Samwise Tech/AI/Robotics Newsletter
Sunday, May 3, 2026
Meta Acquires Robotics Startup ARI to Fuel Humanoid AI Push
Meta has acquired Assured Robot Intelligence (ARI), a startup building foundation models for humanoid robots, for an undisclosed sum. Co-founders Xiaolong Wang — formerly of Nvidia and UC San Diego — and Lerrel Pinto, previously at NYU, will join Meta’s Superintelligence Labs division. ARI was developing AI systems designed to help robots understand, predict, and adapt to human behavior in complex environments. The acquisition follows Meta’s earlier purchase of Manus and signals a deepening commitment to physical AI. Meta now joins Amazon, which acquired humanoid startup Fauna Robotics, and Google in a race to give AI models the physical-world grounding that researchers increasingly believe is necessary for AGI progress.
Source: TechCrunch — 2026-05-01
xAI Launches Grok 4.3 With Aggressively Low Pricing and New Voice Cloning Suite
xAI launched Grok 4.3 alongside a fast voice cloning suite, positioning the model as a cost-efficiency leader rather than a frontier generalist. At $1.25 per million input tokens, it undercuts rivals from Anthropic, OpenAI, and Google significantly. Third-party evaluators Vals AI found Grok 4.3 leads specialized indices including CaseLaw v2 at 79.3% and CorpFin, but it remains below the general frontier set by Claude and GPT-5 variants. The release arrives after months of leadership departures at xAI and just days after Musk admitted under oath that early Grok versions were partly trained using OpenAI model distillation, a revelation that has drawn wide industry attention.
Source: VentureBeat — 2026-05-02
MCP Design Flaw Exposes 200,000 AI Agent Servers to Unauthenticated Command Execution
OX Security researchers have disclosed a critical design flaw in Anthropic’s Model Context Protocol affecting an estimated 200,000 AI agent servers worldwide. The STDIO transport — the default mechanism for connecting AI agents to local tools in every official MCP SDK, including Python, TypeScript, Java, and Rust — executes operating system commands with no sanitization or access controls. OX demonstrated four exploitation paths targeting frameworks including LangFlow, LiteLLM, Flowise, and Windsurf. CVE-2026-30615 in Windsurf required zero user interaction to exploit. Nine days after OX’s initial disclosure in January, Anthropic updated its SECURITY.md file to advise caution but made no architectural changes, calling the behavior a design feature.
Source: VentureBeat — 2026-05-01
SoftBank Forms Roze AI, a Robot-Staffed Data Center Company Eyeing a $100B IPO
SoftBank has established Roze AI, a new company that combines AI-driven data center construction with humanoid robot manufacturing, already targeting a $100 billion IPO by the second half of 2026. The venture represents founder Masayoshi Son’s conviction that physical AI infrastructure will rival digital cloud computing in economic value. Roze plans to operate data centers using robotic labor while simultaneously producing humanoid robots for commercial deployment. The move builds on SoftBank’s $5.4 billion acquisition of ABB’s robotics division, expected to close later this year, and its $500 billion Stargate AI infrastructure partnership with OpenAI and Oracle — one of the largest announced infrastructure investments in history.
Source: TechCrunch — 2026-04-29
Google Rolls Out Gemini to 4 Million GM Vehicles, Replacing Google Assistant
Google announced that Gemini AI is rolling out to cars with Google built-in, starting with approximately four million General Motors vehicles from model year 2022 and newer across Cadillac, Chevrolet, Buick, and GMC brands. The update replaces Google Assistant with a more capable, conversational AI that can handle open-ended queries, route planning with restaurant recommendations from Google Maps, vehicle controls, message summarization, and hands-free responses. Drivers signed into their Google accounts will be prompted to upgrade on their next in-vehicle session. A beta feature called Gemini Live enables open-ended, real-time conversation while driving. The rollout marks Google’s most significant automotive AI push since building Assistant into cars five years ago.
Source: TechCrunch — 2026-04-30
Poolside Releases Laguna XS.2, a Free Open-Source Model for Local Agentic Coding
American AI startup Poolside has released Laguna XS.2, a free open-source coding model designed to run on a single GPU or laptop without an internet connection. The 33-billion total parameter Mixture of Experts model activates only 3 billion parameters at inference time, enabling fast local execution while maintaining competitive agentic coding performance. Licensed under Apache 2.0 and available on Hugging Face, XS.2 is Poolside’s second-generation open model, built on lessons from training its predecessor. The release targets developers who want to fine-tune, quantize, or run coding agents on-premises, entering a competitive field alongside Mistral’s Devstral 2 and OpenAI’s GPT-5-Codex for local and enterprise deployment.
Source: VentureBeat — 2026-04-29
DeepMind’s David Silver Raises $1.1B for AI Lab That Learns Without Human Data
Former DeepMind researcher David Silver, creator of AlphaGo and AlphaZero, has raised $1.1 billion for Ineffable Intelligence, a UK-based AI lab developing models that learn without human-labeled data, reaching a valuation of $5.1 billion. The round was led by Sequoia Capital and Lightspeed Venture Partners, with Google, Nvidia, and Index Ventures also participating. Silver’s work at DeepMind pioneered reinforcement learning and model-based planning — techniques now central to the agentic AI wave. Ineffable is pursuing architectures that go beyond large language models, training systems through self-play and environmental interaction rather than internet-scraped text, an approach that could fundamentally change how frontier AI models are developed.
Source: TechCrunch — 2026-04-27
Top Frontier: Claude Mythos Preview (93.9%) | Claude Opus 4.7 (87.6%) | GPT-5.3 Codex (85.0%)
Top Open Source: DeepSeek V4 Flash Max (79.0%) | Qwen3.6 Plus (78.8%) | DeepSeek V4 Flash (73.7%)
Top Small (15–50B): Qwen3.6-27B (77.2%) | Mistral Small 4 24B (72.1%) | Gemma 3 27B (67.5%)
Top Edge (0–15B): Phi-4 mini 3.8B (62.4%) | Qwen 2.5 7B (58.1%) | Llama 3.2 3B (54.7%)
AI Leaders: NVIDIA $4.60T | Alphabet $3.89T | Microsoft $2.85T
Robotics Leaders: Intuitive Surgical $175B | Fanuc $38.7B | Teradyne ~$18B
Curated by JD · samwise.agency