The Year of the AI Agent That Never Was: Why Your Robot Overlord Is Still a No-Show

If 2025 was supposed to be “the year of the AI agents,” someone forgot to tell reality. While tech giants painted visions of generative AI robots taking over our tasks and running the world, the actual transformation got kicked down the road to 2026—or maybe never.

But what if “never” isn’t just a pessimistic guess? What if mathematics itself says your AI assistant will never be reliable enough to handle anything truly important?

The Mathematical Death Blow to Agentic AI Dreams

In the middle of all the agentic AI hype, a paper appeared with the ominous title “Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models.” Written by Vishal Sikka—former SAP CTO and AI researcher who studied under John McCarthy, one of AI’s founding fathers—and his teenage prodigy son, the paper delivers a mathematical gut punch to the AI agent revolution.

Their conclusion? Large language models are “incapable of carrying out computational and agentic tasks beyond a certain complexity.” Even the newer reasoning models that go beyond simple word prediction won’t fix this fundamental limitation.

“There is no way they can be reliable,” Sikka tells me bluntly. After careers at SAP, Infosys, and Oracle, he now runs Vianai, an AI startup. When I ask if we should forget about AI agents running nuclear power plants, he responds with one word: “Exactly.”

Maybe your AI can file some papers and save you time, but you might have to accept that it’ll make mistakes. Sometimes serious ones.

The Industry Pushes Back—With Math of Its Own

The AI industry isn’t buying this doomsday scenario. After all, AI coding has been a genuine breakthrough, taking off significantly last year. At Davos this week, Google’s Nobel-winning AI head Demis Hassabis reported breakthroughs in minimizing hallucinations, and both tech giants and startups continue pushing the agent narrative.

Enter Harmonic, a startup co-founded by Vlad Tenev (yes, the Robinhood CEO) and Stanford mathematician Tudor Achim. They’re claiming their own mathematical breakthrough with a product called Aristotle that tops benchmarks on reliability.

Harmonic’s approach uses formal mathematical reasoning to verify AI outputs, encoding them in Lean programming language—known for its ability to verify code correctness. Their mission? “Mathematical superintelligence,” starting with coding as a natural extension.

“I would say that most models at this point have the level of pure intelligence required to reason through booking a travel itinerary,” Achim argues, suggesting reliable agentic behavior might not be as impossible as critics claim.

The Uncomfortable Middle Ground

Both perspectives contain uncomfortable truths. Everyone agrees hallucinations remain a persistent problem. In a September paper, OpenAI scientists wrote, “Despite significant progress, hallucinations continue to plague the field, and are still present in the latest models.”

They proved this by asking three models—including ChatGPT—to provide the title of a lead author’s dissertation. All three made up fake titles and misreported the publication year. In their blog about the paper, OpenAI stated grimly that in AI models, “accuracy will never reach 100 percent.”

The truth likely sits somewhere between utopian promises and mathematical pessimism. AI will continue improving at specific, bounded tasks. It might handle your travel bookings or basic coding with increasing reliability. But the dream of truly autonomous AI agents running complex systems—managing power plants, making high-stakes decisions, or genuinely replacing human judgment in critical domains—faces fundamental mathematical barriers that no amount of scaling or engineering may overcome.

Your robot overlord isn’t coming. At least not according to the math.

Tags: AI agents, transformer models, hallucinations, mathematical limitations, formal verification, coding AI, agentic AI, reliability, Vishal Sikka, Vlad Tenev, Tudor Achim, Lean programming, mathematical superintelligence, OpenAI, ChatGPT, Demis Hassabis, Vianai, Harmonic, Aristotle

Viral Sentences:

• “How about never?” might be the real answer to when AI takes over
• Mathematics just punched a hole in the AI agent revolution
• Your robot overlord isn’t coming—at least not according to the math
• The year of AI agents turned into the year of talking about AI agents
• Even reasoning models can’t fix what’s fundamentally broken
• Forget AI agents running nuclear power plants
• Accuracy in AI models will “never reach 100 percent”
• Three different AI models hallucinated fake dissertation titles
• The uncomfortable truth sitting between utopian promises and mathematical pessimism
• Formal methods of mathematical reasoning might be AI’s saving grace
• “Mathematical superintelligence” starts with verifiable coding
• The industry pushes back with math of its own
• Booking travel itineraries might be AI’s reliability ceiling
• The big AI companies promised us the moon, delivered a calculator
• Transformer-based language models have basic limitations that can’t be engineered away
• The teenage prodigy who helped mathematically kill the AI agent dream
• SAP’s former CTO says reliable AI agents are mathematically impossible
• Google’s Nobel winner reports breakthroughs while others report breakdowns
• The uncomfortable middle ground where AI improves but never masters

,