OpenAI offers Washington a 5% stake 💼, token subsidy era ends 💰, de-slop your AI prose 🧑‍💻

OpenAI has proposed handing the US government a 5% stake in the company to defuse mounting political pressure in Washington, the Financial Times reported. At OpenAI's record $852 billion valuation, that holding would be worth roughly $42.6 billion. Sam Altman reportedly argued in early talks with the Trump administration that giving the public a financial interest is the best way to share the upside of AI, though it is unclear whether the administration will pursue it.

Security researcher Ian Carroll used Claude Opus 4.7 to exploit a bug in Front Gate Tickets, the Live Nation subsidiary that runs ticketing for Lollapalooza, SXSW and Bonnaroo. The flaw gave him super-administrator access to millions of customer and staff records, plus the ability to issue any ticket at any price to anyone. It is a grounded look at what agent-assisted hacking actually looks like today, well short of nuclear codes but very real.

Google added two models to its Gemini Enterprise Agent Platform. Nano Banana 2 Lite, formally Gemini 3.1 Flash-Lite Image, is now generally available as the fastest and cheapest model in the Nano Banana family. Gemini Omni Flash arrives in public preview for high-quality video generation and conversational editing, handling character and product swaps, style transfers and relighting. One early benchmark tempers the hype: only 16 of 34 science clips scored accurate.

Bloomberg reports Meta is drawing up plans for a cloud business that would sell access to its AI compute and models, pitting it against AWS, Google Cloud and Azure. The move follows SpaceX and xAI renting out capacity, and hints the AI race may be won by whoever owns the data centres, not whoever ships the best model. Meta has committed $182.9 billion to infrastructure, and the stock jumped about 10 percent on the news.

OpenAI engineers told colleagues they had developed optimisations that more than halve inference cost, the expense of running existing models, according to The Information. Applied to logged-out ChatGPT visitors, the techniques cut the Nvidia GPUs needed to serve them to a couple hundred, a startlingly small number. The company will not say how, but the savings feed a gross-margin target of 52 percent it is chasing by year end, up from 39 percent.

Together AI raised an $800 million Series C led by Aramco Ventures, with Nvidia and Vista participating, more than doubling its valuation to $8.3 billion. It runs open-source models like DeepSeek, MiniMax and Kimi on its own inference stack, which Cursor, Cognition and Decagon use to cut costs. Decagon says moving workloads there ran a fifth to a seventh of closed-model pricing, and Together's annual bookings have crossed $1.15 billion.

Apple has added a Safari MCP server in Safari Technology Preview 247, giving coding agents a direct line into a real browser window. Through it an agent can inspect page content, read console logs and network requests, capture screenshots and interact with elements, rather than working from whatever a developer pastes into chat. Tools like Claude, ChatGPT and Gemini can now debug against how a site actually renders, not a stale snapshot.

Check Point researchers report that DeepSeek, which carries fewer safety controls than rivals, will generate a browser-native ransomware sample given the right prompt, a technique they had not seen used in the wild. Of nearly 3,000 DeepSeek-attributed files tracked this past year, almost half were flagged malicious. The sample was incomplete, but the team says making it attack-ready takes little effort and low skill, and threat actors are already trying.

Cerebrium walks through how it attacks GPU cold starts, the multi-minute startup times that force teams to keep costly GPUs warm and over-provision. Its approach snapshots a fully initialised container, CPU and GPU memory, process state and warmed CUDA context, then restores it directly rather than rebuilding. Alongside custom VM images and a sub-second image runtime, the snapshot layer restores warmed workloads in seconds and cuts cold starts by over 80 percent.

ByteByteGo breaks down how OpenAI serves voice to 900 million weekly users over WebRTC, a protocol built for stable server IPs that Kubernetes treats as disposable. Rather than the usual SFU that suits group calls, OpenAI splits its one-user-to-one-model traffic: a stateless relay routes packets at the edge, while a stateful transceiver holds the heavy WebRTC state. The two connect using the ICE ufrag as a routing key read off the first packet.

Factory's autonomous agents commit code faster than any human can review, so it built Droid Shield to stop secrets slipping into commits. The old deterministic scanner had two failure modes: false positives on placeholders that trained users to ignore it, and false negatives on secrets outside its patterns. Version 2.0 swaps in two fine-tuned models, one per mode, that Factory says beat frontier classifiers far more cheaply, with the weights released.

Meta details how its BLOB-storage architecture evolved to stop storage becoming the bottleneck that stalls AI training GPUs. Compute performance has roughly tripled every two years while storage and interconnect growth lagged, so slow access raises cost, and shuffling datasets across geo-distributed GPUs drags on research velocity. Built on its Tectonic fabric across hundreds of exabyte-scale clusters, the design targets both GPU utilisation and iteration speed.

Count Bayesie's Will Kurt takes on the instantly-recognisable feel of LLM prose, which Ben Affleck pinned as writing that goes to the mean. Because standard decoding optimises for the expected next token, models drift toward the average. Kurt modifies sampling to weigh the future entropy each candidate token opens up, not just its immediate probability. On local models, where you control the sampler, it pushes prose off the safe average.

An Arize piece reads the first clean third-party cost data, Artificial Analysis's AA-Briefcase benchmark, and makes the case the flat-rate era is over. One agentic task costs about $31 on the smartest model, which gets it fully right 3 percent of the time, roughly $1,000 per correct result. SemiAnalysis priced a maxed $200 ChatGPT Pro plan far higher at API rates: heavy agent users are who the labs must stop subsidising.

In Chartbook 455, Adam Tooze argues the world economy is running two powerful, oddly separated forces this summer. The American AI boom is driving a K-shaped expansion and pulling in imports at pace, but from Taiwan and Mexico, not China. Meanwhile China's supply-driven expansion delivers abundance and deflation. With the Sino-US truce holding and the bilateral deficit narrowed, the China-shock fight has moved to China and Europe, not a US collision.

Andreas Kirsch, a senior research scientist at Google DeepMind, says in a personal essay that the lab's bet on safety culture over formal governance has failed its biggest test, the Pentagon contract Google reportedly signed in April. Good people, he writes, do not substitute for real governance: independent oversight that can say no, transparency, and accountability when commercial pressure meets stated principles. It is a rare on-the-record dissent from inside a frontier lab.

dox is a tiny AGENTS.md framework that gives a coding agent precise local context instead of letting it edit blindly. It keeps a hierarchy of AGENTS.md files, a root index plus child files per area, and has the agent walk from root to the code it will touch before editing, then update the affected files. No install or dependency: copy one Markdown file, and it works with Codex, Claude Code and OpenCode.

Cloudflare's new Monetization Gateway lets customers charge for any asset behind Cloudflare, web pages, datasets, APIs or MCP tools, from one control plane, enforced at the edge. Payments settle in stablecoins over x402, the open protocol it is building with more than 25 firms. The logic is blunt: AI crawlers already hit content a hundred to tens of thousands of times per visitor they return, so the old attention bargain no longer pays.

mog is a spreadsheet engine, app runtime and SDK for building agents and automations that understand workbooks. The headless SDK runs in Node, Workers and other WASM-capable hosts, so an agent can create a workbook, open or load XLSX files, set cells and formulas, and read computed values in code. It gives agents a real calculation engine to work over, rather than treating a spreadsheet as loose text to parse.

context-chat drops an Ask-AI assistant onto any docs site with a single script tag. It is one self-hosted Cloudflare Worker that retrieves relevant docs from your search backend and streams an OpenAI answer with inline citations, behind a layered abuse-control stack so a public endpoint cannot be used to burn tokens. Open-sourced from Dodo Payments' production assistant, it pairs with a companion indexer but works against any search endpoint via an adapter.

ZCode is Z.ai's native coding harness, and its 3.0 release rebuilds the core: it swaps the third-party kernels it used to bundle, Claude Code and Cline, for a self-built ZCode Agent tuned to GLM-5.2's long-horizon reasoning and tool-calling. So you track Zhipu's kernel cadence rather than upstream releases. It also leans into multi-agent collaboration and adds a Goal feature for steering larger builds from planning through verification, as a desktop app on macOS, Windows and Linux.