The newest HackerNews stories, desensationalized.

The post presents Bento as a single-file presentation format designed around a predefined slide schema, which trades some raw flexibility for simpler manual editing and shipping. The creator explains that each file contains JSON slide data plus a compact base64-encoded app payload that decompresses in-browser via DecompressionStream, minimizing runtime fetches and keeping the package small. It uses the File System Access API for in-place JSON writeback, with download fallback, and signs updates with ECDSA. The project began from reveal.js and GSAP/Flip, then replaced charting with a custom implementation to avoid size and reliability issues. Collaboration is handled through encrypted CRDT updates via a Cloudflare Durable Objects relay that only sees ciphertext, with invite-based sessions, generated keys, read-only roles, and revocable live access. The author frames the approach as a practical path for local-state software distribution, including education and game-like math/physics content, and suggests this model is becoming more compelling as an alternative to heavier hosted tools.

The page introduces a Hacker News legendary links browser that lists 1,169 Hall of Fame links and 1,994 candidates, with each entry showing historical metrics and ranking context. It features older high-scoring submissions, including a 2002 thread on email limits, Norvig’s 1998 programming guide, the W3C URI guidance post, and Paul Graham’s 'What You Can't Say.' The interface also includes a "Newest class: June 2026" view, where five links recently met Hall criteria, spanning JavaScript history, Rust UI development, rail safety visual design, interview practice, and language-choice tradeoffs. A separate section tracks recently resurfaced links across type buckets such as essays, software, tools, references, media, and organizations. The page also exposes explicit induction rules, requiring sustained multi-submitter activity, multi-year presence, and spaced breakout events. Overall, the update is positioned as an archival and discovery layer that keeps long-running HN material accessible while preserving ranking and qualification context.

Beej Hall argues that his satisfaction comes from directly making things and says AI-generated output often leaves him unable to claim authorship because he did not personally solve, build, and understand each step. He frames the issue as a widening AI dev schism where productivity gains are real but can hollow out the sense of craft, fun, and completion. He gives examples from his own practice, from sci-fi writing and art to woodworking and coding, and says even when AI created the visible result, only hand-led work feels truly his. He contrasts a handcrafted flash-card app with Claude-assisted work and says he can only be proud of the former. He acknowledges prompt crafting requires real human skill in vision, judgment, and communication, yet argues this is not the same as writing the software itself. He tests the boundary with compilers, asking whether translating C to assembly differs from writing that assembly by hand, and says the line remains blurry because machine translation is deterministic and understandable. The article ends with a personal test of agency and the emotional reality that he feels less like a maker when he mostly asks an agent to produce artifacts, even while admitting practical benefits. He presents no hard rule, only a call to live with the ambiguity between asking and making.

The guide presents a practical operations playbook for running Postgres in production at startup scale, arguing that schema design is the most important long-term investment and should be tuned from observed query patterns. It recommends btree indexes for predictable lookups, simple read/write patterns, short transactions, careful lock scope, and migration rules that avoid write-blocking operations unless using concurrent variants. It then frames the query planner as a constrained optimizer that depends on up-to-date statistics, with EXPLAIN and ANALYZE workflows for diagnosing wrong plans, and it highlights that some sequential scans are acceptable or required. For scale, it covers batching writes, connection pooling, and autovacuum as critical controls to prevent dead tuples, bloating, and transaction-id exhaustion. Advanced sections promote FOR UPDATE SKIP LOCKED, partitioning, and trigger-based backfill patterns for large migrations while warning against long transactions. The content emphasizes operational maturity, including avoiding unnecessary complexity in ORMs, but commenters note major gaps: explicit backup/restore planning, stronger monitoring and alerting around failure modes, and a broader set of real-world trade-offs before adopting the recommendations.

The post introduces CrucibleBench, a persistent text-world benchmark inspired by early MUDs, where language models act over 50-turn runs with hidden social objectives, trust-sensitive NPCs, and stateful consequences that persist across actions. The design is framed as a constrained alternative to multimodal simulations or browser tasks: only seven command types, 12 rooms, and 14 items make invalid actions and inefficient behavior easier to detect and quantify. Four social-state NPCs expose trust and suspicion signals, and scoring is split into dialogue-classifier dimensions and judge-minimized dimensions to test sensitivity. The authors report that this split matters because classifier agreement varies widely by model, from 21.7% to 84.8%, while aggregate reliability (κ = 0.04) can miss instability. Removing classifier-dependent dimensions changes six leaderboard positions, including large shifts for leading models, while confidence intervals overlap for much of the top tier. A full table of results from 50 runs per model shows small absolute score differences and varying success rates with notable cost differences, including an expensive frontier model underperforming a lower-cost one in one setting. Main failure modes are dialogue looping (14% to 66% of frontier runs), wrong-room interactions, and exploration paralysis, each captured automatically without human judges from state telemetry. The team emphasizes this is a proof of concept and artifact release, not proof of general social intelligence or real-world deployment readiness, and says LLM judges should be audited per-model with ablation checks. Comments around it note nostalgia for older text worlds and mixed feelings about the benchmark’s retro format, adding a cultural layer to the technical framing.

Airbus announced a new three-year Wing of Tomorrow flight-test campaign using full-scale wing extensions on an A321neo to measure how longer, lighter, slimmer wings affect handling and efficiency in real flight conditions. The extensions, developed and fitted with extensive sensors, are being built in the UK and tested in Toulouse, with digital modeling and wind-tunnel work used beforehand to reduce risk and refine several wing geometries. The company says the programme aims to improve single-aisle aircraft fuel burn and production economics through wing redesign and new manufacturing and assembly processes, building on prior 17-metre ground demonstrators and over 100 related technology trials. Airbus describes the initiative as one of its largest aerodynamics programs and links it to the parallel eXtra Performance Wing demonstrator, which explores shape-changing wings and is expected to fly as a remotely piloted vehicle by the end of 2026. The program has received £227 million in UK-backed funding since 2014 through ATI, Department for Business and Trade, and Innovate UK. In discussion, readers highlight that despite the “folding wing” framing in some outlets, the current campaign tests only the extended configuration rather than fold mechanisms, and they frame gate compatibility as a key unresolved reason to pursue retractable tips eventually.

The article introduces noise as a multi-dimensional function used in procedural generation and shows how 2D noise can be rendered into a texture by sampling a normalized (x, y) coordinate grid. It contrasts raw randomness, such as using a seeded hash-like rand function, with coherent noise that preserves local smoothness while still changing unpredictably at larger distances. The core Perlin method is then explained: choose the surrounding lattice corners, assign each integer lattice point a precomputed unit gradient, compute vectors from those corners to the query point, take four dot products, and blend them by interpolation. To make this practical, the implementation uses a 256-entry gradient table plus a permutation table, duplicated to 512 entries to avoid bounds issues and allow wraparound indexing, while randomizing the permutation via swaps for one-to-one table lookup behavior. The code then computes fractional offsets from the integer cell corner, fetches the four gradient indices, interpolates along X and Y, and maps results into [0, 1] for image output. A final note introduces the smoothing curve used in Ken Perlin’s original algorithm for smoother transitions, after which the article shows a visibly improved result. It closes by acknowledging that improved versions changed smoothing and gradient handling and that deeper updates are deferred. The text is practical but implementation-focused, and it links directly to prior discussion of deterministic noise seeding and an external implementation source.

The content introduces Unlayer as a set of seven reusable primitives for building in-app content surfaces, including email, landing page, popup, and document editors that can be integrated through React, Angular, and Vue components. The platform markets one shared JSON schema, reusable branded templates, image and file management, responsive email rendering across major clients, and export pipelines to HTML, PDF, image, plain text, or ZIP for server-side or client-side flows. It also highlights developer tooling such as an SDK, REST APIs, CLI, agent skills, events, and block-level AI assistance for generating and improving copy. A sample snippet shows basic editor initialization, design loading, and HTML export, signaling a plug-in style workflow. The messaging frames the product as a branded, native experience that keeps users inside the host application and reduces the need for separate third-party tools.

The article investigates the first line of The Wizard’s Castle source code, REM"_(C2SLFF4, from the 1980 Exidy Sorcerer program in Recreational Computing, and why it sits beside PRNG setup code. The author maps Sorcerer BASIC behavior: tokenized program lines are stored as memory nodes, USR() jumps through a trampoline at 260/261, and PEEK(-2049) reads the screen’s lower-right byte. Initial disassembly of the visible characters looked invalid and caused crashes, which initially suggested the line might be a printer or transcription artifact. A later magazine tape version revealed the line also encoded as hex bytes, including a RET, confirming actual machine code existed there. Disassembling those bytes showed the routine read the Z80 R register, rejected zero, stored a value at 0xF7FF, and returned so the subsequent PEEK fed a pseudo-random seed into RND(). The article notes this was meant to emulate RANDOMIZE-style seeding on systems without that command, and that this made map generation less predictable but capped at 127 possible states because R is effectively 7-bit. The investigation also validates that the displayed REM text was not plain ASCII, because high-bit bytes are printed via token rules and can appear misleading in listings. The final assessment is that the magazine listing probably required direct memory entry or special key input to reproduce correctly, and that copying visible text alone may have failed in real typing sessions.

Ghost Cut, introduced in Ishmael, reframes cut as a temporary, reversible move by ghosting selected text on Ctrl/Cmd+X, keeping it in place, restoring it with Escape, and only removing it when pasted. The design is meant to avoid losing clipboard state, reduce layout churn before paste, and make cut-and-move actions atomic with a single-step undo. The author argues that current workflows make accidental clipboard replacement and multi-step recovery common, and presents Ghost Cut as mostly drop-in because keyboard muscle memory stays the same. They also note this behavior resembles file-manager-like cut staging and mention copy-plus-delete as the fallback for users who need classic semantics. Commenters disagree that these are bugs, arguing that cut has long represented copy-plus-delete and that undo is expected to target document state rather than shared OS clipboard state. Several responses also say accidental clipboard replacement is tolerable or even useful, and suggest only OS-level, cross-application changes would make clipboard-restoration behavior coherent.

The article reports that Moonshot AI is alleged to have used a large-scale internal distillation platform to extract Anthropic’s Fable behavior for developing K3, including rotating access paths to evade detection, and that it has deployed GB300-class servers with access in Thailand for model training. It frames this as a U.S.-sanctioned concern: normal distillation for smaller and more efficient models can support innovation, but covert industrial-scale reconstruction of proprietary frontier systems is presented as harmful to U.S. competitiveness and research leadership. Commenters challenge that framing by arguing that distillation is not clearly illegal, is common in the industry, and is often enabled by existing output logs, public datasets, and model access APIs. They also highlight that Anthropic and other large labs have relied on broad web scraping and other ambiguous licensing practices, making moral asymmetry central to the dispute. Participants question whether K3’s timeline could plausibly stem from fresh distillation alone, suggesting Kimi’s core architecture and internal engineering matter at least as much as any copied traces. They further raise practical concerns about detection, legal ownership of outputs from paid interactions, and how terms of service can be enforced across borders. Overall, the discussion shifts the issue from a binary theft claim toward a broader governance problem: whether model capability growth should be limited by anti-distillation policy, competition strategy, or enforceable international norms.

Elena introduces a Lean 4 introduction to formal verification aimed at cryptographers, using the One-Time Pad from Boneh and Shoup as a concrete example. The tutorial defines bitstrings as vectors over ZMod 2, implements XOR component-wise, and proves commutativity, associativity, identity, and self-inverse properties through stepwise tactics. It then models Shannon ciphers in Lean with encryption, decryption, and correctness fields, and instantiates the construction for the one-time pad where encryption and decryption are both XOR. The tutorial emphasizes pedagogical clarity by expanding goals, introducing hypotheses, and using simp, intro, Vector.ext, and lemma composition rather than only terse automation. It also discusses practical naming, namespace hygiene, and how to avoid conflicts with existing Lean xor. The closing section connects the exercise to broader real-world formal verification efforts in blockchain, citing work on Ethereum, zk systems, and Zcash-style verification, while acknowledging modeling and compiler-trust limitations. Readers are encouraged to follow the series for a later proof of perfect secrecy, and comments reinforce that formal methods in Lean can reduce error in security proofs but require tooling maturity and onboarding effort.

The article argues that many once-exclusive human abilities—writing, coding, analysis, translation, design, and research—are becoming commoditized as AI becomes faster and cheaper, making human value increasingly dependent on choosing what to do rather than doing everything. It defines ethics as a practical skill involving moral recognition, conflict reasoning, accountability for downstream effects, and moral courage under pressure, rather than mere personality or rigid rules. AI, it says, amplifies capability but not intention, so harm scales with speed, reach, and distance: decisions made by models can be repeated at industrial scale, deployed faster than institutions can oversee, and diffused across teams that blur personal responsibility. The piece warns that adding guardrails is useful but insufficient because humans still choose objectives, acceptable failure rates, and tradeoffs. It extends this to politics, arguing incentives increasingly reward performance traits linked to deception and domination instead of conscientious restraint. Readers note disagreement that ethics and rules are being conflated and that some aspects of moral guidance could be codified, while others remain preference-laden. Commenters also argue that faster and cheaper does not equal better for meaningful human work, and that deciding what to build and whether to do it is partly a matter of taste, values, and context, not a purely technical issue. Others critique the piece’s tone and attribution, but accept that the core concern is real.

A short complaint claims passkeys were built without considering consumer behavior, replacing familiar password flows with confusing multi-factor workflows. Commenters are split, with one side praising passkeys for reducing phishing risk and simplifying logins through platform-level prompts, biometric unlock, and password managers. Another side says real-world portability is still weak, especially across many devices, browsers, and non-native ecosystems, and asks for clearer guarantees on recovery and cross-device use. Many noted inconsistent website support and fallback paths to passwords, which dilutes security and complicates migration. Shared-account households and family workflows are a recurring friction point because passkeys can block easy sharing. Security-focused critics also warn that practical implementations often reintroduce single points of failure through compromised vaults, master passwords, and proprietary cloud sync. Others argue the core protocol is sound but enterprise and product-level UX decisions are not, creating confusion despite the technical promise. Vendor-lock-in and policy controls around client support raise broader concerns about consumer choice and control. The discussion is less about cryptography and more about adoption quality, where the benefits are real but unevenly delivered depending on provider and app design.

The content presents DeepSQL as a self-hosted platform for Postgres and MySQL that ingests slow-query telemetry from read replicas, indexes schema metadata, and learns workload patterns to recommend optimization actions. It is positioned as a security-conscious deployment model: installations run on Linux hosts, EC2, or containers inside the user VPC with read-only credentials, role-based access, policy controls, and audit logs, while claiming data never leaves local infrastructure. The service adds a natural-language layer for business context such as MRR, active-hotel definitions, and currency rules, and is connected through browser, MCP, or CLI workflows compatible with tools like Claude, Codex, and Cursor. A performance benchmark snippet claims a 40% cost reduction and highlights an ingestion profile showing 17.35 million rows per day from five hospitality customers, comprising 62% of total volume, with daily growth in each of the listed customers. The positioning emphasizes practical operations use cases, including Aurora cost reductions and alternatives to BI tooling, rather than a broad replacement of all database administration responsibilities. The one-line install and quick startup narrative is repeatedly framed around speed, but the deeper value proposition in the copy is continuous query triage and observability-oriented guidance for database engineering teams.

Hologram’s maintainer says the project has moved from a sustainability gap to active development momentum after eight months of sponsorship, citing Curiosum and the Erlang Ecosystem Foundation milestone funding plus GitHub sponsors as the catalyst behind releases from v0.7 to v0.11. Core gains include near-complete client-side stdlib coverage with 150 Erlang functions ported by 49 contributors, JavaScript interop for npm, Web APIs, and Web Components, a larger realtime layer, expanded event and middleware systems, browser-side error handling parity, regex support, and tooling for agent-assisted coding such as llms.txt and agent usage rules. The post links these technical advances to adoption, noting production apps and feedback that describes the framework as smoother than expected and competitive with established ecosystems for full-stack development. It frames the next phase as a local-first roadmap: offline-first behavior, auto-sync, and one Elixir codebase that targets web, mobile, and desktop without duplicating backend and frontend stacks. The roadmap emphasizes developer-time reductions through single-toolchain development, less API contract drift, and AI-agent compatibility from reduced cross-stack complexity. The maintainer says current capabilities are already usable in production and now asks for more company sponsorship to fund auth, forms, testing, security hardening, and standalone mode work before a second EEF milestone round. The post also states that sponsor funds are intended to be directional, with public roadmap visibility so companies can influence priorities while keeping project governance open. The practical message is that the project is positioning itself as a full-stack alternative for teams valuing local-first reliability and smaller hand-rot

Intel said on July 15, 2026 that Panther Lake Core Ultra Series 3 on 18A is in high-volume production using High-NA EUV, meaning notebooks now carry layers printed on ASML’s $380 million tool. High-NA raises numerical aperture from 0.33 to 0.55, enabling finer single-exposure features and replacing some multi-patterning steps. Intel is dual-qualifying those layers on both Low-NA and High-NA on 18A, so shipments can proceed while it collects yield and reliability data before relying fully on the new optics for future nodes. The process still faces higher costs: field size drops to about 429 mm², stitching is required, and one High-NA exposure is modeled at about 2.5 times the cost of Low-NA. ASML’s production EXE:5200B variant is reported around 175 wafers per hour and tighter overlay performance, and the roadmap argument is that this should pay off as nodes demand more pitch reduction. Demand signals are mixed but real, with systems ordered by SK hynix, Samsung, and New York IBM-linked deployments, while TSMC remains cautious through 2nm/A16. The result is that High-NA has crossed into real shipping silicon, but long-term value still depends on whether reduced masks and defects outweigh higher capital, dose, and process complexity.

The accessible content is a Cloudflare access-block page rather than a normal article view. It explains that the site’s security layer blocked the request and advises that suspicious input patterns or malformed data may trigger the restriction, with contact details recommended for appeal. The page lists the Cloudflare Ray ID and IP-related metadata as context for remediation and notes that the protection is provided by Cloudflare. As a result, the source contributes only the block notice and no primary editorial or reporting content from the site.

SWING NYC presents a web-swinging controller layout for a voxel Manhattan environment with separate instruction sets for desktop and touch gameplay. The desktop scheme maps left mouse to aiming and releasing web-swing arcs, WASD steering, contextual space-bar jump or super jump behavior, single air dashes, shift-initiated dives, and rooftop respawn. The touch scheme mirrors the mechanics with a held web-swing input, right-side drag camera control, jump and super-jump behavior, and a hold-to-dive speed option. The listing also attributes the experience to the track “POWERUP!” by Jeremy Blake, suggesting a polished presentation around movement tuning. Commenters note a snapping mechanic that aims to keep traversal safe near nearby buildings, indicating some effort to reduce failure states, but also state the experience can still feel broken, while still describing the game as enjoyable in brief play tests.

A Redfin and Ipsos survey of 4,000 U.S. adults in May 2026 found 53% oppose and 34% support building an AI data center in their neighborhood, with opposition stronger than for other development types such as apartments (39%) or mixed-use projects (32%). The study also found age variation: roughly 65% of baby boomers and 60% of Gen Xers opposed nearby AI centers, while opposition was lower among Gen Z (42%) and millennials (43%). Respondents raised concerns about electricity and water demand, noise, and neighborhood disruption, and cited broader AI anxieties, including a belief among 58% that AI could worsen job and housing affordability. A New York statewide ban on large data centers was referenced as a real-world policy response. In contrast, Redfin’s county-level analysis in Virginia reported substantial tax-driven school gains in Loudoun and Prince William counties, where personal-property revenue linked to data-center equipment rose about 639% and 349% from 2010 to 2025. Education spending per resident rose 77% to about $2,955 in Loudoun and 82% to about $1,589 in Prince William, compared with lower gains in neighboring counties, and local officials said these revenues helped fund schools without increasing homeowner rates. Loudoun’s and Prince William’s residential tax rates also fell in that period, though Loudoun’s higher home values limited the levy per resident from declining. The report credits data-center growth for these fiscal effects but states they do not automatically produce equivalent spending outcomes, and it cites methodology from county financial reports and census estimates across 2010–2025.

Consumer NVIDIA cards no longer support NVLink officially, and bridges remain expensive, but the author benchmarks the practical impact on dual RTX 3090 systems across inference and training workloads. The setup uses an EPYC + Supermicro server platform with full PCIe x16 links and compares layer-split and tensor-parallel inference plus DDP and FSDP training workloads using Llama and Gemma models and multiple model sizes. Initial results suggested NVLink improved tensor-parallel prompt throughput by about 30 percent and nearly tripled FSDP training speed, while layer-split inference and DDP were unchanged. The author then found those non-NVLink numbers were skewed because NCCL was using shared host memory instead of PCIe P2P, since disabling NVLink at the driver level also disabled P2P on consumer GPUs. After disabling the bridge and applying the open-gpu-kernel-modules patch, they confirmed direct P2P and saw prompt-processing performance match NVLink nearly exactly, while FSDP reached just over 90 percent of NVLink performance. Additional PCIe lane throttling showed NVLink has stable performance across bandwidth levels, whereas P2P drops at x8 and x4, with SHM bottlenecking for reasons outside raw PCIe throughput. The findings imply NVLink is most useful on consumer systems where lane count forces GPUs below x16 bandwidth and where Linux patching is unavailable or undesirable. For users on workstation/server platforms with full lanes, upgrading to stronger platform hardware can be cheaper than buying a bridge. The post concludes the bridge is a practical workaround for restricted topologies, especially for FSDP-heavy workloads.

An archival repository republishes the Apollo 11 guidance computer code for Comanche055 in the Command Module and Luminary099 in the Lunar Module, transcribed from MIT Museum scans and presented as a baseline for comparison and correction by contributors. It identifies the CM program as Colossus 2A and the LM program as Luminary 1A, notes assembly with NASA’s yaYUL format, and preserves historical metadata including dates, approvals, and contractual context from March and July 1969 for mission software governance. The source is marked public domain, names digitization contributors, and points to Virtual AGC for compiling and emulation, signaling intent to make historical flight software runnable on modern systems. Commenters on the thread treat this as a rare chance to inspect constrained real-time systems design at scale, including navigation assumptions encoded as time-limited polynomial models and terse inline cautionary comments. They connect the release to prior Hacker News discussions and share supporting resources such as original scans, restoration videos, emulator links, and hardware context, reinforcing broad community engagement with mission-critical coding practices. Multiple voices also highlight operational details like the DSKY interface, contrasting it with modern GUIs and underscoring how software quality had to align with direct manual procedures. The overall framing is historical preservation combined with practical curiosity about running, understanding, and potentially extending the digitized program with contemporary tooling.

Streaming began as always-on, mostly ad-free access that reduced scheduling constraints and once felt cheaper and simpler than cable, but rights fragmentation has since rebuilt scarcity through studio-owned services and shifting exclusives. Disney+, HBO Max, Paramount+, and peers accelerated this model, while regional licensing differences mean the same title can appear on different platforms in different countries or disappear entirely. Subscription costs are rising as users add multiple plans, pay for extra device or high-quality tiers, and increasingly face ad-supported entry levels that make premium access more expensive. Discovery has become more difficult because every service uses separate search, recommendations, and watchlists, and quality settings differ by plan. Viewers report spending substantial time moving between apps, receiving repetitive recommendations, and sometimes switching platforms for sequels or continuity, producing fatigue and abandoned viewing. The essay argues this mirrors some older broadcast habits: fixed-event content, appointment-style sports and live programming, and ad dependency are reappearing. Commenters in the thread note that third-party indexes such as IMDb and JustWatch already improve availability visibility, and some also cite existing unified watchlist tools despite their own limitations. Others stress that expectations for specific must-watch titles and a thinner long-tail catalog reinforce the sense of missing choice even when some alternatives exist. Regulatory and structural critiques also appear, including calls to separate production, distribution, and exhibition roles. Overall, the consensus is that streaming's technical delivery is strong while the business model and interoperability model have eroded convenience and맞.

A developer explored whether Stoffel’s MPC platform can let genomic studies compute useful cohort-level results without ever centralizing raw DNA, and built a local proof of concept around a Monadic DNA-style flow where each client secret-splits six SNP values and sends shares directly to multiple MPC parties. He ran one hundred simulated users with independent identities, computed per-allele counts and a weighted score, and kept only aggregate outputs visible. The run returned expected results for all six target alleles and score, with ~28.26 seconds for infrastructure startup and ~6.17 seconds for the MPC workload on a local machine. The developer emphasized that this setup moved trust from a central database to coordinated infrastructure and showed the privacy boundary clearly through protocol structure. A notable finding was operational rather than cryptographic: simultaneous client starts caused TLS bursts and intermittent failures, while staggering starts by 50 milliseconds resolved the issue. He also found the platform comparatively approachable, noting concise, readable StoffelLang syntax, clear example flow, and smooth local coordinator/party setup as positive signs for version 0.1.0 usability. At the same time, he documented that all four parties ran on one host, inputs were synthetic and deterministic, and no mobile SDK exists, making real deployment concerns explicit. He added that MPC only protects computation, so cohort-size limits, repeated-query defenses, consent, auditing, and secure identity for result readers remain necessary. Remaining work includes cross-organization party operation, real-cluster testing, resilient mobile participation, and larger scale trials before production confidence is earned.

Vasily Grossman’s new NYRB Classics volume, From the Front Line: Stalingrad–Treblinka–Berlin, 1941–1945, is presented as the wartime reporting that fed directly into his major war novels. The review contrasts Stalingrad—a socialist-realist work aligned with wartime state narratives—and Life and Fate, which exposes moral compromise, repression, and state violence after Stalinist ideological collapse. Grossman’s career arc is framed by official favor and censorship: Stalingrad was praised, later denounced during the anti-Semitic turn, while Life and Fate triggered a KGB seizure and near loss of the manuscript before it was smuggled out by collaborators. The editors’ fresh translation work is described as unusually forensic, restoring and cross-checking draft and typeset versions to recover fuller texts. The collection’s chapters trace Grossman across Stalingrad, Ukraine, Poland, and Berlin, showing him as an embedded, physically vulnerable reporter writing under fire with minimal rest. The review emphasizes the Soviet press ecosystem—mass literacy gains, huge military circulation, and millions reading wall-posted or oralized newspapers—and Grossman’s ability to gather trust from soldiers and commanders, which gave his writing both documentary credibility and morale value. His prose is praised for sensory precision and avoidance of self-importance, including vivid industrial and battlefield imagery and intimate vignettes of ordinary troops. The Treblinka dispatch and Holocaust reporting are presented as central, linking his immediate witness work to The Black Book and later Soviet suppression because of its challenge to official narratives of collective suffering. The piece closes by connecting Grossman’s warnings about ideology and state power to contemporary善