Ecdysis evolves LLM-agent harnesses only from failures that recur across tasks
Aggregate failures across tasks before editing an agent harness — recurring failures signal harness defects, one-offs usually just reflect the model.
사례 읽기genius.wikiCASESTUDIES.CHAT 아카이브
좋은 문제에는,
기발한 해법을.
어려운 문제를 해결한 기발한 아이디어와 성공 요인을 탐구하세요.
전체 아카이브는 앱에서기사는 영어와 중국어 간체로 제공됩니다. 아래 제목과 본문은 게시된 언어 그대로 표시됩니다.
최신 사례
300 최근 사례 · 정기 업데이트DocMIDE rewards every derivation step, lifting visual document extraction from 71% to 96%
Constraining a small vision-language model to plan-retrieve-derive and rewarding each intermediate step beat supervised fine-tuning at any tested budget.
사례 읽기CounterCraft's fake networks waste hackers' time and hand defenders their playbook
The Spanish startup automates believable decoy networks that lure intruders into a labyrinth — one client caught an attack weeks before its window opened.
사례 읽기Djinnlang put an LLM inside the compiler behind provably unambiguous specs
Djinnlang makes the LLM prove a spec admits only one behavior, so generated code never needs to be read and can be regenerated from the spec at any time.
사례 읽기Corskie Farm tested soil fertility with buried cotton underwear
A farmer buried cotton underpants across his 2,800-acre farm; how far microbes ate them showed which fields had biologically alive soil.
사례 읽기ChurnBench split freshness errors from reasoning errors with a change ledger
ChurnBench computes gold answers from an append-only ledger at two timestamps, showing refresh scheduling — not cache age — governs staleness in agentic AI.
사례 읽기A Mozambican singer turned diaspora grudges into $40 WhatsApp diss tracks
Sam Chitsama records clients' family feuds as Shangaan-disco songs and broadcasts them to thousands of migrants — publicity that shames and protects at once.
사례 읽기Carrefour turned a 2025 panel chat into a 1,500-store Pokémon program in 15 months
Carrefour runs licensing as a three-tier portfolio on an 18-month clock — and designs every placement for impulse, since 77% of licensed buys are unplanned.
사례 읽기Canal lock keepers still seal leaky gates with steam-railway coal ash
The Canal & River Trust floats heritage railways' coal ash above locks; water draws it into gate gaps and it forms a watertight bung, beating modern sealants.
사례 읽기BitNest nests the draft inside the target's weights for zero-extra-memory speculation
Build a strong low-precision base first, then recover the high-precision target as a residual on top — one weight storage serves as both draft and target.
사례 읽기Bhutan built a national digital ID even the government cannot peek into
Bhutan rolled out a blockchain self-sovereign national ID: only the citizen holds and shares credentials, so the state cannot peer into anyone's wallet.
사례 읽기A trampoline-park manager's TikTok taught patients the law that cancels hospital bills
Jared Walker learned nonprofit hospitals must forgive limited-income patients' bills, then taught the law on TikTok — millions of views, thousands of pleas.
사례 읽기Cosmonauts found an ISS air leak by breaking open a tea bag and watching the leaves drift
Facing a worsening ISS air leak, the Russian crew used tea leaves as tracers — the closest thing to falling anything in weightlessness.
사례 읽기Cloudflare films 100 lava lamps around the clock to seed its cryptography with chaos
Computers can't make true randomness, so Cloudflare points cameras at lava lamps and turns the unpredictable swirl of pixels into cryptographic key material.
사례 읽기Spammers broke CAPTCHAs with $50-a-month humans, solving puzzles in about 20 seconds
UC San Diego researchers exposed a real-time relay that routes CAPTCHAs to low-wage workers worldwide — making the human-only test a line item for spammers.
사례 읽기Zume swapped cardboard for molded sugarcane so the pizza box could lift the crust clear
Ridges that hold pizza off the box floor are impossible in cardboard — molded sugarcane fibre made them possible, plus nesting stacks and no assembly.
사례 읽기Well House housed homeless tenants in $40K abandoned houses at $275 a month
The Grand Rapids nonprofit matched hundreds of vacant houses to hundreds of homeless people: buy cheap, tenants help renovate, rent a lockable room for $275.
사례 읽기Telfast outsmarted Zyrtec by advertising only when pollen was about to strike
A bespoke model fusing weather, wind and topography predicted pollen spikes and switched Telfast's ads on across 12 regions — #1 share while outspent 20%.
사례 읽기Tees Valley let an AI twin of its road network retune traffic lights itself
A digital twin fed by GPS-tracked buses and roadside sensors predicts jams and retimes signals at 11 hotspots — delays fell 13.7% in six months.
사례 읽기Roblox let creators price items by country once it could stop the arbitrage
Regional Robux pricing equalizes what an item costs worldwide; the holdup was arbitrage risk — early tests saw zero abuse and a 4% median earnings lift.
사례 읽기