Controversy about neuro-symbolic AI arose when internal parts of the Anthropic Cloud code were inadvertently leaked.
getty
In today’s column, I examine the recent unintentional leak of various AI components internal to Anthropic, Cloud Code’s widely popular agentic assistant app, which in turn has boosted renewed street cred for proponents of neuro-symbolic AI, though not everyone is sold on what the leaked code does or doesn’t prove.
Back up for a moment to see the bigger picture. Some believe that generic AI and large language models (LLMs) are close to their most viable capabilities. Traditional LLM approaches have gone as far as they could go. An impasse is at hand. An alternative means to advance AI includes hybrid AI that uses not only LLM components, known as subsymbolic AI, but also symbolic AI that is reminiscent of technology used during the expert systems and knowledge-based systems era. It aims to leverage logic-based programming with the use of artificial neural networks (ANNs).
The leaked code of Cloud Code was immediately heralded as a revelation on this complex subject. You see, it turns out that the powerful Cloud Code Agentic AI Assistant consists of a mix of subsymbolic AI and symbolic AI system components. It immediately won praise from those in the hybrid AI camp. If Anthropic is going that route, it certainly confirms the value of combining the semiotic and the symbolic. Others outside that camp were more cautious in drawing any such conclusions and claimed that the hypothesized symbolic AI elements were merely incidental and infallible.
let’s talk about it.
This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, which includes identifying and explaining a variety of impressive AI complexities (see link here).
Agent AI explained
I’ll start by laying a foundation regarding the nature of AI agents. AI agents are the newest area of ​​AI. Prepare yourself because the use of AI agents will become almost ubiquitous in the next year or two. Mark my words.
This is what AI agents are all about.
Imagine you’re using generative AI to plan a vacation trip. You would log in to your Generative AI account traditionally, such as using ChatGPT, GPT-5, GPT-4O, Cloud, Gemini, Llama, Grok, Copilot, etc. Planning your trip will be easier due to the natural language fluency of Generative AI. All you have to do is tell us where you want to go, and then seamlessly engage in a focused conversation about the pros and cons of places to stay and available transportation options.
When it comes to booking your travel, chances are you’ll have to ditch the generative AI and start accessing the websites of hotels, amusement parks, airlines, and other venues to purchase your tickets. Of the leading generator AI available today, relatively few will take the next step on your behalf. It’s up to you to do those micro tasks.
This is where agents and agentic AI come into play.
In earlier days, you would undoubtedly have called a travel agent to make your booking. Although there are still human travel agents, another approach would be to use AI-based agents that are based on Generative AI. The AI ​​has the interactivity you’d expect from generative AI. It is pre-loaded with a series of routines or sets of tasks that underlie a travel agent’s efforts. Using everyday natural language, you interact with Agent AI, who works with you on your planning and can proceed to deal with booking your travel plans.
Agentic AI takes traditional generic AI to a whole new level of use. With agentic AI, the AI ​​acts semi-autonomously rather than simply interacting with you. The AI ​​agent will use one API to connect with an airline and book your flight, and another API to interact with the hotel reservation system and find an available room for you. You don’t need to lift a finger. Overall, action is the paramount basis of AI agents. They can converse, as well as they can act and get things done.
For more information on my discussion of the ins and outs of agentic AI, check out the link here.
Neuro-symbolic AI is another possible future
Switching gears, I’d like to give you the latest on neuro-symbolic AI. Neuro-symbolic AI is a type of two-fer combination, a proverbial two-for-one special combination. You take the popular uses of artificial neural networks (ANNs) that are currently being used at the core of generative AI and LLM, and mix that with rule-based or expert systems (this approach is also known as sub-symbolic AI combined with symbolic AI). The idea is that your goal is to get the best of both worlds. ANNs are primarily data-based methods for introducing AI, while rule-based systems are a logic-based approach.
Many such efforts are already underway; See my discussion on the link here.
Not everyone supports the idea of ​​neuro-symbolic or hybrid AI. A frequent criticism of neuro-symbolic AI is that earlier eras of AI involved rule-based systems – which were later harshly judged as ultimately ineffective or unstable. Critics warn that we should not return to old and now discarded ways of doing things.
A counterargument is that the weaknesses or limitations of rule-based systems can be overcome by incorporating or mixing them with ANNs. Similarly, the limitations of ANNs can be radically transcended by combining them with rule-based systems. The situation is that we should mix both together. This shouldn’t be an all-or-nothing competition.
Thus, instead of throwing away logic-based approaches as an old hackneyed technology, we can still give promising AI approaches a second chance. Of course, some people believe it is reviving something that should have already had a bigger bet made on it. In my view, the synergy of using both capabilities in an integrated manner is very promising. There are strong believers that this is a viable path towards peak AI, such as achieving artificial general intelligence (AGI).
Heated debate about hybrid AI
Within the AI ​​community, there is heated debate about neuro-symbolic AI. Perhaps we are wasting time and effort by exploring neuro-symbolic AI. On the other hand, perhaps we are putting too many eggs in one basket by focusing only on traditional generative AI and LLM. A clear case can be made on both sides of the coin.
There is no doubt that Generic AI and LLM have been quite an attractive form of AI. Billions of dollars have been invested in such AI. The world is well aware of the incredible capabilities of LLM. Furthermore, agentic AI is taking generative AI to a new level of use.
Trying to point to neuro-symbolic AI as a next-generation candidate is challenging because there are not yet outstanding examples that demonstrate the power of hybrid AI. Those in the neuro-symbolist camp always keep an eye out for potential examples that might illustrate the value of hybrid AI approaches.
Well, that day came recently. On March 31, 2026, there was an accidental leak of the source code for some components of Agentic AI, known as cloud code, by Anthropic. Cloud Code is one of the headline-grabbing examples of agentic AI right now. Anyone in the agentic AI space watches cloud code like a hawk, wanting to see what action it can take. Cloud code is a kind of role model.
The source code leak included approximately 500,000 lines of TypeScript spread across approximately 2,000 files. Researchers of all kinds and anyone interested in the inner workings of cloud code pored over the leaked files. They found features that have not yet been activated. He found architectural definitions of how AI was put together. It was like opening a treasure trove of precious gold and gems.
And, within that repository, a file named print.ts contained a series of coding-like logic statements. The list was about a 3,000-line function with about 500 branch points and a dozen levels of nesting. Some insist that this is the smoking gun, providing the requisite proof that symbolic AI is necessary, which should certainly be the case if Cloud Code’s announced agent AI makes use of it.
Heavy encounter between camps
Those in the neuro-symbolist camp were quick to praise Anthropic for its use of symbolic AI techniques. Cloud Code now serves as an excellent and encouraging example for all to see. Without inadvertent leakage, no one except Anthropic insiders would know that a logic-based path was being used. A helpful and encouraging byproduct of the leak was that, finally, there was highly valuable evidence that neuro-symbolism is integral to the progress of AI.
Champagne bottles were opened. Parties were held in the hallways of neuro-symbolic AI researchers. The day when the cloud code was leaked will be recorded in the annals of AI history.
Wow, people on the other side of the dispute responded. At first, logic-based code was cruelly ridiculed as a mess of spaghetti. Is that kind of ethics the cornerstone of the future of AI? No, thanks, came the harsh reply.
Second, this glitch raised fundamental doubts about how the logic-based code was generated. The code may have been written in snippets all at once. It kept expanding and expanding. This implies that there was no coordinated or conscious effort going on. Instead, it was something that was easy to return to, and eventually took on a life of its own. Think of it as a quick solution. Leave aside the notion that this was a cleverly designed route. It was just an afterthought, and nothing more.
Third, and the last nail in the coffin, was that logic-based code dealt primarily with peripheral aspects of the important aspects of AI. Some code does error handling. Some of this does general authentication. Very little, if any, of this code does the AI ​​symbolic work that you might see in a full-rule-based or expert system. In that sense, you cannot reasonably or fairly compare this code to the knowledge-based systems era.
Any such comparison is unfairly equating apples to oranges.
return of the jedi
Not to panic, the neuro-symbolist camp has proposed that code confirms that a logic-based or symbolic approach is needed regardless of what they are doing for an AI system. No matter what kind of mud you throw around, you can’t deny that the code exists, it appears to be an integral part of cloud code, and the developers have chosen to go that route. Whether they did so with great conviction or simply out of convenience, don’t be misled.
Accept reality.
The war of words continues. How will history books portray this situation? It is hard to say. If neuro-symbolic AI takes off and becomes the next great hero of AI, it’s likely that cloud code leaks will get a fair amount of prominence in recounting AI progress. But if neuro-symbolic AI never gets off the ground, the cloud code phenomenon will remain a small footnote somewhere. Very few people will remember the event, and dust will settle on the stories that are now making headlines.
I am hopeful that the former will occur, meaning that neuro-symbolic AI will prove to be the way forward in the best case scenario. Perhaps if the leaked code had contained Darth Vader’s (spoiler alert!) immortal words, “No, I am your father”, it would have brought out the Force’s side in the neuro-symbolic AI camp and given new hope.
Do it or don’t do it. there is no try.
