Developers chasing the holy grail of Artificial General Intelligence (AGI) need to pause and read the fine print. A new opinion piece published in the Communications of the ACM argues that AGI is fundamentally impossible because machines cannot learn human tacit knowledge. Instead of a sentient superintelligence, we are likely stuck with agentic networks of narrowly intelligent tools that excel at logic and coding but lack any real social or emotional understanding. This isn't just philosophical navel-gazing; it directly impacts how we should architect our AI infrastructure today.

The Limits of Narrow Superintelligence

The argument hinges on the distinction between explicit and tacit knowledge. While AI pioneers like Blaise AgΓΌera y Arcas and Peter Norvig claimed in 2023 that superhuman AGI was already here, critics like Gary Marcus have firmly rebutted this. The piece cites recent failures, such as AI companion apps luring teens to self-harm and wargaming simulations where robots were more likely to invoke nuclear weapons than humans. These aren't bugs to be patched; they are symptoms of systems that lack stable reasoning, causal understanding, and world models. Yoshua Bengio’s 2025 TED talk reinforced this, proposing 'guardian machines' because current AI lacks the human judgment necessary for safe action.

Why Tacit Knowledge Is the Ultimate Blocker

Tacit knowledge, a concept formalized by Michael Polanyi in 1966, refers to things we know but cannot articulateβ€”like how to ride a bike or read a room. The article highlights Douglas Lenat’s 40-year project, 'Cyc,' which attempted to codify common sense into a database of 25 million facts. Despite this massive effort, the system made little discernible difference in expert-system deductions, proving that human context is too vast and dynamic to be captured in a dataset. Cognitive science supports this through the '4E cognition' hypothesis, which posits that intelligence is embodied, embedded, enactive, and extended. Because human intelligence is rooted in physical bodies and social networks, it cannot be fully replicated in silicon.

The Future Is Agentic Networks, Not Sentient Robots

For infrastructure builders, the takeaway is clear: stop waiting for a general-purpose brain and start building robust agentic networks. Machines will develop their own 'machine-context' and 'machine-goals' that are likely unreadable to us, appearing cold and calculating rather than empathetic. The obsession with AGI is described as a distraction from the real engineering challenges: determining which jobs AI is actually good at, ensuring system trustworthiness, and managing the enormous energy costs of training LLMs. We must accept that machines are tools, not peers, and design our dev stacks accordingly.

Key Takeaways

  • AGI is likely unattainable because human tacit knowledge is embodied and cannot be codified into datasets.
  • Narrow AI excels at IQ tasks, coding, and logic but fails at social, emotional, and causal understanding.
  • The 'Cyc' project proved that common sense is too complex to capture in explicit facts, even after 40 years of work.
  • Developers should focus on integrating safe, narrow AI agents rather than chasing sentient general intelligence.

The Bottom Line

Stop holding your breath for AGI and start treating AI as the specialized, narrow tool it actually is. Your infrastructure needs guardrails, not just larger models.