
Tech • AI • Robotics • Game
Salesforce introduced Koa, its first in-house reasoning model, inside Agentforce. The system targets long, multi-step enterprise work across sales, marketing and customer support rather than consumer chat use. Salesforce, which reports roughly $38 billion in annual revenue and about 75,000 employees, is positioning AI as software that executes office workflows directly. The launch sharpens the shift from copilots for workers to embedded agents that can absorb parts of white-collar roles.
Nvidia played a central role in the Koa launch through Nemotron 3 Super, an open-weight model used in the system. That extends Nvidia's presence from infrastructure supplier to model-layer competitor alongside major AI labs. The move matters because enterprise buyers increasingly want vertically integrated AI systems, not just accelerators. It also signals a broader race to own the full enterprise AI stack from compute to deployable agents.
Beijing hosted the World Humanoid Robot Games, a five-day event featuring more than 2,000 robots across 51 events. The competition mixed athletics, football and industrial-style tasks with a choreographed opening featuring about 100 robots marching with an orchestra. Beyond engineering demos, the event projected industrial confidence and national ambition. The result was as much a soft-power display as a robotics competition.
State media spotlighted a robot called Lightning, credited with a 100-meter time of about 9.3 seconds, faster than Usain Bolt's world record. Unitree's H1 also stood out, reportedly taking four gold medals in athletics events. Those headline claims delivered viral moments and reinforced the image of fast-moving Chinese humanoid development. At the same time, they raised familiar questions about demonstration conditions, repeatability and competitive framing.
Some of the most widely shared images from Beijing showed humanoids stumbling, hitting barriers or being removed on stretcher-like rigs. Those scenes undercut the polished messaging by revealing how fragile many systems remain outside tightly controlled routines. The contrast between medal tallies and public failures offered a more realistic picture of the field. Humanoid robotics is advancing quickly, but reliability and robustness are still major bottlenecks.
The World Humanoid Robot Games were overwhelmingly dominated by Chinese manufacturers, with no European entrants prominently reported. The concentration reinforced the event's role as a domestic industrial showcase rather than a balanced global contest. It also suggested how much of the current humanoid momentum is being organized through China's manufacturing ecosystem and state-backed visibility. For rivals, the message was that scale, iteration speed and national coordination are becoming strategic advantages.
GrokBot was presented as a network of personal AI assistants built around long-term memory and role-based setup. Each bot can be assigned a name, description and ongoing purpose, then retain context from earlier interactions to become more useful over time. The design points toward AI that behaves less like a single chatbot and more like a persistent household operating layer. That memory-first model is central to making automation practical in everyday life.
One GrokBot use case, Fixer, stores appliance details, paint colors, filter sizes and maintenance dates by ingesting photos and notes. Other assistants connect with Gmail and Google Calendar to coordinate doctor visits for elderly parents, pull prep instructions from emails and keep separate visit histories. The same setup can handle reminders, appointment booking, subscription tracking, bargain hunting and weekend planning. Together, the examples show consumer AI moving from question answering toward delegated household and caregiving logistics.