OpenAI ships its own AI chip 💻, the tokenmaxxing era ends 💼, AI designs radio chips 🧑‍💻

Jalapeño is a custom inference ASIC purpose-built for large language models and agentic workloads, not a repurposed training accelerator. The design pairs a large compute chiplet with HBM memory to attack inference bottlenecks like costly data movement and compute-memory balance, targeting higher performance per watt than today's leading hardware. It is OpenAI's first generation of its own silicon, a strategic move to lessen its reliance on Nvidia.

Qualcomm had a double AI-infrastructure day. It agreed to acquire AI software firm Modular for about $3.9 billion in stock to make inference faster and cheaper across hardware, and Dragonfly C1000, its new data-centre CPU for agentic AI, arrived with Meta as its first Big Tech customer when production starts in 2028. The chipmaker now targets $15 billion in data-centre sales by 2029, a direct swing at Nvidia.

In a letter to US senators, Anthropic accused Alibaba of "brazenly" and "illicitly" running the largest known distillation attack against its models, using thousands of fraudulent accounts to make 28.8 million exchanges with Claude between April and June. The campaign targeted Claude's software-engineering and agentic-reasoning capabilities. Separately, Alibaba sued the US government over its addition to a list of firms linked to the Chinese military.

Google made computer use a built-in tool in Gemini 3.5 Flash, moving it out of the standalone Gemini 2.5 model and into the main Flash model. Developers can now build agents that see, reason, and act across browser, mobile, and desktop environments through the Gemini API and Enterprise Agent Platform. Google points it at long-horizon work like continuous software testing and knowledge tasks across professional applications.

SK hynix filed to raise up to $29.43 billion through a Nasdaq listing scheduled for July 10, with every dollar earmarked for new fabs, an HBM packaging plant, and EUV chipmaking gear. None of the funded projects will produce memory in time to ease the shortage still driving up prices. The company holds about 57% of the HBM market and expects AI demand to keep supply tight until 2030.

After a year of urging staff to max out AI usage, companies are now rationing it. Leaked audio reported by 404 Media captures Accenture trying to stop staff from draining its token reserves on basic tasks like turning PDFs into slides, just after warning that non-users could miss promotions. "AI is becoming material to the cost structure," its agentic-AI lead said, with leaders still asking whether the spend returns value.

Despite layoffs that cite AI and coding tools, SignalFire's analysis of millions of careers across 80 million companies ranks engineering as the most resilient function of 2025. Total hiring at large tech firms fell 25% from 2019 levels, but engineering roles dropped just 11%, and engineers made up 55% of all new hires at the top 12 companies. The on-the-ground data, the firm says, contradicts the replacement story.

Princeton researchers used reinforcement learning and inverse design to generate radio-frequency integrated circuits from scratch, a domain long treated as a hand-crafted "dark art." Diffusion models produced novel or human-interpretable RF layouts that hit record performance while cutting design time sharply. The bottleneck now is data: the authors argue progress needs large, shared chip-design datasets and open ecosystems so models can learn universal electromagnetic behaviour.

A hands-on test ran Z.AI's open-weight GLM 5.2 through four real jobs inside a production Next.js codebase: an architecture audit, a UI redesign, and a 45-minute autonomous bug hunt pulling from Sentry and Vercel logs. Total cost was $3.36 for roughly 6 million tokens, and it matched an existing design system on the first try. The pitch is cost and vendor independence, not just benchmark wins.

Qwen released Qwen-AgentWorld, a pair of language world models (35B-A3B and 397B-A17B) trained on more than 10 million real interaction trajectories to simulate agentic environments across seven domains. Used as a decoupled simulator, it lets teams train agents on thousands of controllable synthetic environments, with gains that surpass real-environment training alone. As a foundation model, the same world-model training acts as a warm-up that lifts performance across seven agentic benchmarks.

Apple runs its most ambitious AI feature in about a gigabyte of memory on the iPhone while running a far larger model on its own servers, and the two share little beyond the word "transformer." It explains why the two diverge on hardware target, training, and architecture, and why the same device-versus-datacentre split repeats at Google, Microsoft, and Meta. They are answers to different constraints, not points on a ladder.

Wing partner Tanay Jaipuria frames the choice now confronting product builders: become the agent your users open, or power the horizontal agent they already live in, like Claude Code, Codex, or Copilot. Building the agent means owning the interface and the usage, but fighting for a habit on tools where the day already starts. He covers how to tell which way you lean and what doing both actually demands.

Robert Englander argues the industry took the wrong lesson from ChatGPT: the breakthrough was natural-language understanding, not conversation. Language-native software treats language as the interface, resolving stated intent into a deterministic operation and falling back to dialogue only when meaning is unclear. Execution, he notes, is the easy part for databases and schedulers; deciding which instruction to run is the hard problem language models actually solve.

MagicSchool, a generative-AI platform for K-12, grew from zero to 7 million users and $10 million ARR in under a year, in a sector VCs treat as a startup graveyard. Founder Adeel Khan, an ex-teacher, reached 1M users in five months organically and now serves districts covering one in five US children. A sales team that is 85-90% former teachers turned an AI wrapper into a category platform.

Walmart regional load manager Leo Garcia used what he learned in an AI course to build tools with Code Puppy, the company's in-house coding agent. One analyses hundreds of available truckloads and flags the best five, getting drivers home sooner while running fewer empty trailers. A former driver himself, Garcia is the kind of non-engineer Walmart is now equipping across the company to build their own software.

DESIGN.md is a Google Labs format spec that describes a visual identity to coding agents, giving them a persistent, structured understanding of a design system instead of re-deriving it each prompt. The project is trending hard, with more than 17,000 stars and a 0.3.0 version landed in mid-June. Point your agent at a DESIGN.md and generated UI should match your tokens and components rather than drifting.

hunk is a review-first terminal diff viewer for agent-authored changesets, built on OpenTUI and Pierre diffs. It mirrors Git's diff and show commands but opens changes in a review UI with sidebar navigation, inline AI annotations, and a watch mode that auto-reloads as the working tree changes. It auto-detects Jujutsu and Sapling too, and installs in one line through npm or Homebrew.