence transduction could be built entirely on attention, dispensing with recurrence and convolution.
The paper introduced the Transformer.
Its breakthrough was architectural, economic, and philosophical at the same time. Architecturally, it replaced step-by-step recurrence with self-attention. Economically, it made sequence modeling far more parallelizable on GPUs and later specialized AI accelerators. Philosophically, it suggested that the right general mechanism might scale across tasks with fewer hand-built assumptions.
A Transformer begins by converting tokens into embeddings: vectors in a high-dimensional space. It also adds positional information because a pure attention mechanism needs a way to know order.
Then each token embedding is projected into three different vectors: Query, Key, and Value. These are produced by multiplying the input by learned weight matrices usually called WQ, WK, and WV.
A Query asks: what am I looking for?
A Key answers: what information do I contain?
A Value carries: what content should I contribute if another token attends to me?
The model compares Queries and Keys across tokens, usually through dot products. If one token’s Query aligns strongly with another token’s Key, the attention score rises. After normalization through softmax, those scores become weights. The model then computes a weighted mixture of the Value vectors.
In plain language, every token can look at every other token and decide how much each one matters.
This creates a flexible graph of relationships inside the sentence. In “the horse ran because it was frightened,” the token “it” can attend strongly to “horse.” In code, a variable can attend to its definition. In a legal document, a clause can attend to a definition many paragraphs earlier. The model learns these relationships through data rather than receiving hand-coded grammar rules.
Multi-head attention makes the system richer. Instead of one attention pattern, the Transformer runs several attention heads in parallel. One head might track syntax. Another might track references. Another might track positional patterns. Their outputs are combined and passed through feed-forward networks. Residual connections and layer normalization stabilize the repeated transformations.
Stack enough of these layers, train them on enough data, and the result becomes a general pattern engine. It can translate, summarize, write code, answer questions, generate images when fused with diffusion systems, and control tools when wrapped in an agent loop.
The Transformer became the AI world’s first truly general-purpose architecture. It was the x86 moment for deep learning: a common computational substrate that could scale across modalities and products.
The Transformer’s Power Came From Scale
The success of the Transformer validated Rich Sutton’s “bitter lesson”: in AI, general methods that leverage computation tend to beat systems filled with human-designed knowledge. The Transformer fit that lesson perfectly. It did not require engineers to encode grammar, facts, logic, or world models by hand. It needed data, compute, optimization, and scale.
This created the modern AI industrial stack.
NVIDIA GPUs became the engines because they excel at matrix multiplication. Hyperscale cloud providers became the factories because training and serving frontier models requires huge clusters, high-speed networking, power infrastructure, cooling, and software orchestration. Data became a strategic asset because models improve when they consume more diverse, higher-quality examples. Enterprise software became the distribution layer because language models could be embedded into workflows.
The Transformer’s potential remains enormous.
It is a universal interface for language. It can compress expertise into an interactive assistant. It can turn unstructured data into structured outputs. It can write code, draft memos, inspect contracts, generate synthetic media, orchestrate tools, and make software more conversational. When connected to retrieval systems, APIs, memory, and user-specific context, it becomes a new work layer above existing applications.
But the same architecture that created the boom also created the bottleneck.
The Shortcomings Are Built Into the Machine
The first limitation is compute.
Self-attention compares tokens with other tokens. In the standard form, the attention matrix grows quadratically with sequence length. Double the context and the attention computation grows much faster than double. Engineers have invented many optimizations, but the basic pressure remains: longer context, larger models, and more users require vast memory bandwidth, power, and serving infrastructure.
The second limitation is data hunger.
A human child learns language from a tiny fraction of the text used to train modern language models. A frontier model often needs enormous pretraining corpora, expensive reinforcement learning or preference tuning, and continuous post-training to become useful. The system appears intelligent after it has absorbed an extraordinary amount of human-generated signal.
The third limitation is static learning.
After pretraining, a model’s core weights are mostly fixed. It can use context. It can retrieve documents. It can store external memories. But updating the model’s internal knowledge safely, continuously, and efficiently remains difficult. Naive continual learning can cause catastrophic forgetting, where new training degrades old capabilities. The result is a strange kind of intelligence: highly capable inside a context window, yet still brittle as a lifelong learner.
The fourth limitation is grounding.
Transformers learn statistical structure. They can infer patterns, simulate reasoning, and generate fluent explanations. They also hallucinate because fluency and truth are different objectives. Retrieval, tool use, verification loops, and agent frameworks help, but the base model still lacks the kind of grounded sensorimotor learning that biological systems use from infancy.
The fifth limitation is energy.
The human brain runs on roughly the power draw of a dim light bulb. Training and serving today’s AI systems require chips, clusters, and data centers at an entirely different scale. The gap is more than an engineering annoyance. It shapes margins, deployment models, geopolitical infrastructure, and the ability to put advanced intelligence on edge devices.
These limits do not mean Transformers are a dead end. They mean the current architecture may represent a powerful local optimum: the best architecture we have found for scaling matrix-based pattern learning on modern hardware, while leaving open the possibility that nature already discovered a more efficient algorithm.
That is where cortical columns enter the story.
Cortical Columns: The Brain’s Repeating Microcircuit
The neocortex is the folded outer sheet of the brain associated with perception, planning, language, and abstract thought. A striking feature of the neocortex is its repeated structure. Across different regions, neurons are organized into layers and column-like circuits. Vernon Mountcastle’s work helped make the cortical column one of the central concepts in neuroscience.
A cortical column can be understood as a small vertical processing unit running through the layers of the cortex. It contains groups of neurons connected across layers, plus lateral links to neighboring columns. The exact definition and function remain debated, but the broad idea is powerful: the brain may use many variations of a repeated computational motif.
Jeff Hawkins, Subutai Ahmad, and Yuwei Cui proposed that cortical columns can learn models of the world by combining sensory input with location signals. In their theory, a single column detects a feature together with its location relative to an object. Through movement and repeated sensing, it builds a model of the object. Multiple columns then collaborate through lateral connections, each contributing partial knowledge until the system reaches a stable interpretation.
This is a very different computational philosophy from today’s language models.
A Transformer learns by absorbing a gigantic corpus and compressing statistical relationships into weights. A cortical-column-inspired system suggests a more embodied style of learning: observe part of the world, know where that part sits relative to an object, move or receive new input, update the model, and let neighboring units reach consensus.
The key words are location, movement, partial evidence, consensus, and continuous updating.
That sounds much closer to how humans learn. We do not need to see every chair ever made to understand chairs. We touch, look, move, compare, and build stable object models from sparse experience. We generalize from a handful of examples because the brain appears to carry strong architectural priors about objects, space, causality, and temporal continuity.
If artificial systems could borrow even a small part of that efficiency, the economics of AI would change.
Why Flourish Might Be the Next Breakthrough
Flourish is interesting because it is making the architectural bet explicit.
According to WIRED, Flourish describes itself as a neuro-AI company trying to solve two of AI’s hardest problems: power efficiency and continuous learning. The company is building a system it calls Cortex AI, designed to match more of the computational capacity, learning efficiency, and power budget of the human brain. Thomas Reardon’s stated ambition is a synthetic AI brain running at 50 watts or less.
The company reportedly raised $500 million at a $2.5 billion valuation, with funding from Jeff Bezos and others. Its team includes neuroscientists and AI researchers working side by side. The focus, according to the WIRED reporting, includes cortical columns, connectomics, hippocampus-inspired memory, and models that can learn continuously with far less training data.
The reason this matters is that the Transformer era has become a resource race. More parameters, more tokens, more GPUs, more power, more data centers. That race has produced astonishing progress, but it also creates an opening for any architecture that can produce useful intelligence with lower energy, lower data requirements, and better continual adaptation.
Flourish is trying to attack the constraints at the root.
If Cortex AI can learn continuously, it would address the static-weight problem. Instead of retraining a model in massive batches, the system could update from ongoing experience. If it can use hippocampus-like memory, it could separate fast episodic learning from slower structural learning, closer to how biological memory appears to work. If it can borrow principles from cortical columns, it could improve grounding, object modeling, and abstraction from sparse data. If it can run on tens of watts, it could move powerful AI from data centers into local devices, robots, wearables, vehicles, and industrial systems.
That combination would directly challenge the economic assumptions of the current stack.
A 50-watt AI system would change deployment. Continuous learning would change personalization. Brain-inspired memory would change agents. Efficient object and world modeling would change robotics. A new architecture that reduces dependence on frontier-scale training clusters would shift value from raw compute accumulation toward algorithmic design, neuroscience-derived architectures, and specialized silicon.
This is the bull case for Flourish: it is chasing the part of intelligence Transformers approximate poorly.
How Cortex AI Claims to Solve the Transformer Problems
Cortex AI should be understood as an ambition rather than a proven public product. Flourish has not disclosed enough technical detail to evaluate the architecture rigorously. The right framing is conditional: if the company can translate neuroscience into working silicon and software, the approach could solve several pain points in today’s AI.
First, it targets energy efficiency.
Transformer models burn power because they perform enormous matrix operations across huge parameter sets and context windows. A cortex-inspired architecture would aim to use sparse, modular, event-driven, or locality-aware computation. Biological brains do not activate every neuron at full intensity for every thought. They route activity through specialized circuits. A successful Cortex AI system would likely need similar selectivity: activate the relevant computational columns, memory systems, and pathways rather than the whole machine for every query.
Second, it targets continual learning.
Transformers can adapt through context and external memory, but their internal learning mostly happens during expensive training runs. Cortex AI appears to be aimed at systems that keep learning after deployment. That matters for agents. A useful agent should become better at a user’s workflow over time, remember the consequences of actions, update procedures, and refine its model of the environment without requiring a full retraining pipeline.
Third, it targets sample efficiency.
The human brain learns from sparse, noisy, embodied data. Cortical-column-inspired learning could give AI better priors about objects, space, and causality. Instead of learning every relationship from internet-scale text, a model could infer more from fewer examples by using a structured architecture closer to how perception and memory work in living systems.
Fourth, it targets grounding.
A language model can describe a coffee mug. A grounded system should understand that the mug has a handle, occupies space, can be grasped, can contain liquid, can fall, can break, and looks different from different angles while remaining the same object. Cortical columns, with their emphasis on features at locations and consensus across partial views, point toward this kind of stable object modeling.
Fifth, it targets edge intelligence.
If advanced models can run on laptop-level power, the market expands. AI can leave the data center and become embedded in devices that learn locally. This matters for privacy, latency, cost, robotics, defense, healthcare devices, and consumer hardware. It also changes who captures value. The winners may be companies that own efficient architectures, specialized chips, embedded operating systems, and high-quality local learning loops.
The Skeptical Case Still Matters
Flourish is a bold bet, and bold bets often fail.
Neuroscience has inspired AI many times. Neuromorphic computing has had cycles of excitement. The brain is poorly understood compared with the precision required to build reliable commercial systems. Cortical columns themselves remain scientifically debated. A useful analogy can mislead if engineers copy biology at the wrong level of abstraction.
The Transformer also has a major advantage: it works. It scales. It has an ecosystem. It runs on existing hardware. It has huge developer mindshare, open-source momentum, cloud support, tooling, benchmarks, and business demand. Any challenger architecture must beat a moving target. Transformers are improving in efficiency, context length, multimodality, reasoning scaffolds, retrieval, tool use, and inference cost.
This means Cortex AI does not need to replace Transformers everywhere to matter. A narrower breakthrough would still be valuable. A memory system that improves agents, a low-power model for edge devices, a continual-learning module, or a more efficient world-modeling architecture could become a major component in hybrid systems.
The future may look less like one architecture replacing another and more like a stack: Transformers for language and broad knowledge, retrieval systems for factual grounding, tool-using agents for execution, and cortex-inspired modules for memory, adaptation, perception, and embodied learning.
The Investment Narrative: From Scaling Compute to Scaling Learning
The first AI trade was scale.
More GPUs. More data centers. More tokens. More parameters. More electricity. This trade has obvious winners: NVIDIA, advanced packaging providers, memory suppliers, networking vendors, hyperscale cloud platforms, and companies that can monetize AI copilots across large installed bases.
The next AI trade may be learning efficiency.
If the industry remains on the current path, intelligence keeps getting better but also more capital-intensive. That favors the richest labs and largest cloud platforms. If a new architecture reduces the amount of compute and data required to learn, the economics shift. The scarce asset becomes less about owning the biggest cluster and more about owning the best learning architecture.
This is why Flourish deserves attention. Its bet sits directly at the fault line of the current AI economy. Transformers proved that intelligence can be scaled through computation. Cortex AI is betting that intelligence can be made more efficient by rediscovering some of the brain’s algorithms.
The story of AI evolution therefore has three acts.
The first act was rules. Humans tried to write intelligence directly.
The second act was weights. Machines learned statistical structure through data and gradient descent.
The third act may be architecture. The industry may discover that the next leap requires better computational primitives: memory systems that update continuously, world models that learn from sparse evidence, and modular circuits that achieve more with less energy.
Flourish may fail. Cortex AI may turn out to be too vague, too biological, too early, or too difficult to commercialize. But the question it is asking is exactly the right one: if the brain can learn continuously, generalize from limited data, ground concepts in the world, and operate on about 20 watts, why should artificial intelligence require city-scale infrastructure to approximate a fragment of that ability?
That question will not go away.
The Transformer gave AI its industrial revolution. Cortex-inspired systems are searching for its biological revolution. The next breakthrough may come from the point where those two histories finally meet: the brute-force lesson of scale, and the quiet efficiency of the cortex.
Risk Note
This is a technology and market-structure analysis, not investment advice. Flourish and Cortex AI remain early and technically unproven based on public information. Transformer-based systems continue to improve quickly, and the incumbent compute ecosystem may remain dominant even if brain-inspired modules become useful. Any investment conclusion should be tested against product evidence, customer adoption, technical disclosures, and competitive response.