Sumary develops an AI-powered accounting platform that automates bookkeeping, financial reporting and compliance tasks, helping accounting firms streamline operations, reduce manual work and focus more on client advisory services.

AI-driven managed security startup Daylight raises $80M on 40-fold revenue growth - SiliconANGLE

Industrial automation startup Path Robotics raises $300M in funding - SiliconANGLE

Physical AI, dalla teoria alla realtà: ALBA Robot sviluppa veicoli autonomi per aeroporti e ospedali. La sfida della mobilità intelligente.

La startup italiana porta l'intelligenza artificiale nelle chat per gestire appuntamenti, scadenze e documenti legali. Il servizio debutta in cinque Paesi europei, mentre la società supera i 30 milioni di ricavi

.NET 11 is here with C# 15, faster developer tools, runtime and library improvements, and new capabilities for web, mobile, desktop, and AI apps.

As participants talk, Claude can draft the plan being discussed, map out a timeline on a whiteboard, answer questions in the meeting chat, and access connected work apps to retrieve information, such as last year's figures.

Physical AI is the next frontier. But data is a huge bottleneck.

The big challenge is that robots need to see how people work with their hands. That data is hard to collect and label.

Memories Upload from @memories_ai turns everyday first-person video into robot training data.

Impressive paper from Robbyant on physical AI.

(bookmark it)

Robot hands use a range of tactile sensors, and a model trained on one sensor can struggle to transfer to another.

That's a big challenge in robotics right now.

LingBot-Touch is a foundation model to address this challenge. It was pre…

A Blog post by QVAC on Hugging Face

Zyphra’s ZONOS3-Mini is an open-source text-to-speech model with 100M active parameters, cloning and real-time on-device speech generation.

Learn to build a custom harness, folks.

This Microsoft work builds a harness to improve long-context reasoning without synthetic data or post-training.

(bookmark it)

They show that Llama3-8B gains 64.18 points on average across long-context benchmarks with no training at all. The gain comes from…

Sparxity, a German innovator in sparse-native computing, today announced the close of its Series A bringing cumulative funding to €80 million ($90

The deal comes a few weeks after Oura postponed its up to $2.2 billion IPO indefinitely, citing “uncertainty in the IPO market.”

Secondo il rapporto BEI 2026, l'AI generativa è usata dal 48% delle imprese italiane, contro il 67% dell'Ue: l'Italia è terzultima. In più, il 33% non ne prevede l'adozione nei prossimi 12 mesi. E tra le aziende che già la utilizzano, soltanto il 19% vede i benefici principali

Soma presenta Metis, il tutor digitale che migliora l’apprendimento nelle università tramite intelligenza artificiale.

Keep your coding agent up to date with the latest web best practices - GoogleChrome/modern-web-guidance

Adobe is taking on Figma with a new prototype builder tool in Firefly.

Brainstorm on a shared whiteboard, build a spreadsheet, and share dashboards with your team, all with Claude. Available today on all paid plans.

Leaker Aaron Perris has discovered evidence of a new Mac app in the works at Apple that makes it easy to run local AI models.

A Blog post by TNG Technology Consulting GmbH on Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

MiniMax has published MiniMax-H3-Image on Hugging Face, with API, Hailuo and MiniMax Code access, a community licence, and 16-megapixel, 16-reference specs.

Sparse autoencoders (SAEs) are commonly evaluated by variance explained and ablation damage. We show how massive activations undermine both evaluations, and propose a correction. Massive activations cause up to 99.9% of activation variance to lie along the first principal component (PC1). Because the PC1 component has no unique sparse decomposition, each SAE run rebuilds it from a different set of features. These features dominate importance rankings: an average of 47 of the 50 most important features are unmatched in a separate run, compared with about 27 of a random 50. A dictionary can also report 99.96% of variance explained while leaving 33% of the variance outside PC1 unexplained. We bypass PC1 by training the SAE on activations with PC1 removed, then adding PC1 to the reconstruction. In affected models, the top-50 feature sharing rate rises from 4-39.5% to 73-99.5%. At the same TopK sparsity, the resulting SAE degrades the model up to 44% less and reduces content reconstruction error by up to 33.9%.