Wolf Digest — Saturday, June 6, 2026

Blackstone-backed operator AirTrunk said Friday it will invest thirty billion dollars in India by 2030 to develop five gigawatts of new data center capacity, one of the largest single commitments to the country's digital infrastructure. India's capacity is projected to climb from about one and a half gigawatts today to as much as eight gigawatts by 2030, per Bernstein, helped by New Delhi's offer of tax exemptions through 2047 on overseas services run from Indian data centers. AirTrunk, which entered India via its Lumina CloudInfra acquisition, is planning a three-gigawatt campus at the Raigad Pen Growth Center worth roughly twenty-one billion dollars, on top of a six-hundred-megawatt pipeline across Mumbai, Chennai, and Hyderabad. The announcement followed a meeting between chief executive Robin Khuda and Prime Minister Modi; power availability remains the cited bottleneck.

Emil Michael, the undersecretary of defense for research and engineering and the Pentagon's chief technology officer, told the Washington Post's Building America Summit that AI companies bear responsibility for the "weaponization potential" of their models, singling out Anthropic's Mythos. He tied the point to a White House executive order issued Tuesday that creates a voluntary "AI cybersecurity clearinghouse," letting firms give the Defense Department roughly thirty days to scan covered systems for software vulnerabilities, for instance in power grids or hospitals, before public release, and credited OpenAI, Anthropic, and Google for agreeing. Michael noted the unresolved friction with Anthropic, which has been excluded from some Pentagon deals and sued the administration over a supply-chain-risk label after restricting military use of Claude. He added that federal monthly AI users jumped from about eighty thousand to one and a half million in six months.

France will field its homegrown AI command system, Arcadia, at NATO's Coalition Warrior Interoperability Exercise in Poland from June 8 to 26, positioning it as a sovereignty-driven alternative to Palantir's Maven Smart System, which the alliance adopted in August 2025. Built with Mistral AI, Safran.AI, Thales, and Airbus and already tested in the Dacian Fall and Orion 26 exercises, Arcadia uses a highly decentralized mesh of field-deployed servers rather than a central cloud for battlefield resilience, and Gen. Patrick Justel pitched it as compliant with NATO's Federated Mission Networking standards, a claim Palantir disputed. France has also built a staff-officer large language model named Berthier that synthesizes information and drafts courses of action while leaving decisions to commanders. Justel said several European partners have expressed interest.

Google Research detailed a cross-corpus agentic RAG system now in public preview on the Gemini Enterprise Agent Platform. It decomposes multi-hop queries across a Root, Planner, Query Rewriter, RAG, and Synthesis agent, with the key differentiator being a Sufficient Context Agent that detects what information is still missing and keeps searching rather than guessing or returning "not found." On the FramesQA benchmark of eight hundred twenty-four queries over two thousand six hundred seventy-six PDFs, it answered ninety point one percent correctly even when the planner had to choose the right corpus from four options, with latency within three percent of the single-corpus version, and claims up to thirty-four percent higher factuality than standard RAG. The launch lines up with Google's stated surge in Gemini Enterprise demand that prompted its SpaceX compute deal.

The Information reports that SpaceX's AI lab, xAI, has had a chaotic year, churning through leaders and staff and bringing in outside help to close the coding gap with Anthropic, while playing what the piece calls a long-running cat-and-mouse game with that same rival. The headline's twist, that xAI now ends up "powering" Anthropic, points to the compute arrangement in which Anthropic rents the SpaceX-owned Colossus cluster originally built by xAI. The body is paywalled, but the framing captures the strange inversion of a would-be model competitor becoming its rival's landlord.

Code2LoRA trains a hypernetwork to emit repository-specific LoRA adapters, injecting project context (imports, APIs, conventions) with zero inference-time token overhead, unlike long-context RAG or per-repo fine-tuning. A static variant converts a single snapshot into an adapter for stable codebases, while an evolving variant maintains an adapter as the repository changes, targeting the brittleness of existing methods under software evolution. The framing is a direct answer to the cost of repository-scale context for code language models.

VideoKR is positioned as the first large-scale training corpus aimed specifically at knowledge- and reasoning-heavy video understanding, with three hundred fifteen thousand reasoning examples over one hundred forty-five thousand newly collected, CC-licensed expert-domain videos. A human-in-the-loop, skill-oriented pipeline generates examples of progressively deeper difficulty while controlling diversity and the reliability of chain-of-thought rationales. The contribution is the data and curation methodology rather than a new architecture.

JPMorgan, Bank of America, and Citi are among banks planning to join a new settlement network, slated to launch next year, that would let them move money between one another on a blockchain, according to The Information. The system, to be run by the bank-owned consortium The Clearing House, would handle clearing and settlement of tokenized deposits. Details beyond the lede are paywalled, but the move signals incumbent banks building shared on-chain rails rather than ceding the function to crypto-native players.

A ten-month study by a CSIS and Foundation for Defense of Democracies commission lays out a blueprint for a standalone US Cyber Force, and its most provocative recommendation is a force with no enlisted personnel, roughly thirty thousand uniformed and civilian members including twenty thousand active-duty officers and warrant officers. The rationale is bluntly economic: the enlisted pay scale, the study argues, cannot compensate elite cyber operators competing with industry salaries. It estimates upward of eleven billion dollars in initial cost from reallocated funds, about two point seven billion for personnel, an initial operating capability in eighteen months, and lateral entry on a managerial-and-technical two-track model. Several former cyber officials pushed back, warning that discarding the enlisted cadre would sacrifice a hands-on culture and talent pipeline instead of fixing enlisted pay.

Simon Willison released an alpha PyPI package, micropython-wasm, that runs Python safely inside a WebAssembly sandbox by compiling MicroPython to WASI and executing it through wasmtime, chosen over Pyodide because Pyodide only runs in a browser or Node. It meets his sandbox criteria: clean install, wasmtime-enforced memory limits, CPU limits via wasmtime's "fuel" concept (a roughly twenty-million default he is still tuning), controlled file and network access, and host functions implemented in seventy-eight lines of C compiled into a three-hundred-sixty-two-kilobyte WASM blob. A blocking host function keeps interpreter state resident across calls so variables persist between runs, and it powers a new datasette-agent-micropython plugin. Largely built with GPT-5.5 Pro and Codex, it is explicitly an alpha he would not yet trust for high-stakes use, though he has so far failed to make the model break out of it.

On-policy distillation supervises only output probabilities, which suffers from Monte Carlo KL sampling variance over large vocabularies and discards the teacher's intermediate states. OPRD lifts distillation into hidden-state space by aligning the student's representations with the teacher's, rather than treating the teacher as a black box after the language-model head. The method targets the two structural limits of output-only on-policy distillation directly, with the aim of more stable and information-rich transfer.

Prior approaches to translating low-resource languages, continued training or stuffing a grammar book into context, tend to overfit specific languages with weak zero-shot transfer. This work argues models must instead acquire the meta-skill of using in-context linguistic knowledge, and uses reinforcement learning to elicit it so the model generalizes to languages unseen at training time. The goal is scalable translation of extremely low-resource languages without per-language memorization.

"Benchmark Everything Everywhere All at Once" introduces Benchmark Agent, a fully autonomous agentic system for building benchmarks, attacking the labor cost of construction and the rapid saturation that erodes a benchmark's ability to discriminate between frontier models after release. The pitch is sustainable, reusable evaluation generation rather than another fixed static set, addressing the treadmill of benchmark obsolescence.

Unified video generation-and-editing models typically rely on 13B-plus parameters and condition on source video by concatenating tokens, which doubles sequence length and quadruples self-attention cost. LoomVideo targets that overhead while still interpreting interleaved multimodal inputs, aiming for a unified framework that avoids the prohibitive compute of token concatenation for editing. The emphasis is efficiency in the unified generation-editing setting.

At SXSW London, UC Irvine psychologist Gloria Mark argued that we have largely lost control of our attention, citing her research showing average attention spans on a screen falling from about two and a half minutes in 2003 to seventy-five seconds in 2012 to roughly forty-seven seconds in work conducted from 2014 to 2020, with faster task-switching directly correlated to higher measured stress. She extends the concern to generative AI: offloading writing, summarizing, and evaluating to chatbots removes the depth of processing needed to learn and retain, risking atrophy of critical-thinking and, through sycophantic synthetic companions, emotional intelligence. Mark concedes the evidence on social media's effects on children remains formally inconclusive even amid major lawsuits, and prescribes deliberately reintroducing friction, reading the book, meeting in person, skipping the GPS.

A Build Small Hackathon field report runs a live five-agent economy on Qwen2.5-3B, each woodland creature trading five goods for "pebbles," served with vLLM on Modal and decided in a single batched GPU call per turn. The thesis is that small models make many-agent, many-step simulations feasible because frontier models are too slow and costly per tick, and that the real engineering is bridging a 3B model's perfect formatting against its weak reasoning: it emitted valid JSON on every call but showed poor economic judgment, fixed with sharper prompting and a tolerant parse-and-repair layer. A "Wood Legends" mechanic reskins Tulip Mania, the South Sea Bubble, and 1929 bank runs as live shocks; in one unscripted run a reskinned bank run drove an owl to dump honey and crash its price from ten to three. Full reasoning traces are released as an open dataset.

Real planning problems disclose world and user constraints gradually through interaction rather than up front, a regime existing benchmarks underexplore. AdaPlanBench is a dynamic interactive benchmark testing whether agents can adaptively plan and re-plan as dual constraints are progressively revealed, isolating adaptive planning from one-shot planning. It targets a concrete gap in how agentic planning is currently measured.

Household robots are usually scored on task completion, but everyday settings involve value conflicts where the right action prioritizes human autonomy, efficiency, or social appropriateness over finishing the task. RobotValues offers ten thousand value-conflict scenarios to probe a robot planner's value preferences, filling a gap where no benchmark evaluated these tradeoffs. It reframes embodied evaluation around normative judgment rather than pure success rate.

Role-playing agents should let a character's values and behavior evolve with the story rather than holding a fixed persona, yet existing benchmarks test factual recall at a chapter, not alignment with the character's psychological trajectory in unexplored scenarios. ArcANE (Arc-Aware Narrative Evaluation) is an automatically constructed benchmark spanning seventeen novels and eighty principal characters, segmenting each into a character arc to test temporally appropriate behavior.

Agents usually act only on explicit requests, surfacing just the problems a user already noticed while many others hide in the broader context. TIDE frames discovering multiple hidden problems as its own task, where coexisting issues, unknown in number, must be uncovered, grounded in supporting evidence, and paired with concrete fixes, using template-guided iteration. It pushes agents from reactive assistants toward proactive auditors of their context.

Vision-language-action models inherit broad knowledge from pretrained vision-language backbones, but the mismatch between semantic VLM spaces and embodied control policies hampers precise perception-to-action mapping. AffordanceVLA inserts structured affordance forecasting as a task-oriented intermediate representation to bridge that gap for instruction-following manipulation. The affordance layer is the proposed mechanism for tighter grounding between perception and control.

Diffusion editors follow natural-language instructions with strong visual fidelity but mostly operate at surface instruction-following, ignoring implicit contextual constraints and producing plausible yet logically inconsistent edits. RE-Edit evaluates image-editing systems across five complementary reasoning dimensions to expose this gap between visual quality and logical correctness. The benchmark isolates reasoning, not just rendering, in evaluating edits.

Video generators synthesize compelling footage but stay confined to the virtual domain; Dream.exe asks how well that footage reflects physical law by testing whether depicted motion translates into executable robot behavior. Robotic manipulation becomes a concrete, measurable probe of whether a model has internalized physics, turning visual realism into an actionable benchmark. It is an evaluation of physical grounding rather than a new generator.

Converting past-interaction experience into reusable parametric capability is a promising route to continual learning, but where prior work studied single-iteration transfer, this paper finds that under repeated multi-iteration experience learning existing methods suffer progressive capability collapse rather than compounding gains. The authors dissect the failure across vital dimensions of experience to rethink internalization for self-evolving agents. The negative result, not a fix, is the headline.

Frustrated that the Army never delivered promised cost tradeoffs and timelines for its year-old Transformation Initiative, the House Armed Services Committee used its NDAA markup to legally require annual reports by February 15 and briefings by March 15 detailing what the service is buying and cutting. A flashpoint is aviation: the Army requested just one Black Hawk and five Chinooks, which the House raised to seven and twelve over supply-chain concerns, even as the Army argues fewer legacy aircraft make sense ahead of the MV-75 Cheyenne II. The friction is compounded by Defense Secretary Hegseth ordering a fresh review of the initiative, apparently news to Army Secretary Driscoll.

KITScenes Multimodal addresses gaps in sensor fidelity, map completeness, and geographic diversity in existing driving datasets, offering a synchronized European suite of high-resolution global-shutter cameras, lidar beyond four hundred meters, 4D imaging radar, and redundant GNSS/INS localization. Its HD maps are claimed to be the most complete of any sensor dataset, validated through autonomous driving. The contribution is sensor and map quality plus regional coverage.

This work studies a conversational assistant that can access a user's personal camera roll and retrieve relevant photos to answer queries ranging from simple factual ones to open-ended recommendations across years of images. The challenge is retrieval and reasoning over thousands of personal photos, a setting distinct from standard VQA. It targets the practical personal-assistant case of querying one's own photo library.