You've asked three experts the same question and gotten three different answers. One says it's impossible, one says it's already solved, and you're left with nothing to act on.
Here's what you get: four questions that tell you when it's worth testing something yourself, even outside your field, the story of a research firm that did exactly that with scorpion antivenom, and what to check before you spend a cent.
Start with the basic idea. Some questions can't be settled by asking around, because nobody actually has the answer yet. Keep collecting opinions and you just get more opinions, not more certainty. Once the cost of testing it yourself drops low enough, the fastest route to a real answer can be doing work completely outside your field, just to get a straight answer to the question that is in your field.
Four questions before you jump in
Each question checks something different. If one of them fails, testing it yourself is the wrong move, not a shortcut.
- Do experts actually disagree, or have you just not found the right one?If answers range from "it's never worked" to "it'll solve everything", nobody really knows. If most experts say roughly the same thing, you just haven't found the one who knows. Keep looking instead of testing.
- Is the cost of testing small next to the decision riding on it?Compare what the test costs against what you're about to spend or commit based on the answer. If the test is a fraction of that, it's justified.
- Can someone other than the AI confirm the result?A result only the model vouches for isn't a result. You need an independent, human check against the real world.
- Will the answer actually change what you do next?If your decision stays the same either way, skip it. Only test questions where the answer picks your next step.
Citrini Research, a stocks and macro research shop, ran into exactly this. It wanted to know whether AI drug discovery was a real investment theme, and couldn't get a straight read: some experts said AI has never produced an important drug, others said it would cure everything within five years. Instead of asking one more expert, it used Claude to design proteins that bind scorpion toxins, and sent the results to an independent lab.
Claude ran 300 sub-agents that generated 318,000 protein candidates in 64 hours. Of those, 211 went to the lab, and 17 were confirmed to actually bind four different scorpion toxins. The whole thing cost under $200,000, less than a single senior hire for a year.
Wait, before you get excited. Binding a toxin isn't the same as neutralizing it, and that's exactly what critics pointed out: proteins that bind these toxins have existed in the literature for over 15 years, and the step that actually proves a drug works hasn't happened yet. Citrini said as much itself: 64 hours of digital design turned straight into months of human work, DNA synthesis, protein production, real toxins, and another 4 to 7 years before anything could reach a patient. AI moves fast at inventing, not at proving the invention is real. That bottleneck hasn't moved.
Next ten minutes: write down one question where you keep getting contradictory answers from experts, and run it through the four questions above. If it passes all four, that's what you test first this week.