How Does AI Think? Dr. Idan Blank Explores What Happens Beneath the Surface

When you ask a chatbot to summarize a paper or draft an argument, it feels almost magical. The answer comes out smooth, coherent, as if the machine were actually “thinking.” But what’s really going on inside those massive neural nets? At a packed hall on September 10, Dr. Idan Blank from UCLA tried to peel back the illusion.

Blank’s talk, titled “Understanding ‘Understanding’ in Language Models,” wasn’t about praising AI as a new kind of mind. Instead, he wanted to test a very specific claim: that the same next‑word prediction machinery that powers ChatGPT can, under certain circumstances, mimic human‑like comprehension.

“All models are wrong… but transformers give us a computational representation of virtually any sentence,” he told the audience, chuckling a little. “That alone makes them worth studying, even if they turn out to be… not great.” He stressed that, rather than judging AI only by the cleverness of its prompts, his lab actually opens the black box and watches the gears turn.

His first experiment was simple, almost playful. He fed the model a sentence that makes grammatical sense but stumps common sense: “The garden in the cat chased a butterfly.” The verb “chased” clearly expects an animate subject, yet the sentence forces the garden into that role. In a classic transformer, attention heads act like tiny spotlights, linking words together. Some heads specialize in connecting verbs to subjects. If they cared only about syntax, the model would dutifully bind “chased” to “garden.”

Surprisingly, the model didn’t. When the subject–verb pairing was bizarre, the attention shifted toward the more plausible “cat.” Blank pointed out that humans do something similar: we never read grammar in a vacuum; meaning nudges our parsing in real time. It’s not proof that the AI truly “gets” the garden, but it does show that a system trained purely on next‑word prediction can develop a human‑like interaction between grammar and semantics.

The next test dug deeper into meaning. Blank compared three sentences: “A doctor pulled a prisoner,” “A prisoner was pulled by a doctor,” and “A prisoner pulled a doctor.” The first two differ in word order but describe the same event; the third flips the roles. Human listeners instantly see the first pair as closer in meaning than either is to the third.

What the transformers did was almost the opposite. Their internal vectors clustered more tightly around grammatical structure than around the underlying event. Some attention heads still singled out the agent and patient, but that information was buried, not central to the overall representation. As Blank summed up, “You can recover the full syntactic tree with reasonable accuracy, yet that doesn’t guarantee a hidden representation of meaning.”

So, a model can store the pieces of a story without assembling them the way we do. That finding matters because many AI applications assume that a good grasp of syntax automatically yields understanding of content—an assumption the data now call into question.

Blank didn’t stop at pure text. His current project blends pictures with words, asking whether multimodal models can infer why a speaker chose a particular expression. He showed participants three objects—a big hammer, a tiny hammer, and a large cork. When asked to pick the “big hammer,” humans instantly factor in both size and object type. Some language‑vision models seemed to do the same, predicting “big” only when a clear size contrast existed.

Yet there’s a cautionary footnote: those images also contained more hammer‑shaped pixels, which could be the real cue. In other words, the model might be responding to raw visual statistics rather than genuine pragmatic reasoning.

“At this point we have like ten different ways to explain the same phenomenon,” Blank laughed, gesturing to a whiteboard crowded with arrows. “The challenge now is to design experiments that can finally tease apart the “meaning” component from all the clever shortcuts the network has learned.”

Overall, Blank’s work paints a nuanced picture. Transformers are undeniably powerful—far better at handling language than any previous computational model. But they often arrive at sensible answers through routes that look nothing like human reasoning. Understanding those routes, Blank argues, is essential if we ever hope to build AI that truly collaborates with us rather than simply mimics us.