Why Can't Prompting Make AI Smarter?

According to The Register, the core issue is that large language models (LLMs) are deterministic code—they have a fixed architecture, training data, and weights. Prompting can guide the model to use its existing knowledge more effectively, but it cannot add new reasoning capabilities or expand its factual knowledge beyond what was encoded during training. The Register reported that "prompt engineering is about interface design, not intelligence enhancement." This is a fundamental constraint that no amount of clever phrasing can overcome.

What Does This Mean for Enterprises That Bought Into Prompt Engineering?

Enterprises that have built internal prompt libraries and hired prompt engineers as a core competency face a difficult pivot. The Register's analysis suggests that these investments yield diminishing returns. For example, a company that spent $500,000 on prompt optimization for its customer service chatbot may find that performance plateaus after a few percentage points of improvement. The real bottleneck is the model itself. According to a discussion on Hacker News following the article, developers reported that "switching from GPT-4 to Claude 3.5 gave a 40% improvement in task completion, while six months of prompt tuning on GPT-4 only yielded 5%." This data point underscores that model selection matters far more than prompt craftsmanship.

Who Benefits From This Reassessment?

The winners are companies that focus on model fine-tuning, retrieval-augmented generation (RAG), and architecture innovation. For instance, Anthropic and Google DeepMind, which invest heavily in model improvements rather than prompt wrappers, stand to gain. The losers include the growing ecosystem of prompt engineering consultancies and tools that promise 'prompt optimization as a service.' The Register's piece implicitly calls into question the entire business model of companies like PromptBase and HumanFirst, which sell prompt templates and optimization services.

Approach	Cost	Capability Ceiling	Long-Term Viability
Prompt Engineering	Low (labor only)	Fixed by model	Short-term workaround
Fine-Tuning	Medium (compute + data)	Improves domain performance	Medium-term
Model Selection	Variable (API costs)	Highest raw capability	Best immediate ROI
Architecture Innovation	High (R&D + compute)	Unlimited (new capabilities)	Long-term solution
Verdict	Model selection and fine-tuning outperform prompt engineering for sustained gains. Architecture innovation is the only path to fundamentally smarter AI.

What Should Developers Do Now?

The practical takeaway is clear: stop treating prompt engineering as a silver bullet. Instead, developers should benchmark multiple models for their specific task, invest in RAG for up-to-date knowledge, and consider fine-tuning for domain-specific tasks. The Register's article suggests that the marginal cost of a better model is often lower than the cumulative cost of prompt optimization. For example, upgrading from GPT-4 to GPT-4 Turbo might cost 2x more per token but deliver 3x better accuracy on complex tasks—a far better ROI than hiring a prompt engineer.

Is This the End of Prompt Engineering as a Discipline?

Not entirely. Prompt engineering still has value for rapid prototyping and for models that cannot be fine-tuned easily. But its role should be downgraded from 'core capability' to 'tactical tool.' The Register's analysis is a wake-up call for anyone who believed that better prompts alone could unlock AGI. The future of AI progress lies in code—better architectures, larger datasets, and more efficient training methods—not in better phrasing.

1. By Q1 2027, at least two major prompt-engineering consultancies (e.g., PromptBase or HumanFirst) will pivot to fine-tuning or RAG services, or shut down.
2. By Q2 2027, enterprise AI spending will shift from prompt optimization tools to model evaluation platforms like LangSmith and Weights & Biases, which focus on model selection and fine-tuning.
3. By Q3 2027, at least one major LLM provider (likely Anthropic or Google) will release a model that explicitly downplays prompt engineering in its documentation, emphasizing code-level improvements instead.

• Prompt engineering is a tactical tool, not a strategic capability. The model is the product.
• Enterprises should benchmark models first, optimize prompts second.
• The prompt engineering industry is a bubble that will deflate as the market wakes up to architectural constraints.
• Fine-tuning and RAG offer better ROI than prompt optimization for domain-specific tasks.
• The next leap in AI capability will come from code changes, not prompt changes.