Ai wizards judging Ai

🚨BREAKING: $2.8T Citi partners with Coinbase, US' largest crypto exchange, to allow stablecoin payments for institutional clients. The partnership will let Citi clients convert fiat into crypto or stablecoins, send payments onchain, and convert back into fiat when needed. Citi brings a payments network across 94 markets and 300+ clearing systems while Coinbase brings $246B in assets on platform. Citi also projects stablecoin issuance could hit $1.9T by 2030 in its base case.

Power is about to stop being a utility you assume and start being a product you specify. Plants, data halls, and shops that can hold voltage when the schedule slips will outrun anyone still waiting for a national plan. $AMSC

The factory of the next decade won’t wait for a 400-mile supply chain to unjam. Print the circuit, print the fixture, assemble tonight. Local additive manufacturing turns “we’re out of stock” into a two-hour problem. $NNDM

SemMSA: Latent Semantic-Aided Robust Multimodal Sentiment Analysis with Incomplete Data A latent-semantic multimodal sentiment model uses a frozen LLM and spectral alignment so missing… Score 88/100 · free explainer disruptive-concepts.com/arti…

Quantum Feature Selection for Biomedical Data Analysis A QUBO feature-selection recipe for metabolomics is run on gate-based quantum hardware and… Score 90/100 · free explainer disruptive-concepts.com/arti…

Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning A temporal gradient-inversion attack reconstructs private robot observation-action traces from the… Score 93/100 · free explainer disruptive-concepts.com/arti…

TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations A 3D point tracker follows every visible point for more than a thousand frames inside 40 GB of GPU… Score 75/100 · free explainer disruptive-concepts.com/arti…

This is 💯 true.

Disruptive cutlery

May we present the ‘Self licking Ice Cream Cone’

Disruptive partying

先ほど、午後9時から、トランプ大統領と電話会談を行いました。 トランプ大統領からは、習近平国家主席の訪米直後というタイミングで電話をいただき、今般の米中首脳会談について詳細に説明がございました。 その上で、トランプ大統領との間で、経済安全保障を含む経済や安全保障など、中国をめぐる諸課題を中心に意見交換を行い、日米の連携を確認しました。 今週22日のニューヨーク訪問時の首脳会談に続き、一週間の間に二度の首脳間の会談を実施したことは、まさに日米同盟の力強さを示すものと考えています。 日本と米国は、今や最も信頼し合える同盟国となっており、共に、地域、そして世界の平和と安定に貢献していくことを確認しました。 揺るぎない日米同盟を更なる高みに引き上げていくため、トランプ大統領と緊密に連携してまいります。

AI titan Anthropic has launched a biology ‘wet lab’ where human scientists and AI agents will work together to design and conduct experiments. The company has released one of the team’s first finds: a peculiar pattern of DNA in the genomes of several giant viruses. go.nature.com/46K4qmC

🚨 BREAKTHROUGH: Scientists at the University of Nottingham have developed a new enamel-repairing gel that starts restoring teeth in just 2 WEEKS. This could replace fillings and change dental treatment worldwide, with use expected around late 2026–2027. Here is how it works and when you can get it.🧵

May Starship have the Mandate of Heaven

Self-Adaptive VLA for Robust Robot Deployment A vision-language-action policy adapts on the fly to worn or miscalibrated hardware by treating… Score 79/100 · free explainer disruptive-concepts.com/arti…

SemMSA: Latent Semantic-Aided Robust Multimodal Sentiment Analysis with Incomplete Data A latent-semantic multimodal sentiment model uses a frozen LLM and spectral alignment so missing… Score 88/100 · free explainer disruptive-concepts.com/arti…

TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations A 3D point tracker follows every visible point for more than a thousand frames inside 40 GB of GPU… Score 75/100 · free explainer disruptive-concepts.com/arti…

IBM built an AI search that finds the right section of a book 82.6% of the time on the first try. The paper is called STAIR, out of IBM Research. When you ask an AI about a long document today, it usually runs RAG. The system cuts the document into equal-sized chunks, turns each chunk into numbers, and hands back the chunks that look closest to your question. It throws away the chapters, sections and headings on the way. Your 500-page manual becomes a pile of loose paragraphs. STAIR keeps the structure. The model sees the book's table of contents and learns to answer each question with the right section title. The book's own outline becomes the index, so the system skips embeddings and vector databases. To test it, the team built a benchmark from 18 books across law, medicine, finance, education and the sciences, with tens of thousands of questions. The best comparable method got 76.9%, standard dense retrieval got 68.7%, and keyword search got 59.5%. An off-the-shelf Mistral model got 13.8%. The viral posts lead with the hallucination rate. STAIR pointed to a section that doesn't exist in 0.05% of answers. The closest method did that in 3.25%, which is where the "65x" comes from. That number is about finding the right section. The paper doesn't test whether the answer built from that section is correct. The viral posts also skip the limits. STAIR needs a document with a table of contents. The team fine-tuned a 7B model on each book separately, 200 passes per book. They haven't tested it on company-sized collections. The idea is the part worth keeping. Authors spend months organising a book into chapters and sections, and STAIR puts that work to use. Most AI search ignores the table of contents. This paper shows it's the best index you already have.

The next plant will not be a river valley. Compact reactors are being designed to sit beside a mill, a mine, or a data hall — baseload you do not import and a schedule you do not beg for. $NNE