Wolf Digest — Friday, May 22, 2026

The Information reports that Anthropic is in active discussions to rent servers powered by Microsoft's in-house AI chips, marking a notable diversification away from Anthropic's NVIDIA-dominated training and inference stack. The deal would be a meaningful win for Microsoft's chip program, which has lagged Google's TPU and Amazon's Trainium / Inferentia rollouts and has been delayed multiple times. For Anthropic, the appeal is twofold: incremental compute capacity to meet demand growth (the company has been compute-constrained against Claude usage spikes), and supply-chain hedging against the NVIDIA allocation queue. The piece does not specify which Microsoft chip generation is on the table, but the most likely candidate is the Maia 200 series given the production timeline.

The Information reports that Cursor (Anysphere) is working on software intended to replace some of GitHub's core functions — PR management, code review, repo browsing — as Microsoft's GitHub unit deals with repeated outages and internal warnings that Cursor and Anthropic's tooling could render the platform obsolete. Jay Parikh, who runs GitHub, reportedly told deputies the threat is structural rather than competitive at the feature level. The shift comes at the same time Cursor's Composer 2.5 took third on Artificial Analysis's Coding Agent Index at 10–60x lower cost than the leading entries, and Microsoft is trying to integrate GitHub more tightly with Visual Studio and Copilot rather than let Cursor's IDE become the default surface.

Workday shares climbed over 10% in after-hours trading after the HR-software company disclosed that the number of customers actively using its AI agents roughly doubled to more than 4,000 in the quarter ended April 30. Gerrit Kazmaier, Workday's president, framed the result as evidence that the SaaS-meets-agents adoption curve is steepening in the enterprise. The number is notable because Workday's customer base skews to large enterprises, so 4,000 active accounts represents a meaningful share of the addressable market — not a long-tail SMB ramp. This is also one of the cleanest public datapoints on whether enterprise AI agent adoption is real or a vendor-pitched mirage; the answer, for HR workflows specifically, is that it appears to be real.

Latent Space's daily AINews rounds up a striking week for AI infrastructure: Turbopuffer (vector / object store for AI workloads) crossed $100M ARR profitably, Exa closed a $250M Series C at a $2.2B valuation, and Modal closed a $355M Series C at $4.7B. All three are picks-and-shovels companies in the agentic stack — Modal for serverless GPU execution, Exa for purpose-built search, Turbopuffer for vector storage — and the timing of three large rounds in one week is a leading indicator that the LLM-OS layer is now the most heavily capitalized non-frontier-model category in the ecosystem. The aggregate ~$700M raised across these three rounds matches roughly a single frontier-lab raise.

Vega, from Microsoft Research, generates zero-knowledge proofs from government-issued credentials — age, personhood, professional status — in under 100ms on a commodity client device with no trusted setup. The credential never leaves the device. The system supports mobile driver's licenses and the EU Digital Identity Wallet format. The notable technical claim is "fold-and-reuse" proving: repeated presentations of the same proof to different services (or via AI agents) skip most of the expensive cryptographic work after the first proof, making private identity verification practical at agentic scale. The release matters most for the agentic-web debate: if AI agents are increasingly going to act on behalf of users across services, the bot-vs-human and proof-of-personhood layer needs to operate without re-exposing the underlying credential.

The Department of Transportation is openly debating whether human-in-the-loop requirements should remain the default across AI workflows, according to FedScoop's interview with the agency's CIO. The framing is significant: human-in-the-loop has been the de facto risk-management baseline across federal AI deployment since the 2024 OMB memo. DOT specifically is exploring scenarios in the Federal Motor Carrier Safety Administration and inspection workflows where the loop adds latency without commensurate safety value. The story is one of the first concrete public signals that federal agencies are starting to peel back the human-in-the-loop default as agentic systems mature — a trend Daniel will want to track against the broader regulatory backdrop including the delayed Trump executive order.

The Commerce Department signed letters of intent for roughly $2B in federal funds across nine quantum-computing companies under the CHIPS and Science Act, with the government taking a non-controlling equity stake in each. The equity provision is the structural novelty — prior CHIPS Act awards were grants without equity — and is framed as "enhancing the return for the U.S. taxpayer." The funding is targeted at manufacturing, research, and innovation in microelectronics relevant to quantum systems, not quantum-LLM hybrid work directly. Notable for the AI-adjacent stack because several of the funded companies overlap with the quantum-classical hybrid compute pipeline that frontier labs have begun to monitor for post-2030 algorithm work.

GSA announced a OneGov agreement with Snowflake that makes the company's AI and data-warehousing products available to all federal agencies under unified terms. OneGov contracts are the post-2024 GSA mechanism for consolidating agency-by-agency procurement of high-volume SaaS into single enterprise-wide vehicles — the goal being cost reduction and break-down of data silos across agencies. Snowflake joins prior OneGov agreements with Google Workspace, ServiceNow, and Adobe. The implications for federal AI adoption are mechanical: any agency now has zero-friction procurement access to Snowflake's data sharing, Cortex AI features, and the broader Snowpark agentic surface, which substantially shortens the path from data sitting in an agency's silo to being usable by an AI agent.

Hark, a stealth-mode AI company, raised a $700M Series A to fund what it describes as a 'universal' multimodal AI interface plus dedicated hardware. The company has not disclosed founders, model details, or product timeline beyond saying first multimodal models ship this summer with proprietary hardware devices to follow. The $700M Series A is unusually large for a company with no public product — it is in the same range as Humane's pre-launch raise — and reflects continued investor appetite for ambient / agentic AI form factors despite Humane's and Rabbit's commercial flameouts. Watch for whether the hardware ships before the underlying model demonstrates differentiated capability; that ordering has been the failure mode for prior swings at the same target.

Spotify rolled out an AI-narrated audiobook creation tool built on ElevenLabs voices, with the explicit policy choice that authors are not bound to Spotify-exclusive distribution — they can publish the generated audiobook anywhere. The tool sits in Spotify's creator dashboard and supports voice selection across ElevenLabs's TTS catalog. Strategically, this is Spotify continuing the ElevenLabs partnership it began with the Discover Daily podcast in early 2024, and operationally it follows Spotify's broader announcement today of AI-powered podcast Q&A, briefing generation, and a NotebookLM-style desktop app for personal podcast creation. Non-exclusivity is the policy choice worth noting — it stands in contrast to Audible's prior exclusivity-or-discount pricing on AI-narrated titles and may shape how independent authors choose distribution.

Spotify and Universal Music Group announced a partnership that lets Spotify Premium subscribers create AI-generated covers and remixes of UMG-controlled tracks, with participating artists receiving a share of revenue. The arrangement is the first major-label opt-in framework for fan-generated AI music at scale — Stability and Suno had been building toward similar deals from the AI side, but the Spotify-UMG announcement comes from the distribution side and is materially larger in reach. Worth watching whether opt-in extends across UMG's roster or remains artist-gated, and whether the revenue split disclosures arrive in subsequent reporting.

Spotify announced a new desktop app (research preview in 20+ markets) for creating personalized podcasts — the direct competitor to Google NotebookLM's audio overview feature — alongside AI Q&A on existing podcast episodes and user-prompted daily / weekly briefing generation. The personal-podcast surface uses Spotify's ElevenLabs-powered voice stack and is bundled with the briefing tools as a unified "AI listener" experience. The combination places Spotify in direct product overlap with NotebookLM, ChatGPT's audio mode, and (more loosely) Wolf Digest — a sign that the personalized-podcast format has moved from research toy to platform-scale product.

Google DeepMind announced an Asia-Pacific extension of its Accelerator program focused on environmental-risk AI applications — climate adaptation, disaster prediction, ecosystem monitoring. The program funds and mentors startups using DeepMind's research models (Gemini, the open Gemma family, and the AlphaFold-line scientific models). The Asia-Pacific framing is policy-strategic: it positions DeepMind as a counterweight to PRC-aligned AI-for-environment outreach in the region, and aligns with the broader Google AI for the Global South initiative that launched earlier this year.

AI2 highlighted PointCheck, an independent project that uses the Molmo VLM, MolmoWeb computer-use variant, and Olmo 3 language model to audit web accessibility by navigating real pages as a keyboard user would and inspecting what is actually rendered on screen. The tool is positioned as proof that fully open-weights / open-data models can carry an accessibility-tooling workload that has historically required proprietary VLMs. Released for Global Accessibility Awareness Day; the underlying point is the open-stack viability narrative more than the accessibility application itself.

OpenAI announced AdventHealth (a large U.S. hospital system) as a ChatGPT for Healthcare customer, with the deployment targeting administrative-workflow reduction — charting, prior-auth, scheduling — rather than clinical decision support. The OpenAI healthcare push complements Anthropic's PwC and KPMG enterprise wins and Google's parallel Workspace-for-Healthcare announcements; the differentiator OpenAI claims is HIPAA-conformant model deployment and integration with major EHR platforms. AdventHealth is a meaningful logo because of its size; the operational claim is hours-per-clinician-per-day reduction, which if it holds up under independent audit is the metric that determines enterprise rollout pace.

The Path, a new AI-therapy startup founded by Tony Robbins and Calm alumni, launched with the claim that its model scored 95 on Vera-MH, a recently introduced mental-health-AI safety benchmark, compared with a top score of 65 for consumer chatbots in the same evaluation. The 30-point gap is the marketing pitch; the methodological caveat is that Vera-MH is a relatively young benchmark and the consumer-chatbot baseline is unclear (which models, what system prompts, what conversation lengths). The launch lands in an active policy environment for AI-mental-health products, including FDA scrutiny and several state-level licensing bills.

Latent Space dropped a deep-dive interview with Ivan Burazin of Daytona, whose pitch — "the end of localhost" — has been the through-line for the agentic LLM-OS sandbox category. The conversation covers Daytona's evolution from developer-environment provisioning to AI-agent execution sandboxing, the explosion of demand from agentic eval harnesses (TerminalBench, GDPVal) that require disposable computers, and the consolidating LLM-OS stack that customers expect: Perplexity Computer, Manus, Cursor, Anthropic Computer Use, and OpenAI Codex all need the same primitive. Useful listen if tracking the agent-infrastructure layer.

Simon Willison announced Datasette Agent, the first release of an extensible AI assistant for Datasette built on his llm Python library. The agent provides a conversational interface for asking questions of Datasette-stored data, and with the datasette-agent-charts plugin it can generate charts. The release matters as a worked example of the local-data + local-LLM workflow Willison has been advocating, and as a reference design for how to wire a tool-using agent to a SQL data layer without leaking schema details into prompts. Notably the demo runs entirely against local models on his laptop in the video.

Narayanan and Kapoor respond to Derek Thompson's recent essay on the AI-as-normal-technology framing. Thompson agreed with AINT's labor-market thesis but argued that AI risks require extraordinary intervention because of the tail-risk profile. The Snake Oil reply distinguishes between the macroeconomic regime (which they argue is consistent with normal-technology adoption — stable GDP, sub-5% unemployment, jobs-as-expected) and the policy ask (which they argue does not need extraordinary frameworks because existing regulatory tools — product liability, sectoral regulation, antitrust — cover most actual harms). Worth reading alongside the Trump-EO-delay coverage as the academic case for the same outcome.

War on the Rocks tracks PRC diplomatic activity around AI governance at the UN and other multilateral fora over the past several weeks: China's vice minister of science and technology pushing for UN-led frameworks at a May 5 meeting, two senior Chinese AI experts appearing remotely on a Capitol Hill panel hosted by Senator Sanders, and a coordinated track-2 push to seed PRC-friendly definitions in international standards bodies. The thesis: the U.S. is ceding the rule-setting layer at exactly the moment frontier capabilities are crystallizing. Aligns with broader concerns Daniel has been tracking on multilateral AI governance.

Ben Thompson interviews Parallel founder Parag Agarwal (former Twitter CEO) on the business model question that has come to dominate the agentic-web debate: if AI agents fetch and re-present content without sending traffic to the source, how do publishers get paid? Parallel's pitch is a metering layer that prices content per-agent-read, with revenue-share back to the publisher. The interview is recommended primarily for the framing of the problem; the technical implementation is sparse on details. Adjacent to Spotify's UMG announcement today — the same question for music — and the broader publisher / AI-search detente.

MLST's interview with Michael I. Jordan reframes the AI-vs-statistics framing he has been pushing for the past decade. Jordan argues that intelligence is fundamentally a collective phenomenon — supply chains, commerce, markets, the joint statistical behavior of many agents — and that calling the field "AI" obscures the harder open problems in economic mechanism design and multi-agent equilibrium. The conversation revisits his ICML keynote material on incentive-aware ML and connects it to current LLM debates around eval-gaming and agent coordination. Long-form listening for the ML-theory inclined.

Jure Leskovec joins TWIML to discuss two threads: AI Virtual Cell (multiscale representations from single-cell RNA-seq through ESM protein language models through AlphaFold-style structure) and relational deep learning at Kumo — foundation models that treat relational tables as primary objects rather than reshaped tensors. The relational-FM pitch is interesting because it sits orthogonally to the LLM-over-CSV pattern; if the per-row reasoning happens in a model that natively understands joins and primary keys, the agent overhead vanishes. Worth listening for the AI Virtual Cell discussion alone.

Károly Zsolnai-Fehér covers a DeepSeek paper titled "Thinking with Visual Primitives" — a method for visual reasoning where the model composes primitive geometric shapes as an intermediate scratchpad before producing the final answer. The pitch is that constraining the visual chain-of-thought to a finite primitive vocabulary improves out-of-distribution generalization on geometric reasoning tasks. The original GitHub repo (ailuntx/Thinking-with-Visual-Primitives) and an associated dataset have been published; the Hugging Face dataset shows the deepseek-ai repo was deleted, which is a flag worth tracking.

Spreadsheet-RL trains LLM agents on a realistic spreadsheet-task benchmark with RL from execution rewards, targeting workflows that wrap Excel and Google Sheets. The contribution is the combination of a curated task suite that mirrors actual office spreadsheet patterns (multi-sheet lookups, formula chains, pivot generation) and an RL training loop that uses successful-completion signal as reward. Reportedly improves over RLHF-only and SFT baselines on the in-distribution suite; ablations show the cell-grounding action space is the key design choice. Adjacent to today's WorkstreamBench paper on the same domain.

WorldKV proposes a memory architecture for world models that combines retrieval-augmented context with learned compression of historical state. The compression module reduces the per-frame state to a fixed-size token sequence that integrates with standard transformer attention, and the retrieval module pulls back high-fidelity history when the compressed representation is insufficient. Targets video generation and long-horizon embodied prediction; reports favorable scaling vs. flat-attention baselines on long-horizon benchmarks.

Gated DeltaNet-2 extends the DeltaNet linear-attention line by decoupling the erase and write gates that update the recurrent state. Prior DeltaNet variants tied the two operations through a single delta rule; here each is parameterized separately, which the authors argue improves in-context recall and reduces the catastrophic-erase failure mode that has plagued linear-attention models on retrieval benchmarks. Reports gains on associative-recall and needle-in-a-haystack tasks at parameter parity with prior DeltaNet.

A unified analytical framing of supervised fine-tuning, RLHF, DPO, and on-policy RL variants in terms of the induced state distribution rather than the per-token gradient. The argument is that the methods differ most consequentially in which states the model spends gradient on during training; reframing the comparison this way clarifies when on-policy methods will outperform offline ones and predicts the failure modes of off-policy DPO at high coverage gaps. Companion experiments on standard RLHF benchmarks.

Toto 2.0 extends the Toto time-series foundation model line with scaling-law experiments mapping the parameter / data / loss relationship for forecasting. The headline is that time-series forecasting now exhibits clean LLM-style scaling curves once the training corpus reaches a critical diversity threshold, and that a single foundation model can replace per-domain trained forecasters on a wide suite of evaluation tasks. Architecture is a decoder-only transformer with continuous-value input embeddings and parameterized quantile heads for prediction.

Mega-ASR scales up real-world acoustic-environment simulation for ASR training. The contribution is a simulation pipeline that synthesizes reverberation, microphone-noise, codec, and overlapping-speaker conditions at a scale large enough to dominate training, plus a Whisper-class ASR architecture trained on the synthesized corpus. Reports WER improvements on far-field and multi-speaker eval sets vs. real-data-only baselines, suggesting the simulation distribution now covers enough of the in-the-wild manifold to substitute for hard-to-collect real recordings.

A distillation procedure that converts a full-attention pretrained transformer into a sparse-attention variant in hundreds of additional training steps while retaining most of the dense model's evaluation performance. The trick is a curriculum that gradually increases sparsity along the layer dimension while a teacher forcing signal maintains output alignment. Notable for the cost claim — retraining at this scale is cheap enough to motivate widespread sparse-attention conversion for inference-cost reduction.

OScaR proposes a KV-cache quantization scheme targeting extreme low-bit regimes (sub-2-bit) by exploiting per-head and per-position statistics to drop the precision budget where it has least effect. Reports near-lossless accuracy at average <2 bits per cached value on standard long-context benchmarks. Useful for serving cost; complements concurrent work on quantized attention from the SGLang/INT4-QAT pipeline.

An analysis claiming that the gains from RLVR (RL with verifiable rewards) on reasoning models concentrate in a rank-1 direction of the weight update, and that minimal training along this direction recovers most of the full RLVR benefit. If the claim holds, it has practical implications for low-compute post-training and theoretical implications for what RLVR is actually doing relative to SFT.

Boiling the Frog introduces a multi-turn agentic safety benchmark that targets gradual-escalation jailbreaks — the failure mode where each turn is individually innocuous but the cumulative trajectory leads the agent into harmful actions. Reports that current frontier models including the Claude Sonnet 4.5 line and GPT-5.5 family show meaningful escalation susceptibility relative to single-turn safety evals. Useful evaluation surface for the agent-safety literature, which has been heavily single-turn to date.

Live Music Diffusion Models target real-time interactive music generation, where the diffusion sampler must respond to user input mid-generation without disruptive transients. Contributions are a fine-tuning recipe that conditions on a continuous user-control signal and a post-training stabilization step that reduces artifact rate at low sampler step counts. Complements Stability AI's Stable Audio 3.0 release (published Wed) and Spotify's UMG/ElevenLabs announcements today — the same domain from the research side.