
Tech • AI • Robotics
Anthropic said an unreleased Claude-based research system improved a long-stalled bound tied to the Riemann hypothesis from 41.6% to 67.2%. The run reportedly used about 60 parallel sub-agents, generated 31 million output tokens, executed 2,400 commands, and reviewed 54 arXiv papers across 36 hours. Internal mathematicians checked the result before outside verification by Brian Conrey and Dan Goldston. The advance is being framed as machine-assisted synthesis rather than a proof of the Riemann hypothesis itself.
The reported breakthrough did not hinge on wholly new mathematics. Instead, the system combined two previously published papers from different subfields that had not been connected in this way. That makes the result notable as an example of AI-driven literature synthesis and search over large technical spaces. It also sharpens the distinction between formal theorem proving and broader research assistance.
From August 14, Claude Code is set to make Auto mode the default for Pro, Max, and Team users. Anthropic is replacing repeated permission prompts with a classifier-based safety layer intended to preserve protections while reducing friction. The shift signals growing confidence that coding agents can be trusted with more autonomous execution in mainstream workflows. It also raises the competitive bar for developer tools that still rely on constant user approvals.
Unitree unveiled a humanoid robot called Superman and said it can hit a 2 meter standing vertical jump and 12.66 meters per second top speed. Those figures would exceed the reported human standing-jump mark of roughly 1.8 meters and edge past Usain Bolt’s peak speed near 12.42 m/s. The company said the machine was built in just over three months, with hardware still described as unfinished. If validated, the demonstration would mark a major leap in actuator power, balance, and high-speed control.
The headline numbers come from company demonstration footage and have not been independently verified. Engineers note that the jump’s landing phase is as important as the lift, because absorbing a 2 meter drop without instability is a severe control and durability test. For now, the footage is better read as evidence of extreme hardware performance than of immediate field readiness. That caveat matters as humanoid robotics marketing grows more aggressive.
Robotics startups have reportedly raised more than $18 billion since January, already surpassing all of 2025 while the year is only halfway through. Agility Robotics is said to be preparing a public listing around $2.5 billion, while Figure is valued near $39 billion. A German robotics company also reportedly secured nearly $1.5 billion in a single round. The financing surge suggests investors increasingly see physical AI as the next major commercial platform.
Developers are increasingly abandoning the idea that a single large model can serve as a robot’s entire brain. The emerging architecture is layered: separate systems handle planning, world prediction, action, monitoring, and verification in sequence. That reflects hard lessons from earlier waves of brittle factory, warehouse, and humanoid automation. The new thesis is that progress comes from combining stronger hardware with modular multimodal intelligence, not from dropping one frontier model into a body.
Grockbot cut entry pricing from $200 to $60 a month through Cursor Pro and added a limited free tier. The product is being positioned as a multi-agent workspace, with specialized agents for tasks such as inbox triage, meetings, research, investing, and brand partnerships. It also touts a connector marketplace spanning tools including Gmail, Google Calendar, ClickUp, Granola, and X. The pricing move broadens access as AI workspaces compete to become the default automation layer for knowledge workers.