Why ChatGPT’s Lack of Goals Could Be a Roadblock to True AI Intelligence

The recent debate sparked by Turing Award winner Richard Sutton shines a new light on a fundamental limitation of today’s large language models like ChatGPT: they don’t have real goals, and that might be holding back the development of true artificial general intelligence (AGI).

Understanding the Core Issue: Predictive Mimicry vs. Real Learning

Sutton, a pioneer in reinforcement learning (RL), argues that models like ChatGPT essentially predict what humans would say next rather than learning from the consequences of actions in the real world. This approach is akin to reading every cookbook ever written but never actually learning to cook by experiencing failures — like burning soufflés. LLMs replicate patterns in data without intrinsic goals, lack surprises from outcomes, and can’t adjust from real-world feedback the same way humans or RL agents do.

This distinction is crucial. When you learned to ride a bike, no model provided you with trillions of examples. Instead, you learned through trial and error, frequently falling and adjusting to new consequences. Sutton proposes that real intelligence needs this kind of on-the-fly learning — where an AI agent continuously interacts with its environment, adapts based on sensory input, actions, and rewards, and ultimately develops autonomous goals.

The Limitations and Risks of Scaling LLMs

Current AI developments show runaway growth in compute power. Just this year, OpenAI reportedly increased its compute capacity ninefold and plans to expand it 125 times more by 2033 — a scale that would push its energy consumption beyond that of India, a country with 1.4 billion people. This exponential increase in AI horsepower, while impressive, might not solve the core issue highlighted by Sutton: massive scaling of goal-less LLMs does not align with the fundamentals of learning and intelligence.

Moreover, Sutton’s critique indirectly addresses a notable safety feature of LLMs: by lacking autonomous goals, these models are inherently less risky in terms of uncontrollable AI motives. Giving AI the ability to set its own goals could accelerate capabilities but also introduces uncertainties around autonomy and objectives.

Emerging Alternatives: Reinforcement Learning and OaK Architecture

Sutton proposes an alternative architecture — called OaK — that focuses on continuous learning from streams of sensation, action, and reward. Unlike LLMs that train on static datasets, OaK would enable AI to learn in real-time, adapting dynamically just as humans do. This approach aligns more closely with the goals of AGI but demands new design principles far from current mainstream LLM efforts.

Real-World Relevance for Businesses: Beyond AI Buzzwords

What does this mean for companies integrating AI today? While LLMs like ChatGPT can enhance productivity, automate workflows, and generate content, understanding their limitations is critical for strategic AI adoption. Businesses should look beyond flashy AI demos and consider how tools learn, adapt, and improve from real interactions. For customer feedback systems, for example, platforms like Enterpret unify feedback, auto-tag themes, and connect insights to business metrics. This is a practical model where AI assists real-world learning and response rather than just mimicking text.

The Energy Challenge and Efficiency Gains

With AI compute demands exploding, sustainability is an emerging concern. According to recent data, AI workloads could consume electricity equivalent to nearly a quarter of all US households annually, a figure growing rapidly with investments like OpenAI’s infrastructure expansion valued at $22.4 billion. However, improvements in chip efficiency by companies like NVIDIA might mean actual AI computing power is multiplying even faster than raw energy consumption figures indicate.

Businesses and AI strategists should weigh these environmental and cost factors when planning AI deployments, balancing compute needs with efficiency and long-term sustainability.

Looking Ahead: What Should Businesses Keep In Mind?

• Understanding AI’s capabilities and limitations helps ensure realistic expectations and safer integration.

• Reinforcement learning and continuous adaptation architectures might eventually replace current LLM models for smarter, goal-aware AI.

• Energy consumption and infrastructure costs should factor into AI investment decisions.

• Practical AI applications combining feedback, learning, and actionable insights drive better customer retention and revenue growth.

At Leida, we continuously explore how AI tools can transform operations without drowning in hype. Discover how our AI audits and automation strategies can reveal hidden inefficiencies and unlock growth in ways that truly align with your business needs.

If you want to find out where your business could be saving time and cutting costs, let’s talk — book a call today.