“Find an angle nobody else would think of.” Give that instruction to an AI and it will happily produce a long list: collateral auctions, redemption queues, stranded assets, unusual exchange deposit routes, claims on dissolved organizations, and tokens representing baskets of assets.
It sounds like a treasure map. Then comes the awkward question: if humans cannot think of the opportunity, how exactly is an AI supposed to?
Telling it to “be even more creative” might eventually produce a plan to redeem spaceship fuel on a blockchain. Great pitch. Where is the cash flow?
The practical answer is to use creative methods to widen the search, then let real prices, contracts, deadlines, and redemption rules determine which anomalies deserve investigation.
Why Asking for Originality Can Produce the Same Old Ideas
Language models learn from enormous amounts of writing, but their answers often cluster around familiar patterns. A 2026 ICML study tested a technique called Verbalized Sampling, which asks a model to state several answers with associated probabilities. In creative-writing tasks, output diversity rose by 1.6–2.1 times compared with direct prompting. That is a creativity result, not a measured increase in trading returns. [1]
Another 2026 preprint searched for solutions through analogies to structurally similar problems in other scientific fields. It reported improvements of 90–173% in particular diversity measures. [2]
These techniques help AI venture off the beaten path. But an unusual idea and a profitable trade are still two separate achievements.
The Gap Between a Brilliant Pitch and a Working Result
At ICLR 2025, a study involving more than 100 NLP researchers found that AI-generated research proposals received higher novelty scores than human proposals in blinded reviews, with somewhat weaker feasibility scores. [3]
A follow-up execution study assigned ideas to 43 expert researchers. Once those ideas were implemented and evaluated, the scores of AI-origin ideas declined more sharply. [4]
Neither experiment was a DeFi profit test. The transferable lesson is simpler: “I have never heard of this” is not the same as “I can actually collect the money.”
Researchers are also developing ways to compare new ideas against prior literature and revise duplicates, or to make specialized AI agents debate and refine competing hypotheses. [5][6] This is a good reason to separate idea generation from verification.
What Counts as DeFi “Dumpster Diving”?
Decentralized finance, or DeFi, refers to blockchain-based systems for exchanging, lending, borrowing, and redeeming assets. “Dumpster diving” here is a metaphor for looking at overlooked, small opportunities and complicated claims, not a recommendation to exploit people or bypass rules.
Seven useful hunting grounds are collateral auctions; small redemptions; winding-down protocols; new markets with unfamiliar technical tools; exchange deposit pathways; claims and DAO dissolution proceeds; and basket-backed tokens or scheduled market windows. These categories may overlap. What differs is why other participants might consider the opportunity too annoying.
A hypothesis might be: “Could a token be cheap because holders have not read the redemption paperwork?” Or, “Could a venue show a higher price that is unreachable because deposits are restricted?”
Crucially, an exchange accepting a deposit does not automatically create a legal redemption right. Identity, supported networks, service terms, withdrawal restrictions, and beneficial ownership must all match.
Let AI Find the Strange Number Before Inventing a Story
Imagine a hypothetical token priced at $97 on one venue while a documented process appears to return $100 of value to eligible holders. The displayed gap is $3.
A brainstorming agent shouts, “Arbitrage!” A discovery agent asks: Who is eligible? How long will settlement take? What will fees cost? Is the $100 quoted in comparable units? Can you buy and exit at the stated size? Could the redemption be suspended?
Only after those questions does the price difference become a credible candidate. Net profit requires subtracting fees, execution costs, and expected losses from price movements. Tail risks need separate treatment.
A repeatable workflow is: collect market and official-document data; normalize assets, rights, units, and timestamps; flag mismatches; investigate why competing bots have passed; then independently test the failure conditions and, where permitted, run a small verification.
Find first, explain second. The unexplained observation becomes the seed of the new hypothesis.
Seven Scouts Should Not Become Seven Cheerleaders
Seven agents can still make one mistake seven times. If they all read the same summary and vote “promising,” the vote proves little.
Give them different jobs: unusual candidates, prior-art checks, official rights verification, executable prices and depth, exchange restrictions, adversarial risk review, and reproducibility. Promote an idea only when independent evidence converges, not when agents agree with one another.
Compare this approach against ordinary brainstorming using the same time budget. Count new executable opportunities per 100 candidates, then track actual net profits rather than attractive narratives.
The Punchline: A Broom Can Beat a Crystal Ball
The aim is not to make AI look like a genius in a meeting. It is to uncover one overlooked discrepancy that survives contact with reality.
Flashy high-speed trading may resemble Formula 1. This work looks more like the lost-property counter after closing time: unglamorous, full of paperwork, and surprisingly difficult. The important question is whether the item can actually be claimed.
Research sources: [1] Verbalized Sampling, ICML 2026 · [2] Analogical Reasoning, 2026 preprint · [3] AI vs human ideas, ICLR 2025 · [4] Execution gap, 2025 preprint · [5] SciMON, ACL 2024 · [6] Co-Scientist, Nature 2026
