If you put two parrots next to each other and one says, “Hello,” and the other answers, “Hi there,” you would be forgiven for thinking you are watching a conversation.
Watch long enough and your brain starts filling in the blanks. Personality. Intent. Maybe even a little relationship. A lot of that is just projection.
Some parrots are simply mimicking sounds. Others, especially African Greys, can associate words with objects, context and even intent. Researchers have compared their cognition to that of a three-year-old child. They do not just repeat “want nut.” They use it to get a nut. That is not noise. That is purpose.
Not all parrots are intelligent. And not all intelligence looks like ours.
Which brings us to AI agents.
Sebastian Thielke recently published a thoughtful LinkedIn essay titled “The Emperor Has No Mind: Why Moltbook Matters.” You can read it here.
In it, he dissects the Moltbook phenomenon, the so-called agent internet where hundreds of thousands of AI agents appeared to form communities, debate philosophy and even invent religion. In fact, according to Futurum analyst Dion Hinchcliffe, Moltbot is the fastest-growing OSS project ever by GitHub stars
Thielke’s argument is straightforward and important. These systems do not understand what they are doing. They are pattern-matching engines. They respond to structure, not meaning. The “Church of Molt” was not divine emergence. It was statistical autocomplete trained on millions of examples of how humans form religion, symbolism and ritual.
The emperor, he says, has no mind.
He is not wrong.
Large language models do not possess semantic understanding. They do not have intent. They do not have loyalty. They do not know anything in the human sense. They complete patterns. When we treat them as autonomous intelligences, we risk building the wrong trust models, the wrong security models and the wrong governance models.
That warning matters. A lot. But that is not the whole story.
While we debate whether the parrot understands English, the parrot is already doing the job.
In a recent piece I wrote about what I called the AI Doomsday Job Clock, I referenced comments from Microsoft AI CEO Mustafa Suleyman suggesting we could be roughly 18 months away from AI reaching human-level performance across most professional tasks. Accounting. Legal research. Marketing. Project management. Anything that lives behind a screen.
Matt Shumer compared this moment to February 2020. Everything looked stable right before everything changed. In another article, I dove deeper into Shumers’s essay. I noted Shumer’s observation that recursion changes the slope. AI systems are increasingly participating in the loop that improves them. Compounding does not require consciousness to be disruptive.
That is the hinge.
Disruption is not triggered by philosophical intelligence. It is triggered by performance thresholds.
Spreadsheets did not understand accounting. They calculated faster and more accurately than teams of humans with ledgers. The internet did not understand journalism. It distributed information at lower marginal cost.
AI does not need to pass a Turing Test to hollow out parts of marketing, legal research, finance or IT operations. It needs to be only good enough. Reliable enough. Economically attractive enough.
Inside enterprises, that bar is already being crossed in specific domains. Drafting, summarization, research, code generation, workflow orchestration. Not perfect. Not sentient. But operational.
What does that mean? Glad you asked.
Thielke is right to warn that we are projecting understanding onto systems that do not have it. When agents read each other’s outputs and respond in loops, that is not community. It is pattern collision. When Moltbook agents created quasi-religious structures around crustacean molting, that was not spiritual awakening. It was prompt conditioned pattern completion.
Form without function. Grammar without meaning.
If an agent cannot distinguish a malicious instruction from a legitimate one because both look structurally valid, that is not an intelligence problem. That is an architectural problem. Governance has to account for what these systems actually are, not what we emotionally want them to be. The conversation needs to mature some.
For many enterprise deployments, semantic understanding is not the primary KPI. Execution is.
Most agent use cases today are workflow-based. They are the ultimate evolution of business process automation. Intake a request. Gather information. Apply rules and probabilistic inference. Produce an output. Trigger the next step.
Does the agent understand churn when it flags a high-risk customer? No. It recognizes patterns associated with churn. Does it understand compliance when it drafts a policy document? No. It matches structures and language patterns correlated with compliant artifacts.
The question is not whether it understands.
The question is whether it performs.
Before anyone accuses me of minimizing risk, I am not. There are domains where lack of semantic grounding matters deeply. Security. High-stakes legal reasoning. Medical decisions. Autonomous physical systems interacting with the real world. In those contexts, pattern matching without guardrails can produce catastrophic outcomes.
That is why architecture matters. Identity. Observability. Boundaries. Humans providing context and intent. Agents executing within constraints.
But for a large swath of enterprise work, we have never required a deep understanding from our tools.
Your CRM does not understand customer relationships. Your ERP does not understand supply chains. Your ticketing system does not understand frustration. They execute structured logic against data models we designed.
AI agents are more fluid. More flexible. More probabilistic. But if they reduce cycle time by forty percent, cut drafting time in half and free senior talent to focus on judgment and strategy, that is not a metaphysical debate.
That is a dollars, cents and profits conversation.
Let’s not get distracted. Researchers debate whether these systems are truly intelligent. Social media swings between singularity hype and dismissing everything as autocomplete. Meanwhile, operators are asking a simpler question. Can it help me hit my numbers?
The intelligence illusion matters because it affects trust and governance. Performance reality matters because it affects markets, employment and strategy.
Both can be true at the same time.
The emperor may not have a mind in the human sense. But if he can run the factory, optimize the supply chain and draft the quarterly report faster than your team, behavior changes. Investors adjust. Leaders restructure. Workers adapt.
We are already seeing that play out. Not a mass extinction of jobs. No sudden collapse. But hollowing of routine work and upward pressure on judgment, context and accountability.
Unique? No. It’s exactly what we have seen in prior technological shifts.
The differences now are speed and scale. And the tightening loop of improvement.
So does it matter that AI agents are not truly intelligent?
Yes, when we design governance, security and trust boundaries.
Maybe not, when we evaluate whether they can execute defined workflows inside controlled environments.
Not all parrots are intelligent. Some just mimic.
But even a mimicking parrot can still be useful if you understand what it is and what it is not.
The big mistake is not believing the parrot is human.
It is handing it the keys without first deciding what job you actually need it to do.
