We have a "Hallucination" problem. Or at least, that’s what everyone keeps saying. We complain that ChatGPT makes up facts. We mock Gemini for citing court cases that don't exist. We treat Large Language Models (LLMs) like faulty search engines or drunk interns. But what if I told you that we are using the tool wrong? We are obsessed with using AI as an Oracle—a machine that spits out the one true answer. But a fascinating new perspective from the AI research blog argmin.net suggests we should be doing the exact opposite. Oracle argmin.net We shouldn't be asking AI to solve our problems. We should be asking AI to give us better problems. We should be asking AI to give us better problems. This is the era of "Prompts for Open Problems," and it’s about to change how we do science, coding, and engineering. "Prompts for Open Problems," The "Jeopardy" Trap Here is the current state of AI usage: User: "Write a Python function to sort this list."
AI: "Here is def bubble_sort(arr): ..." User: "Write a Python function to sort this list." User: AI: "Here is def bubble_sort(arr): ..." AI: def bubble_sort(arr): ... This is boring. It’s a "Homework Problem." The answer already exists on StackOverflow; the AI is just retrieving it. But real innovation doesn't happen by retrieving known answers. It happens by asking questions that nobody has asked before. Ben Recht, the author behind argmin.net, argues that finding a good problem is often mathematically harder than solving it. argmin.net Checking if a sorting network works? Easy. A computer can do that in milliseconds.
Finding a novel sorting network that is faster than anything humans have discovered? Extremely hard. Checking if a sorting network works? Easy. A computer can do that in milliseconds. Checking Finding a novel sorting network that is faster than anything humans have discovered? Extremely hard. Finding This is where the hallucination "bug" becomes a feature. ![Image Description: A stylized illustration of a robot tossing thousands of crumpled paper balls (ideas) into a hoop. A human stands by the hoop, uncrumpling only the ones that made it in. The background is a blackboard with complex math formulas.] (Caption: Let the AI generate the chaos. You verify the brilliance.) The "P vs NP" of Creativity To put this in developer terms: Creativity is an NP problem. Creativity is an NP problem. In computer science, P problems are easy to solve. NP problems are hard to solve, but easy to verify. P NP AI is terrible at truth because truth is strict. But AI is amazing at generating plausible-sounding noise. If we flip the script, we can treat the AI as a High-Speed Hypothesis Generator. amazing High-Speed Hypothesis Generator Instead of asking: "What is the optimal matrix multiplication algorithm?" (The AI will likely hallucinate or recite an old one). "What is the optimal matrix multiplication algorithm?" We ask: "Generate 1,000 weird, non-standard variations of matrix multiplication steps." "Generate 1,000 weird, non-standard variations of matrix multiplication steps." Then, we write a simple script (a Verifier) to test them. 999 will be garbage. But 1 might be a breakthrough. Verifier This isn't theoretical. This is exactly how Google DeepMind’s FunSearch worked. It didn't "know" the answer to the Cap Set problem. It hallucinated millions of code snippets until one of them happened to solve the math problem better than any human ever had. FunSearch happened Why "Prompting for Problems" is the New Meta If you are a developer or researcher, you need to pivot your workflow. Stop using AI as a documentation lookup tool and start using it as a Search Engine for the Unknown. Search Engine for the Unknown. 1. The Fuzzing Revolution Don't ask the AI to fix your code. Ask the AI to break your code. break "Here is my API schema. Generate 50 malicious edge-case JSON payloads that might cause a race condition or a memory leak." "Here is my API schema. Generate 50 malicious edge-case JSON payloads that might cause a race condition or a memory leak." "Here is my API schema. Generate 50 malicious edge-case JSON payloads that might cause a race condition or a memory leak." The AI is great at imagining disaster scenarios you didn't think of. It generates the "problem" (the exploit); you solve it (the patch). 2. The "Counter-Example" Hunter In research, we often look for a pattern. "Generate a graph structure that satisfies condition X but violates condition Y." "Generate a graph structure that satisfies condition X but violates condition Y." "Generate a graph structure that satisfies condition X but violates condition Y." The AI can sweep through the "search space" of possibilities much faster than you can sketch on a whiteboard. It proposes candidates; you check the math. 3. Design Space Exploration "Propose 10 alternative architectures for this system that prioritize latency over consistency, even if they look unconventional." "Propose 10 alternative architectures for this system that prioritize latency over consistency, even if they look unconventional." "Propose 10 alternative architectures for this system that prioritize latency over consistency, even if they look unconventional." Most will be bad. One might uncover a trade-off you hadn't considered. ![Image Description: A digital landscape showing a 'Search Space'. Most of the terrain is dark fog. A spotlight beam (labeled 'AI') is frantically darting around, illuminating hidden gems in the darkness.] (Caption: AI is a flashlight in the dark forest of possibility.) The Death of the Oracle, The Rise of the Muse We need to lower our standards for AI accuracy and raise our standards for AI diversity. accuracy diversity If you treat AI as an Oracle, a single mistake is a failure. If you treat AI as a Muse (or a Stochastic Generator), a single hit is a massive success. The argmin article puts it perfectly: We are moving from a world where we prompt for solutions to a world where we prompt for open problems. argmin solutions open problems The next Nobel Prize won't be won by an AI. It will be won by a human who used an AI to ask a question so strange, so novel, and so "hallucinated" that it turned out to be the missing piece of the universe. So, stop correcting your AI. Let it get weird. The bugs might just be features in disguise. 5 Takeaways for Developers: Hallucination is a Feature: Use AI’s randomness to explore search spaces (ideas, code paths) you wouldn't touch.
Verify, Don't Trust: The workflow is AI Generates -> Script Verifies -> Human Adopts.
Reverse Prompting: Instead of asking for the code to solve X, ask for 10 edge cases where X fails.
Cheap Generation: Generating 1,000 bad ideas costs pennies. Finding 1 good idea is priceless.
The Shift: Move from "How do I do this?" to "What are the possibilities?" Hallucination is a Feature: Use AI’s randomness to explore search spaces (ideas, code paths) you wouldn't touch. Hallucination is a Feature: Verify, Don't Trust: The workflow is AI Generates -> Script Verifies -> Human Adopts. Verify, Don't Trust: AI Generates -> Script Verifies -> Human Adopts Reverse Prompting: Instead of asking for the code to solve X, ask for 10 edge cases where X fails. Reverse Prompting: Cheap Generation: Generating 1,000 bad ideas costs pennies. Finding 1 good idea is priceless. Cheap Generation: The Shift: Move from "How do I do this?" to "What are the possibilities?" The Shift: Liked this perspective? Smash that clap button and follow me for more deep dives into the philosophy of AI. Liked this perspective? Smash that clap button and follow me for more deep dives into the philosophy of AI.

This story contains AI-generated text. The author has used AI either for research, to generate outlines, or write the text itself. 

Google

Stop Asking AI for Answers: Why the Future of Research is "Problem Generation"

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Google Just Killed LeetCode: Why AlphaEvolve is the End of the "Algorithm Interview"

My Unorthodox Tech Stack and How a Marketing Degree Makes Me a Better React Developer

The Coming Shift: Why Multi-Agent Systems Will Redefine Work in 2026

Progressive Context Disclosure and the Token Problem

Generative AI: The Future of Creativity and Problem-Solving is Here

The Noonification: Use This 7-Step McKinsey Framework to Solve Any Problem (1/10/2023)

Google Just Killed LeetCode: Why AlphaEvolve is the End of the "Algorithm Interview"

My Unorthodox Tech Stack and How a Marketing Degree Makes Me a Better React Developer

The Coming Shift: Why Multi-Agent Systems Will Redefine Work in 2026

Progressive Context Disclosure and the Token Problem

Generative AI: The Future of Creativity and Problem-Solving is Here

The Noonification: Use This 7-Step McKinsey Framework to Solve Any Problem (1/10/2023)

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps