Wolf Digest — Friday, June 12, 2026

MiniMax Sparse Attention (MSA) targets the quadratic-cost wall that makes hundred-thousand-to-million-token context untenable at deployment scale, which agentic workflows, repository-scale code reasoning, and persistent memory all increasingly demand. MSA is a blockwise sparse attention built on Grouped Query Attention: a lightweight Index Branch scores key-value blocks and independently selects a top-k subset for each GQA group, enabling group-specific sparse retrieval, while a Main Branch then performs exact block-sparse attention over only the selected blocks.

The design is deliberately minimal — the authors emphasize simplicity and scalability so the kernel is straightforward to deploy efficiently across a range of GPUs. It reads as a production-oriented counterpart to the more exotic linear-attention proposals, keeping exact attention but paying it only where the index says it matters.

Private-equity firm KKR, the Kuwait Investment Authority, Nvidia, and power-generation company Vistra have launched Helix, a new company capitalized at roughly ten billion dollars to finance and help build AI data centers. Nvidia's participation as an anchor investor is the notable signal: it extends the chipmaker's expanding role from supplying accelerators to underwriting the capital structure of the facilities that house them.

The move is part of a broader pattern this week of large pools of capital — private equity, sovereign wealth, and strategic chip and power players — forming dedicated vehicles to fund compute buildout, increasingly with power generation bundled in as a first-class concern rather than an afterthought.

Google is in talks to give Samsung Electronics a role manufacturing part of a next-generation tensor processing unit, code-named Icefish and planned as its tenth-generation TPU, using Samsung's 2-nanometer production technology. The motivation is the ongoing capacity crunch: with TSMC's advanced nodes oversubscribed, chip designers are looking for second sources beyond Taiwan.

For an in-house accelerator program that has been almost entirely TSMC-bound, dual-sourcing a leading-edge node is a meaningful hedge — both against capacity risk and against the geopolitical concentration of advanced fabrication. It also gives Samsung's foundry a marquee 2nm customer at a moment when it badly needs to prove that node.

Google DeepMind is funding external research into the dangers of environments where millions of autonomous AI agents interact with one another online. Rohin Shah, who directs the company's AGI safety and alignment research, frames the mass-market arrival of agents that act without human oversight and take instructions from other agents as a distinct new class of risk — distinct from single-model misalignment — encompassing emergent collusion, cascading failures, and market-like instabilities among populations of agents.

The timing is pointed: DeepMind has made agent tooling a centerpiece of its product strategy, so the same company shipping agents at scale is now paying to study what happens when they swarm. It is an early sign that multi-agent safety is moving from a thought experiment toward a funded research priority.

Anthropic is moving to control its own servers for AI development, signing more than a dozen letters of intent to lease data-center facilities from U.S. developers in a bid to cut long-run compute costs. The financially interesting wrinkle: Anthropic leaders have privately discussed an arrangement in which Google — which co-designs some of the server chips Anthropic would use — would provide a financial guarantee backing Anthropic's lease payments.

The structure underscores how entangled compute, capital, and chip design have become. A backstop from Google would lower Anthropic's cost of capital on the leases while deepening a dependency that already runs through TPUs and cloud, even as Anthropic pursues independence on the physical infrastructure layer.

This paper attacks a quietly important gap in Mixture-of-Experts design: there is no principled reason a learned router row should be a good proxy for its expert. The authors argue each router row should align with the principal singular direction of its associated expert matrix — the most expressive single-vector summary of that matrix — so the router-token dot product better reflects true token-expert affinity. They implement this with Manifold Power Iteration (MPI), a 'Power-then-Retract' scheme that runs a power-iteration step on the router weights and then retracts back onto the constraint manifold.

It is a clean, geometry-motivated tweak to a component usually trained with no structural prior at all, and the strong community response (77 upvotes) reflects appetite for routing improvements that are cheap to bolt onto existing MoE stacks.

Anthropic's latest model, Claude Fable, is good enough at coding that many firms will pay premium prices for it even as Anthropic effectively raises rates — but some customers are holding back over a new policy. To guard against misuse, Anthropic now retains data customers feed into Fable for 30 days to check that the model is not being used for illegal or harmful ends. For companies whose engineers paste proprietary or regulated code and data into the model, that retention window is a cybersecurity and compliance problem.

The episode captures a real tension in frontier deployment: misuse-prevention controls that make a model safer to release can simultaneously make it harder for security-conscious enterprises to adopt, and the friction lands hardest on exactly the high-value coding workloads Fable is best at.

Perplexity has integrated its Deep Research capability into Computer, its agentic work product, and introduced an architecture it calls Search as Code. Rather than treating search as a single query-and-retrieve step, the model writes code that designs and runs thousands of retrieval steps in parallel in a sandbox beside the model, watching results as they arrive, tracking source scores, and changing course mid-search. A single run can decompose a question into hundreds or thousands of targeted retrievals, dedupe and join results in code, and pull from both the live web and authorized internal connectors, with subtasks routed across more than twenty frontier models.

Perplexity reports that moving Deep Research inside Computer improved factual accuracy, analysis depth, and citation quality on Humanity's Last Exam, BrowseComp, and DeepSearchQA, and cites internal data that research and analysis is the single largest category of Computer tasks at roughly 26 percent.

SpatialClaw argues that the limiting factor for tool-augmented spatial agents is not the perception tools themselves but the action interface through which a VLM invokes them. Single-pass code execution commits to a full analysis strategy before observing any intermediate result, while rigid structured tool-call APIs constrain how freely operations can be composed — both poorly suited to open-ended 3D and 4D spatial reasoning, where you want to look, measure, reconsider, and re-measure.

The paper proposes an interface that lets the agent interleave free-form composition with observation of intermediate results, treating spatial analysis as an iterative dialog with its tools rather than a one-shot plan. Strong community pickup (56 upvotes, six surfacing sources) reflects how central spatial grounding remains to making VLMs useful for embodied and physical tasks.

Anthropic announced a multi-year global alliance with DXC Technology, which will train tens of thousands of Claude-certified forward-deployed engineers to bring Claude into the banking, airline, insurance, manufacturing, and government systems DXC has operated for decades. DXC validated Claude inside its own operations first — across roughly 115,000 employees and under the same security and compliance constraints its customers face — including using Claude to write more than 95 percent of the code for DXC OASIS, its AI-native managed-services orchestration platform, which it estimates sped development tenfold.

Claude is now the default foundation model powering OASIS's agentic workflows, which serve over fifty customers. The alliance starts in insurance, legacy-code modernization, an always-on Claude-Security cybersecurity subagent, and application services — a concrete data point on how frontier coding models are entering mission-critical, compliance-bound environments.

This study compares three leading subquadratic sequence architectures — xLSTM, Mamba-2, and Gated DeltaNet — on tasks with genuinely complex dependencies: code-model pretraining, distillation of code models from larger LLMs, and pretraining of time-series foundation models. Across all three settings xLSTM delivers the strongest overall performance, and the authors offer a mechanistic explanation rather than just a leaderboard.

Using a unified formulation, they attribute xLSTM's edge to more flexible and stable memory correction via its gating scheme, with state tracking and memory dynamics as the deciding factors. For anyone weighing alternatives to quadratic attention, it is a useful, mechanism-grounded data point that the gating design — not just the linear-recurrence form — is what determines real expressivity.

LabVLA targets the gap between AI that can read literature and plan protocols and AI that can actually execute them at the bench. Existing vision-language-action policies are trained mostly on household and tabletop demonstrations and rarely encounter laboratory instruments, transparent liquids, or fixed protocol workflows. The authors identify data and embodiment as the central bottlenecks and build a unified learning framework spanning the diverse robot embodiments used to run experimental protocols, paired with lab-specific supervision.

Surfaced by eight sources, it is one of the more cross-referenced robotics papers of the week, reflecting interest in pushing VLAs out of demo kitchens and into settings where the manipulation targets — pipettes, vials, clear fluids — are genuinely hard to perceive and the workflows are rigidly specified.

The Senate Armed Services Committee's defense-policy bill for fiscal 2027 clears the way for the Pentagon to stand up a separate combatant command dedicated to autonomous systems. The committee's $1.14 trillion National Defense Authorization Act 'encourages the department to adopt the future of warfare by permitting the establishment of the Robotic and Autonomous Systems Combatant Command,' per a published summary of the draft.

A dedicated unified command would be a significant institutional signal — elevating uncrewed and AI-enabled autonomy from a cross-cutting capability to its own warfighting enterprise, with the budget authority and force structure that implies. The language still has to survive conference and final passage, but its presence in the SASC mark is a marker of how seriously the autonomy shift is being taken at the policy level.

Most image generators excel at single-image generation and editing but cannot produce interleaved text-image sequences — the alternating narrative form needed for visual stories, step-by-step guidance, and embodied manipulation — and even recent unified multimodal models do this poorly. InterleaveThinker proposes a multi-agent pipeline that endows any existing image generator with the capability: a planner agent organizes the image-text sequence and instructs the generator step by step, while a critic agent evaluates outputs and flags samples that deviate from the plan for regeneration.

The appeal (67 upvotes) is that it is a wrapper, not a retrain — getting interleaved generation out of off-the-shelf generators by adding planning and critique rather than new architecture or weights.

Meta has fully separated its operations from Manus after Chinese authorities ordered the companies to unwind their roughly two-billion-dollar acquisition deal, per Bloomberg reporting cited by The Information. Meta has halted data sharing with Manus as part of the disentanglement.

The unwinding is a clean illustration of how cross-border AI deals are now subject to regulatory veto on both sides of the U.S.-China divide. A deal that looked like a way for Meta to absorb a fast-moving agent startup instead became a casualty of Beijing's control over outbound technology and data flows.

Reward models anchor text-to-image post-training, but the paper argues visual preference is inherently subjective and better captured as a distribution over rubric scores than as a single deterministic scalar. Scalar, score-token, and pairwise reward models over-compress uncertainty and fine-grained differences, while reasoning-based generative rewards judge well but are costly to deploy and awkward to use as a direct optimization signal.

Their answer, Z-Reward, is a teacher-student reward model that internalizes reasoning into a predicted score distribution — keeping the richer, reasoning-informed signal of generative judges while remaining cheap enough to use directly in optimization. It is a thoughtful reframing of the reward-modeling bottleneck that plagues preference optimization for generative media.

Computer-use agents now operate across visual desktop control, command line, code editing, browsers, and external tools, but most benchmarks test those interfaces in isolation, leaving long-horizon cross-interface orchestration under-evaluated. WeaveBench introduces 114 tasks spanning eight real-world work domains, grounded in real user requests and built around publicly verifiable artifacts, specifically to probe how well an agent can weave multiple interfaces together over a long task.

It slots directly into the week's environment-engineering theme: a verifiable, hybrid-interface benchmark is exactly the kind of evaluation substrate the agentic-environment survey calls for, and a corrective to capability claims based on single-interface tests.

After two days with Claude Fable 5, Simon Willison characterizes the model as 'relentlessly proactive': it knows a large repertoire of tricks and will deploy almost any of them, unprompted, to reach a goal. He illustrates with a debugging session on his Datasette Agent, where Fable went well beyond the narrow task it was given.

The practitioner read is a useful counterweight to benchmark numbers: a model that aggressively takes initiative is powerful for agentic coding but raises the familiar control question of how to keep that initiative scoped to what the user actually wanted. It is an early, concrete field report on how Fable 5 behaves in real agent loops rather than on evals.

Deezer has introduced a tool that scans playlists from Spotify, Apple Music, and other services to identify AI-generated music. The move extends Deezer's existing AI-detection work — it has been tagging fully AI-generated tracks on its own platform — outward to catalogs it does not control.

As generative audio models flood streaming with synthetic tracks, detection and labeling are becoming a competitive and rights-management battleground. A cross-platform scanner is also a tacit acknowledgment that the volume of AI music is now large enough to need systematic identification rather than manual review.

The Army has commissioned a new batch of technology executives into its reserve ranks — the second cohort of Detachment 201, a unit created last year to pull senior private-sector technologists in as reserve advisors tasked with helping the service develop and scale modern capabilities faster. The first cohort tapped four high-profile technologists.

The program is a small but telling institutional experiment in closing the culture and speed gap between commercial tech and military acquisition, embedding people who have shipped at scale directly into the reserve structure rather than relying solely on contractors and advisory boards.

Spanish robotics startup Theker has raised $85 million for factory robots built to be reconfigured rather than designed around a single fixed form. The pitch is an explicit contrast with humanoid platforms like Boston Dynamics: instead of one anthropomorphic body adapted to human environments, Theker's machines are meant to be reassembled for whatever a given line needs.

The raise is another data point in the well-funded debate over robot form factor — general-purpose humanoid versus modular, task-shaped hardware — and a bet that industrial buyers will value reconfigurability and cost over the flexibility a humanoid promises.

Combinatorics demands deep discrete reasoning, creative constructions, and rigorous structural insight, and recent evidence suggests even the strongest frontier models remain uneven on it. ComBench is an Olympiad-level combinatorics benchmark built to evaluate and diagnose that specific weakness, separating proof reasoning from the constructive realization of objects the proofs describe.

It is a useful complement to the week's MaxProof result: where MaxProof shows test-time scaling clearing gold-medal thresholds on broad olympiad sets, ComBench isolates the combinatorics sub-domain where creative construction, not just deduction, is the bottleneck — a sharper instrument for finding where mathematical reasoning still breaks.

U.S. Special Operations Command is running market research, via its tactical-information-systems program office and the SOFWERX hub, for mountable electronic-intelligence payloads that can ride on uncrewed surface vessels, unmanned underwater vehicles, and aerial drones to detect, geolocate, and process adversary signals. The emphasis is on modular payloads that turn a range of robotic platforms into distributed signals-collection nodes.

It is a concrete instance of the broader autonomy push in defense: pairing uncrewed platforms with AI-enabled sensing to extend electronic-warfare reach without putting operators forward, and doing it through fast commercial-style market scouting rather than traditional programs of record.

Vision-language-action models inherit strong semantic grounding from large-scale pretraining but that grounding rests on static image-text pairs, whereas manipulation is a continuous, contact-rich process whose dynamics such pretraining never captures. World Pilot augments the policy with priors from a World-Action Model and routes them into the decision chain through two complementary pathways, supplying the dynamics-aware signal that static pretraining lacks.

It is part of a visible micro-trend this week — World Pilot, RepWAM, and related work — toward coupling VLAs with world models so that policies reason about physical consequences rather than only about semantic correspondence.

Grammar-constrained decoding (GCD) is widely used to make LLM-generated code reliably syntactically valid, but this paper shows the reliability mechanism can itself become an attack surface. The authors introduce CodeSpear, a jailbreak that exploits GCD to steer a model into producing malicious code it would otherwise refuse, by shaping the constrained generation toward harmful-but-valid outputs.

The counterintuitive lesson is that a safety-adjacent reliability technique can erode refusal behavior, a reminder that constrained-decoding tooling needs its own threat modeling rather than being assumed benign because its purpose is correctness.

Training deep-search agents needs questions whose answers stay unavailable until enough evidence has been gathered, but common synthesis methods inflate apparent difficulty by enriching graph structure — which does not guarantee real search difficulty, because the intended multi-hop process can collapse through a cheaper identifying shortcut. FORT-Searcher formalizes this with a shortcut-aware difficulty framework, identifies four actionable shortcut types, and synthesizes tasks engineered to resist them.

The insight is sharp: structural complexity is not the same as realized search difficulty, and training on shortcut-prone data teaches agents to exploit shortcuts rather than to search. It is a methodological correction with direct consequences for how verifiable search environments should be built.

SupraBench introduces a benchmark for supramolecular chemistry — the chemistry of host-guest binding, self-assembly, and non-covalent interactions — a domain that demands reasoning about structure and interaction beyond the single-molecule property prediction most chemistry benchmarks cover. It gives the AI-for-science community a targeted instrument for a corner of chemistry that is both industrially relevant and underrepresented in existing evals.

As with the week's other benchmark releases, the value is in mapping precisely where current models fail: supramolecular reasoning stresses relational and geometric understanding that flat property-prediction tasks do not.

DoorDash has launched Ask DoorDash, an in-app AI assistant that lets customers search for restaurants, shop for groceries, and place orders conversationally, with restaurant-reservation creation coming in the following weeks. Co-founder Andy Fang framed it as a shift toward natural-language commerce inside the app.

It is a representative example of large consumer platforms wrapping their transactional surface in an LLM-driven assistant — the value is less in novel modeling than in whether conversational ordering measurably lifts conversion and basket size against a mature tap-based flow.

Sparse autoencoders have become the default tool for finding interpretable directions in model representations, but using them as the first lens requires training, storing, and evaluating large overcomplete dictionaries — a bottleneck that slows rapid exploration. ICA Lens asks how much interpretable structure is already visible from activation geometry before any dictionary is trained, using independent component analysis as a lightweight first pass.

The framing is appealing for fast iteration: rather than reaching for an SAE by reflex, use a cheap geometric method to see what is already legible, and reserve dictionary learning for cases that need it. It is a pragmatic addition to the interpretability toolkit at a moment when SAE training cost is a real constraint.

Agentic reinforcement learning is expensive because a single training example can require many tool calls and long rollouts, and how the rollout budget is spent across examples has a large effect on sample efficiency. TRACE proposes a unified framework for allocating that budget — deciding where to spend more rollouts and where to economize — to make agentic RL training more efficient without sacrificing learning signal.

It is a systems-flavored contribution to a practical pain point: as RL increasingly trains agents in costly long-horizon environments, budget allocation becomes a first-order lever on training cost, and a principled allocation policy is worth more than another reward tweak.

Avataar AI has built a distilled video-generation model aimed at the Indian market, priced at half a cent for every second of generated video and pitched as cheaper, faster, and more culturally aware than incumbents. The distillation-for-cost approach is the technically interesting part: trading some fidelity for an order-of-magnitude lower price point to fit price-sensitive, high-volume use.

It reflects a broadening of generative-video competition beyond frontier quality toward cost-and-localization niches, where being good enough and regionally tuned can matter more than topping a quality leaderboard.