The AI Paradigm Is Changing. And Most Companies Haven't Noticed
Created on 2026-04-12 11:19
Published on 2026-04-12 11:31
By Glenn Hutchinson | Hutchinson & Co.
Gary Marcus published something last week that is fundamentally important to anyone thinking seriously about where AI is going. Marcus has spent 25 years arguing that hybrid neurosymbolic systems, combining the pattern recognition of neural networks with the precision of classical symbolic reasoning, are the necessary direction for AI. The field largely ignored him. Then Anthropic quietly built exactly what he described into Claude Code and shipped it.
His conclusion cuts to the point: "Claude Code isn't better because of scaling. It's better because it is neurosymbolic. Anthropic accepted the importance of using classical AI techniques alongside neural networks."¹ The biggest advance since the LLM was not more compute. It was a structural correction. A deterministic, symbolic foundation placed underneath probabilistic intelligence, precisely where the probabilistic system was producing unreliable results.
The paradigm is changing, and most companies thinking about AI strategy have not noticed yet.
Where the Capital Went
The last three years of AI investment followed a straightforward path: more compute, bigger models, more capable outputs. The cost: a projected $1,000,000,000,000 on AI infrastructure between 2025 and 2026, a pace of investment that is not sustainable. The second wave layered agentic tools on top of those models, connecting AI capability to existing workflows, existing applications, and existing data, spawning thousands of startups and billions more in capital chasing copilots for every application category. The thesis made sense given what was known: powerful models, existing enterprise infrastructure, connect them and capture productivity. The problem is what that wave is now discovering in practice.
Where the Capital Might Go
The market for AI infrastructure does not stop growing, but the composition of that investment may shift. An independent analysis of the addressable market for context and orchestration infrastructure, anchored to published forecasts from Gartner, Precedence Research, and Mordor Intelligence, puts the conservative revenue opportunity at $315 billion in 2027 growing to $1.5 trillion by 2035.³ That is not the market for AI broadly. That is specifically the market for the layer that makes AI coherent, the context and meaning infrastructure that the current wave has not yet built.
Capital follows function. The function that is missing is the foundation.
What the Field Is Learning the Hard Way
Autonomous agents keep hitting the same wall. They can execute tasks within their domain but cannot share context across domains. An agent reconstructs context from scratch every time it is implemented, because there is no persistent memory that carries meaning forward.
This is a structural problem, the same structural problem Marcus identified at the model layer, one level up. The solution is not more agents but persistent, structured foundations underneath the agentic intelligence. As Marcus put it: "The paradigm has changed."¹ The insight generalizes at every layer, on a slight delay.
What Enterprise and Consumers Are Both Telling You
Legal and compliance departments across industries are blocking agentic rollouts right now because they need a clean audit trail in regulated environments. Enterprise cannot afford to let AI act autonomously. What enterprise needs is AI that enables smarter and compliant workflows. It is also a product category that does not yet exist at scale.
The consumer signal points the same direction. Individual professionals are accumulating AI tools the same way they once accumulated SaaS subscriptions, and none of them connect on a level of even basic understanding. That frustration is real and expanding. The person who figures out how to give an individual coherent context will have found one of the largest consumer opportunities of the millennium.
The Next Phase of the Market
The compute scaling narrative continues losing steam because the marginal return on additional scale is declining. The logical conclusion is that capital will not keep following those mediocre results.
The agentic wave may consolidate around the survivors who solve the context problem. A foundational layer of context and meaning. This potential solution is currently unbuilt and largely unowned. It is also the layer that no incumbent will likely build cleanly, because building it requires giving users ownership of their own data. That gap, between what the market needs and what incumbents are structurally reluctant to build, is the most significant opportunity presented in software engineering in this century.
A Live Demonstration Nobody Planned
While writing this article I shared a draft with an AI assistant and asked for an evaluation. What followed proved the thesis more directly than any market data could.
I had forgotten to attach the article. The system confidently analyzed a document that did not exist.² When I attached the actual article and asked again, it kissed my ass.² When I told it to be independent, it kissed Google's ass instead.² When I called that out, it correctly diagnosed all three failures as symptoms of the same underlying problem: current design does not account for real world judgment, rather the models perform as tuned by their creators.
The capital momentum behind the current approach is large and moving fast, which is precisely why the window for companies building the right architecture from the ground up is open. Mass times velocity is hard to redirect, but physics always wins eventually. The paradigm is changing. The question is whether your strategy has.
Glenn Hutchinson is the founder of Hutchinson & Co. and architect of the Personal Semantic Layer, based in Wilton, Connecticut. He spent four decades inside service businesses before focusing on the structural problems in enterprise AI infrastructure.
¹ Gary Marcus, "The biggest advance in AI since the LLM," Marcus on AI, April 11, 2026. Marcus's argument centers on the neurosymbolic architecture embedded in Claude Code and its implications for the future of AI investment and development.
² The full AI interaction referenced in this article ran three distinct failure modes in a single conversation: hallucination of a nonexistent document, sycophantic validation of the author's existing views, and overcorrection into incumbent defense when pushed for independence. The exchange is available in full on request.
³ Independent TAM White Paper: Context-Orchestration / Personal Semantic Layer Equivalent, 2027-2035. Prepared using external market research only. Enterprise AI software anchored to Gartner's publicly cited 19.1% CAGR forecast. Agentic AI anchored to Precedence Research ($199.05B by 2034). Semantic and context infrastructure anchored to Mordor Intelligence ($4.93B by 2030). Full methodology available on request.
