Meta AI’s Mac beta reads screens and dictates across apps
The beta uses Meta’s Muse Spark model to answer questions from a shared screen.
Modelos, agentes, aplicações: a inteligência artificial posta a trabalhar em problemas reais.
Aqui, a inteligência artificial é julgada pelas provas: o que faz, para quem, a que custo. A secção segue os modelos e os seus progressos mensuráveis, os agentes que executam tarefas reais e, sobretudo, as implantações — um hospital que rastreia mais doentes, uma rede elétrica mais bem gerida, uma administração que responde mais depressa. A IA-espetáculo, as rondas de financiamento sem produto e as profecias não têm aqui lugar.
Preferimos os casos de uso documentados, com os números à vista: quantos exames, que taxa de erro, que poupança. Os atores não ocidentais — laboratórios chineses, indianos, dos Emirados — são tratados como os americanos, porque boa parte da implantação real acontece aí. Cada artigo cita as suas fontes e distingue o protótipo do serviço em produção.
The beta uses Meta’s Muse Spark model to answer questions from a shared screen.
Hierarchical CoAtNet 1 recognized yoga poses with more than 93% accuracy during testing.
The Southampton team found excess and enlarged centrosomes are separate defects that can occupy different tumor regions.
It improved accuracy by 7.9% over its proprietary base model and 9.8% over evaluated open-source models.
IPO-Mine splits filings into sections and extracts charts for machine analysis.
Alipay has released the AHA multi-agent interoperability protocol, enabling agents on different devices to discover one another and pass along tasks.
The model combines 2.8 trillion parameters with a 1-million-token context window.
The opt-in beta can find previously viewed pages, preview them and group duplicate tabs.
The custom harness raised Claude Opus 5 from 30% to 100% on an interactive reasoning benchmark.
The study tested an active-learning system in a live Instagram campaign for an outdoor activities company.
Xiaohongshu’s dots3-note preview activates 16 billion parameters from a 280 billion-parameter MoE model.
An automated score found more changes after treatment than the traditional Lund-Mackay method.
Chinese models accounted for 34.25 trillion of 69 trillion tokens used globally in one week.
Wan-Animate-2 runs at 24 frames per second with weights developers can run locally.
The US-only API offers models from OpenAI, Anthropic, DeepSeek and other providers.
The Google app will let users describe the topics and content they want in Discover.
OpenAI has launched an Apple Messages connector for ChatGPT on Mac, allowing it to search chat histories and draft and send messages.
Startpage can be selected directly as the default search engine on desktop in five European markets.