★ Top Story
TechCrunch AIMeta has released Muse Spark Open, which it claims is the most capable US open-weight model to date, featuring a multimodal mixture-of-experts architecture. In a notable cross-platform partnership, Meta and Apple collaborated to make the model run locally on Apple silicon via the MLX framework, requiring two M5 Ultra Mac Studios. The open-weight release marks a significant escalation in the race between proprietary and open frontier models.
Read the original →
TechCrunch AIAnthropic is allowing independent AI safety organization METR unprecedented access to its internal systems, equivalent to what its own employees can see. Crucially, Anthropic has contractually surrendered the right to block or suppress METR's published assessments of its risks, incidents, and practices. The move sets a new benchmark for third-party AI safety auditing and could pressure other frontier labs to follow suit.
Read more →
Ars TechnicaThe Federal Trade Commission has launched investigations into both Anthropic and OpenAI, examining whether security incidents at the two leading AI labs caused harm to American consumers. The probe comes despite the Trump administration previously dismissing AI safety concerns, and regulators are reportedly considering fines comparable to those levied against companies that delay disclosing data breaches. The investigations signal a potential shift toward more aggressive federal oversight of the AI industry.
Read more →
NVIDIA BlogNVIDIA has released Nemotron 3.5 Super VL, a 120-billion-parameter open model targeting specialized always-on agentic workloads, alongside two new edge-optimized Bolt models designed to run on local hardware with as little as 4GB of memory. The release spans the full deployment spectrum from data center to desktop, reinforcing NVIDIA's push to make its model family competitive across every tier of the AI stack. The Bolt models are specifically tuned for local agentic tasks where latency and memory constraints are critical.
Read more →
The VergeOpenAI has upgraded its Google Chrome extension to allow ChatGPT to view a user's currently open browser tabs when explicitly tagged in a conversation. The feature effectively replaces the scrapped ChatGPT Atlas browser project, bringing contextual browsing awareness directly into the existing extension. The capability raises fresh questions about browser privacy and the expanding ambient reach of AI assistants into users' daily workflows.
Read more →
Hugging FaceIBM has released Granite Switch 4.1, a new model that consolidates chat, structured decision-making, and speech capabilities behind a single API endpoint. The unified architecture is designed to simplify enterprise deployments by eliminating the need to orchestrate multiple specialized models for different modalities. The release continues IBM's strategy of positioning Granite as a practical, enterprise-grade open model family.
Read more →
The VergeNVIDIA's RTX Spark platform has officially launched, representing a new class of Windows PCs co-optimized with Microsoft specifically for local AI workloads, creative tasks, and gaming. Microsoft has made targeted changes to Windows itself — including updates to core apps like Notepad — to take advantage of the platform's on-device AI capabilities. The launch marks one of the most significant hardware-software co-design efforts for consumer AI PCs to date.
Read more →
TechCrunch AIBluesky has broadened its AI-powered Attie platform beyond custom feed creation to support what it calls "social vibe coding" — letting users build interactive games, quizzes, polls, and full websites that can be shared directly within the Bluesky social network. The expansion positions Bluesky as a platform for lightweight social app creation, not just content consumption. CEO Toni Schneider framed it as a way to make the open social web more participatory and creative.
Read more →
TechCrunch AIAmazon-owned Bee has unveiled its first new hardware since being acquired, including a refreshed wearable and a credit-card-sized recorder that clips onto a smartphone. Both devices are designed to passively capture and summarize conversations, generate to-do lists, and integrate with Alexa and Amazon's shopping ecosystem. The launch intensifies competition in the ambient AI wearable space and raises ongoing consent and privacy concerns about always-on recording devices.
Read more →
TechCrunch AISpeaking at TechCrunch Disrupt, Amazon's head of devices Panos Panay outlined a vision where ambient AI hardware anticipates user needs in the background, making the traditional app paradigm obsolete. Panay argued that the next generation of AI devices will shift from reactive tools that wait for commands to proactive systems that act on context. The comments signal Amazon's broader hardware and AI strategy as it competes with Apple, Google, and Meta for the ambient computing future.
Read more →
TechCrunch AIYouTube Music has started applying labels to tracks that are fully or partially generated by AI, including songs where AI contributed to composition or vocals, with the labels appearing directly on the player page. The move comes as streaming platforms face a growing deluge of AI-produced content that is difficult for listeners to distinguish from human-made music. Competitor Deezer also announced it will roll out an opt-in filter to hide AI-generated music entirely by year's end.
Read more →
The VergeA growing wave of criticism is targeting Meta's Muse AI personal assistant and similar products for ingesting data about third parties — friends, colleagues, bystanders — who never agreed to be part of any AI system. Critics argue that the design of these "personal AI" products structurally externalizes privacy costs onto people who have no say in the matter. The controversy is fueling a broader debate about whether consent frameworks built for individual users are fundamentally inadequate for ambient AI.
Read more →
TechCrunch AIStartup Ricursive Intelligence is developing AI systems that can autonomously design semiconductor chips, learn from each design iteration, and progressively improve subsequent generations — with the explicit goal of compressing the chip design cycle. The company is targeting the structural bottleneck in AI scaling: the slow, expensive, and human-intensive process of designing the hardware that AI itself runs on. If successful, the approach could create a feedback loop where AI accelerates its own compute infrastructure.
Read more →
Hugging FaceNVIDIA and collaborators have published details of a large-scale pretraining effort for Indic languages, built on a 2.8 trillion token dataset alongside an open data generation pipeline. The release addresses a significant gap in multilingual AI, where South Asian languages spoken by hundreds of millions of people have been systematically underrepresented in frontier model training. Making the pipeline open-source allows other researchers to extend and replicate the work for other low-resource language families.
Read more →
TechCrunch AICybersecurity startup Apate.ai has taken an unconventional approach to fraud detection: deploying AI agents that deliberately engage with scammers in order to map their tactics, infrastructure, and scripts in real time. The intelligence gathered is then used to help organizations proactively defend against fraud before it reaches real victims. The strategy flips the typical defensive posture of AI security tools into an active, intelligence-gathering operation.
Read more →
TechCrunch AIFlock Safety CEO Garrett Langley publicly claimed the company's surveillance microphones detect only gunshots, but the company's own website describes broader audio detection capabilities, contradicting his statements. The discrepancy has intensified scrutiny from the ACLU and privacy advocates, and a separate lawsuit filed by a Georgia man alleges that Flock's license plate reader network was used to enable warrantless tracking and retaliation by a sheriff's office. The cases are converging into a significant legal and public relations crisis for the AI-powered surveillance company.
Read more →No stories from this source today.