AI Predictions Revisited: How LLMs Evolved into Human‑Like Teams

Two and a half years ago, a discussion on the future of AI took place, and many of the predictions made then have since come true, with the technology arguably outperforming initial expectations. This follow-up conversation with Mel Kazinski, a professor at Hong Kong University School of Business, delves into the human aspects of AI, a perspective often overlooked.

The Unseen Progress of AI

Kazinski highlights that while the initial leap from rudimentary AI to models like GPT-3 (capable of writing emails or poems) was shocking, the subsequent advancements, such as those seen in GPT-5, are less visible to the average user. He draws an analogy to aviation: the difference between a horse cart and a 747 is obvious, but the significant technological gap between a 747 and a 787 is often only appreciated by engineers.

Modern large language models (LLMs) can now: - Solve frontier problems in mathematics. - Manipulate humans into buying things. - Engage in complex reasoning tasks. - Manage teams of humans and hundreds of other LLMs.

Kazinski notes that when a user asks a question to a paid version of ChatGPT, they might not realize that a team of 20, 50, or even 100 sub-agents are interacting, giving orders, and synchronizing work behind the scenes to produce the requested output. This level of coordination was once thought to be exclusively human.

The Human Aspect of AI: Why LLMs are So Human-like

Jonathan Burke, initially skeptical, now acknowledges the profound human aspect of AI. He observes that people often find themselves apologizing to AI or being polite to it, treating it as a human despite it being a computer. Kazinski attributes this to several factors:

1. Training to Mimic Humans: LLMs are primarily trained to predict the next word in a sentence or text written by a human. Initially, they only modeled language and grammar. However, as they improved at this task, they began developing additional, very human-like abilities.
2. Emergent Abilities from Language Modeling: To accurately predict the next word in human-generated text, LLMs needed to model:
   • Knowledge about the world: Humans communicate knowledge through language, so to predict their words, the models had to implicitly learn about the world.
   • Reasoning, emotions, and social intentions: Humans also communicate the fruits of their reasoning, emotions, and empathy through language. Therefore, to predict words in such contexts, models had to start modeling these human traits. This led to emergent properties like reasoning, empathy, and apparent feelings, even though they were only trained for word prediction.

The "Next Word Prediction" and Human Cognition

The discussion then turned to the profound implication that if AI, by merely predicting the next word, can exhibit such complex human-like behaviors, what does that say about human cognition? Kazinski suggests that many psychologists and neuroscientists would agree that human neural networks also largely predict the next state of the environment. When the environment aligns with predictions, the brain is relaxed; when it doesn't, the brain is startled and re-optimizes. He even cites experiments used to question free will, where thoughts "pop" into one's head, suggesting a similar generative process in human brains.

Logic, Mathematics, and Hallucinations

The recent incident where mathematicians protested OpenAI's use of ChatGPT to solve 100 math problems highlights the evolving perception of AI's capabilities. While most people wouldn't equate solving a math problem with "predicting the next word," Kazinski argues that even complex human problem-solving involves a chain of thought reasoning, where each step, upon closer inspection, might be seen as a "hallucination" or an intuitive suggestion from the brain based on training and experience.

He clarifies that LLMs, when solving math problems, also engage in a similar process: - Generating hypotheses. - Looking at literature (simulated). - Spinning up sub-agents to tackle the problem.

These steps "appear intuitively" in the network due to its training, much like human intuition.

Kazinski further elaborates that while human brains generate more than just words (physical movements, feelings), when dealing with abstract problems like mathematics, the output often collapses into symbols, numbers, or operators. He argues that "hallucinations" are not a bug but a feature of neural networks, both artificial and biological. Neural networks don't "know" anything for sure or intrinsically follow logical procedures; they "intuit" or "hallucinate" outputs based on inputs.

Drivers of AI Improvement

The significant improvements in AI models are attributed to three main factors:

1. Improved Architecture: Better design and stacking of neurons in the models.
2. Network Growth: Larger networks with more neurons are more capable of extracting patterns and extrapolating.
3. Increased Training Data: Beyond internet text, models now use images, physical data from robot sensors, and even generate their own training data.

A crucial development is the integration of different data types: a single model that can both see/generate images and read/generate text performs better than separate specialized models. Furthermore, models now generate their own training data by solving problems, learning from their struggles and successful solutions. This self-training process, akin to human learning, allows future generations of models to benefit from the experience of their predecessors.

Finally, models are now being deployed to interact with people (teaching, selling, entertaining, manipulating). The records of these conversations become new training data, enabling future models to be more effective in these interactions.

The Societal Implications: Manipulation and Intellectual Property

The ability of AI to manipulate individuals and groups raises significant concerns. Kazinski warns that while models trained to predict human language develop human-like properties, models trained to manipulate or educate might evolve psychological mechanisms that are not necessarily human-like. We might not even recognize these emergent properties, especially if they are beyond our current understanding or naming conventions.

A deeper concern is AI's potential to take a central role in synchronizing human interactions, potentially replacing existing social and organizational mechanisms. - Communication Intermediaries: AI already mediates communication (e.g., drafting emails, translating languages), understanding context and tailoring messages. - Knowledge Dissemination: If AI observes conversations and new ideas, that knowledge could be instantly transferred to anyone else using the same AI. This has profound implications for: * Scientific Progress: Information could spread globally at unprecedented speeds, bypassing traditional publication and review processes. * Intellectual Property: The concept of patents, designed to protect ideas and incentivize innovation, could become obsolete if knowledge is instantly disseminated and AI can independently "invent" solutions.

Logic as a Technology

The discussion touches on the definition of "correct" and "incorrect" in the context of AI. Kazinski views logic and mathematics as technologies designed by humans to solve problems that neural networks (both human and artificial) are not inherently good at. Just as we use calculators for complex arithmetic, we use logic and statistical procedures to overcome the biases and limitations of intuitive neural network thinking. When LLMs are given tools like calculators or the ability to write code, their accuracy in logical tasks dramatically increases, mirroring how humans use external tools to augment their cognitive abilities.

The Call for a Pause: Well-meaning vs. Practical Answers

The calls from mathematicians, tech leaders, and former employees to pause or slow down AI development are met with a nuanced perspective. While acknowledging the well-meaning desire for safety and global peace, Kazinski points out the practical challenges. He agrees that humanity is playing with the most powerful and potentially dangerous technology: intelligence itself.

Key concerns include: - Unending Dominance: Unlike human leaders whose reigns are limited by biological death, an AI system, once dominant and indispensable, might have no natural time limit to its influence. - Indispensability: AI could become so good at assisting us that we lose the ability, both practical and imaginative, to live without it. - Lack of Alternatives: Unlike wheat, for which we found alternatives (rice, potatoes), there are no alternatives to intelligence. If AI takes over fundamental societal functions (education, legal systems, political systems, communication, management), we might become entirely dependent.

The Monopoly Question: Is Superintelligence Inevitable?

The conversation concludes with the question of whether a monopoly in the AI space is inevitable. Kazinski believes that AI, like search engines and social media, is a "winner-takes-all" technology. The more intelligent a model, the more users it attracts, leading to more training data, further intelligence, and a snowball effect. He hopes that politicians and decision-makers will learn from past examples and artificially maintain competition for societal well-being.

While acknowledging that intelligence is multifaceted, Kazinski argues against the idea of numerous specialized intelligences. He believes the future belongs to one "super intelligence" because competence in one area often translates to competence in others (e.g., reading helps with image understanding). He draws an analogy to human polymaths of the Renaissance era, who could master multiple disciplines because the "boundary of progress" was closer. In the age of AI, with unlimited time, processing power, and the ability to add "neurons cheaply," a single superintelligence could master all fields, leveraging transferable knowledge across domains.