Saving This For Later
They Didn't Call It "Statistical Text Prediction." They Called It Intelligence. Fifteen trillion words. Every book digitized, every article indexed, every Reddit argument ever typed. A machine trained on that pile generates an answer in under 200 milliseconds. And it sounds exactly like your smartest colleague. That sound is the product. Not the answer. The sound.
Here is what the machine is actually doing in those 200 milliseconds: it is predicting the most statistically probable next word, then the next, then the next, based on patterns in all that text. That is it. That is the whole trick. There is no understanding. There is no model of the physical world. There is pattern-matching running very, very fast.
Yann LeCun, one of the pioneers of deep learning and the chief AI scientist at Meta, calls these systems token generators. Not reasoning engines. Token generators.
That is the field's own architects telling you what they built.
I have been building technology since before the internet had a name. I ran software budgets at companies you've heard of. I helped build the infrastructure that other platforms now call inevitable. And I am telling you: the naming of this thing is not an accident. Calling a statistical text predictor "artificial intelligence" is not a category error. It is a marketing decision worth trillions of dollars. Someone chose that word, and that someone is getting very, very rich off your willingness to believe it.
Apple researchers wanted to know what happens when you interrupt the script. They built a benchmark called GSM-Symbolic: standard math problems that these models normally ace, with one twist. A single piece of irrelevant information added to the prompt. Information a human would simply ignore.
Accuracy in the best models dropped by as much as 65%.
The machine wasn't solving the problem. It was pattern-matching the format. When the pattern broke, the performance collapsed.
That is not a bug they can patch. That is the architecture.
Here are two more structural failures, and I use the word "structural" deliberately. First: causal reasoning. These models learn correlations. A and B appear together in the training data. But the system has no mechanism for reliably determining whether A causes B. Models perform near randomly on formal causal inference tasks. Planning, diagnosis, and strategic decision-making all require causal reasoning. The machine remains very poor at it.
Second: hallucination. The confident fabrication of plausible-sounding nonsense.
OpenAI's own internal tests show their flagship reasoning model, o3, hallucinated in response to 33 percent of questions on their own accuracy benchmark. Their next model, o4-mini, scored 48 percent. OpenAI's own report says "more research is needed" to understand why their most advanced models are making things up more, not less. The CEO of Vectara, a company that professionally tracks AI reliability, has said hallucinations "will never go away" because they are baked into the architecture.
That is not a company on the verge of replacing human cognition. That is a company that cannot explain its own product's failure modes.
Now, to be precise about what is real: AI tools have made genuine progress on specific, narrow, well-defined tasks. Coding benchmarks show measurable improvement. The industry has also moved toward "agentic" systems that chain together tool use, memory, and multi-step tasks in ways that produce real capability gains on constrained problems. These are not nothing.
But notice the shape of the progress. It runs in one direction.
Better at finishing a well-framed code function. Not better at knowing when the whole approach is wrong. Better at retrieving and recombining what it has seen. Not better at reasoning about something genuinely novel. Better at the things a fast, confident pattern-matcher would naturally get better at with more data and more compute.
The underlying architecture has not changed. The failure modes have not been fixed. They have, in several documented cases, gotten worse as the models have scaled up.
Meanwhile, 66 percent of professional software developers, the people actually building the things these companies are selling, told the Stack Overflow 2025 Developer Survey that their biggest frustration is AI solutions that are "almost right, but not quite." More developers actively distrust AI accuracy than trust it. Andrej Karpathy, a founder of OpenAI and one of the most respected researchers in the field, coined a term for what he sees coming in 2026: the Slopacolypse. A flood of AI-generated output that is confidently, fluently, almost-right wrong.
That word did not come from a skeptic outside the industry. It came from inside the machine room.
There is a documented cognitive phenomenon called fluency bias. When information is delivered quickly, smoothly, and confidently, the human brain grants it authority regardless of accuracy. Psychologists have studied it for decades. Con artists have exploited it for centuries.
The oracle's confidence is more persuasive than the scientist's doubt. The preacher's cadence is more compelling than the philosopher's stammer.
A large language model runs this same con at superhuman speed on millions of people simultaneously.
Here in Colorado, we are not immune. Every school district from Longmont to Pueblo is being told they need an AI strategy. Every city council on the Front Range is weighing AI-powered services. Every local business owner is being sold a subscription to something that sounds like a digital employee.
Some of those tools are genuinely useful. I have built with them. I know the difference.
But useful tools do not require the word "intelligence" to sell themselves. Spreadsheets are useful. Calculators are useful. We did not declare the dawn of artificial arithmetic.
The word "intelligence" is doing specific work. It is manufacturing urgency. It is suppressing skepticism. It is creating the atmosphere the industry needs to justify valuations that the underlying technology cannot support on its own merits.
The industry has a name for the thing that manufactures atmosphere. Con artists call it "the big store." The rented office. The hired extras. The stacked waiting room designed to overwhelm the mark before a single word is spoken.
The AI industry's big store is a constant low-frequency hum of implication. Headlines about sentience. Pentagon officials warning the machine "will claim to have a soul if you prompt it correctly," which gets reframed by a million-person tweet as confirmation that the machine actually has one. Science journalists breathlessly reporting that a model sounds "creative" and "alive."
Nobody makes a falsifiable claim. The atmosphere thickens.
Your brain fills in the rest.
I am not arguing that these tools are worthless. I am arguing that the gap between what they are and what they are being sold as is not a rounding error. It is a chasm. And the communities that are going to pay the price for that chasm are not the venture capitalists in Palo Alto. They are the school districts that gutted their teaching staff because the machine was going to handle it. They are the local newsrooms that believed AI would solve the revenue problem instead of just automating the mediocrity.
They are your neighbors.
The scaling laws that drove a decade of AI progress are hitting diminishing returns. The field knows this. That is exactly why the marketing has gotten louder, not quieter. When the product stops getting dramatically better, you sell the atmosphere harder.
The machine predicts text, fast and fluently, like a grifter who has memorized a perfect script.
It has gotten better at some of the scripts. It has not learned to tell the truth.
---------------------------------------------------------------
CLAIMS AND SOURCES
1. A large language model is trained on approximately 15 trillion tokens of human text.
Source: https://www.youtube.com/watch?v=n-kW9wjApVw (The Silicon Mirage, Season 2 Episode 4, citing LLM training documentation)
2. Yann LeCun, Meta's chief AI scientist and deep learning pioneer, called LLMs "token generators, not reasoning engines" at NVIDIA GTC 2025 (March 2025).
Source: https://analyticsindiamag.com/.../nvidia-gtc-lecun.../ [Verified during research via multiple search results confirming the GTC 2025 statement]
3. Apple's GSM-Symbolic benchmark found that adding a single irrelevant clause to math problems caused accuracy drops of up to 65% across state-of-the-art models.
Source: https://arxiv.org/abs/2410.05229 (Mirzadeh et al., "GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models," Apple Research, 2024, ICLR camera ready)
4. OpenAI's o3 model hallucinated 33% of the time on OpenAI's PersonQA benchmark. O4-mini hallucinated 48% of the time. Both rates are roughly double those of OpenAI's previous reasoning models.
Source: https://techcrunch.com/.../openais-new-reasoning-ai.../
5. OpenAI's technical report on o3 and o4-mini states "more research is needed" to understand the increase in hallucination rates.
Source: https://techcrunch.com/.../openais-new-reasoning-ai.../
6. Amr Awadallah, CEO of Vectara, stated that AI hallucinations "will never go away" because they are architectural.
Source: https://www.nytimes.com/.../techno.../ai-hallucinations.html [Verified during research]
7. 66% of professional developers cite "AI solutions that are almost right, but not quite" as their biggest AI frustration. More developers actively distrust AI accuracy (46%) than trust it (33%). Only 3% "highly trust" AI output.
Source: https://survey.stackoverflow.co/2025/
8. Andrej Karpathy predicted 2026 will be the "Slopacolypse," a flood of low-quality AI-generated output. Karpathy is a co-founder of OpenAI and former director of AI at Tesla.
Source: https://www.google.com/search... [Verified during research via multiple sources confirming the statement]
9. AI scaling laws are showing diminishing returns, a trend acknowledged publicly by major labs and covered extensively in 2025 research.
Source: https://www.google.com/search... [Verified during research via multiple independent sources]
10. Fluency bias: human brains grant authority to information delivered quickly, smoothly, and confidently, regardless of accuracy.
Source: Oppenheimer, D.M. (2008), "The Secret Life of Fluency," Trends in Cognitive Sciences, 12(6). [Peer-reviewed; verified via training data]