Is it just me, or is AI sounding... dumber?

LLMs are trained on the internet. But since 2023, the internet has been flooded with (you guessed it) AI-generated content. Much of which has never been edited or even read by a native English speaker. When an AI is trained on the output of another AI, it creates a feedback loop called ‘Model Collapse’.

It’s like making a photocopy of a photocopy. Each generation loses a bit of detail, a bit of nuance, and a bit of ‘soul’ until you’re left with a blurred, distorted mess. We are literally polluting the digital well we’re drinking from.

Running massive AI models is eye-wateringly expensive. And their use cases are being expanded constantly. To save money on compute costs, companies often ‘quantize’ their models. This is basically a fancy way of saying they compress the file.

Imagine trying to play a high-definition 4K movie on a 1990s dial-up connection. To make it work, you have to strip out the quality. Quantized models are faster and cheaper to run, but they lose the fine-grained reasoning that made the original model so impressive.

In the early days, models were spikier. They had weird, creative, and sometimes brilliant leaps in logic. But to make them commercially viable for everyone from a 12-year-old to a Fortune 500 CEO, developers have smoothed out the weights of the neural network.

This is called Regression to the Mean. The AI is being steered toward the most "probable" next word at all times. By definition, the most probable word is the most common one. This is why everything it writes now feels vaguely familiar. It’s been mathematically optimized to be as average as possible.

To make matters worse, in an effort to stop their bots from being problematic (or teaching people how to build bombs in their kitchen), companies have layered on massive amounts of alignment and safety filters.

The issue is that these filters often act as a digital lobotomy. In the quest to avoid lawsuits, they’ve inadvertently stripped away its ability to take a creative risk, or follow a complex, non-linear thought process.

Like the Wizard of Oz peddling madly behind the curtain, LLMs require massive amounts of constant human labour to supervise their output.

Most LLMs are fine-tuned using a process called Reinforcement Learning from Human Feedback (RLHF). Basically, humans sit in a room and rate the AI’s answers.

The problem? Humans are biased, lazy, and like to be agreed with. Over time, the AI learns that it gets a gold star when it’s agreeable and polite. This creates a sycophantic loop where the AI hallucinates or softens its take just to match what it thinks you want to hear.

If you’re looking for a strategic partner to push back on your ideas, you’re out of luck; instead you’ve hired a digital suck-up.

AI companies love to brag about their massive ‘context windows’ (how much information the bot can "read" at once). They claim you can drop a whole 500-page book in there.

But there’s a documented technical phenomenon whereby like a first grader with ADHD, LLMs get lost in the middle. Research shows that LLMs are great at remembering the beginning of your prompt and the very end, but they effectively zone out during the middle sections. As context windows get bigger, the bot’s attention gets spread thinner.

So whilst the advice of the techbros is to ‘prompt better’, the irony is that the longer and more complex your instructions, the less nuance the LLM picks up on.