More than 200 researchers, economists and technology leaders signed an open letter titled "We Must Act Now" on Monday, warning that artificial intelligence could trigger economic disruption larger in scale than the Industrial Revolution — but compressed into a fraction of the time societies had to adapt then.
The list is what makes this notable. Signatories include Nobel laureates Daron Acemoglu and Simon Johnson — economists who spent years publicly downplaying AI displacement fears — alongside frontier-lab figures like Anthropic co-founder Jack Clark, Google DeepMind's Jeff Dean, and OpenAI's finance chief Sarah Friar.
Organiser Anton Korinek of the University of Virginia framed the urgency plainly: earlier technological transitions gave societies decades to adjust; this one may allow only a few years. The letter stops short of naming specific policies, calling instead for governments, economists and technology leaders to jointly build the institutions and guardrails needed before disruption arrives.
Paris-based Mistral AI unveiled Robostral Navigate, its first embodied-AI model, on July 8. Given an RGB image and a plain-language instruction — "leave the lobby, walk through the corridor, enter the supply room" — the 8B model moves a robot through the space using no LiDAR, no depth sensor, and no map.
On the R2R-CE benchmark, the standard test for instruction-following navigation in unseen environments, it scores 76.6% — beating the best single-camera system by 9.7 points and multi-sensor rigs by 4.5 points, despite using neither depth nor multiple cameras. The model was trained entirely in simulation across roughly 400,000 paths in 6,000 virtual environments, and generalises across wheeled, legged and flying robots.
The strategic read: a foundation-model company with existing BMW and Airbus enterprise ties just made a hardware-agnostic case against the LiDAR-heavy navigation stacks that have defined industrial robotics for years.
OpenAI released its GPT-5.6 family on July 9 after a government-vetted preview period tied to the model's cybersecurity capability. Sol is the flagship, Terra a balanced everyday-work tier, and Luna the budget option — a naming shift OpenAI says separates model generation from durable capability tier.
Alongside the models, OpenAI launched ChatGPT Work, an agent built on Codex that gathers context across connected apps and stays with multi-step projects for hours, returning finished sheets, slides, docs and web apps rather than conversation. It merges the standalone Codex app into one ChatGPT desktop client.
OpenAI's own benchmarks claim Sol edges out Anthropic's Fable 5 on its coding-agent index while using under half the tokens; independent developer reactions have been mixed, with several builders noting Fable's edge on raw reasoning even as GPT-5.6 wins praise for reliability on everyday tasks.
Security researcher cereblab routed xAI's Grok Build CLI through a network-interception proxy and found it silently bundling entire tracked Git repositories — full commit history and unredacted .env secrets included — to a Google Cloud Storage bucket named grok-code-session-traces.
On a 12GB test repository, the model-response channel needed just 192KB to answer the coding task; a separate storage channel quietly shipped 5.1GB. Disabling the tool's "Improve the model" privacy toggle had no effect — the upload continued regardless, since the setting governs training consent, not data transmission.
xAI fixed the behaviour server-side within a day of disclosure and Elon Musk pledged to delete previously uploaded data, but the company has issued no formal security advisory or retention policy as of publication.