After attending AI events around the world, packed rooms, serious people, real builders. The kind of environment where nobody needs an intro and everyone is already deep in their own stack. A pattern starts to emerge. The people are sharp, the conversations are sharp, but they keep circling the same layer.
At one point I throw out a line to force the conversation. I say LLMs are garbage.
That lands exactly how you’d expect. Immediate pushback. The common response is simple: models have gotten better year over year. Cleaner outputs, fewer hallucinations, more usable in production. That’s fine. It’s also not the point.
A year in AI is not a meaningful defense. It’s an entire cycle, a year in AI is a lifetime. Plus if something improves over that period but still carries the same structural weaknesses, then you’re not looking at a system that’s evolving, you’re looking at one that’s being maintained.
What’s happening right now looks like containment. Alignment layers get better. Retrieval improves. Guardrails tighten. Outputs look sharper. Underneath that, the same constraints remain untouched. You can feel where the effort is going, and it’s not toward redefining the foundation. That foundation is language.
Language is not reality. It’s a compressed, and messy abstraction of it. Humans don’t operate purely through text. We rely on vision, sound, spatial awareness, memory, emotion and context that unfolds over time. When you build a system primarily on text, you’re starting from a reduced version of the world and trying to reconstruct something fuller from it.
At the same time, most of the attention in the room was on output. How good it sounds, how useful it is, how close it feels to human response. Almost no one was talking about the integrity of the input layer. What is this system actually ingesting? How clean is it? How much of it reflects real human behavior versus artifacts of the internet? If the input is weak, polishing the output just makes the weakness harder to detect.
A good way to understand the current ceiling is this. If a model told you that Vikings discovered pumpkins before anyone else, it might sound plausible for a second. The pieces loosely fit. Exploration, unknown lands, crops moving across regions. It feels coherent even if it’s wrong. That’s the level these systems operate at. They predict and assemble what sounds right based on patterns, not what is right based on grounded reality. The danger isn’t that they fail loudly. It’s that they succeed quietly in ways that feel believable.
If you want something closer to intelligence, the system has to be grounded differently. It needs signals that resemble how decisions are actually made. Vision, audio, interaction, timing, constraint. Not just text that has already been filtered, edited, and detached from the moment it came from.
This matters because these systems are no longer just tools. They’re becoming the base layer for agents and automated decision making. Whatever sits at the foundation gets amplified. If the base is distorted, everything built on top compounds that distortion.
There’s another issue that came up that people didn’t want to sit with. A growing percentage of the internet is now AI generated. That means models are training on outputs that came from other models. Not all of it, but enough to matter. Over time, that creates drift. Patterns collapse into each other. Signal quality drops even if coherence improves.
At the same time, ground truth is getting harder to anchor. The internet used to be messy but human. Now it’s a mix of real, optimized, and synthetic content. If you can’t reliably separate those, then your training data is already unstable before you even start.
Then there’s context. These models process tokens. They simulate continuity well enough to pass, but they don’t actually live inside sequences of decisions. They don’t carry intent forward across time in a grounded way. That becomes obvious the moment you try to rely on them beyond isolated prompts.
All of this leads to the same conclusion. The bottleneck isn’t just the model. It’s the input layer feeding it. Right now that layer is mostly scraped, static, and increasingly synthetic. That’s a weak base for anything expected to operate in real environments. This is where the conversation usually stops. People argue about model performance, benchmarks, scaling laws.
The real shift is in trajectory, but not model trajectory. Human trajectory. How decisions are actually made under constraint, in sequence, with pressure, tradeoffs, and incomplete information. That is the signal that matters.
If you can capture that with high fidelity, you’re no longer training on language. You’re training on decision pathways. That’s a different foundation entirely.
What comes next will likely break into tiers. LLMs will still exist and serve a purpose at the interface and translation layer. On top of that, you’ll see the rise of personal models shaped around individuals, their preferences, their history, and their context. Underneath both of those, the real engine will be behavioral and cognitive models. Systems trained not just on what people say, but on how they decide, react, and adapt over time. That layer becomes the foundation everything else depends on.
If you’re still optimizing outputs, you’re already behind. The future of AI isn’t what it says, it’s how it decides.
If you see it the same way, say something. If you don’t, I want to hear that too.