🧠 Why AI’s Still Dumber Than a 3-Month-Old

Sign Up | Home Page | Meet The Team

In today’s sunday special:

Read time: 7 minutes

📜 THE PRELUDE

As you scoot between the booth and the table, your elbow knocks over a cup, sending it tumbling over the table’s edge. You grit your teeth, bracing for the impact. It lands with a barely audible “clink!,” completely intact. Your split-second reaction to the falling cup wasn’t planned. It was an instinctive response.

Believe it or not, humans experience millions of distinct social interactions throughout their lives. Over time, the human brain compresses them into a mental model equipped with predictive coding that continuously anticipates reactions to specific actions.

Today, AI’s ability to grasp reality is severely limited because it doesn’t possess lived experience. To achieve superintelligence, it must directly interact with the underlying laws that govern the physical world, whether it’s gravity, friction, or collision, to understand the forces and flows of the physical universe and supercharge everything from humanoid robots to scientific discoveries.

🥕 INFANTS HAVE PHYSICAL INTUITION

In 1991, Canadian American research psychologist Renée Baillargeon tested whether young infants understand object permanence: an object continues to exist even when it can no longer be seen.

Since young infants can’t explain what they’re thinking, she measured how long they looked at different events. This method of measurement is rooted in the fundamental belief that babies tend to look longer at events that are inconsistent with what they understand should happen.

In one experiment, 3.5-month-old infants watched a short carrot and a tall carrot slide along a track. The track’s center was hidden by a screen with a large window in its upper half. The short carrot was shorter than the large window’s lower edge, and it expectedly didn’t appear in the large window when passing behind it. The tall carrot was taller than the large window’s lower edge, and it unexpectedly didn’t appear in the large window when passing behind it.

The 3.5-month-old infants looked significantly longer at the tall carrot event than at the short carrot event, suggesting that they held a mental model of the existence, height, and trajectory of each carrot behind the screen and expected the tall carrot to appear in the large window and were surprised when it didn’t.

💭 FROM PREDICTING TO PROBLEM-SOLVING

The world’s most popular LLMs, like ChatGPT, Claude, and Gemini, are statistical systems designed to predict the probability of a sequence of tokens. In simple terms, they’re essentially sophisticated autocomplete machines trained on the entire internet to process diverse inputs and generate plausible outputs that sound human.

For context, tokens are units of text that enable LLMs to understand humans. They bridge the gap between human input and machine output. In English, one token generally corresponds to four characters. This translates to roughly ¾ of a word, meaning 1,000 tokens are roughly equivalent to 750 words.

To train an LLM, it’s shown millions of sequences of text and is prompted to predict the next token. This foundational training method is referred to as autoregressive next-token prediction. For example, when given: “I feel anxious when I speak in front of a {BLANK}!” LLMs ask themselves, given the tokens so far, what’s the most likely next token? In this instance, it might predict: “{CROWD}!”

Each time the LLM makes an inaccurate prediction, it adjusts billions of weights, or numerical parameters, that help it recognize which patterns of tokens are more important for making better predictions in the future. These numerical parameters control how tokens relate to each other within the Neural Network (NN), which mimics the human brain by processing diverse inputs through hundreds of transformer layers comprised of millions of “nodes.” The core components of a transformer layer include:

LRMs are designed to spend more time thinking before responding. They achieve this by leveraging Test-Time Compute (TTC), which dedicates additional compute to AI inference: everything that happens after you enter your prompt. Imagine asking ChatGPT: “Summarize this article in three bullet points!” AI inference is the reasoning ChatGPT performs to generate that bulleted summary. The core components of AI inference include:

🌍 WORDS TO WORLDS IS AI’S NEXT FRONTIER?

On Sept. 13th, 2024, World Labs raised a $230 million Series B funding round at a $1 billion post-money valuation to advance LWMs, which attempt to emulate the way the human brain builds mental models of everyday life. Dr. Fei-Fei Li, widely recognized as the “godmother of AI,” described world models and spatial intelligence as AI’s next frontier. On Feb. 18th, 2026, World Labs secured a $200 million strategic investment from Autodesk as part of a larger $1 billion growth-stage funding round.

Imagine asking ChatGPT: “The dog fetched the {BLANK}!” In this instance, it might predict: “{TENNIS BALL}!” ChatGPT excels at understanding and generating tokens. It can tell you what happens, but not why or how it happens. ChatGPT can’t understand, simulate, and predict the dynamics of the physical world like we do. If we toss a tennis ball into the air, we know gravity will pull it back down. This kind of physical intuition is something we naturally develop by observing how gravity, friction, and collision influence the physical world around us.

We rely on our physical intuition every day to predict how the physical world might react to our actions. For example, knowing how much force to use when opening a car door. It’s not some magic ability. It’s our brain constantly building a mental model of how the physical world works based on our lived experience.

So, what if language models and reasoning models could also autonomously experience the forces and flows of the physical universe? That’s where world models come in. They aim to help ChatGPT, Claude, and Gemini evolve beyond text, images, audio, video, and code by learning to uncover the underlying laws that govern the physical universe.

🔑 KEY TAKEAWAY

“A four-year-old child has seen 50x more data than the biggest LLMs,” said Yann LeCun, who famously served as Meta’s Chief AI Scientist. “Text is simply too low bandwidth and too scarce a modality to learn how the world works.” As young infants, we begin to build mental models of how the physical world functions. In contrast, LLMs predict tokens, and LRMS reason about and act on tokens. Without experiencing the underlying laws that govern the physical universe, both struggle to truly navigate and comprehend unseen, real-world environments.

📒 FINAL NOTE

{{rp_personalized_text}}

Share your unique referral link: {{rp_refer_url}}