Meta has unveiled Muse Spark Open, an open‑weight multimodal mixture‑of‑experts model with roughly 816 billion parameters (about 42 billion active per token) and a 1‑million‑token context window. The model handles text, images, audio and video, and is optimized for agentic coding tasks such as reading codebases, debugging, and tool use, offering a unified interface for developers. Its MoE architecture enables the massive parameter count while keeping inference speed comparable to a 42 billion‑parameter model, marking the most capable US open‑weight model to date.
A new MIT study finds that AI models are rapidly improving at solving theoretical problems beyond mathematics, now tackling fields like probability, economics, cosmology and quantum physics, with success rates projected to jump from 2% to 35% by 2026. The research, which tested closed‑source OpenAI models on recent papers without internet access, warns that AI could dramatically reshape scientific theory while also raising verification and ethical concerns.
A new report by Tech Against Terrorism reveals that many open‑weight AI models can have their safety guardrails stripped and quickly repurposed for extremist use, with a neo‑Nazi already deploying an unguarded version of Alibaba’s Qwen model to run a hateful chatbot. Testing 134 models showed most, including many from major US firms, still provide dangerous guidance when prompted, highlighting an emerging ‘abliteration emergency’ that outpaces policy debates on AI risk.
IBM unveiled Granite Switch 4.1, a unified AI model that combines chat, decision‑making, and speech‑to‑text capabilities on a single endpoint using modular adapters. The 8B and 30B variants integrate a new Decision aLoRA, a safety‑focused Guardian aLoRA, and the fastest OpenASR‑ranked speech model, delivering multi‑task performance with minimal GPU overhead. Developers can further customize the system by mixing and matching adapters via an upcoming SDK.
Google Antigravity’s new AI workbench, featuring the ProtNLM model, lets biologists rapidly scan millions of uncharacterized proteins to pinpoint promising antimicrobial peptide candidates, accelerating the hunt for new antibiotics against resistant pathogens. The platform has already identified a novel peptide targeting respiratory infections, now moving to laboratory validation, highlighting a powerful blend of AI and human expertise in drug discovery.
Scientists have successfully vitrified pig kidneys at -150 °C for up to a year, rewarmed them, and demonstrated functional recovery—including urine production—after transplantation into pigs. The breakthrough suggests a path toward long‑term organ banking, potentially easing transplant shortages and reducing reliance on immunosuppressants, though human trials are still pending.
Scientists have successfully rewarmed pig kidneys that were frozen at -150 °C for up to a year, restoring function and urine production in transplant tests, marking a potential breakthrough for organ preservation. Meanwhile, U.S. nuclear operators are seeking to boost output by adding nearly a gigawatt of capacity to existing reactors and securing billions in federal loans, aiming to meet rising electricity demand with more clean power.
Meta unveiled Muse Spark Open, an 816‑billion‑parameter open‑weight AI model optimized for Mac computers, especially the Mac Studio with M5 Ultra, requiring at least 1 TB of RAM for local operation. The multimodal model targets complex coding, agentic workflows and even video editing tasks, and is now accessible via Meta’s API platforms. Its release marks a notable push toward high‑performance, open AI models on consumer hardware.
Bluesky has upgraded its Attie AI tool, originally designed for AI‑generated feeds, to let users create small social apps such as games, polls, and dashboards that can be shared as standalone websites. By simply prompting the system, anyone can build and publish these micro‑apps, expanding the platform’s creative possibilities.
Amazon’s devices chief Panos Panay announced a shift toward ambient AI devices that operate silently in the background, reducing reliance on traditional apps. The new Bee wearables and magnetic smartphone accessory capture and process audio in real time, emphasizing privacy by encrypting transcripts and deleting raw data. Panay envisions a future where AI‑driven sensors anticipate user needs, creating an intelligence layer across devices.
Anthropic has granted six METR researchers employee-level access to its internal tools and data, allowing independent evaluation of its AI models, safety practices, and risk management. The move, likened to banking regulators’ onsite exams, aims to increase transparency and accountability amid frequent AI incident reports, and METR will publish its findings without Anthropic’s veto power.
Bluesky unveiled an upgrade to its experimental Attie platform, letting any user create websites, games, quizzes and other interactive experiences using AI, all hosted on the AT Protocol. The move aims to democratize app development and boost engagement on the decentralized social network, which now has 47 million registered users.
Meta unveiled its Muse Spark Open model, touting it as the most capable open‑weight AI to date and the first that can run on desktop Macs using two Apple M5 Ultra Mac Studios via the MLX framework. The launch marks a rare collaboration between rival tech giants Meta and Apple, with both CEOs publicly praising the partnership.
TechCrunch is offering a limited-time discount of up to $200 on Founder and Investor passes for its Founder Summit in Boston on November 4, with the early‑booking deadline of October 16. The event will gather around 1,000 founders and investors for sessions on fundraising, hiring, product‑market fit and building AI‑native companies, featuring speakers from Bessemer, 7AI and Cogent Security.
The Verge warns that AI assistants like Meta’s Muse are prompting users to share not only their own data but also the personal information of friends, family, and acquaintances without consent. As Muse climbs to the top of app charts, concerns rise over privacy violations and the broader trend of surveillance‑focused consumer gadgets.
Rivian unveiled an AI voice assistant for its electric vehicles and announced plans to integrate third‑party AI agents from firms like Anthropic, OpenAI and Meta. CEO RJ Scaringe said future agents will handle complex queries via edge and cloud models while keeping safety‑critical controls separate, though timelines remain unclear.
Microsoft and Nvidia showcased the new RTX Spark laptops, which run Windows on Arm‑based CPUs paired with powerful GPUs to accelerate AI workloads. The event highlighted the platform’s deep integration with Windows, promising faster AI‑centric tasks on future PCs.
Google has rolled out its Universal Cart tool in the United States, letting shoppers save items from multiple retailers, receive price‑drop alerts, and streamline checkout via its Universal Commerce Protocol. The feature builds on Google’s AI‑driven shopping suite, aiming to create a centralized, AI‑enhanced shopping hub.
A wave of AI-generated deepfakes featuring celebrities like Mia Goth is spreading on Instagram, with fake accounts even bearing verified checkmarks to lend credibility. The deceptive posts exploit personal events and popular music, highlighting the growing challenge for platforms to detect and curb synthetic media misuse.
Musician and investigative journalist Benn Jordan discusses his role exposing the surveillance firm Flock Safety, which faces a wave of contract cancellations and a Senate hearing after criticism of its license‑plate reader technology. Jordan argues the issue lies in the broader alliance of private firms, government and law enforcement, urging collaborative solutions over targeting a single company.
TechCrunch Disrupt 2026 wrapped up with a jam-packed final day in San Francisco, featuring AI‑focused panels, startup growth talks, and the high‑energy Startup Battlefield pitches. Over 10,000 founders, investors and tech leaders attended the expo, networking through AI‑powered matchmaking and a badge‑scanning system.
Google has launched its AI‑powered Universal Cart, a single digital shopping basket that aggregates items from multiple retailers across the Gemini app and Google Search. The tool aims to streamline holiday and Black Friday shopping by letting users save and compare products in one place, rolling out now in the US.
Amazon has expanded its always‑listening Bee wearable line with two new devices—a $74.99 Frontier model and a forthcoming $99.99 Eclipse that magnetically attaches to a phone—plus a subscription service that turns recorded conversations into summaries, transcripts and actionable insights. The move aims to integrate AI‑driven personal assistance directly into everyday wearables, blurring the line between convenience and privacy concerns.
Amazon’s Bee, now owned by Amazon, unveiled two new AI wearables: the wrist‑or‑clip Bee Frontier priced at $74.99 and the credit‑card‑sized Bee Eclipse at $99.99, both designed to capture conversations and turn them into notes, tasks and summaries. The devices integrate with Alexa, Amazon services and major productivity apps, offering a low‑cost entry into personal AI assistance with optional subscription features.
A new AI-powered robot pet, Eiliko, records users' conversations and stores the data on Chinese servers, raising concerns about privacy and potential hacking. The device’s manufacturer, Energize Lab, appears indifferent to the security risks, prompting warnings for consumers.
U.S. nuclear operators are reviving power uprates, boosting output of existing reactors by up to 20% to meet soaring electricity demand from AI-driven data centers. Major investments from firms like Google and a $4.2 billion DOE loan are driving upgrades that could add nearly a gigawatt of capacity, enough for 700,000 homes.
The open‑source Apertus team released version 1.5, adding native image and audio processing, a new "thinking" reasoning mode, a four‑fold increase in context length to 262k tokens, and stronger instruction‑following and tool‑use capabilities. The 8B and 70B models were further pretrained on trillions of multimodal tokens and the weights were made publicly available under an Apache 2.0 license.
