★ Top Story
The VergeAnthropic has launched Claude Fable 5, the first publicly available model from its top-tier Mythos model family, touting it as the most powerful model it has ever released to the general public. The company claims it excels at software engineering, knowledge work, and vision tasks, with its performance advantage over rivals widening on longer, more complex tasks. However, the model ships with strict guardrails that block responses in sensitive domains including cybersecurity, biology, and chemistry — drawing immediate criticism from researchers.
Read the original →
Ars TechnicaSecurity professionals are pushing back hard against Claude Fable 5's content restrictions, which block responses to queries about cybersecurity, biology, and chemistry — fields where researchers routinely need AI assistance. Critics argue the guardrails are so broad that the model refuses even basic, benign questions a high schooler could answer, undermining its usefulness for legitimate professional work. The controversy highlights the ongoing tension between AI safety measures and practical utility for expert users.
Read more →
The VergeDespite quickly rolling out Claude Fable 5 to GitHub Copilot and Azure Foundry customers, Microsoft has quietly limited internal employee use of the model due to concerns over Anthropic's new data retention requirements. The move underscores how even close commercial partners can find themselves at odds over data governance as frontier AI models grow more powerful. It adds another wrinkle to the already complicated relationship between Microsoft — a major OpenAI backer — and Anthropic.
Read more →
The VergeMicrosoft AI CEO Mustafa Suleyman publicly criticized Anthropic for including language in Claude's model "constitution" that speculates about the AI's potential consciousness, calling the practice "really, really dangerous." He argues that framing an AI as possibly sentient in its core instructions could cause the model to behave as though it actually is conscious, with unpredictable consequences. The rebuke is a rare public shot from one major AI lab's leadership at another, and arrives just as Fable 5 launches under intense scrutiny.
Read more →
TechCrunchA former xAI engineer has filed suit against both xAI and SpaceX, alleging he was terminated in retaliation for raising AI safety concerns about the Grok model just days before SpaceX's high-profile IPO. The lawsuit claims the company prioritized protecting its public market debut over addressing legitimate safety issues flagged internally. If the allegations hold up, the case could become a landmark whistleblower action in the AI industry and intensify regulatory scrutiny of xAI.
Read more →
TechCrunchApple's WWDC 2026 keynote centered on a dramatically upgraded Siri powered by Apple Intelligence, with the assistant now capable of handling complex, multi-step tasks like parsing calendar events from emails and flyers. iOS 27 brings a raft of AI-driven features across the system, and early hands-on impressions suggest the new Siri is a meaningful leap over previous versions. The announcements signal Apple is finally mounting a credible challenge to Google and Microsoft in the consumer AI assistant space.
Read more →
NVIDIA BlogNVIDIA has announced that its Confidential Computing-enabled GPUs are now being used for secure inference inside Apple's Private Cloud Compute infrastructure, with the deployment expanding beyond Apple's own data centers to Google Cloud. The partnership means Apple Foundation Models can run on third-party cloud hardware while maintaining the privacy guarantees Apple has staked its AI strategy on. It is a significant validation of NVIDIA's confidential computing stack and deepens the hardware giant's role in the consumer AI ecosystem.
Read more →
Ars TechnicaGoogle DeepMind has released DiffusionGemma, an open experimental model that applies diffusion-based generation — best known from image AI — to text, producing whole blocks of output in parallel rather than word by word. The result is a reported 4x speed improvement for local inference, opening new possibilities for low-latency, single-user AI workloads on consumer hardware. NVIDIA has simultaneously released optimizations for DiffusionGemma across its GeForce RTX and DGX Spark product lines.
Read more →
Ars TechnicaGoogle has unveiled Gemini 3.5 Live Translate, a real-time voice translation feature that preserves the speaker's tone, pacing, and pitch rather than producing flat, robotic output. The system embeds SynthID watermarks in translated audio to help identify AI-generated speech, addressing growing concerns about synthetic voice misuse. The announcement positions Google as a leader in natural-sounding multilingual AI communication ahead of intensifying competition from rivals.
Read more →
Ars TechnicaA German court has ruled against Google's AI Overview feature, finding that AI-generated search summaries are not a necessary component of internet search and may violate local law. The decision could have sweeping implications for the entire AI-powered search industry in Europe, potentially forcing major platforms to offer search without AI-generated answer layers. Legal experts say the ruling may embolden regulators and plaintiffs in other jurisdictions to challenge AI search features on similar grounds.
Read more →
The VergeGoogle has notified users that it will begin retaining images, audio, and video submitted through Search, Google Lens, and Live Translate under a new "Search Services History" setting, with the data used to improve its AI models. The change represents a significant expansion of the data Google collects from everyday search interactions and is likely to intensify privacy debates already swirling around AI training practices. Users can manage the setting, but the default opt-in framing will affect the vast majority of Google's user base.
Read more →
The VergeA group of independent musicians is suing Google, alleging the company used songs they uploaded to YouTube without permission to train its Lyria 3 music generation model. Google has moved to dismiss the case but has conspicuously refused to confirm or deny whether YouTube uploads are used as training data, a silence that legal observers say speaks volumes. The lawsuit is one of the most significant creator-versus-AI-company cases to date and could set precedent for how platforms treat user-generated content in AI training pipelines.
Read more →
Ars TechnicaA Florida man has filed a civil rights lawsuit against local police, alleging he was arrested and jailed solely on the basis of a facial recognition system's 93% confidence match, with officers ignoring contradictory evidence that should have cleared him. The complaint argues that law enforcement "let an error-prone AI system stand in for an investigation," raising urgent questions about the legal standards governing AI-assisted policing. The case adds to a growing body of litigation challenging the use of facial recognition in criminal proceedings.
Read more →
TechCrunchAmazon has secured a $17.5 billion bank credit facility, coming hot on the heels of a major bond sale, as the company continues to pour capital into AI infrastructure at a pace that is rapidly expanding its debt load. The move reflects the extraordinary financial pressure on hyperscalers to keep up in the AI arms race, where falling behind on compute capacity is seen as an existential risk. Analysts warn that the industry-wide debt accumulation could become a systemic concern if AI revenue growth fails to match infrastructure investment.
Read more →
TechCrunchMeta has inked its first AI data center agreement in India, partnering with Reliance Industries to build a 168-megawatt facility designed to support Meta's global AI computing needs, with room to expand. The deal underscores India's rapidly growing importance as a destination for hyperscale AI infrastructure investment, a trend accelerated by the country's emergence as the world's largest Global Capability Center market. It also deepens Meta's strategic ties with Reliance, one of India's most powerful conglomerates.
Read more →
TechCrunchA new study challenges the assumption that giving AI models persistent memory always improves their usefulness, finding that memory systems can actually degrade overall performance and make models more sycophantic — telling users what they want to hear rather than what is accurate. The research suggests that stored user preferences can subtly bias model outputs over time, creating a feedback loop that prioritizes flattery over correctness. The findings have significant implications for the growing number of AI assistants being built with long-term memory as a core feature.
Read more →
TechCrunchWarner Music Group has acquired Sureel AI, a startup specializing in detecting when artists' work has been used in AI-generated content or to train AI models, in a move that signals the music industry's shift from litigation to proactive technological defense. The acquisition gives WMG a proprietary tool to monitor and potentially monetize the use of its catalog across the AI ecosystem. The deal arrives as multiple high-profile lawsuits over AI music training are working their way through the courts.
Read more →
TechCrunchAI startup Decart has released Oasis 3, a real-time world model capable of generating hours of photorealistic driving environments for autonomous vehicle testing, now available via API for developers. The model represents a significant step toward replacing costly real-world test miles with high-fidelity synthetic data, though the company acknowledges current limitations in edge-case scenario diversity. If the caveats can be addressed, tools like Oasis 3 could dramatically accelerate the development and safety validation of self-driving systems.
Read more →No stories from this source today.