Wolf Digest — Tuesday, May 19, 2026

NVIDIA's CEO used the Dell Technologies World keynote to describe enterprise AI demand as "going parabolic, utterly parabolic," framed around the Vera Rubin NVL72 delivering agentic-AI inference at one-tenth the cost per token versus prior generations. The slide deck emphasized Dell's role as the on-prem distribution partner for the new Vera CPU and the OpenAI Codex enterprise rollout, tying together a same-week trio of announcements (Vera CPU shipments, Dell-Codex on-prem partnership, and the Google-Blackstone TPU cloud) into a single "capacity is being committed years out" narrative.

Google and Blackstone are creating a joint cloud-computing company that will rent Google's tensor processing units to outside AI developers, according to The Information's source with direct knowledge. The structure positions Blackstone as the capital and real-estate balance sheet for new TPU-only data centers while Google supplies the silicon and the JAX/XLA toolchain, broadening TPU reach beyond GCP. The deal is read internally as Google's response to the perception that TPUs remain underutilized relative to their compute efficiency because the developer surface area outside Google's first-party tooling is thin — by carving out a co-branded provider, TPUs can be sold to AI shops that won't touch GCP's broader contract structure.

OpenAI announced a partnership with Dell to deliver Codex into hybrid and on-premises enterprise environments — a notable concession to enterprises whose source code, regulatory posture, or data-residency rules forbid sending repository contents to OpenAI's hosted endpoints. The deal pairs Codex's agentic coding harness with Dell's PowerEdge servers and the broader Dell AI Factory stack, with deployment topology that keeps model weights on-prem for sensitive customers and routes through Dell-managed inference for hybrid customers. The move closes a gap that Anthropic-via-AWS has been exploiting in regulated verticals and signals that the "cloud-only" era of agentic coding tools is functionally over.

Latent Space ran a two-hour special with guest host Noah Smith and Brandon Anderson interviewing Yaroslav Azhnyuk, the PetCube founder now running The Fourth Law — one of the most active AI-guided FPV-drone companies operating in Ukraine. Topics span the modern drone tech stack (vision-language target acquisition, sub-second guidance under jamming, autonomy under contested GPS), the empirical economics of drone-versus-armor exchanges, why Western procurement timelines are structurally incompatible with the Ukrainian iteration cadence, and what an AI-saturated battlefield implies for Western readiness. Azhnyuk argues the West remains anchored to expensive exquisite platforms while Ukraine has functionally built a software-defined air force at $500-per-airframe price points.

The authors show that post-trained MoE models can drop roughly half of their experts at inference time without meaningful quality loss, via a self-distillation procedure that compresses expert-routing diversity into a compact subset of always-on experts. The recipe runs a short distillation pass where the dense model teaches its pruned twin to match logits under reduced expert budgets, then re-tunes the routing head. Reported results across Mixtral- and Qwen-MoE-class models show <1% degradation on MMLU, GSM8K, and HumanEval at 0.5× expert count, with proportional FLOPs and KV-cache savings during decoding.

LongLive-2.0 is an NVFP4-native serving and training stack for long-video diffusion models, with pipeline and tensor parallelism redesigned around Blackwell's 4-bit floating-point microscaling. The authors report end-to-end throughput improvements on 30-second 720p video generation at fixed quality versus FP8/INT8 baselines, with attention-bias reweighting and a custom KV-cache layout to absorb the dynamic-range loss. The infrastructure piece — rather than the architectural innovation — is the contribution: it's a reference implementation showing the NVFP4 path is viable for diffusion workloads, not just decoder-only LLMs.

Ben Thompson argues that local opposition to AI data centers has legitimate grounding — noise, grid stress, groundwater draw, and the perception that benefits accrue to distant shareholders while costs land on host counties — and that the only durable solution is direct payments to affected residents. The post lays out the political economy of why permitting fights have intensified through 2025-2026, why hyperscalers have so far relied on tax abatements that flow to local governments rather than residents, and why a more honest "pay the impacted households" structure would shorten timelines for the next round of multi-gigawatt builds.

SandboxAQ is exposing its large quantitative models (LQMs) for drug discovery — including its binding-affinity prediction, ADMET, and molecular-dynamics surrogate models — through Anthropic's Claude as agent-callable tools. The pitch is that access, not raw capability, is the binding constraint in computational drug discovery, and the Claude integration lets non-specialist medicinal chemists run the workflows in conversation rather than via custom Python. SandboxAQ's models compete with Chai Discovery and Isomorphic Labs on the model side; the bet here is that distribution through a general-purpose agent surface beats a specialist UI.

MIT Tech Review's pre-I/O take argues Google enters this year's conference as a clear third in the foundation-model race after spending Gemini 2.5 Pro's high-water mark of mid-2025 watching Claude Opus 4.5, GPT-5.1 Codex Max, and Gemini 3 trade the crown. Expected announcements: a Gemini 3.5 update, deeper agent integration across Workspace, an expanded TPU roadmap addressing the Blackstone partnership, and a hardware demo for Project Astra successor. The piece sets up the framing for whatever Google ships tomorrow as a make-or-break moment for the "Google can't ship" narrative.

MIT Tech Review's deep dive on the Anduril-Meta defense headset partnership covers the Lattice-on-Aria-glasses integration, IVAS competition dynamics, and the Army's selection criteria heading into the next phase of soldier-borne AR. The piece quotes Anduril and Meta on how computer-vision pipelines, low-latency sensor fusion, and Lattice's threat-tracking layer are being adapted from the company's drone-targeting stack into wearable form factors with the thermal and power constraints that distinguish soldier kit from consumer Ray-Ban Stories.

Hugging Face and IBM Research launched the Open Agent Leaderboard, an evaluation harness that runs agent loops across τ²-Bench Telecom, GAIA, SWE-Bench-Verified, and a new IBM-contributed AppWorld task set, with a reproducible Docker harness so open-weights models can be benchmarked under the same conditions as closed providers. Day-zero leaderboard shows Claude Sonnet 4.6 and GPT-5.4 leading the closed track; Qwen3.6, DeepSeek V4 Pro, and GLM-5.1 dominating the open track. The harness's contribution is methodological: standard tool definitions and execution environments so that "ran the agent in our own way" numbers cannot be cherry-picked.

NVIDIA's Cosmos team and Hugging Face published a recipe for LoRA and DoRA fine-tuning of Cosmos Predict 2.5 on robot-trajectory video data, enabling teams to specialize the world-model on a target robot embodiment with relatively small per-task adapter footprints rather than full re-pretraining. The recipe covers VAE conditioning, action-token encoding, and the new DoRA decomposition variant that NVIDIA recommends for cross-embodiment transfer where rank-only LoRA underperforms. Reference notebooks include data preparation for the Galaxea G1 and Unitree H1.

PaddleOCR 3.5 has been ported to run natively on the Transformers backend rather than only PaddlePaddle's runtime, broadening the model's deployment surface to anywhere transformers / accelerate / vLLM are supported. The blog covers the document-parsing pipeline (layout analysis → OCR → table-structure recognition → key-value extraction) and benchmarks against MinerU, dots.ocr, and Qwen-VL on the OmniDocBench suite — PaddleOCR 3.5 leads on Chinese tables and complex layouts while losing on math-heavy academic papers to Qwen-VL.

U.S. Special Operations Command leadership warned at the SOFIC conference that the pace of multi-agent autonomy development — drone swarms, autonomous boats, autonomous resupply — remains too slow for the operational gap SOCOM sees opening with peer adversaries. The piece quotes SOCOM's acquisition executive arguing the bottleneck is not algorithms but the data, simulation environments, and live-flight test capacity for AI-mediated coordination. SOCOM is asking industry for collaborative autonomy at edge networks under contested communications, not just single-platform autonomy.

DefenseScoop reports the War Department's reorganization of the cyber acquisition portfolio — announced earlier in May — is already running into integration friction between the new Cyber Acquisition Executive and the legacy service cyber commands. Programs cited as at-risk include the joint software factory consolidation and the Cyber Operational Test and Evaluation framework that was supposed to standardize AI-enabled defensive cyber tool acquisition across services.

A bipartisan House bill would require federal civilian agencies to use NIST's AI Risk Management Framework as the default reference for procurement and operational governance of AI systems. Supporters frame the bill as codifying what's already de-facto practice across CISA, OMB-led pilots, and the DoD's responsible-AI rollout, while opponents argue NIST's framework is too high-level to be operationally binding without significant additional guidance. The bill cleared subcommittee with bipartisan votes and is expected to head to full committee within weeks.

The piece argues Army Aviation spent the 2015–2025 decade pursuing the Future Vertical Lift programs while letting cheap autonomous rotorcraft slide to industry competitors. The author draws explicit lessons for the current drone integration push: don't repeat the "single-platform-of-record" mistake; institutionalize software-defined modularity so that the AI/autonomy stack can be replaced independently of the airframe; and accept that the right unit cost is closer to the FPV economy than the FVL economy.

Analog Devices is in advanced talks to acquire Empower Semiconductor for approximately $1.5 billion, per The Information's sources. Empower's vertical-power-delivery chips are increasingly used alongside high-power AI accelerators to manage current at the package level — a design constraint that's become acute on Hopper, Blackwell, and especially Vera Rubin where transient draw demands are no longer comfortably handled by traditional VRMs. The deal would expand Analog Devices' AI-adjacent silicon portfolio at a moment when the power-delivery layer is being recognized as a distinct strategic bottleneck.

The Information reports internal Microsoft tension over GitHub Copilot's eroding share against Cursor, Claude Code, and Anthropic's growing developer footprint. Executives quoted in the piece point to the time-to-feature gap on agentic coding (Cursor and Claude Code ship rapidly; Copilot's agent surface lags), the IDE-versus-CLI split, and the awkward position of competing with the OpenAI Codex Microsoft is also selling. Internal memos cited in the article call for a faster Copilot iteration cadence and a reorganization that moves agentic-coding leadership closer to the GitHub product team.

Meta is reorganizing several thousand engineers into newly-spun-up AI product groups while signaling layoffs are coming for the unmoved population. The reorg follows the Muse Spark launch in April and is aimed at consolidating Meta's frontier-model effort under a tighter command structure — partly in response to the Llama-4 instability and the year-of-departures Meta absorbed in 2025. The signal value to the industry is that Meta is still betting big on superintelligence-scale models even as it trims teams that don't sit inside the new structure.

CiteVQA introduces a Doc-VQA evaluation that scores not just the final answer but the element-level bounding-box citation that supports it, exposing a failure mode in current MLLMs: arriving at the correct answer while grounding it in the wrong passage — a critical problem in legal, financial, and medical document workflows. The benchmark asks models to return both answer and bbox citations, and reports a substantial drop in joint accuracy versus answer-only accuracy across Claude, GPT, and Qwen-VL document agents.

PhysBrain 1.0 proposes a data engine that converts large-scale human egocentric video into structured Q-A physical-commonsense supervision, then transfers those priors to VLA policies via a capability-preserving adaptation stage. The team reports gains on cross-embodiment manipulation benchmarks where robot-trajectory-only training plateaus, arguing that the human-video distribution is a complement to robot data rather than a substitute.

The paper argues that "tool use" framed as JSON-structured function calling is strictly weaker than "code as harness" — letting the agent emit Python that orchestrates tools, control flow, and intermediate state. They benchmark on scientific-discovery and software-engineering tasks where multi-step coordination dominates, and report agents written as code-emitters outperform JSON-tool-call agents by 7–14 points on success rate at matched inference cost. Aligns with the broader Claude Code / Aider / Devin trajectory.

A position-paper-plus-survey on auto-research agents — agents that propose hypotheses, design experiments, write code, run them, and iterate. Maps the current landscape (Sakana's AI Scientist, Prime Intellect's auto-research benchmark, Google's AI co-scientist, FunSearch-class systems) onto a six-stage maturity ladder. Argues current systems clear the rung-1 "execute a written plan" bar but stall at rung-3 "propose a non-obvious hypothesis worth running." Useful as a synthesis even if the framework is not load-bearing.

The paper builds linear probes targeted at extracting manifold structure (not just point features) from polysemantic neurons under superposition, extending the Toy Models of Superposition framework into the regime where features form continuous low-dimensional manifolds rather than discrete directions. Validates on Anthropic's Pythia-class checkpoints, showing geometric structure (loops, lines, simplex faces) recoverable from features that single-direction probes mis-identify. Relevant for SAE methodology — suggests that the standard L1-on-codes recipe systematically under-recovers manifold features.

Generalizes DPO/IPO to arbitrary preference families parameterized by a learned scoring function, rather than the Bradley-Terry log-likelihood baked into DPO. Reports gains on Reward-Bench, mixed-objective alignment, and pluralistic preference modeling tasks where the BT assumption is known to be poorly calibrated.

The authors argue that "skill packages" for visual agents need to encode multimodal procedural knowledge — recognizing state, interpreting visual progress, deciding next action — rather than text-only prompts or code. They formalize what a multimodal skill package is and benchmark visual-agent task success after equipping models with MMSkills versus textual-only skill libraries; substantial gains on the agent-arena visual subset.

EnvFactory automatically synthesizes executable tool-use environments from natural-language task specifications, then trains agents inside them with a robust-RL objective that adds adversarial environment perturbations during rollouts. The resulting policies show better out-of-distribution generalization than agents trained on fixed environments, with improvements concentrated on tasks involving multi-API state.

Qumus is a wet-lab autonomous experimentalist for quantum-material characterization (Raman, transport, magnetometry), integrating a planning agent, a vision-language perception stack, and a robotic manipulation layer. The paper reports an autonomous campaign that mapped a portion of the phase diagram of a 2D vdW material, with experiment cycles dropping from human-paced days to autonomous hours.

Amazon is rolling out an Alexa+ feature that generates conversational podcast episodes from arbitrary topic prompts, leaning on the same TTS+LLM stack as NotebookLM's Audio Overviews but with on-device streaming and Echo-show integration. The feature is positioned as a daily-briefing competitor to NotebookLM and to ElevenLabs's Discover Daily and is the first generative-audio user feature shipped under the Alexa+ banner since the Anthropic-powered Alexa+ relaunch.