Why ChatGPT’s Lack of Goals Is a Critical Challenge for AI’s Future

The rapid growth of artificial intelligence has ushered in incredible advances, yet an important debate looms large: Is the current trajectory, centered on large language models (LLMs) like ChatGPT, truly the path to genuine artificial general intelligence (AGI)? Reinforcement learning (RL) pioneer and Turing Award winner Richard Sutton argues that it’s not. Despite their remarkable language capabilities, LLMs fundamentally lack goals, an absence that might make them a dead end for achieving true intelligence.

The Compute Explosion Behind AI Models

OpenAI recently announced that it has increased its computational capacity ninefold just this year, with plans to ramp it up by 125 times again by 2033—an energy footprint predicted to surpass that of India, a nation with 1.4 billion people. This staggering scale highlights the resource intensity of training and running large models like GPT-4 and beyond. Moreover, while more efficient chips from NVIDIA help, actual AI horsepower could expand well beyond this estimate.

This dramatic growth raises sustainability concerns, with some studies estimating that AI could consume as much electricity annually as 22% of all US households by mid-decade. The question begs: Is scaling compute and data alone the solution to building meaningful AI, or do fundamental model architectures need rethinking?

Richard Sutton’s Critique: Mimicry vs. True Learning

Richard Sutton, a leading figure in reinforcement learning, argues that LLMs like ChatGPT are fundamentally about sophisticated mimicry. They predict the next word based on learned patterns from text rather than developing real-world understanding or goals. For Sutton, this is akin to reading every cookbook but never actually learning to cook.

Key points from Sutton’s critique include:

• No real goals: LLMs optimize for next-token prediction, not meaningful objectives.

• No learning from consequences: They cannot be surprised or adapt based on real-world outcomes.

• Lack of autonomy: They mimic intelligence instead of building it through active learning.

Sutton contrasts this with how humans learn skills like riding a bike—through trial, error, and continuous adaptation driven by real consequences.

The Alternative Vision: Continuous Learning Agents

Sutton proposes a new AI architecture, OaK (Online adaptive Knowledge), where an agent learns continuously from a live stream of sensation, action, and reward. This agent wouldn’t rely solely on massive offline training runs but would adapt and learn on the job, developing its own motivations and goals.

This approach aligns closely with RL principles and stands in contrast to how LLMs operate. If Sutton's argument holds, simply scaling models like GPT-6 infinitely won’t get us to AGI. Instead, we need architectures that allow AI to set and learn from its own goals.

The Safety Implications of Goal-less AI

Interestingly, Sutton points out that the lack of autonomous goal-setting in LLMs might actually be a critical safety feature. Without self-set goals, these models remain passive and controlled, reducing risks associated with autonomous AI agents pursuing unpredictable objectives.

Lessons for Business and AI Strategy

For companies adopting AI, this insight underscores that while LLMs offer powerful tools for tasks like customer service, content creation, and data analysis, they are not yet at the stage of independent decision-making agents.

At Leida, we focus on harnessing AI’s existing strengths—such as analyzing customer feedback efficiently with tools like Enterpret, which can unify and interpret massive volumes of data to improve business strategies without requiring autonomous AI decision-making.

Bringing It Home: Navigating the Future of AI

The current AI boom is breathtakingly powerful and fast-moving. However, the debate raised by Sutton reminds us that the next leaps in AI may come not from bigger and bigger models but from fundamentally different ways of learning and interacting with the world.

In the meantime, practical use cases that combine AI’s pattern recognition with human goals will continue to deliver real value. Staying informed on these evolving capabilities ensures you can leverage AI strategically as it matures.

If you’re curious how AI could uncover hidden bottlenecks in your workflows, book a call with our team below.