Desk
Broad artificial intelligence coverage.
NVIDIA is open-sourcing a shared way to describe Kubernetes AI workloads, a small-sounding fix to a costly problem that also deepens its reach into the operations layer.
With Codex multiplying branches and pushes, OpenAI’s source control reliability at times fell below 90%. Pando, a Rust Git backend now serving its whole monorepo, cut median push latency nearly 80% by dropping repository-wide snapshots. The lesson: agent output stresses everything around the code.
OpenAI says future compute campuses will cover their infrastructure costs and curb demand when the grid is strained, starting with a planned Effingham County, Georgia campus. The pilot isn’t running yet, the savings figures are utility forecasts, and the dispatch terms decide whether it’s real.
OpenAI told users it will combine Chat and Work in the desktop app, testing with select users first. The less noticed change is that ChatGPT and Codex will draw from one token pool, which may mean more frequent resets for heavy users as agentic tasks grow.
The metric is right: judge AI coding spend by merged work, not tokens. Just remember this case comes from a vendor that sells the on-premises fix.
LifeVal, built from time-use surveys and a 3,995-person poll, tests whether models can handle everyday chores. OpenAI says it has used it for a year to push ChatGPT toward taking actions. Without public tasks or grading details, outsiders can’t check the high scores.
NVIDIA is betting that the hard part of data agents isn't writing SQL, it's knowing what the numbers mean, and it has open-sourced the layer that's supposed to hold that meaning.
A solo developer compiled a tiny working GPU into a small language model's weights, and while it's far too slow to use, it's a sharp argument that models can compute exactly, not just guess.
Driscoll takes a newly created role advising OpenAI leadership and partners on AI for national and cyber security, with an emphasis on US allies. The hire signals a deliberate push into defense; the release names no limits on what OpenAI will build or who reviews deployments.
Qodo feeds Codex company rules and PR history before it writes code, then reviews the result in the same session to catch cross-repository breakage. The builder-plus-reviewer design is sound. The OpenAI customer story offers no measured results.
Axios reports OpenAI’s Chris Lehane sees the midterms as a preview of a 2028 AI election, with ideas including AI speed limits and the rename. Opposition to data centers is about bills, water and secrecy, so a new name without commitments will read as spin.
Nace's top decision model just left the cloud. Routing, triage and policy checks can now run on a well-equipped Mac, and that changes who gets to audit them.