OpenAI and Broadcom unveiled Jalapeño, a custom inference processor built from scratch in nine months to run current and future LLMs. Early tests show it delivers substantially higher performance per watt than existing accelerators, positioning OpenAI to reduce dependence on external chip vendors and scale its AI services.
Anthropic says Alibaba ran a distillation campaign that siphoned 28.8 million Claude interactions across 25,000 fake accounts, the biggest known AI model theft to date. The alleged attack could accelerate Chinese rivals toward Anthropic’s Mythos capabilities and highlights a looming security gap in frontier AI models.
A high‑school developer known as slqnt rewrote Valve’s classic FPS in just three months, using WebGL and WebAssembly so the full game loads and plays inside a normal browser. It proves the web platform can handle demanding 3‑D titles, but the demo may be short‑lived.
The post argues that AI agents should output GitHub pull requests instead of chat replies, turning model output into a reviewable artifact. PRs embed code changes, CI checks, and approval workflows, letting humans focus on validation rather than manual copy‑paste. This shift promises tighter integration, clearer accountability, and safer AI‑driven development.
Google’s Gemini 3.5 Flash now embeds native computer‑use tools, letting agents see, reason, and act across browsers, mobile and desktop. This integration improves long‑horizon automation like continuous testing and knowledge‑work, while safety layers guard against prompt‑injection risks. Developers can access it via the Gemini API and Enterprise Agent Platform.
Linear’s React apps are leaving styled‑components for StyleX, a build‑time CSS solution. The move slashes runtime style generation, restores strong component encapsulation, and avoids the maintenance‑mode limits of styled‑components, which never adopted React 18’s useInsertionEffect. The migration also future‑proofs the UI stack for modern React patterns.
In April Anthropic released a postmortem revealing three separate changes that broke Claude Code’s performance, high effort mode latency, Agent SDK issue, and Cowork glitch, leading to UI freezes and reduced usage limits. All were fixed by April 20 and future defaults switched to medium effort to prevent recurrence.
Ford discovered its AI‑driven quality systems missed critical defects, prompting the automaker to rehire over 350 experienced engineers, including former staff and supplier experts, to restore institutional knowledge and train junior teams. The move highlights the limits of AI when data and tacit expertise are insufficient, underscoring the need for human oversight in high‑risk manufacturing.
A Rust rewrite of the 300k‑line MediaInfoLib and ExifTool stack reproduces their XML output byte‑for‑byte while cutting runtime overhead and eliminating unsafe code. The project proves you can replace legacy C++ and Perl media parsers with a single, cargo‑installable crate, simplifying builds and tightening security.
A federal judge denied Workday’s bid to toss a class‑action alleging its AI screening tools discriminate against Black, older, and disabled applicants. The ruling keeps claims under California’s Fair Employment and Housing Act and the ADA alive, signalling heightened legal risk for vendors using automated hiring.
AI labs like DeepMind and Anthropic now employ dozens of in-house philosophers to shape model behavior, tackle bias, and define value alignment. Their presence signals that ethical reasoning is becoming a core part of AI development, not an afterthought, and opens a new career path for humanities graduates.
You’re caught up.
That was all 11 stories in the 25 June issue.