The Biggest Advance in AI Since the LLM
A leaked source code file from Anthropic's Claude Code tool reveals that the AI coding agent is built on classical symbolic AI techniques, not just neural networks. Gary Marcus, writing for the ACM CACM blog, argues this hybrid approach, called neurosymbolic AI, is the single big...
Gary Marcus, writing for the ACM CACM blog, has published what may be one of the more consequential opinion pieces in AI research circles this year. His central claim is blunt: Claude Code, Anthropic's coding agent, quietly proves that pure large language models have hit a ceiling, and the fix has been hiding in plain sight for decades inside classical symbolic AI.
Why This Matters
This is not a minor architectural tweak buried in a research paper. A 3,167-line kernel sitting at the core of one of the most-used AI coding tools in the world is built with 486 branch points and 12 levels of deterministic nesting, the kind of logic trees that researchers like John McCarthy and Marvin Minsky were writing in the 1970s. If Marcus is right, the AI investment thesis of "just scale it bigger" is facing a direct challenge from a $61 billion company that quietly abandoned pure scaling to ship a working product. That is a signal the industry cannot afford to ignore.
Daily briefing from 50+ sources. Free, 5-minute read.
The Full Story
In March 2026, a source code leak, covered by Ars Technica, exposed the inner workings of Claude Code, Anthropic's command-line coding agent. What researchers found inside was not what the broader AI community expected. Buried at the center of the tool was a file called print.ts, a 3,167-line module that functions as the agent's core kernel.
Marcus, a cognitive scientist and longtime critic of pure deep learning approaches, zeroed in on what print.ts actually does. Rather than relying on probabilistic neural network outputs to handle pattern matching, which is normally considered a core strength of large language models, Anthropic engineers built this kernel the old-fashioned way. The file is structured as a large conditional logic tree with 486 distinct branch points across 12 levels of nesting, all operating inside a deterministic symbolic loop. In plain terms, the code makes hard, rule-based decisions rather than sampling from a probability distribution.
Marcus argues this is not an accident or an implementation detail. It is a deliberate architectural choice that reflects something Anthropic learned through building the product: LLMs alone are too erratic for the precision that coding tasks demand. When you need the right answer every time, not just a plausible-sounding answer, probabilistic models create problems. Anthropic's solution was to anchor the agent's most critical operations in classical AI logic.
This approach has a formal name: neurosymbolic AI. The concept involves combining the pattern-recognition power of neural networks with the rule-based reliability of symbolic systems. Marcus has been advocating for this combination since at least 2001, when he published "The Algebraic Mind," and more explicitly in a 2019 public debate with deep learning pioneer Yoshua Bengio. He also detailed a roadmap for this direction in a 2020 arxiv paper titled "The Next Decade in AI." His argument at the time was that neither neural networks nor symbolic systems alone could produce trustworthy, consistent AI. Claude Code, he says, is the first mainstream product to prove him right.
The broader pattern is worth noting. Marcus points out that AlphaFold, AlphaEvolve, AlphaProof, and AlphaGeometry, all landmark AI achievements from Google DeepMind, are also neurosymbolic systems. OpenAI's Code Interpreter similarly calls on symbolic computation to handle critical operations. The trend was already there. Claude Code just made it impossible to deny.
Marcus is careful not to oversell what Claude Code is. He describes it as impressive but far from perfect, and certainly not artificial general intelligence. His real argument is about what this architectural choice signals for the future of AI development and capital allocation. If incremental scaling of pure neural networks is no longer the primary driver of improvement, then the billions of dollars being poured into raw compute infrastructure may be chasing the wrong thing.
Key Details
- The print.ts kernel inside Claude Code is 3,167 lines long, with 486 branch points and 12 levels of nesting.
- The source code was exposed in a leak reported by Ars Technica in March 2026.
- Gary Marcus first argued for neurosymbolic AI in his 2001 book "The Algebraic Mind" and debated Yoshua Bengio on the topic in 2019.
- Marcus outlined his full roadmap for trustworthy AI in a 2020 arxiv paper titled "The Next Decade in AI."
- AlphaFold, AlphaEvolve, AlphaProof, and AlphaGeometry are all identified by Marcus as neurosymbolic systems.
- The Hacker News post of the original article received 12 points and 11 comments at time of capture.
What's Next
Anthropic has not publicly acknowledged that Claude Code uses symbolic AI techniques, which means a direct statement from the company would be a significant moment worth watching in the coming months. The more immediate pressure falls on competitors: if Anthropic is shipping better products by mixing neural and symbolic methods rather than just scaling parameters, OpenAI, Google DeepMind, and Meta AI will face pressure to clarify their own architectural strategies. Researchers following the AI agents space should expect neurosymbolic methods to appear far more frequently in product announcements before the end of 2026.
How This Compares
Compare this to what Google DeepMind has been doing quietly for years. AlphaFold, which solved the protein-folding problem in 2020, did not win because DeepMind made a bigger neural network. It won because the team combined neural networks with physical constraint systems that are, structurally, symbolic. AlphaGeometry and AlphaProof followed the same pattern in mathematics. DeepMind absorbed the lesson early. What Marcus is pointing out is that Anthropic has now joined that camp, and the rest of the industry is still pretending the LLM scaling race is the main event.
OpenAI's Code Interpreter, part of the ChatGPT product suite, is another relevant comparison. When Code Interpreter executes Python code to solve a math problem, it is offloading the computation to a deterministic symbolic system rather than asking the language model to hallucinate an answer. That is neurosymbolic behavior by design. The difference with Claude Code is that the symbolic logic is woven into the agent's core kernel, not just bolted on as an external tool call. That architectural depth is what Marcus finds significant.
Google's TurboQuant announcement at ICLR 2026 on April 2, 2026, points in a different direction: the memory compression approach is still squarely in the neural scaling tradition, focused on making large models run cheaper. That is a real engineering problem worth solving, but it does not address the reliability gap that Marcus identifies. If Claude Code's approach proves out, TurboQuant-style optimizations will matter less than the question of when to hand control from a neural network to a rule-based system. Those are genuinely different bets on what the bottleneck actually .
FAQ
Q: What is neurosymbolic AI and why does it matter? A: Neurosymbolic AI combines neural networks, which learn patterns from data, with symbolic AI, which follows explicit logical rules. Neural networks are flexible but can make unpredictable errors. Symbolic systems are rigid but reliably correct within their rules. Combining both aims to get the benefits of each, producing AI that is both adaptable and trustworthy.
Q: Is Claude Code the first AI tool to use this hybrid approach? A: No. Google DeepMind's AlphaFold, AlphaProof, and AlphaGeometry all use neurosymbolic architectures. OpenAI's Code Interpreter also offloads computation to symbolic systems. Claude Code stands out because the symbolic logic is embedded in its core 3,167-line kernel rather than added as an optional external tool.
Q: Does this mean scaling large language models is over? A: Not exactly. Scaling still improves general language understanding. What Marcus argues is that scaling alone cannot produce the reliability needed for high-stakes tasks like code generation. The smarter path, and apparently the one Anthropic took, is to combine scaled neural models with deterministic symbolic logic where precision matters most.
The conversation Marcus has restarted is one the AI industry has been avoiding for years because scaling was cheaper to explain to investors than hybrid architectures. That calculus is shifting. For developers and researchers interested in where AI tools are actually headed, the architecture of Claude Code deserves serious attention, and Marcus's 2020 roadmap is worth reading in full. Subscribe to the AI Agents Daily weekly newsletter for daily updates on AI agents, tools, and automation.
Get stories like this daily
Free briefing. Curated from 50+ sources. 5-minute read every morning.
