Slugs kept shorting Stretford's traffic lights — seals and copper tape ended the outages
A year of signal failures at a Trafford junction traced to slugs seeking warmth; sealing the controller and copper tape ended the outages.
Read the caseIngenious ideas that solved difficult problems. Explore what made them work.
The full archive, in the appA year of signal failures at a Trafford junction traced to slugs seeking warmth; sealing the controller and copper tape ended the outages.
Read the casemacOS 27 removed the Apple Intelligence off-switch, so one developer reached for Apple's own configuration-profile mechanism to remove it reversibly
Read the caseIn a Rest of World interview, co-founder Anirudha Jalan describes a waste marketplace keeping recyclers supplied via on-site checks and moisture data.
Read the caseReinforcement learning from prediction feedback makes clinical-note summaries keep decision-relevant evidence that generic, fluency-tuned summaries drop.
Read the caseAfter Gulf states banned under-18 camel jockeys in 2005, Qatar-funded RAQBI got owners to actually use robots: cheap, faceless, trivially copyable.
Read the casePlurPO has the model simulate the stakeholders affected by a user's plan and trains it on responses they could all accept — no ground-truth labels needed.
Read the caseNurse Gillian Taylor embedded sensors in a standard transfer board, so every trolley-to-bed move weighs the patient — no extra step before critical drugs.
Read the caseCookware maker Our Place copied the lotus leaf's microstructure so food can't stick — a PFAS-free pan whose nonstick is the metal itself, not a coating.
Read the caseCompression and recall needs are already encoded in a model's internal representations before each action, so memory management can be steered proactively.
Read the caseMinimum Risk Training, from 2016 machine translation, optimizes sequence-level metrics directly, lifting Qwen2.5-7B-Instruct from 16.20% to 68.95% accuracy.
Read the caseIn 114%-inflation Argentina, the e-wallet invests stored money in funds by default and credits returns daily — so users never forfeit gains when they withdraw.
Read the casePretraining the ViT through report generation with a frozen LLM closes the semantic-interface gap left by CLIP pretraining and transfers across downstream LLMs.
Read the caseA Proposer picks and configures explanation tools; an Actor is optimized against faithfulness rewards, turning tool outputs into natural-language explanations.
Read the caseAfter 24 surgeries, industrial designer Ubiratan Bizarro fixed the mechanics instead of the neurology — and João Carlos Martins played all ten fingers again.
Read the caseBinary masks over FFN dimensions, one semantic domain at a time, turn a pretrained dense LLM into a 64-expert MoE running 2-2.5x faster.
Read the caseAn informal tutor economy sells $6-$30 AI classes to South Asian shopkeepers and freelancers through closed WhatsApp groups, undercutting formal courses.
Read the caseA graph-based memory stores reusable strategies and retrieves only the relevant subgraph per query, so token cost stays bounded while the memory grows.
Read the caseExpert-graded metrics scaled by prediction-powered LLM judging feed a REJECT/CONDITIONAL/SCALE gate; a global-bank pilot made reviewer catch rate measurable.
Read the caseA 38x smaller self-supervised model running at 14.7 FPS on a CPU alone produced better costmaps than a deployed foundation-model system on 23 of 24 field logs.
Read the caseSeparate the continuity-bearing substrate (identity, memory, body) from replaceable execution bindings, so an agent survives model, harness and host swaps.
Read the case