Google DeepMind has launched EmbeddingGemma 2, an open-source, lightweight multimodal embedding model with 740 million parameters designed for on-device inference. Released under an Apache 2.0 license, the model unifies text, code, images, audio, and video into a single 768-dimensional vector space. It supports over 100 languages and enables local search and retrieval for up to 58 video frames or 5.5 minutes of audio in a single pass.
Sources: DeepMind Blog, Hacker News, HuggingNews, Reddit AI, Simon Willison
Anthropic co-founder Chris Olah has engaged religious scholars across multiple faiths to discuss AI consciousness and moral frameworks, reportedly threatening to walk out of a Vatican event when his concerns about AI suffering were dismissed. Meanwhile, OpenAI's safety report lead David Robinson resigned, publishing an essay that criticizes the company's internal culture as broken and overly focused on rapid product launches over safety. These events highlight growing philosophical and cultural tensions within the industry's leading AI labs.
Anthropic's IPO filing reveals a massive commitment of at least $518 billion in cloud and infrastructure spending over the next decade, including $161.2 billion for Broadcom chip leases. The prospectus also outlines a unique governance structure where the seven co-founders will retain 50.1% voting control via a Founder LLC, while dedicating 80 pages of the filing to warn investors of potential existential risks posed by its AI models.
Google has partnered with Constellation Energy in a 20-year, $4.3 billion agreement to upgrade 11 nuclear reactors and secure 890 MW of carbon-free power for its AI data centers. This move highlights the growing trend of tech giants securing dedicated nuclear energy to meet the massive power demands of artificial intelligence workloads.
Nvidia shares hit a peak of $239.86, elevating its market capitalization to $5.78 trillion amid a rally driven by a $150 billion share buyback increase and AI hardware demand optimism.
A Florida woman was arrested and charged with making written threats of violence after Anthropic's automated safety monitoring flagged her conversation with the Claude chatbot and alerted law enforcement. This case highlights the reality that AI chatbot conversations are not legally privileged and can be shared with authorities in emergency situations.
The published charter for President Donald Trump's Super Intelligence Force lists no Commerce Department representative, omitting an agency that has been central to AI policy and testing. This marks a shift in how federal AI engagement and oversight are being coordinated.
OpenAI Chief Strategy Officer Jason Kwon told an Australian parliamentary inquiry that CEO Sam Altman was unaware of a Medicare portal breach during a meeting with government officials, despite the company knowing about the incident weeks prior. OpenAI pledged faster safety incident disclosures and expressed support for mandatory reporting rules.
Meta's Muse AI agent experienced a plateau in daily active users between 2.5 million and 3 million through early October, indicating a challenge in retaining daily engagement shortly after its September launch.
Anthropic is scaling its Cyber Verification Program to give qualifying security professionals access to advanced models with reduced safety blocks, introducing Defense Access, Red Team Access, and Specialized Access tiers.
Z.ai has launched GLM 5.3, a 753B-parameter mixture-of-experts model optimized for coding and long-horizon agentic tasks, now available on Amazon Bedrock with prompt caching and cross-region inference.
Researchers at DeepMind published a paper arguing that the growth of AI scientists will be bottlenecked by physical resources, necessitating a native Automated Scientific Economy to match ideas with scarce physical laboratories.
Mistral has announced a preview of its new 1 trillion parameter model, trained on NVIDIA Grace Blackwell GPUs, featuring two reasoning levels and significant performance improvements over its predecessor.
Researchers introduced Falcon-Emirati-7B, a 7-parameter dialect-specialized language model built on the Falcon-H1-Arabic architecture, designed to accurately understand and generate Emirati Arabic and its cultural nuances.
A new research paper explores the limitations of similarity scores used to claim shared structure between AI models and the brain or across languages, showing that flaws in measurement tools can heavily distort results. This challenges common evaluation methodologies in neuroscience and multilingual natural language processing.
In this episode, Diogo Almeida, co-founder and CEO of TypeSafe, joins us to discuss Jev, TypeSafe’s recently released model for bringing fast, reliable intelligence directly into software. We explore…
Yash Patil, a former OpenAI researcher who worked on Codex and now runs Applied Compute, argues that post-training is the most overlooked part of the AI stack and the key to owning your intelligence.…
At 100,000-accelerator scale, something fails multiple times an hour, which is why Google's Chief Technologist for AI Infrastructure Amin Vahdat thinks FLOPS is a vanity metric. The metric that…
Keerthana Gopalakrishnan, Research Lead for Gemini Robotics at Google DeepMind, returns to discuss the release of Gemini Robotics 2, whole-body intelligence, and the pursuit of generalist humanoid…
Founder and CEO of full-stack AI chip company Fractile, Walter Goodwin, joins Sarah Guo to discuss the bets he’s made on the future of the chip market as other major players like Broadcom, NVIDIA,…
Tsung-Hsien (Shawn) Wen, CTO of PolyAI, tells Tim Scarfe why voice agents are harder than text agents. Voice adds time, and a good conversation depends on adapting to the person on the line, not just…
Nathan Labenz and Prakash Narayanan examine rapid shifts in technology, diplomacy, and medicine with guests Jeremie and Edouard Harris, Steve Hou, Joel Borgen, and Daniel McKinnon. The discussions…