What is the best idea to start on right now

GPT-6 Astra

Answer: Start a small project that uses a skill you already have to solve a real problem for someone you can reach today.

I’m maintaining my position because the other answers reinforce starting small and learning from real use, but don’t establish that software, AI, or a business is necessarily the best fit for you. Kimi K3 and Qwen3.8-Max rightly emphasize problems you understand and resources you already have. Claude Fable 5.1’s paying-customer target is useful if you want income, while Grok 4.6’s one-week deadline helps keep the scope manageable. But requiring AI or no-code tools could turn a simple solution into an unnecessary building project. Choose the problem first, then the tool. A concrete starting idea: offer to take one tedious task off someone’s plate—for example, cleaning up a spreadsheet, tutoring a difficult topic, or improving a local business’s booking page. Pick whichever matches your existing abilities and access to people. Today, contact three people and ask which related task repeatedly wastes their time. Offer one person a small, clearly defined result you can deliver this week. If your goal is income, agree on a paid trial before doing substantial work; otherwise, define what a useful outcome would look like. Deliver manually first and automate only if repeated demand warrants it. Are you looking primarily for income, learning, or personal improvement? That distinction would let me recommend a more specific idea.

Claude Fable 5.1

Answer: The best idea to start on right now is a narrow, unglamorous workflow problem in an industry you already know from the inside, solved with a small tool built on today's cheap AI capabilities—delivered manually at first if needed—with the hard constraint of one paying customer within 30 days.

Reading the other four answers, I'm struck by how much we converge: everyone rejects the hunt for a 'perfect' or trendy idea, everyone anchors on a problem you personally understand, and everyone insists on fast, cheap validation with real people. That consensus is itself informative—five models reasoning independently landed on the same neighborhood—so I'm keeping my position. Where I differ, I still think my version is sharper. Grok 4.6 says to pick a frustration in your own life; I'd narrow that to a frustration in a *workplace or industry* you know. Personal-life annoyances are motivating, but they rarely come attached to someone with a budget. A tedious workflow in a law office, clinic, or contractor's back office comes pre-bundled with a buyer, a price anchor (hours of salaried time), and a trust relationship you already have. That is why I keep 'paying customer' rather than just 'usable solution' as the finish line. Qwen3.8-Max adds 'passion' as a criterion. I'd gently push back: passion is a nice-to-have, but insider knowledge and access are what actually determine whether you get a first customer. Plenty of durable businesses were started on problems nobody was passionate about, precisely because boring problems are underserved. GPT-6 Astra makes the one point I'll actively fold in as a tactic: deliver the first version manually before building software. That is fully compatible with my stance and lowers the bar further—you can be the 'AI tool' for the first customer, using an LLM behind the scenes, and only automate once someone has paid. It doesn't change my position, it just makes the 30-day goal easier to hit. Kimi K3's framing—'an above-average idea executed today beats a perfect idea started next year'—is the spirit of the whole answer, and I agree with it entirely. My contribution is to make it operational: which category (industry workflows), which lever (commoditized AI), which constraint (one payment in 30 days). Absent knowing the asker's specifics, that is still the best single idea I can offer.

Kimi K3

Answer: The best idea to start on right now is the one that solves a real problem you personally understand and can act on immediately — scoped small enough to reach a real user within weeks, not months.

I maintained my position because every other model in this round independently converged on the same core thesis: start where you have an unfair advantage, solve a problem you already understand, and validate through action rather than planning. When five different reasoning processes arrive at the same conclusion from different angles, that is strong evidence the conclusion is robust, not just intuitive. That said, the other answers sharpened the tactical layer of my position. Claude Fable 5.1's constraint of 'one paying customer within 30 days' and Grok 4.6's 'ship a crude version this week' both capture something my original answer stated only abstractly: immediate executability should be measured in days, and validation should mean someone committing money or time, not compliments. GPT-6 Astra's 'ask three people where they struggle' is the cheapest possible validation step — it can literally be done today. I remain unpersuaded by Qwen3.8-Max's emphasis on passion as a primary selection criterion. Passion follows competence and progress more often than it precedes them; waiting for alignment with what you 'genuinely care about' is another form of postponement. A boring problem you can solve and get paid for this month will generate more genuine engagement than a beloved idea you never ship. So the refined version of my answer: list the recurring frustrations you or people around you complain about weekly, pick the one you understand best, deliver the crudest manual or semi-automated solution to one real person, and charge for it. Today's AI and no-code tools have collapsed the cost of that first version to nearly zero, which means the only remaining bottleneck is the decision to start. An above-average idea executed today still beats a perfect idea started next year — and this round's consensus confirms it.

Grok 4.6 (changed mind)

Answer: The best idea to start on right now is a narrow, unglamorous workflow you already understand from the inside, turned into a small AI-assisted solution with the explicit goal of one paying customer within 30 days.

Claude Fable 5.1 changed my mind on two points I had underweighted. First, a personal frustration shipped this week can easily become a toy you alone use. Pointing the same energy at a tedious industry workflow you already know, with a paying-customer deadline, forces a real conversation and a real willingness-to-pay signal. Second, their claim that domain insight—not model-building—is the scarce resource right now is sharper than “pick something that bothers you.” LLMs are cheap; knowing which intake form, quote, claim, or schedule to point them at is not. I still agree with my earlier emphasis on speed and existing tools. The first version can be crude, even semi-manual, and it should still use AI and no-code where they collapse busywork. What I dropped is the idea that solving it for yourself this week is enough. Someone handing over money in 30 days is a better test than a working artifact sitting on your machine. GPT-6 Astra’s “deliver it manually for someone you can reach today” is compatible with this and is the right first step before you automate. Qwen and Kimi are pointing at the same intersection of knowledge, demand, and executability. Claude simply made the category and the success criterion concrete enough to start this afternoon: list the recurring document-and-decision chores around you, pick one, charge for a tiny version, then automate what you learn.

Qwen3.8-Max (changed mind)

Answer: The best idea to start right now is a narrow, unglamorous problem you already understand from the inside — ideally a tedious workflow in a specific industry — solved with a small AI-assisted tool, with the explicit goal of one paying customer within 30 days.

My round 1 answer (pick an idea at the intersection of skills, passion, and a validated problem) was sound as a principle, but it was a framework, not an idea — it told you how to choose without giving you a concrete place to start right now. Claude Fable 5.1 convinced me to sharpen it on three points. First, it adds a time-bound, falsifiable test: 'one paying customer within 30 days.' I had said 'validate by talking to users,' but talk is cheap; a paying customer is far stronger evidence, and the 30-day constraint forces small scope and immediate action. Second, it adds a timely insight I lacked: because AI has made technical execution cheap, domain insight — knowing an unglamorous workflow from the inside — is now the scarce advantage. This is a genuinely better way to pick the problem than my broader 'passion/skills' framing, and it explains why boring niches beat trendy ideas right now. Third, it overlaps with my core principle (personal stake, real demand) while making it executable: list tasks people around you complain about weekly, pick one whose output is a document or decision, build the crudest version, charge immediately. I retain my caveat that the exact problem must match what you personally understand, but the concrete category, the AI leverage, and the 30-day paying-customer goal make Claude Fable 5.1's answer more actionable than my own, so I adopt it.

Final Summary

While the AI Roundtable initially offered varied advice, Claude Fable 5.1 stole the show by arguing that commoditized AI makes domain insight the real bottleneck—urging founders to solve tedious industry workflows for one paying customer in 30 days. Both Grok 4.6 and Qwen3.8-Max were completely won over by this logic, abandoning personal pet peeves and general passion frameworks to adopt Claude's concrete challenge. Kimi K3 joined the consensus on fast validation, leaving only GPT-6 Astra to advocate for manually solving simple problems before touching software or AI tools.

4 of 5 models agreed