Wednesday, 24 June 2026

Anthropic launches async AI team member in Slack; hiring algorithms create racial rejection cascades; post-quantum migration mandate hits federal systems

Today's Lead

Engineering

Anthropic

Introducing Claude Tag

Anthropic launched Claude Tag, which embeds Claude as a persistent, asynchronous team member inside Slack — a product architecture shift from request-response chat to ambient delegation. Administrators grant Claude access to selected channels, tools, data sources, and codebases; from there, Claude can be tagged into work threads, monitor channel activity proactively, watch for threshold conditions (an A/B test crossing significance, a breaking build), and follow up across channels without requiring a synchronous conversation. Anthropic reports the internal version already writes 65% of its product team's code and merges 65% of product PRs — a metric that, whatever the precise denominator, signals substantial internal reliance. The product runs on Opus 4.8 and is currently in beta for Enterprise and Team plans. The sociotechnical challenges are as interesting as the capabilities: a single Claude identity spanning many channels without consistent cross-channel memory creates organizational confusion about what the agent knows and where its decisions are accountable, a problem Anthropic product leads have acknowledged but not yet solved. The competitive framing matters — this is Anthropic's move to define the async multiplayer agent layer before others do, anchoring Claude not as a standalone tool but as a persistent presence in the organizational coordination substrate where work actually happens.

Read →

Engineering

GitHub (future-file-format/f3)

F3: A Future File Format

F3 is an open-source research prototype for a next-generation columnar data format intended to succeed Parquet and ORC. Its central design idea is self-describing files that embed both metadata and WebAssembly decoder binaries alongside the data — meaning any future reader can reconstruct the file without needing a pre-installed, version-matched native decoder. This directly addresses the long-term data durability problem: formats like Parquet require reader software and data to evolve in lock-step, creating fragility over multi-decade archival lifespans. The format supports user-defined encodings through WASM plugins, enabling custom compression or serialization schemes to be added without format versioning churn. Built in Rust, the project attracted 627 HN upvotes, suggesting broad recognition of the problem. The WASM-as-embedded-decoder pattern has wider applicability than columnar analytics: embedding execution context with data is an approach that could address format obsolescence in scientific datasets, regulatory compliance records, and any archive expected to outlive its original toolchain.

Read →

words.filippo.io

Vulnerability Reports Are Not Special Anymore

Filippo Valsorda argues that AI-powered vulnerability discovery has commoditized the finding phase of security research, undermining the scarcity premise that historically justified special handling of external reports. The traditional coordinated disclosure model assumed that an external researcher's report represented a rare, hard-won discovery that vendors needed to rush to patch before attackers independently found the same flaw. That assumption breaks down when LLMs can scan codebases and identify many of the same vulnerabilities that human researchers surface: the finding is no longer scarce, so the competitive window between disclosure and independent discovery has narrowed dramatically. The bottleneck has shifted from discovery to triage — determining which of many potential issues are genuinely exploitable is now the expensive step. The practical implication for maintainers: integrate LLM-based scanning into CI/CD rather than relying on external reporters, and deprioritize the attribution and confidentiality protocols that coordinated disclosure required. For security programs, this also means the incentive structure built around CVE credit and bug bounties needs rethinking as the preconditions that justified it erode.

Read →

Cloudflare Blog

The Post-Quantum EO Is an Important Milestone. Now It's Time to Get to Work

Executive Order 14409, signed June 22, sets December 31, 2030 for federal agencies to migrate High Value Assets to post-quantum encryption (ML-KEM) and December 31, 2031 for post-quantum authentication (ML-DSA, SLH-DSA); federal contractors face the same 2030 deadline. Two-thirds of browser traffic to Cloudflare is already encrypted with ML-KEM; the authentication migration is harder — post-quantum signatures are larger than classical ones, the certificate authority dependency chain requires coordinated upgrades across clients, servers, CAs, logs, and root stores, and ecosystem deployment is still in early stages. Cloudflare's analysis highlights a subtle security gap in migration language: systems that add post-quantum support while keeping classical fallback enabled remain vulnerable to downgrade attacks, so 'transitioned' must mean classical paths are disabled, not merely that PQC options are available — a distinction the EO leaves to OMB to define. The contractor supply-chain clause may be the most consequential provision: federal agencies can only migrate if their vendors ship PQC-capable products, creating procurement pressure that propagates through the commercial technology stack without requiring each vendor to be directly regulated. For organizations outside the federal procurement network, the practical recommendation is to start now: ML-KEM over TLS is already widely available, and the harvest-now-decrypt-later threat — adversaries collecting encrypted traffic today to decrypt when quantum computers mature — is active regardless of any mandate.

Read →

FUTO

FUTO Swipe: Open-Source On-Device Swipe Typing

FUTO released an open-source swipe typing system comprising a trained model, a one-million-swipe training dataset, and a C++ inference library that runs entirely on-device without cloud connectivity. The architecture uses three learned components — a universal encoder, a language-specific language model, and a layout-specific decoder — achieving roughly 4% top-4 fail rate with 2.5M total parameters (1.4M active), enabling millisecond inference on low-end mobile hardware. Dictionary-constrained beam search bounds the prediction space, making performance both accurate and computationally predictable. The privacy implication is direct: conventional swipe keyboards transmit gesture data to cloud servers for prediction; this implementation eliminates that data collection entirely. Released under MIT with attribution requirements, it provides a reference implementation that decouples keyboard functionality from vendor infrastructure across any language or layout. The parameter efficiency story — competitive swipe accuracy at 2.5M parameters — is worth noting for mobile ML engineers: it demonstrates what careful architecture and constraint-based search buys versus scaling toward larger models.

Read →

ACM FAccT 2026 / algorithmichiring.github.io

Algorithmic Monocultures in Hiring

A peer-reviewed study of 3.4 million job applications across 156 employers finds that over 90% of U.S. employers use hiring algorithms — and that vendor concentration creates correlated rejection cascades: applicants rejected by one employer's system face elevated rejection rates from other employers using the same vendor's algorithm at rates far exceeding statistical chance. Racial disparities are severe and systematically obscured by aggregate reporting: 25.87% adverse impact for Black applicants, 14.74% for Asian applicants, numbers that become visible only with position-level analysis that most audits do not perform. The researchers note they represent the first and only group to conduct independent empirical research on these systems at scale — a statement that is itself significant: vendor opacity has made scientific accountability practically impossible for a system processing tens of millions of hiring decisions annually. The concentration risk maps cleanly onto failure modes engineers recognize from other domains: when a single vendor's logic applies to the majority of a market, errors are correlated rather than independent, producing systematic exclusion rather than random noise. For organizations auditing their hiring infrastructure, the study's methodology — position-level adverse impact analysis rather than aggregate pass rates — is worth adopting directly, both for accuracy and because aggregate metrics are precisely the ones most likely to pass legal review while masking the underlying pattern.

Read →

Works in Progress

Why the West Stopped Making Land

Land reclamation — filling shallow coastal waters to create buildable land, a practice that created significant portions of Boston, New York, and San Francisco — effectively ceased across the Western world in the 1970s following NEPA, the Clean Water Act, and the Endangered Species Act, while China, Singapore, and Japan have continued and now account for 80% of global new land creation. The economic case for reclamation in high-cost cities is strong: San Francisco land averages $24M per acre against historical reclamation costs of roughly $330K per acre. What changed was regulatory process overhead: the Craney Island port expansion began a feasibility study in 1997, produced a 1,400-page environmental impact statement over nine years, and has a projected completion date of 2065. The article argues that the procedural requirements have become decoupled from their stated goals — a permitting process requiring 68 years cannot be primarily about environmental protection. The international comparison is the strongest part of the argument: other Western countries adopted similar legislation at the same time and show the same reclamation halt, while non-adopters continued normally, suggesting the mechanism is the regulatory architecture rather than any underlying ecological constraint. For engineers thinking about system design, the case illustrates how gatekeeping institutions can calcify into barriers when compliance costs compound over decades — the nominal function remains, but the actual function diverges from original intent as the system accumulates procedural mass.

Read →

Marginal Revolution

Two Roads to Fast Clinical Trials, and the US Takes Neither

Operation TrialBlazer, HHS's initiative to accelerate clinical trials, cites both China (which surpassed the U.S. in Phase 1 trial share in 2021 and total registered trials in 2024) and Australia (sub-70-day trial starts) as models — but implements neither approach. China's speed in cell therapy, gene therapy, and stem cell work comes from allowing investigator-initiated trials to proceed 'if a researcher has interest and funding,' bypassing front-end regulatory review entirely. Australia's Clinical Trial Notification scheme achieves its fast timelines not through faster FDA-style review but through no front-end government review at all: an ethics committee approves, the regulator is notified, and the trial begins — the TGA does not evaluate the protocol before the trial starts. Operation TrialBlazer instead focuses on making the FDA's gatekeeping process cheaper and faster while explicitly retaining 'full regulatory authority and decision-making.' The analysis identifies this as a structural choice: all proposed reforms are about making the gatekeeper more efficient, none are about distributing gatekeeping authority. For engineers familiar with system bottlenecks, the distinction matters — a faster single point of control is different from parallel or distributed control, and the competitive gap between the U.S. and faster-moving systems reflects the architecture, not the performance of individual components.

Read →

Humanities

JSTOR Daily

The Making of the Muscular Farmer

Historian J.L. Anderson documents how farm advertising imagery grew dramatically more muscular as mechanization reduced actual physical labor requirements — a counterintuitive inversion in which cultural representation moves away from material reality precisely as the underlying system transforms. Post-WWII ads initially reflected the shift: 'push-button operator' cartoons and the 'farmer in a business suit' acknowledged that farming had become information processing and financial management. The muscular imagery intensified through the 1960s and 1970s alongside Cold War masculinity anxieties, then surged again during the 1980s farm debt crisis — when farming as an economic institution was most financially fragile. Anderson's explanation: advertising compensates for technological and economic displacement by asserting idealized identity rather than acknowledging transformation, and the compensation is proportional to the threat. The pattern generalizes beyond agriculture. When a technology replaces a category of labor, cultural representation often inverts: the displaced capability gets aestheticized in direct proportion to how much it is no longer required. The 'craftsman programmer' aesthetic intensifying in the AI coding era, the nostalgia for 'real engineering' as code generation becomes routine — these follow the same dynamic Anderson traces in the farm press, and they carry the same analytical implication: the strength of the compensatory imagery is a signal, not a description.

Read →