Elon Musk says Grok 4.8 is a 2.5-trillion-parameter model trained with xAI’s new C++ software stack and will start reinforcement learning this week. We check the claim against every parameter count and stack post Musk has made, the delayed Grok 4.7, Grok 4.6’s published benchmarks, and what a 2.5T model means for memory, speed and price.
A humanoid robot learned to sprint, throw spin kicks and land aerial cartwheels with one training recipe. We read the BeyondMimic paper from UC Berkeley and Stanford behind the headline, split its 70.8% humanlike score into walking and running, show that 15 of about 150 minutes of motion ran on hardware, explain the motion capture behind the obstacle demos and the licence on its training motion, and set it against the sprint records at the World Humanoid Robot Games.
Meta AI and the University of Illinois have published EvoHarness-RL, a framework that trains an agent to run its own memory, progress and experience stores instead of following a hand-written harness. On the ALFWorld benchmark it took Alibaba’s open-weight Qwen3-8B from 47.9% to 96.9%, a whisker past Claude Opus 4.5’s unaided 96.4%. This breakdown covers what the framework actually trains, the full results table rather than the two rows that travel well, why the same harness pushes Opus 4.5 to 98.5%, why Claude Opus sits inside the training loop as teacher and consolidator, what harness annealing and harness evolution mean for latency, and what any of it is worth if you are shipping agents commercially.
Barret Zoph announced on 26 August 2026 that he is joining Google DeepMind as vice president of research, working on reinforcement learning and post-training for Gemini. It closes a twenty-two month sequence that took him from OpenAI to co-founding Thinking Machines Lab with Mira Murati, to a disputed firing on 14 January 2026, to five months leading enterprise business back at OpenAI. This breakdown covers what the new role is and is not, both accounts of the Thinking Machines termination, his 2016-2022 Google Brain record, the 2026 departures that emptied the seat he is filling, where Thinking Machines stands now, and what any organisation buying AI should take from it.
Hugging Face and its Pollen Robotics subsidiary opened pre-orders on 27 August 2026 for Microduck, a 25 cm bipedal robot duck priced at $399 before taxes and shipping, with first deliveries targeted before Christmas 2026. The machine carries 15 motors, a wide-angle camera, an 8×8 time-of-flight LiDAR array, two IMUs and an articulated beak that works as a gripper, on a Rockchip RK3566 with 1 GB of RAM and 32 GB of storage running its neural policy at 50 Hz. Seven behaviours ship on the device — walking, rolling on skates, grasping, kicking, standing up after a fall, sitting and a generated audio identity — and every one was trained in MuJoCo with proximal policy optimisation and transferred to hardware. The control stack and the full training pipeline are published under Apache 2.0, but the hardware is not open and Pollen Robotics says it has no plans to change that. This breakdown covers the pack pricing, the complete published specification, the Open Duck Mini project the design grew out of, how it compares with the Reachy Mini line, and the battery, payload and compute ceilings that decide who should actually buy one.
Progressive Robot: Your Gateway to Comprehensive IT Solutions — Specializing in Web Development, Mobile App Development, and Expert IT Services.
© All Copyright by Progressiverobot.com
VAT Number ( 506152326 )
