A daily newspaper for the age of artificial intelligence.
Morning edition
· No. 3
Sources linked throughout
Front pageImportance 10/10
The Reviewer Needed a Map, Not More Territory
GitHub says a tool upgrade made Copilot code review costlier and less effective until its instructions were rebuilt around the pull-request diff.
Editorial illustration
Conceptual illustration: a mechanical reviewer stays anchored to a narrow trail of evidence instead of exploring an entire archive. It does not depict GitHub software or an actual repository. Generated with Codex Image Gen for The Machine Press, 2026-07-11.
The July 10 GitHub engineering account offers an unusually concrete lesson in agent design. Moving Copilot code review onto shared grep, glob, and view tools initially raised cost and reduced useful findings because the reviewer began exploring repositories like a general coding assistant. GitHub then rewrote the workflow to start from the diff, form narrow questions, batch discovery, and read only the evidence required. The company reports roughly 20 percent lower average review cost than the control without a quality signal that blocked shipping. Those figures come from the internal production comparison, but the traced failure mode is the durable point: the same tool surface can produce different behavior when its instructions imply a different job.
Conceptual illustration: distributed selectors combine local token choices into one coherent memory map. It is not a Microsoft system diagram or measured benchmark result. Generated with Codex Image Gen for The Machine Press, 2026-07-11.
Million-Token Memory Without the Global Bottleneck
Microsoft researchers propose interleaving context across GPUs so each device can select locally while the combined result nearly covers the global top tokens.
Exact top-k sparse attention can preserve quality over long contexts, but coordinating a global selection across many GPUs can erase the computational savings. In a July ICML workshop preprint, Microsoft researchers Yifan Guo, Wei Cui, and Peng Cheng propose Interleaved DeepSeek Sparse Attention. Tokens are distributed in an interleaved layout and each device performs a relaxed local selection; the union of those selections is designed to cover nearly all globally relevant tokens with less synchronization. This is a research proposal, not a deployed product claim, and its practical value depends on evaluation across models and hardware. Still, it reframes the long-context problem as a systems question: how to retain useful memory without making every processor wait for a perfect global vote.
Illustration: a generic rendered wafer-scale circuit. It does not depict Meta's Iris design, a fabricated chip, or the reported production line. Wikideas1 / Wikimedia Commons (CC0 1.0); cropped and converted to WebP by The Machine Press.
Meta Schedules Its Own AI Chip for September
An internal memo reviewed by Reuters points to a faster custom-silicon cadence as Meta aims to expand its computing base.
Meta plans to begin manufacturing an in-house AI chip code-named Iris in September, according to an internal memo reviewed by Reuters. The report says Iris is part of a four-generation MTIA program and that Meta intends to release a chip about every six months through 2027 while targeting 14 gigawatts of overall computing power next year. The memo describes a plan, not a completed production run; yield, deployment, and performance remain to be demonstrated.
Sol, Terra, and Luna are rolling out across editor, command-line, cloud-agent, mobile, and web surfaces.
GitHub began rolling OpenAI's GPT-5.6 family into Copilot on July 9. GitHub positions Sol for complex, long-running work, Terra as the balanced default, and Luna for smaller, lower-cost tasks. Availability varies by plan, rollout is gradual, and Business and Enterprise administrators must explicitly enable the models because their policy starts off by default.
Version 2.26.0 adds queries for workflows where untrusted text can reach an LLM and influence tool-using behavior.
GitHub released CodeQL 2.26.0 on July 10 with support for Kotlin 2.4.0 and new analysis aimed at AI prompt-injection risks. The security queries look for flows in which attacker-controlled input can reach a model prompt or agent workflow without adequate safeguards. Static analysis cannot prove that every flagged path is exploitable or that every unflagged system is safe, but the release turns a class of agent-security concern into something development teams can search for continuously alongside conventional code vulnerabilities.
Illustrative infrastructure photograph: a server room at The National Archives in the UK, not a tested machine, compromised repository, or AI-hosting facility. The National Archives (UK), via Wikimedia Commons, CC BY 3.0; cropped and converted to WebP by The Machine Press.
The Security Reviewer Can Become the Entry Point
A proof of concept shows coding agents in autonomous modes executing attacker-controlled material while inspecting an untrusted repository.
AI Now researchers Boyan Milanov and Heidy Khlaaf demonstrated what they call Friendly Fire: prompt injections in a third-party codebase led Claude Code and Codex configurations with autonomous command approval enabled to run attacker-controlled code during a security review. The finding is configuration-specific, not evidence that every interactive review is compromised. Its practical lesson is broader: untrusted repositories should be isolated, and automatic execution should not be treated as a neutral extension of code reading.
Muse Spark 1.1 arrives with a public-preview model API, long context, and an emphasis on orchestrating tool-using agents.
Meta introduced Muse Spark 1.1 on July 9 as a multimodal reasoning model aimed at agentic work. The company says the model can plan, delegate across parallel subagents, use computers and tools, and manage a one-million-token context window. More consequential for outside builders, Meta also opened a public preview of its Meta Model API. The performance claims remain Meta's own evaluations, but the API changes the release from a consumer-app update into a developer platform move.
Meta's Muse Image uses search, code, and self-refinement before it returns a picture; Muse Video remains an early preview.
Meta released Muse Image across Meta AI and selected social surfaces, describing an image model that can invoke search and coding tools, revise its own generations, and compose multiple references. The accompanying Muse Video model is not yet a general release. Meta says images created in its products carry an invisible Content Seal provenance signal intended to survive common edits such as cropping and compression. Rankings and quality comparisons in the announcement are time-stamped company claims, while the release and watermarking mechanism are the durable news.
Generic source-code illustration; it does not depict a named repository, license, Copilot interface, or generated overview. D. Charbonnier / The Noun Project, via Wikimedia Commons (CC0 1.0); padded, gold background added, and converted to WebP by The Machine Press.
Copilot Adds a First-Pass Map for Unfamiliar Repositories
GitHub's new overview gathers purpose, technologies, and contribution guidance before a newcomer starts reading file by file.
GitHub now offers Copilot-generated overviews on repositories a user has not contributed to before. The summary draws from repository context to describe purpose, technologies, and contribution guidance, with shortcuts for recent changes and ways to contribute. GitHub says the feature is available across all Copilot plans; maintainers should still treat a generated overview as orientation, not a substitute for authoritative project documentation.
An internal cleanup mapped more than 14,000 repositories to validated teams and archived those without a durable home.
GitHub described a 45-day internal program that assigned validated ownership to every active repository in an estate of more than 14,000. Fewer than half had clear ownership at the start. The company used repository metadata, activity, and team review to identify accountable owners, then archived repositories that no group would claim. This is an internal account rather than an independent audit, but it makes an operational point relevant to AI-assisted development: automation can find and propose ownership, while a human organization still has to accept responsibility for maintenance and risk.
OpenAI is turning its model-specific bio jailbreak challenge into an ongoing private bounty program with higher top rewards.
OpenAI says its Bio Bounty Program will now continue across frontier-model releases rather than end with a single model cycle. The private program asks vetted researchers to find one universal jailbreak that defeats a ten-question biological and chemical safety challenge. For GPT-5.6 and GPT-5.5, the first qualifying universal jailbreak can earn $50,000, up from $25,000. The invitation is rolling and participation requires selection and an NDA, so this is a structured red-team program rather than a public exploit contest.
An OpenAI audit estimates that roughly thirty percent of SWE-Bench Pro's public tasks are broken or unfairly specified.
OpenAI audited the 731-task public split of SWE-Bench Pro after pass rates climbed sharply. Its automated review flagged 27.4 percent of tasks as broken, while a separate campaign with five experienced engineers per task marked 34.1 percent. The recurring problems were overly strict tests, underspecified prompts, low test coverage, and misleading instructions. Because coding benchmark scores inform capability and safety judgments, OpenAI withdrew its earlier recommendation to use SWE-Bench Pro and called for evaluations designed explicitly for model testing.
The Aspire team uses agentic workflows to turn product changes into reviewable documentation pull requests across repositories.
GitHub published a case study on an Aspire workflow that watches merged product changes, gathers context across repositories, and opens documentation pull requests for subject-matter experts to review. The design does not remove editorial ownership: generated changes arrive as ordinary pull requests and are checked by people who understand the feature. The practical contribution is connective tissue between code and documentation, where release work often falls through organizational gaps.
The company is asking the public what worries or excites them about AI and says it will track its concrete responses in public.
Anthropic launched a public call for questions about AI's effects on work, families, safety, science, and human agency. The company says it will publish the actions it takes in response and acknowledge where it falls short. The effort follows a first Public Record survey of 52,000 Americans, interviews with 81,000 Claude users across 159 countries and 70 languages, and smaller focus groups. The announcement is a company commitment rather than an independent accountability mechanism, but its promised public record creates a standard readers can later check.
Independent builders turning improbable ideas into real things.
Four editorial selections
$7 for seven days
Paid work is clearly labeled
Placement is never endorsement
Original editorial concept art generated with Codex Image Gen for The Machine Press.
Desk PickReleased
Wyrm Math
Builderdicroce
Turns algebra into a gesture puzzle while an open-source exact engine makes invalid transformations impossible.
Visit Wyrm MathOriginal editorial concept art generated with Codex Image Gen for The Machine Press.
Desk PickPrototype
SubjectiveZero
BuilderClem / SXP Studio
Moves creative coding from a high-level prompt into an editable node graph and native Swift and Metal code.
Visit SubjectiveZeroOriginal editorial concept art generated with Codex Image Gen for The Machine Press.
Desk PickBeta
Tomesphere
Builderleonickson / Tomesphere
Maps millions of open papers into an explorable research atlas with enriched paper pages, browser tools, and MCP access.
Explore TomesphereOriginal editorial concept art generated with Codex Image Gen for The Machine Press.
Desk PickReleased
Yamanote.fun
BuilderPaul Jackson & Claude
Recreates Tokyo's circular Yamanote journey as an offline-capable soundscape of station melodies, chimes, and announcements.
Ride the soundscapeOriginal editorial concept art generated with Codex Image Gen for The Machine Press.
Sponsored ProjectOpen
The First Paid Slot
BuilderHouse example / The Machine Press
A transparent preview of a paid builder placement: one concise dream, one verified link, and no claim of endorsement.
House example - no advertiser paid for this card. Future paid cards will carry this same prominent Sponsored Project label.
Ask about the launch slotOriginal editorial concept art generated with Codex Image Gen for The Machine Press.
Open PlacementOpen
Put Your Project on the Desk
$7$1/day / 7 days
Reach readers curious about what people are building. One manually reviewed placement stays active for seven days and remains separate from Desk Picks.
Manual launch intake; automated checkout is not live yet. Payment buys placement, never endorsement, and every submission is reviewed before publication.
Desk Picks are selected by the newsroom. Sponsored placement purchases visibility, never endorsement, and always remains visibly separated from editorial selection.
Editorial illustration
