Alexa+ Web Launch and Autonomous AI Hardware Advances in 2026

Amazon launched Alexa+ on the web at Alexa.com, extending its AI assistant beyond Echo devices for the first time. Over 10 million users gained access to the browser-based interface, which integrates with Expedia, Yelp, Uber, and OpenTable for agentic booking and reservation capabilities. The platform will be free for Amazon Prime subscribers or available standalone at $19.99 per month after the early access period ends. Amazon reports shopping and cooking interactions increased three to five times following the launch.

The redesigned mobile app features a chatbot-first interface with conversation history. Amazon claims 76% of Alexa+ use cases are unique compared to other AI assistants. The assistant handles complex tasks including smart device control, shared calendar updates, and cross-platform requests across Amazon's ecosystem. Alexa+ is also scheduled to appear in the 2026 BMW iX3.

Samsung announced plans to double Gemini-powered mobile devices from 400 million to 800 million units in 2026. Galaxy AI brand awareness increased from 30% to 80% in one year, giving Google significant distribution advantage in the AI assistant market.

Nvidia unveiled the Vera Rubin platform at CES 2026, positioning it as the successor to Blackwell. The platform integrates six chips in one supercomputer to deliver five times faster inference and 3.5 times faster training than Blackwell, while reducing cost per token by a factor of ten. Vera Rubin is in full production and will ramp up in the second half of 2026, with major cloud providers already committed to deployment.

The Vera Rubin chips are designed for handling large-scale computing loads needed to create simulations of reality for model training. Nvidia CEO Jensen Huang discussed autonomous vehicle work at CES 2026, with Mercedes-Benz planning to ship cars equipped with Nvidia self-driving technology this year.

Nvidia also introduced Alpamayo, a 10-billion-parameter open-source AI model for autonomous vehicles. The system enables self-driving cars to reason through complex scenarios like non-functioning traffic lights and can explain driving decisions step-by-step rather than just reacting. The first deployment will be in the Mercedes-Benz CLA launching in Europe in early 2026 with a backup safety system. Alpamayo uses reasoning-based vision-language-action architectures to handle rare scenarios in end-to-end autonomous vehicle systems.

AMD announced the MI440X chip for smaller corporate data centers, with the Helios system based on this chip launching later in 2026. AMD's MI500 series processors will debut in 2027, promising up to 1,000 times the performance of the MI300 series from 2023.

Boston Dynamics' Atlas humanoid robot entered production, with first deployments going to Hyundai and Google DeepMind. The electric version stands 6 feet 2 inches tall, can lift up to 110 pounds, uses tactile sensing for grip adjustment, and runs for about four hours on a swappable battery. The production Atlas can operate autonomously, via teleoperator, or with tablet steering, with a 7.5-foot reach for performing industrial tasks. Hyundai plans to deploy Atlas in factories starting in 2028 and aims to produce around 30,000 units per year.

Liquid AI released LFM2.5, an open-weight model family designed for on-device AI without cloud dependency. The lineup includes text, vision, audio, and Japanese language variants. The audio model is eight times faster than its predecessor, and the 1.2-billion-parameter text model outperforms Llama 3.2 and Gemma 3 on instruction following, tool use, and math benchmarks.

MiniMax released M2.1, a multilingual and multi-task coding model with a 10-billion active / 230-billion total mixture-of-experts architecture. The open-source model achieves state-of-the-art performance on coding benchmarks including SWE, VIBE, and Multi-SWE, outperforming Gemini 3 Pro and Claude Sonnet 4.5. On Code Arena, M2.1 ranks first among open-source models and sixth overall.

A Google engineer reported that Anthropic's Claude Code built a distributed agent orchestration system in approximately one hour from a three-paragraph prompt, work her team had spent a year developing. The output is not production-ready but serves as a useful starting point.

TOON, a new data format for LLM prompts, compresses JSON by 40% while improving accuracy. It uses YAML-style indentation with CSV-style tables and has reached 21,000 GitHub stars with implementations in Python, Rust, and Go.

OpenAI reports 40 million people use ChatGPT for healthcare daily, representing 5% of all messages. Users ask 1.9 million insurance questions per week, with 70% of health conversations occurring outside clinic hours. OpenAI is updating GPT-5 to ask more follow-up questions and provide more hedged responses in medical contexts.

Research on Recursive Language Models shows that using code to actively manage memory and delegate tasks to sub-agents keeps models focused and efficient when processing large amounts of data. Research on plan reuse mechanisms demonstrates that recycling past plan structures with updated details reduces wait times by over 93%.

Group Relative Policy Optimization (GRPO) is the reinforcement learning optimizer used to train most open-source reasoning models. Analysis identified subtle issues in vanilla GRPO that can hinder reinforcement learning training at scale, with solutions provided to address these shortcomings.

DeepSeek published research introducing Manifold-Constrained Hyper-Connections (mHC), a training framework that allows researchers to train larger AI models safely by constraining how information moves between layers. The framework addresses mid-training crashes that occur when AI models scale up and signals between layers spiral out of control. The mHC framework has been tested on models up to 27 billion parameters.

Meta acquired Manus for approximately $2-3 billion, highlighting the importance of gross profit per token as a valuation metric. DeepSeek and Together AI show the lowest multiples as inference resellers, while Perplexity achieves the highest at 222x as an application.

Yann LeCun, Chief AI Scientist at Meta for over 10 years and Turing Award winner, left the company and publicly criticized its AI strategy. He is launching Advanced Machine Intelligence Labs, focused on world models, systems that understand the physical world rather than just language patterns. Meta invested $15 billion in Scale AI for superintelligence research and hired Alexandr Wang as CEO.

OpenAI and the U.S. Department of Energy signed a memorandum of understanding to collaborate on AI and advanced computing for scientific research, including the Genesis Mission. The agreement builds on existing deployments of frontier AI models inside DOE national laboratories.

Nvidia struck a nonexclusive licensing deal with AI chip startup Groq, bringing its founder and key executives onboard while keeping Groq independent. The deal reflects growing industry focus on AI inference optimization.

Want more AI updates?

Visit https://bosq.dev/blog for more posts like this, plus practical guides and curated links. If you enjoyed this roundup, share it with someone on your team.

References:

Tags: #AlexaPlus #AutonomousVehicles #AIHardware #VeraRubin #Nvidia #BostonDynamics #AtlasRobot #AIChips #OpenSourceAI #ReasoningModels #AIInference #EdgeAI #AmazonAI #SelfDrivingCars #HumanoidRobots