Samwise Tech/AI/Robotics Newsletter
Tuesday, September 22, 2026
Xiaomi’s MiMo-V2.6-Pro Debuts as World’s Top Open-Weights AI Model
Xiaomi has released MiMo-V2.6-Pro alongside a lighter MiMo-V2.6-Flash model, with the Pro variant claiming the top position on the Agentic Ability (AA) Index among open-weights models — surpassing DeepSeek V4 Pro. The release marks a significant entry by the Chinese consumer electronics giant into the frontier AI competition, where open-weights challengers have been pushing the pace in 2026. MiMo-V2.6-Pro earns an AA Index score of 46, demonstrating strong performance on coding and reasoning tasks. The Flash variant offers a lower-cost option for teams needing capable but more affordable agentic AI. Both models are available to download now.
Sources: VentureBeat
OpenAI’s AI Resolves 100+ Open Math Problems; Advisory Group Formed to Manage Pace
OpenAI announced Monday that the same internal model that solved the Navier-Stokes Millennium Prize problem has now resolved more than 100 additional open problems across mathematics. To address growing tensions with the math community, the company also launched the Advisory Group on Mathematics and Artificial Intelligence, hosted at Princeton’s Institute for Advanced Study. Nine prominent mathematicians will assess the significance of new AI-generated results and coordinate their release — but the group explicitly lacks power to slow OpenAI’s internal research. Twenty-five Fields Medal winners had signed an open letter warning that AI labs risk disrupting mathematical careers and intellectual priority.
Sources: TechCrunch
xAI’s Grok 4.7 Brings Coding Gains but High Token Use Clouds Real-World Economics
xAI has released Grok 4.7, delivering notable improvements in coding performance while maintaining the model’s existing affordable pricing tier. The update extends the company’s push to position Grok as a cost-competitive alternative to premium frontier models. However, analysts warn that Grok 4.7’s gains come with significantly elevated token consumption, raising concerns about real-world return on investment for teams running the model at scale. While the per-token price remains unchanged, higher call volumes required to complete complex tasks could offset the apparent cost advantage. The release continues xAI’s rapid iteration cadence as competition among frontier coding models intensifies.
Sources: VentureBeat
Boston Dynamics Opens Robot Training Facility Inside Hyundai’s Georgia Auto Plant
Boston Dynamics has opened the Robotics Metaplant Application Center (RMAC), a permanent facility embedded in Hyundai Motor Group’s Metaplant America in Georgia, dedicated to training Atlas humanoid robots on real automotive manufacturing tasks. The facility — dubbed a ‘Robot Behavior Factory’ — uses teleoperation, reinforcement learning, and Universal Manipulation Interface devices to teach Atlas to handle complex parts logistics and sequencing ahead of assembly. Boston Dynamics says early results confirm its production Atlas design assumptions were largely correct. The company plans to expand RMAC tenfold into a new building next year, with component assembly training targeted for 2030.
Sources: The Robot Report
Autonomous Racing Achieves Historic First: AI Drivers Overtake on Road Course at Laguna Seca
The Indy Autonomous Challenge completed its latest university racing event at Laguna Seca on September 3, marking the first time autonomous vehicles have successfully overtaken each other on a road course in IAC competition. Italy’s Unimore Racing from the University of Modena and Reggio Emilia took the win in one-on-one heats against Purdue AI Racing. Nine university teams from North America, Europe, and Asia competed, with all AI driving decisions made by software during heats. In another first, teams were required to qualify via simulation — using a digital twin of Laguna Seca in D-Space’s software-in-the-loop environment — before reaching the physical track.
Sources: The Robot Report
Google Open-Sources EnvHarness to Dynamically Adapt AI Agent Training Environments
Google Cloud AI Research has open-sourced EnvHarness (Apache 2.0), a programmable framework that wraps static training environments and automatically adapts task difficulty to an agent’s specific weaknesses. The system’s EnvRigger module runs an automated Observe-Diagnose-Write-Validate loop to generate targeted training data. Three composable components — Stage, Contract, and Chain — allow practitioners to modify starting conditions, constrain action spaces, and sequence tasks. Testing across SWE-bench Verified, ALFWorld, WebArena, OfficeQA, and SpreadsheetBench showed agents improving by up to nine points; on SWE-bench Verified specifically, a base agent improved from 47.67 percent to 54.79 percent. The code is available on GitHub at google-research/envharness.
Sources: VentureBeat
Recycling Humanoid Robots Is a Multi-Billion-Dollar Problem Nobody Has Solved Yet
As humanoid robots race toward mass production, recycling firm Re-Teck warns the industry has overlooked a critical challenge: decommissioning these machines is far more complex than conventional scrapping. A single humanoid contains 10,000 to 15,000 components, up to 4 kilograms of rare-earth neodymium magnets — exceeding the amount in an EV — and battery packs prone to thermal runaway if crushed. Memory systems may hold sensitive proprietary and biometric data. Traditional bulk shredding destroys rare-earth material value, while safe disassembly requires skilled technicians working in hazardous conditions. Re-Teck argues humanoid OEMs must adopt design-for-recycling principles with modular cartridges before production scales further.
Sources: The Robot Report
⚡ Tech Pulse
Top Frontier Models: GPT-5.6 Sol (89.8% SWE-Verified) | Claude Sonnet 5 (89.4%) | Claude Fable 5 (80.0% SWE-Pro)
Top Open Source Models: DeepSeek V4 Pro (80.6% SWE-Verified) | Qwen3.6-27B (77.2%) | MiMo-V2.6-Pro (AA Index 46, new today)
Top Small Models (15–50B): Qwen3.8-27B (61.7%) | MiniMax M3 (est. 59.0%) | Phi-4 Medium (est. 52%)
Top Edge Models (0–15B): Gemma 3.5 4B (est. 42%) | Phi-4 Mini (est. 40%) | Qwen3-8B (est. 39%)
AI Leaders: NVIDIA ~$5.1T | Alphabet ~$4.5T | Microsoft ~$3.7T
Robotics Leaders: Intuitive Surgical ~$162B | Agility Robotics (SPAC, pre-public) | Figure AI ~$39.5B (private est.)
Every newsletter preserved and searchable
Curated by JD · samwise.agency