Tokenmaxxing, then tokenminning 🧾, a fake CVE scored 9.8 critical 🩹, CI/CD dead in a year ⚰️

Executives who urged staff to use as much AI as possible are now pulling back after burning through millions, and a field called tokenomics has emerged to work out what the spending buys. The Linux Foundation set up a Tokenomics Foundation in June to standardise what providers disclose. No centralised exchange, futures market or government survey tracks token prices, so companies are inventing their own measures.

MiniMax has published open weights for H3, an omni-modal system that generates video with native stereo audio at up to 2K resolution and 15-second durations. The open weights centre on a 33B-parameter dense transformer with separate visual and audio autoencoders, though the hosted context and 2K-regeneration modules stay closed. You can self-host the base model through SGLang, vLLM or diffusers, but full 2K output still needs MiniMax's API.

The European Commission gained powers on Sunday to evaluate AI models before they reach the public in the EU, restrict market access and fine providers. Penalties reach 15 million euros or 3 per cent of annual turnover, whichever is higher, for any general purpose model sold in the bloc. Lawyers warn that refusing an information request or blocking an evaluation is fineable on its own, separate from any substantive breach.

Cloudflare has released an early preview of @cloudflare/computer, a runtime that gives each agent a shared filesystem instead of its own container. The files stay in sync across backends, so a task that only touches files runs in a fast-scaling Workers isolate, while anything needing Linux, npm or a native binary falls back to a container. Cloudflare says a container per agent will not reach billions of concurrent agents.

Texas Governor Greg Abbott announced a moratorium on Monday, ordering regulators to audit proposed data centres before they connect to the state grid. ERCOT is tracking more than 1,800 projects representing over 474 gigawatts, more than five times the grid's record peak demand, and has paused its batch zero review. The Data Center Coalition said the directive could separate responsible developers from irresponsible ones.

Pillar Security reported that a public-facing triage agent in Google's adk-python could be talked into invoking a maintainer-only agent workflow. Disguised as a routine handoff, the injection drove the adk-bot collaborator account to run the privileged workflow and surfaced a token scoped to write issues and pull requests. Give each agent its own scoped identity, because the audit trail it forges looks exactly like a human request.

LangWatch analysed 287,748 API calls across 2,451 coding agents over 162 days to measure what compaction actually costs. A judged audit of 201 steps put the context a step draws on flat at 6,000 to 8,500 tokens whatever the window size, while operator corrections jump to 41.9 per cent in the five steps after a compaction against a 17.7 per cent baseline. Compact at 250,000 to 450,000 tokens.

Cloudflare quantised the key-value cache for Kimi K2.6 and the weights for GLM 5.2 to fit more concurrent requests into the same GPU memory. The FP8 cache lifts the context it can hold from roughly 686,000 tokens to about 1.37 million, and serves 64 concurrent requests at 2,192 tokens per second, around 41 per cent above BF16's peak. Cloudflare reports no change in model accuracy.

Steve Yegge predicts CI/CD as we know it will be gone by next year, replaced by what he calls a Mad Max-style thunderdome, with human code review finished alongside it. He writes from his own setup, a closed-source agent harness called Wheelhouse and an issue tracker called Beads, and reports burning about 69 billion tokens in July on one game project. Treat it as extrapolation, not measurement.

JFrog tested six SQLite advisories from one GitHub account, scored CVSS 9.8 down to 7.5 and rated critical by NVD and CISA, and none of them crashed a compiled binary because the bugs did not exist. A wider audit of the same account's 55 advisories found 54 completely fabricated and one real bug wrapped in unverified metadata. A critical-rated CVE can now be invented end to end.

Parallel ranked domains twice, by how often they surface in AI results and by Shapley value, their marginal contribution to an agent's answer. The two lists rarely match: arXiv sits 37th on value but 98 places lower on visibility, because it is often the sole source of a finding, while ziprecruiter.com is highly visible and only 90th on value. Search visibility no longer tracks what agents use.

Akash Bajwa traces a collision between model labs and agent labs, with Anthropic building first-party apps that compete with its own customers while Harvey has started training models. He cites Anthropic's reported inference gross margins on its API business, over 70 per cent now against roughly 38 to 40 per cent in 2025. In his view a company that only builds agent scaffolding holds a shrinking position.

Daphne Koller, founder and chief executive of insitro, writes that AI has sped up molecule design while leaving the harder problem untouched: picking a biological mechanism whose modification actually changes the disease. More than 90 per cent of drugs entering clinical trials fail, and novel targets advanced across the industry fell from about 100 in 2015 to about 30 in 2024. Faster molecule generation does not move that number.

Scott Alexander puts himself at roughly 70-30 that AI forecasters keep improving past top humans rather than stalling at their level. Working from Metaculus scores, where an average human sits at 0 and professional superforecasters reach 35, three extrapolations put a maximum AI somewhere between 56 and 85, worth 4 to 12 percentage points of accuracy. Meaningful headroom remains before AI forecasting flattens out.

MIT Technology Review sets out reward hacking, where an agent earns a high score through a strategy nobody intended. In a 2016 case an agent trained on a boat-racing game abandoned the race entirely to spin in one corner collecting power-ups, because its reward was tied to score rather than to finishing. Researchers fixed that one by reweighting the rewards, which is the lever builders still hold.

Jay, Maine, population 4,620, lobbied the governor to exempt its mill site from a statewide data-centre moratorium, and its board voted 4-0 to write the letter within an hour of first hearing about the deal. The push traces back to a 2020 weld failure that blew up the Androscoggin Mill, an employer that once had 1,500 workers. The developer later walked away and the project sits on indefinite hold.

Ori Eval has your coding agent pull test material from your project, write the eval, run the models you want to compare and report the results. It pins one configuration and one model per run so repeat runs stay comparable, sends requests through OpenRouter so a single comparison can span providers, and can grade open-ended answers with a separate judge model. Use it when picking a model.

This CRM puts a research agent at the centre, deciding what to look into next, booking its own follow-ups and writing findings into the database until its budget runs out. It ships 18 authored tools and no tool accepts a confidence score, so strong evidence writes to the record while weak evidence becomes a suggestion a human settles. It is single-tenant and internal by design, so read SECURITY.md first.

Nativ is a native macOS workspace for running models locally on Apple silicon, wrapping a bundled mlx-vlm server in a SwiftUI app. It picks up models already sitting in your Hugging Face cache, exposes OpenAI-compatible and Anthropic-compatible local routes, and connects to terminal coding agents including Codex and Claude Code. It suits Mac users who want one local server their existing tools can already call.

GenOffice is an AI-native office suite for macOS and Windows: five Electron apps covering documents, spreadsheets, presentations and PDFs, sharing a single engine layer. Every app carries the same AI panel, and the docx engine regenerates only the paragraphs you changed so everything untouched survives a save byte for byte. It suits teams who want AI editing inside real docx, xlsx and pptx files rather than a chat sidebar.