the AGI race became a prompting race

We were promised AGI. Machines that think. Systems that reason. The next step in intelligence.

What we got instead is a competition to write better system prompts.

The Great Pivot

Somewhere around 2023-2024, the product narrative shifted hard toward “capabilities.” Labs did not literally abandon AGI—OpenAI still puts it in its charter—but the stuff users could buy was increasingly framed around instruction following, tools and measurable tasks.

And honestly? That's not a bad thing. But let's be honest about what happened.

What AGI Was Supposed To Be

The working definition I had in mind:

• General reasoning - Apply intelligence to novel problems without specific training
• Transfer learning - Skills in one domain naturally extend to others
• Autonomous goal pursuit - Set a goal, figure out the steps, execute
• Self-improvement - Get better without human intervention
• Understanding - Not just pattern matching, but genuine comprehension

This was the dream that launched a thousand research papers.

What We Actually Built

What we have now:

• Autoregressive language models - Predict the next token, then add post-training and inference systems
• In-context learning - Follow examples in the prompt
• Instruction following - Do what the system prompt says
• Tool use - Call APIs when prompted to

These are genuinely impressive. Whether they add up to AGI depends on whose definition you use; they do not match the version I had in mind above.

The System Prompt Wars

Here's where it gets interesting. The model still matters—a lot—but product behavior increasingly comes from the whole harness around it: prompts, tools, retrieval, memory, evals and permissions.

Same model, different harness, completely different product.

Claude, GPT and Gemini share transformer ancestry, but they are not the same base model. The product magic is in the weights and post-training plus everything wrapped around them:

You are a helpful assistant that...

That little block is not the entire moat. It is just the most visible part of a much larger system.

The Prompt Engineering Industrial Complex

A new discipline emerged:

• Prompt and eval specialists writing and testing instructions
• System prompt libraries being treated as trade secrets
• A/B testing prompts like landing pages
• Prompt injection attacks as a security category
• Jailbreaking as a competitive sport

We didn't build thinking machines. We built really good instruction followers. And now we're competing on who can write the best instructions.

The Capability Illusion

Modern LLMs create behavior that is easy to over-interpret:

They can code - Sometimes brilliantly, sometimes by reproducing familiar patterns without a stable model of the system

They can reason - But a fluent chain-of-thought is not proof that the visible explanation matches the mechanism that produced the answer

They can learn in-context - But that temporary adaptation is not the same thing as updating model weights

They can use tools - But only through interfaces the surrounding system exposes

The impressive demo is usually a model-plus-system result, not the model acting alone.

Why This Happened

The pivot makes sense when you understand the economics:

1. Scaling laws worked - More compute produced predictable improvements.
2. RLHF and instruction tuning worked - Post-training made models much more useful.
3. Products shipped - ChatGPT proved you could monetize this.
4. AGI is hard - True reasoning is unsolved. Prompting is a workaround.

The path of least resistance was clear: stop trying to solve general intelligence, start making really good chatbots.

The Uncomfortable Truth

Here's what nobody wants to admit:

We don't know how to build AGI. My bet is that scaling alone will not settle it, and better prompts definitely will not settle it by themselves.

Current AI is deeply brittle. Change the prompt slightly and watch the model break. That's not intelligence - that's sensitivity to input formatting.

Benchmarks get optimized against. Contamination, narrow tuning and repeated leaderboard pressure can make a score less representative of real use. That is why I care more about task-level evaluation systems like lvl than one magic aggregate number.

What This Means For Builders

If you're building with AI:

1. The harness is part of your product - Invest in prompts, tools, evals and failure handling accordingly
2. Don't assume intelligence - Your model is following instructions, not thinking
3. Build for failure modes - Every edge case is a prompt failure
4. Stay model-agnostic - Today's best model is tomorrow's commodity

The Path to Real AGI

What might actually move the needle:

• New architectures - Transformers might not be the final answer
• World models - Systems that understand cause and effect
• Active learning - Models that seek out information they need
• Compositional reasoning - Building complex thoughts from simple primitives

But these are research problems, not prompt engineering problems.

Closing Thoughts

The AGI race became a prompting race because prompting works well enough for products. And products make money.

That's not cynical - it's practical. We build with the tools we have.

But let's stop pretending that better system prompts alone are steps toward AGI. They are steps toward more useful products, and that is still worth doing.

Just don't confuse utility with intelligence.

The machines aren't thinking yet. They're just really good at following instructions.

And the humans writing those instructions? They're the real intelligence in the system.

At least for now.

For the builder version of this argument, Simplicity the Saviour gets into why fewer tools and clearer context often beat a giant prompt.