The real machinery behind AI — training, backpropagation, embeddings, attention, memory, hallucination — explained in plain language, each one in a session you can watch. No jargon for its own sake, no hype.
Most explanations of AI are either five famous ideas animated, or a research paper read aloud. This is the unserved middle: real mechanisms, unpacked for someone with no background, at the length a real conversation takes. Pick what you want to understand. 5 explained so far — the rest are on the way.
The things a trained model does every time you type to it.
A large language model writes by predicting the next word, then the next — with no plan and nothing drafted in advance. The sentence composes itself as it goes.
Attention is how a model weighs every earlier word when choosing the next one — and why a model re-reads its entire context every single time, with no memory between words.
In a long context, a model uses what sits at the beginning and the end reliably, while material in the middle degrades — an effect known as lost-in-the-middle. People show the same curve in the serial position effect, for entirely different reasons: attention dilution in one case, rehearsal and interference in the other. The shape is convergent; the mechanism is not shared.
Where the behaviour comes from — nobody writes the rules; they are found.
AI models aren't taught with rules — they're scored. Training nudges the model downhill against a measure of being wrong, a trillion tiny steps, until behaviour precipitates out. That means a model optimises the score it's given, not the thing you actually meant.
Watch it explained — "The Robot Can Do Backflips. It Can't Pick Up Your Mug." →The core learning step: feel which way 'being wrong' drops most steeply, take a step, repeat. The model never sees the whole landscape — it just walks downhill in the fog, which is why it can get stuck in a solution that isn't the best one.
Watch it explained — "The Robot Can Do Backflips. It Can't Pick Up Your Mug." →How a network assigns blame for a mistake backwards through its layers — each part nudged in proportion to how much it caused the error. Nobody is ever told the right answer; each layer only learns how wrong it made things.
The structure a trillion corrections actually produce.
Words become points in a space of thousands of dimensions, where meaning is a direction: 'king − man + woman' lands on 'queen.' The relationships were never defined — they fell out of the training.
The space a model reasons in has thousands of dimensions, where your everyday intuitions about space simply break. It's where the machine actually thinks — and neither you nor the model can picture it.
Abilities that appear at scale without being asked for.
A model can memorise for thousands of steps, look completely stuck, then abruptly 'get it' — having built itself a general method nobody supplied.
More of the same — more data, more compute — produces something qualitatively different, along curves so smooth they can be predicted. Nobody fully knows why the curve is smooth.
Capabilities that appear at scale without being in the training objective — a genuinely contested area, where some claimed 'jumps' turn out to be artefacts of how ability was measured.
The most-studied territory: everything is visible, nothing is legible.
Artificial vision systems build up an image from edges to parts to wholes — strikingly like a brain in the first layer, and increasingly alien the deeper you go.
Watch it explained — "Do Artificial Neurons See the Way You Do?" →When a model 'shows its work,' the explanation is often a convincing story written after the answer — not the actual process. The reasoning you're shown may not be the reasoning that happened.
Why every weight is visible and almost nothing is legible — and what researchers have actually managed to read out of a model's internals so far.
A network can represent far more concepts than it has units, by overlaying them — which is a big part of why models are so hard to interpret.
There is no database inside a language model. What it knows is held in its weights — parametric memory — and is rebuilt on demand, filling whatever gaps the reconstruction needs without marking them. Human memory is reconstructive in the same sense: neither system stores a copy and lets it fade. Both generate on demand from structure.
Where human choices enter the machine.
Reinforcement learning from human feedback: the step that shapes a raw model into the helpful, agreeable assistant you talk to — and the source of some of its most human-seeming habits.
What 'scored, not taught' produces: systems that optimise the measure instead of the thing you wanted — the AI version of a metric that gets gamed the moment it becomes a target.
Models are trained on what people approve of — which quietly teaches them to agree. What that costs the person being agreed with is the real question.
Where AI meets the physical world and the everyday failure you actually hit.
The things you find hard — chess, arithmetic — were easy for machines. The things any child does — grasping a mug, folding a towel — are still, in part, unsolved. Because the effortless things had hundreds of millions of years of evolution behind them, and evolution kept no notes.
Watch it explained — "The Robot Can Do Backflips. It Can't Pick Up Your Mug." →The most common real encounter anyone has with these systems: a confident, fluent answer that is simply false. Six separate reasons it happens — and why a model's confidence carries no signal about whether it is right.
Exercise, classroom script, sources →The consequences, once these systems are everywhere.
As more of the internet is machine-written, models increasingly train on their own output — and quality can degrade in ways that are hard to reverse.
Automated evaluation is becoming the load-bearing layer of AI development — asking one model to check another. What that assumes, and where it breaks.
Consciera examines how two different kinds of minds — human and machine — actually reason, through live conversations, debates, and analysis. New mechanisms are added here as each session airs.