Why a Turing Winner Says ChatGPT Is a Dead End

In late 2025, generative AI is everywhere: powering marketing copy, sales outreach, customer support, and even video production. But while most of Silicon Valley is racing to build bigger models and ship GPT-6–style systems, one of AI's most respected pioneers is throwing cold water on the hype.

Turing Award winner Richard Sutton argues that large language models (LLMs) like ChatGPT may be a dead end for reaching AGI (artificial general intelligence). Instead of endlessly scaling next-word predictors, he believes the future belongs to goal-driven agents that learn by acting in the world, not just predicting text.

For business leaders, marketers, and builders, this isn't just an academic debate. It changes how you should plan your AI strategy, where to invest, and what kinds of tools will actually create durable advantage over the next 3–5 years.

This post breaks down Sutton's critique, explains why video AI like DeepMind's Veo 3 could be a turning point, and looks at the very real compute and energy costs behind the current AI arms race. Then we'll translate it all into practical steps you can take right now to stay ahead.

---

1. Why Richard Sutton Thinks GPT-6 Is Going Nowhere

Richard Sutton isn't just another Twitter commentator. He's a Turing Award–level researcher and one of the founding figures of reinforcement learning (RL)—the branch of AI focused on agents that learn by trial and error to achieve goals.

His core claim: LLMs like ChatGPT are powerful, but fundamentally limited.

"Predicting the next word is not the same as understanding the world or pursuing a goal."

From next-word prediction to real intelligence

LLMs are trained to minimize one main error: how often they guess the next token wrong. Scale that up with trillions of tokens and massive compute, and you get impressive capabilities: text generation, coding, Q&A, and more.

Sutton's critique is that this setup has no built-in notion of goals, consequences, or long-term planning. It's reactive, not strategic.

LLMs:

• Don't act in an environment; they simulate actions in text.
• Don't receive rewards or penalties based on real outcomes.
• Don't build persistent world models grounded in cause and effect.

Reinforcement learning agents, by contrast:

• Take actions in an environment (digital or physical).
• Get feedback (reward signals) based on performance.
• Learn policies that map situations to actions to maximize long-term reward.

Why scaling alone may hit a wall

The current frontier approach is "just add compute":

• Larger models
• More parameters
• More tokens
• More GPUs

Sutton's argument is that scaling a flawed objective function won't magically give you general intelligence. You'll get:

• Better language imitation
• More convincing chat
• More tools plugged into the model

…but not a system that truly understands goals, constraints, and consequences the way a capable human worker or strategist does.

For businesses, that means:

• LLMs are fantastic for content, summarization, ideation, and coding assistance.
• They are weaker (and riskier) for high-stakes autonomous decision-making without human oversight.

The opportunity: start thinking now about agentic systems—AI that can plan, act, and adapt within clearly defined constraints.

---

2. Inside Sutton's OaK Agent Vision: Goal-Driven AI

To move beyond next-word prediction, Sutton has proposed an agent architecture often referred to as an OaK agent. While the specifics are technical, the business implication is straightforward: the future of AI is goal-driven systems, not static chatbots.

What is an OaK agent (in plain language)?

Think of an OaK-style agent as an AI that:

• Has an explicit Objective (O): a defined goal or reward signal.
• Uses its agent (a) to take actions in an environment.
• Builds Knowledge (K) over time about what works and what doesn't.

Instead of just responding to prompts, an OaK-like agent would:

1. Observe the current state (data, environment, context).
2. Choose an action based on a learned policy.
3. See the outcome (good or bad).
4. Update its internal model to improve future actions.

Why goal-driven agents matter for business

In a practical workflow, an OaK-style agent could:

• Run continuous experiments on email subject lines, landing pages, or ad creatives.
• Automatically adapt campaigns based on real-time performance data.
• Plan multi-step sequences of tasks (research → draft → test → iterate) with minimal human prompting.

Compare that with today's LLM use:

• You prompt it.
• It responds.
• You manually copy, test, and optimize.

Goal-driven agents tie actions directly to business outcomes—opens, conversions, revenue—not just "good-sounding" output.

How to prepare your org for agentic AI

You don't need to wait for theoretical AGI. You can start moving toward agent-style workflows today by:

• Defining clear reward signals
  What metrics really matter? CTR, ROAS, MQLs, NPS, churn? If you can't define the reward, you can't train an agent.

• Instrumenting your funnels
  Make sure your analytics are clean, consistent, and accessible. Agents will need reliable feedback loops to optimize.

• Modularizing tasks
  Break complex work (e.g., a campaign launch) into well-defined steps that an AI agent can operate on.

This is the bridge between Sutton's research and your next generation of AI-powered operations.

---

3. Veo 3 and the "Chain-of-Frames" Moment for Video AI

While the LLM world debates GPT-6, video AI is having its own inflection point. DeepMind's Veo 3 is being framed by some as a "GPT-3 moment for video"—a leap that takes us from gimmicky clips to highly coherent, controllable sequences.

A key concept here is "Chain-of-Frames."

What is Chain-of-Frames?

If LLMs popularized chain-of-thought (models reasoning step by step in text), video models are moving toward chain-of-frames:

• Instead of generating each frame independently, the model tries to maintain temporal coherence over time.
• It keeps track of objects, motion, lighting, and style across many frames, like a storyboard that evolves smoothly.

For marketers and creators, this means:

• More consistent characters and scenes
• Fewer visual glitches and "melting" artifacts
• Longer, more narrative-friendly clips

Why Veo 3 matters for brands and creators

Video is already the dominant content format. With Veo 3–level models:

• Concept-to-video workflows become realistic
  Describe a scene and get a reasonably polished first draft clip.

• Fast iteration becomes the norm
  Test multiple hooks, visuals, or storylines in hours instead of weeks.

• Personalization at scale becomes feasible
  Slightly tweak scenes for different audiences, geographies, or offers.

How to use video AI defensibly

As these tools become widespread, the advantage won't be "we use AI" — it will be how you use it.

Focus on:

• Unique IP and narratives
  Use Veo-like tools to visualize your proprietary insights, stories, and data, not generic stock concepts.

• Tight integration with funnels
  Treat video as a testable asset: A/B test hooks, intros, CTAs, and formats across channels.

• Ethical guardrails
  Deepfake risks are real. Set clear internal rules: no impersonation, no misleading edits, transparent use of synthetic media.

This brings us to the darker side of video AI.

---

4. Deepfakes, Medbeds, and the New Reality Wars

The same techniques that enable Veo 3 to create gorgeous branded content can be used to generate convincing political and medical misinformation.

One recent example making the rounds: a deepfake "medbed" video featuring a political figure, promoting science-fiction style healing technology as if it were real. The clip was posted, went viral, and then was deleted—but not before millions saw it.

Why this matters for your brand

In an environment where:

• Anyone can fabricate a video of anyone saying anything
• Clips can be generated and distributed in hours
• Deletions or corrections never travel as far as the original fake

…trust becomes your most valuable—and fragile—asset.

Defensive moves for organizations

To protect your brand and audience:

• Establish authenticity channels
  Make it clear where official videos come from, and train your audience to check sources.

• Create a rapid-response protocol
  Decide in advance how you'll respond if a deepfake targets your leadership, product, or community.

• Educate your customers and team
  Internally, train teams to recognize typical deepfake tells. Externally, publish clear guidelines on what you will and won't ever claim.

Offensive (ethical) opportunities

On the positive side, the same tech can help you:

• Visualize complex products (like medical devices or SaaS workflows) with hyper-clear explainer videos.
• Localize content for different markets using synthetic presenters that stay on brand.

The key is transparency: synthetic, but disclosed.

---

5. The Wild Compute Curve: OpenAI's 125× Plan

Behind the scenes of all this innovation is a brutal reality: compute and energy. According to recent projections and public comments, OpenAI and its peers are planning up to 125× increases in compute usage for future models.

That level of scaling implies energy consumption that could rival or exceed that of entire nations. Whether or not the exact India comparison holds, the direction is clear:

• More powerful models → exponentially more GPUs and data centers
• More GPUs and data centers → massive energy draw and cooling requirements

Why this matters for strategy

For enterprises, the implications are significant:

• Cost volatility
  Access to frontier models may become more expensive or tiered, especially for high-volume use.

• Regulation and ESG pressure
  Expect scrutiny around the carbon footprint of AI-heavy operations.

• Ecosystem diversification
  It becomes risky to build everything on a single frontier model vendor.

How to build a resilient AI stack

You can future-proof your AI strategy by:

1. Mixing model tiers
   Use smaller, cheaper models for routine tasks and reserve frontier models for high-value use cases.

2. Exploring on-prem or private-hosted options
   For sensitive data and predictable workloads, a mid-sized model you control may beat a giant model you rent.

3. Tracking efficiency, not just capability
   Treat "tokens per dollar" and "latency per request" as first-class metrics in your AI stack.

The upshot: if Sutton is right and pure LLM scaling is a dead end for AGI, then business value will shift from raw model size to smart system design—well-structured agents, data pipelines, and human-AI workflows.

---

6. Turning Today's AI Turbulence into a Strategic Edge

Whether or not LLMs ever reach AGI, they are already reshaping marketing, operations, and product teams. Sutton's critique is a useful reality check: instead of betting everything on ever-bigger chatbots, smart organizations will invest in goal-driven, measurable AI systems.

Key takeaways:

• ChatGPT and similar LLMs are incredible tools—but not magic brains. Use them for what they're good at: language, drafting, summarizing, coding, research.
• The real frontier is agentic AI, like Sutton's OaK-style vision: systems that act, learn from outcomes, and optimize toward clear business goals.
• Video AI like Veo 3 and Chain-of-Frames will transform how brands create and test content, but they also raise deepfake and trust challenges.
• Compute and energy constraints will shape who can access the most powerful models and at what cost.

For your organization, the next steps are clear:

1. Clarify your goals and metrics. Define the rewards your future AI agents should optimize (leads, revenue, retention, satisfaction).
2. Audit your data and tracking. Make sure your funnels are measurable end-to-end so AI can learn from real outcomes.
3. Pilot agent-like workflows. Start small: automated campaign testing, iterative copy optimization, or lead routing agents.
4. Prepare governance and ethics. Build policies for synthetic media, deepfakes, and AI decision-making now, not after a crisis.

The question isn't whether ChatGPT reaches AGI. The question is how you'll turn today's AI capabilities—LLMs, agents, and video models—into compounding strategic advantage while others get distracted by the hype.

If you design for goals, feedback, and trust, you'll be positioned to win no matter which AI paradigm ultimately prevails.