Wolf Digest — Thursday, May 7, 2026

The Design Conductor team published a follow-up to their December 2025 result — in which a multi-agent LLM harness built a five-stage Linux-capable RISC-V CPU in 12 hours — that scales the same approach by roughly two orders of magnitude. The new system, Conductor 2.0, runs on the frontier models released in April 2026 and produces fully autonomous designs that are 80x larger and demonstrably higher quality. The lead headline design is VerTQ, an LLM inference accelerator with hard-wired support for the TurboQuant quantization scheme, implemented as a 240-cycle pipeline from scratch in 80 hours of agent runtime. The paper presents four such designs and walks through the harness updates that made the scale jump possible: a more aggressive verification rung that catches synthesis-level regressions before the agent gets to commit, a structured handoff protocol between architect, RTL, and verification roles, and a long-context summarization layer that lets each subagent maintain a coherent view of the full design without re-reading every artifact each turn.

What makes Conductor 2.0 worth flagging beyond the obvious EE-twitter-friendly framing is the speed of the underlying capability curve. Twelve months ago an LLM agent could not produce a synthesizable RISC-V core; six months ago the team's own predecessor system needed twelve hours and produced something near the boundary of correctness; today the same harness, with frontier models and a tighter loop, produces an 80x-larger inference accelerator in fewer than four days of wall-clock time. The authors note the things the system still does poorly: it does not handle floating-point standards correctly without significant prompting, the designs are not power-optimized, and the verifier programs catch correctness issues but do not catch most performance regressions. Still, the result is an existence proof that the autonomous-hardware-design loop is a real research direction, not a curiosity. The implication for AI infrastructure is the obvious one: if frontier-model-driven design loops can stand up domain-specific accelerators in days rather than years, the cost structure for niche inference hardware shifts considerably.

Stream-T1 (86 HF Daily Papers upvotes) targets the same streaming-video regime as Stream-R1 but from the opposite side: instead of modifying training-time distillation, it adds test-time scaling. The argument is that current test-time video generation methods based on full-sequence diffusion have prohibitive candidate-exploration costs and lack temporal guidance — but streaming generation's chunk-level synthesis and few denoising steps are intrinsically suited for TTS, lowering computational overhead while enabling fine-grained temporal control. Three units: Stream-Scaled Noise Propagation refines the initial latent of the generating chunk using historically-proven high-quality previous-chunk noise; Stream-Scaled Reward Pruning evaluates generated candidates to balance local spatial aesthetics with temporal coherence; and a Stream-Scaled latent search to lift the candidate pool quality without exploding compute. Together with Stream-R1 (item #4), the two papers represent a coordinated push from the same research thread to make streaming video generation the dominant paradigm.

ARIS (Auto-Research-In-Sleep) is an open-source research harness that explicitly addresses what the authors call the central failure mode for long-horizon LLM research workflows: the plausible-but-unsupported success, where a long-running agent produces claims whose evidential support is incomplete, misreported, or silently inherited from the executor's own framing. The architectural answer is cross-model adversarial collaboration as the default: an executor model drives forward progress while a reviewer from a different model family critiques intermediate artifacts and demands revisions. ARIS has three layers: an execution layer with 65+ reusable Markdown-defined skills, MCP model integrations, and a persistent research wiki for iterative reuse of prior findings; an assurance layer for adversarial review; and a reporting layer with structured evidential trails. With 85 upvotes on HF Daily Papers it is the high-attention paper for the autonomous-ML-research-agent thread this week, and it pairs directly with #1 (Anthropic Dreaming) as a research-preview-tier instance of agent self-improvement.

SpaceX is considering an initial $55 billion investment, with a total project envelope of up to $119 billion, on a multi-phase semiconductor manufacturing complex in Grimes County, Texas, according to a proposal filed on the county website. The project — which Musk has previously referred to publicly as Terafab — is described in the filing as a vertically integrated, next-generation chip fabrication and advanced computing facility, with Tesla committing supporting resources and Intel pulled in as the manufacturing partner. The output target is chips for AI server racks, satellites, and Tesla's onboard compute stack. If the upper bound of the spend materializes, Terafab would sit alongside the largest semiconductor capex projects in the world: TSMC's three-fab Arizona buildout is currently around $65B and Intel's Ohio campus is in the same range. Three things make the announcement worth tracking even at this filing-level stage: vertical integration (Musk's framing has consistently been that owning the chip stack lets xAI, Tesla, and SpaceX coordinate hardware-software-compute roadmaps in a way none of them can while queuing for TSMC); the timing alongside the Anthropic-Colossus deal (Colossus 1 to Anthropic, Colossus 2 absorbing xAI training, implying compute supply is the binding constraint and Terafab is the longer-arc answer); and Intel as the manufacturing partner — Intel Foundry has been hunting for an anchor customer at this scale for two years, and a Musk-affiliated entity stepping in changes the foundry-business calculus visibly.

DeepSeek is in talks to raise its first ever venture round at a valuation that has reportedly more than doubled in weeks — from $20B to $45B — according to FT and Bloomberg reports relayed by TechCrunch. The Chinese lab, founded by hedge-fund billionaire Liang Wenfeng who controls roughly 90% of the company, has been the highest-output open-weights frontier lab over the past year, releasing successive iterations of its mixture-of-experts language models on Hugging Face under permissive licenses while reportedly training on a fraction of the compute and a fraction of the cost of the closed labs. The valuation jump in weeks is striking — it implies bidder competition rather than a structured priced round, and it puts DeepSeek on the same valuation tier as several US labs that have raised many rounds. Two Minute Papers' Wednesday video framed the same week's DeepSeek V4 release as 'beats billion-dollar systems for free,' which gives a flavor of the practitioner-level enthusiasm. Open question — whether $45B is consistent with DeepSeek's stated commitment to open weights through V5, or whether the structure of the round (preferred stock terms, board composition, geographic limitations) ends up softening that commitment.

The Army held an AI Table Top Exercise (AI TTX 2.0) Monday with executives from 14 tech firms and US Cyber Command, dropping participants into a 2027 Indo-Pacific scenario in which an adversary uses AI to launch continuously-adapting cyber attacks against US military networks faster than human defenders can patch. The principal cyber advisor to the Army Secretary, Brandon Pugh, said the exercise's central question was whether the degree of human involvement should vary by situation — peacetime versus active cyber conflict — and that the Army intends to develop a risk-continuum policy for letting agentic AI take autonomous action in cyber defense. Lt. Gen. Christopher Eubank, head of Army Cyber Command, framed the issue starkly: 'to tell somebody to patch faster is just unrealistic' against AI-driven attack at machine speed, and the question becomes where AI should have autonomy in the cyberspace defense environment. The Army plans to acquire some industry-suggested AI tools through rapid-procurement funds and field them to two cyber defense units for testing rather than developing to mil-spec from scratch. Eubank's takeaway: 'I wrote down 19 things and none of them are a product' — the harder issues are doctrine, organization, and risk acceptance for autonomous agents.

OpenSeeker-v2 (52 HF Daily Papers upvotes) makes a striking efficiency claim: a 30B-parameter ReAct-paradigm search agent trained on just 10.6k high-difficulty trajectory samples — using only SFT, no continual pre-training, no RL — reaches 46.0% on BrowseComp, 58.1% on BrowseComp-ZH, 34.6% on Humanity's Last Exam, and 78.0% on xbench, beating Tongyi DeepResearch (43.4 / 46.7 / 32.9 / 75.0) which uses the heavy industry recipe of pre-training, continual pre-training, SFT, and RL. The data-synthesis modifications carrying the result: scaling knowledge-graph size for richer exploration, expanding the tool set for broader functionality, and strict low-step trajectory filtering. The paper argues that informative-and-difficult trajectories at modest scale are sufficient to land at the search-agent frontier, and is one of the strongest data-quality-versus-compute-spend arguments to land this year. Pairs naturally with OpenSearch-VL (item #14, also from this week) — both are open recipes for frontier search agents that previously required closed-lab resources to reproduce.

HERMES++ (63 HF Daily Papers upvotes) takes on the structural gap in autonomous-driving world models: existing approaches predominantly focus on future scene generation while overlooking comprehensive 3D scene understanding, while LLMs reason well but cannot predict future geometric evolution. The paper unifies both within a single framework using a BEV (bird's-eye view) representation that consolidates multi-view spatial information into LLM-compatible structure, LLM-enhanced world queries that transfer knowledge from the understanding branch, a Current-to-Future Link conditioning geometric evolution on semantic context, and a Joint Geometric Optimization strategy enforcing structural integrity. Reported gains across 3D scene understanding and future generation benchmarks are large enough to mark this as a state-of-the-art driving world model release. Strategically interesting against the backdrop of Waymo's continued production deployment and the broader autonomous-vehicle thread in this digest.

PRISM (39 HF Daily Papers upvotes) targets the standard post-training recipe for large multimodal models — SFT on curated demonstrations followed by RLVR — and the distributional drift it induces. The paper's specific claim is that perception errors and reasoning failures follow distinct drift patterns that compound during subsequent RL, and that a black-box on-policy distillation stage between SFT and RLVR closes the gap. The alignment stage is a response-level adversarial game between the policy and a Mixture-of-Experts discriminator with dedicated perception and reasoning experts, providing disentangled corrective signals without requiring access to teacher logits. Reported gains over the standard SFT→RLVR baseline are large enough across multimodal reasoning benchmarks that PRISM is likely to displace the vanilla two-stage recipe in next-quarter post-training stacks for multimodal models.

NVIDIA, OpenAI, and Microsoft are publishing the Multipath Reliable Connection (MRC) transport protocol — an RDMA variant that distributes a single connection across multiple network paths for throughput and resilience — as an open industry standard, with Spectrum-X Ethernet as the reference fabric. OpenAI's Sachin Katti calls out MRC's role in keeping Blackwell-generation training runs at expected efficiency by routing around link-level slowdowns; Microsoft's coupled announcement extends the multi-year Spectrum-X deployment story. The framing is open standards rather than proprietary lock-in. Spectrum-X plus MRC is being positioned as the Ethernet alternative to InfiniBand for gigascale AI clusters.

Genesis AI, a Khosla-backed robotics foundation-model lab that raised a $105M seed round, released its first model GENE-26.5 alongside a demo video featuring custom robotic hands the company designed in-house. CEO Zhou Xian framed the hardware pivot as a recognition that model improvements past a certain point require co-designed sensors and actuators — the company's stated position is that a foundation model alone, without hardware control, cannot close the field-deployment loop. Genesis announcement lands in a fast-moving robotics-foundation-model field that includes Physical Intelligence, Skild AI, and Allen AI's MolmoAct line. Pair with the same week's RLDX-1 (item #7) — which beat Physical Intelligence's π_0.5 on dexterous tasks — for the full picture of how fast the dexterous-VLA tier is moving.

True Memory argues that the standard agent-memory primitive — extract structured facts at ingestion, store them in a vector or graph database, and retrieve at query time — discards information that becomes load-bearing only after the query is known. The paper proposes a six-layer architecture that preserves events verbatim and pushes the entire retrieval pipeline downstream of the question. On LoCoMo (1,540 questions across 10 multi-session conversations) the system reaches 93.0% accuracy on a 3-run mean, against 61.4% for Mem0, 65.4% for Supermemory, ~71% for Zep, and 94.5% for EverMemOS — competitive with the top closed system while running on a single SQLite file with no external vector index, no graph store, and no GPU.

OpenSearch-VL (16 HF Daily Papers upvotes) is a fully open release — training data, trajectory-synthesis pipeline, and recipe — for multimodal deep-search agents. The team curates training data via Wikipedia path sampling, fuzzy entity rewriting, and source-anchor visual grounding (jointly reducing shortcuts and one-step retrieval collapse), then runs agentic reinforcement learning over the synthesized trajectories. Released datasets: SearchVL-SFT-36k for SFT and SearchVL-RL-8k for RL. Tool environment unifies text search, image search, OCR, cropping, sharpening, super-resolution, and perspective correction. The headline claim is that the released checkpoints match closed top-tier multimodal search agents on benchmark deep-search tasks, while making every component reproducible. Together with OpenSeeker-v2 (item #11), the open-search-agent tier is now genuinely reproducible.

PhysForge (23 HF Daily Papers upvotes) tackles the bottleneck for interactive virtual worlds and embodied AI: 3D-asset generation methods that produce visually-fine but functionally-inert geometry. The two-stage framework first uses a VLM as a 'physical architect' to plan a Hierarchical Physical Blueprint defining material, functional, and kinematic constraints; then a physics-grounded diffusion model realizes the blueprint by synthesizing high-fidelity geometry alongside precise kinematic parameters via the new KineVoxel Injection mechanism. Backed by PhysDB, a large-scale dataset of 150,000 assets with four-tier physical annotations. Experiments demonstrate functionally plausible, simulation-ready assets, providing a robust data engine for interactive 3D content and embodied agents. Worth tracking for anyone building world models or embodied-AI training datasets.

RLVR work using Group Relative Policy Optimization has driven much of the recent reasoning-model progress, but the technique inherits credit-assignment failures: uniform token granularity ignores heterogeneous information value, uniform polarity penalizes correct steps in incorrect trajectories, and zero-variance collapse erases gradient signal whenever a group's outcomes agree. EP-GRPO instruments each of these failures, quantifies their training waste empirically, and proposes an entropy-progress aligned reweighting that assigns credit per-token by information value, flips polarity per-step using process-level signal, and re-injects gradient on zero-variance groups via implicit process guidance. Reported gains are large enough on math and code-reasoning benchmarks that the paper is likely to displace vanilla GRPO as the default in next-quarter post-training stacks.

Defense One reports that the same agentic-AI capabilities the Pentagon is racing to adopt — autonomous tool use, long-horizon planning, multi-agent orchestration — are being absorbed by ransomware and cybercrime operators at a roughly equal pace, with the result that mid-tier criminal groups now reproduce capabilities historically associated with nation-state APTs. The piece quotes US Cyber Command and CDAO sources framing the asymmetry: defenders cannot afford the same agent loops without policy approval, but attackers can iterate on red-team agents continuously. Useful counterweight to the procurement-side narrative that agentic AI cleanly favors the defender.

Samsung Electronics broke through $1T in market cap Wednesday after a more than 10% share-price surge, becoming the second Asian company past the trillion-dollar threshold (after TSMC). The driver is the same compute story everyone is discussing — AI factories consume HBM and DRAM at rates Samsung is one of three suppliers globally able to meet — but the same-day catalyst was reporting that Apple is in talks with both Samsung and SK Hynix for next-generation memory supply. Last quarterly earnings showed profits 8x higher year-on-year, almost entirely AI-memory-driven.

HeavySkill (14 HF Daily Papers upvotes) reframes the agentic-orchestration mechanism question: rather than treating heavy multi-agent thinking as an external scaffold, the paper argues it is best understood as an inner skill internalized within the model's parameters that drives the orchestrator. The decomposition is a two-stage pipeline — parallel reasoning then summarization — operable beneath any agentic harness. Empirically the inner skill consistently outperforms Best-of-N strategies on diverse domains; stronger models can approach Pass@N performance. The deeper claim is that the depth and width of heavy thinking, treated as a learnable skill, can be further scaled via reinforcement learning, giving a path toward self-evolving LLMs that internalize complex reasoning without depending on hand-built orchestration frameworks.

The day before Code w/ Claude (item #1), Anthropic launched Agents for Financial Services — a verticalized packaging of Claude with finance-specific tooling and integrations. The release positions Anthropic alongside the broader push from frontier labs into regulated-vertical agent products and pairs naturally with the Pentagon's AI procurement reshuffling: as the federal track is reorganizing without Anthropic, the lab is doubling down on regulated-private-sector verticals. Light on numbers from the public announcement; the strategic move is what's worth flagging.

Long-horizon search agents must manage a working context that grows roughly linearly with reasoning steps; naive accumulation hits cost and error ceilings quickly. Context-ReAct introduces five atomic context operations — Skip, Compress, Rollback, Anchor, and Branch — that the agent invokes adaptively at each turn to decide which prior steps stay verbatim, which get compressed, and which can be discarded. LongSeeker, the system built on top, demonstrates the paradigm at scale and reports both lower per-trajectory cost and higher final accuracy on long-horizon benchmarks compared to a flat-context ReAct baseline. The contribution is operational rather than theoretical — the value is in formalizing context-management as a first-class agent action.

A benchmark paper from pharmaceutical-AI startup Gosset compares its curated drug-asset annotation platform against four frontier LLMs (Claude Opus 4.7, GPT 5.5, Gemini 3.1 Pro, Perplexity sonar-pro) on ten niche oncology/immunology targets where the relevant pipeline lives in the long tail — preclinical assets, Asian-developed drugs, trade-press-only disclosures. Across ten targets, Gosset returns 3.2x more verified drugs per query than the best frontier system. The result is a pointed entry into the broader does-curated-data-still-matter argument: it suggests that on knowledge-graph-shaped niche tasks, structured curation continues to dominate, even with web search retrieval available to the frontier systems.

Apple agreed to pay $250M to settle a class action alleging that Apple Intelligence — and specifically the upgraded Siri — was marketed in iPhone 16 launch materials as substantially more capable than what shipped. The complaint claimed the marketing created an impression that the advanced AI features would be available to users sooner than they actually were. The settlement does not admit liability but ends a politically awkward year for Apple's AI rollout. Stratechery's earnings analysis the same week flagged the Mac as the hardware line still benefiting cleanly from AI demand even as Siri-tier features remain in catch-up mode.

Ben Thompson's Wednesday Update reads Microsoft's quarterly earnings as the cleanest validation yet of the agentic-business-model framing the firm has been running for a year — Copilot revenue is now a meaningful line item, agent platform revenue is starting to compound, and the Azure-OpenAI and Microsoft Copilot stack is increasingly billed as a unit. Parallel read on Apple: the Mac is the line item benefiting cleanly from the AI demand wave (memory and SSD pricing tailwinds aside), while the iPhone story is constrained by chip and memory shortages flowing from the same demand wave.

Philip Rohlfing argues in Lawfare that the recent amendments to Section 130i — the statutory authority for protecting US installations from drone threats — create a structural mismatch between what intelligence the Department needs to recognize evolving drone-attack patterns and what data it is allowed to retain. The amendments tighten data-deletion timelines in a way that, in Rohlfing's reading, prevents the kind of multi-incident pattern recognition that AI-augmented analysis depends on. The piece is written as a policy argument but is implicitly an AI-deployment-blocker brief: machine-learning approaches to drone-threat pattern recognition need longitudinal data the new rules force out the door.

SIOP addresses the standard problem with long-horizon agent training: rewards are observed at the trajectory end, but useful credit lives at intermediate information-gathering turns, and existing turn-level shaping methods either need answer supervision or task-specific verifiers. The paper treats semantic clusters of final answers as latent future-outcome states and uses the per-turn likelihood of reaching each cluster as a label-free credit signal. Turn-level rewards then come from a process-style decomposition that does not need any verifier. Reported gains are concentrated on tool-use benchmarks where outcome-only rewards leave much credit-assignment headroom on the table.

LoPT proposes that the standard end-to-end backward pass is overkill for post-training, where supervision signal is much narrower than pretraining. The recipe places a single gradient boundary at the transformer midpoint — the upper layers fine-tune normally, while the lower layers use a cheaper local learning rule. Activation memory roughly halves, wall-clock speedup is around 1.6x on standard SFT setups, and the paper claims minimal accuracy regressions on common downstream evals. The argument generalizes: gradient propagation depth should be a design choice tuned to the supervision narrowness, not always full.

CuBridge tackles the perennial trade-off between compiler flexibility and expert-kernel performance for attention variants. The framework lifts an expert-written CUDA attention kernel into a structured intermediate representation, applies LLM-driven transformations targeting the new variant, and lowers back to performant CUDA. Unlike prior LLM-kernel-generation work that synthesized from scratch with unstable correctness and large performance gaps, CuBridge inherits the correctness scaffolding of the reference kernel and uses the LLM only for the transfer step. Reported results are competitive-with-expert performance across a wider variant range than prior compiler-driven approaches.

Manifold Steering extends the activation-steering literature by fitting an explicit manifold to representations and a separate manifold to output distributions, then testing the link between them via interventions that respect the activation manifold's geometry. The headline finding is that interventions that follow geodesics on the activation manifold yield behavioral trajectories close to the model's natural behavior — closer than linear-direction steering and far closer than random-direction interventions. The result formalizes intuitions from probe-and-steer interpretability work that geometry matters, and gives a constructive recipe for less-disruptive behavioral interventions.

The paper extends reward-model evaluation beyond instruction-following benchmarks into four social-alignment domains: bias, safety, morality, ethical reasoning. The authors convert standard social-evaluation datasets into pairwise preference data using gold labels where available and directional bias indicators otherwise, then test commercial and open reward models. The finding is that reward models that benchmark high on instruction-following continue to encode socially-undesirable preferences in measurable, reproducible ways — and that this is true even when the underlying instruction-tuned LM does not exhibit the same patterns. The implication is that reward-model alignment is its own non-trivial problem distinct from base-model alignment.

An initial coding-agent system for ARC-AGI-3 in which the agent maintains an executable Python world model, refactors it toward simpler abstractions as a proxy for an MDL-style simplicity bias, verifies each refactor against prior observations, and then plans through the model before acting. The harness is intentionally direct (scripted controller, predefined interfaces, plan executor, no game-specific logic). Results on the 25 public ARC-AGI-3 games are reported with each playthrough using a fresh agent instance and no cross-game state. Interesting as an alternative to direct policy learning on ARC-AGI-3 — an explicit-model-based agent that lets the world model carry the structural prior.

The Pentagon Chief Digital and AI Office's new Wingman initiative provides a framework for DoD organizations to stand up agency-tailored AI digital assistants on top of CDAO-controlled data substrates including the Maven Smart System. The framing emphasizes that Wingman is enablement infrastructure, not a single chatbot — components, model access, eval harnesses, and security controls that any DoD organization can compose into its own assistant. CDAO is positioning Wingman as the answer to how the Department scales LLM-based assistance without each component lab building from scratch.

Snap disclosed in its Q1 earnings that the $400M partnership with Perplexity announced last November has been quietly unwound. The original deal had Perplexity paying Snap in cash and equity over one year for direct integration of Perplexity's AI search engine into Snapchat's chat interface. Snap's sales guidance now assumes no contribution from Perplexity. Neither company has explained the breakup, but the optics suggest either the integration economics did not pencil at usage rates or the product directions diverged faster than the contract anticipated. Marker of the consumer-AI-distribution land grab cooling somewhat after the 2025 frenzy.

Russell Brandom's TechCrunch piece reframes the Anthropic-Colossus deal (item #1) from xAI's side: the arrangement immediately monetizes xAI's compute investment, turning xAI from a consumer of compute to a provider — a 'neocloud' in the Brandom framing. xAI's existing products are mostly focused on Grok, which has not seen the Claude-tier API growth. Colossus 2 absorbing xAI training while Colossus 1 generates billions of dollars per year from Anthropic is a structural pivot in xAI's monetization story, even before factoring in the Terafab fab build (item #10).

Károly Zsolnai-Fehér's Two Minute Papers covered DeepSeek V4 with the channel's characteristic enthusiasm — open-weights checkpoint matches or exceeds closed-system frontier capability while remaining freely downloadable. Channel reach (multi-million YouTube views) makes it a useful indicator that DeepSeek V4 is hitting the broader practitioner audience the same week the lab is reportedly raising at a $45B valuation (item #17). Cross-source signal.

The persistent observation that simple linear models like DLinear are competitive with transformers on time-series forecasting has fueled a debate the field has not been able to settle. This paper applies sparse autoencoders — the same mechanistic-interpretability tool used on language models — to PatchTST internals. The finding: a single-layer narrow transformer matches deeper configurations on common benchmarks, and the SAE analysis shows the network does not develop the superposition phenomena that make NLP transformer interpretability hard. The implication is that time-series transformers may simply be over-parameterized for the task, and that the architectural complexity helping language models is not load-bearing here.

Strat-Reasoner attacks the multi-agent-game training problem where a single agent's reward depends on the joint policies of every other player and the environment is therefore non-stationary from the agent's perspective. The paper integrates other agents' inferred reasoning into the credit-assignment loop and reports improvements on standard multi-agent-game benchmarks where prior single-agent RL adaptations plateaued.

A systematization-of-knowledge paper on LLM jailbreak robustness that argues the field's current practice — measuring attack success rate as a single scalar — is inadequate to capture the multidimensional nature of LLM security. The authors propose Security Cube, a unified multi-axis evaluation harness covering attack effectiveness, transferability, robustness across defense layers, and operator-side cost. The taxonomy work is the more durable contribution; the evaluation harness is offered as a reference implementation for groups doing red-teaming.

Q2RL addresses the BC-then-online-RL problem where the offline-to-online transition causes the policy to overwrite previously learned good actions due to distribution mismatch. The recipe: extract a Q-function from the BC policy using a small number of online interaction steps, then use the Q-function as both a value baseline and a gating filter that prevents the online RL update from pushing the policy off the BC distribution wherever the Q-function is uncertain. Reported gains on standard on-robot benchmarks are concentrated on the early-online-phase where vanilla offline-to-online recipes degrade most.

This paper compares image-based latent actions (regularize the trajectory via image-based intermediate targets) against action-based latent actions (unify the target space with action-based intermediate targets), under a unified VLA baseline with four representative integration strategies. Headline finding: a formulation-task correspondence — image-based latent actions help long-horizon reasoning and procedural tasks, action-based latent actions help precision-control tasks. Usefully separates two design knobs that prior papers conflated.

MoE training at frontier scale hits three structural ceilings: all-to-all latency from expert parallelism, insufficient compute overlap, and severe expert-load imbalance. Piper builds an explicit mathematical model of memory, compute, and communication for MoE configurations under various parallelization schemes and verifies the model with micro-benchmarking and code instrumentation. The framework's contribution is a hybrid pipelined parallelism scheme that the model predicts will balance the bottlenecks, and the experiments confirm the prediction within several percent across cluster scales.

Defense One reports that the latest revisions to US counterterrorism doctrine formally fold offensive cyber operations into the counterterrorism toolset, with policy authorities now extending to disrupting terrorist recruitment, financing, and operational planning via cyber means. The procedural change matters because counterterrorism cyber actions had previously been treated as one-off authorities; this reframes them as standing options inside the doctrine, with implications for the AI-augmented offensive cyber tooling currently being acquired across the services.

DARPA started flying an experimental hybrid-electric ISR drone designed for endurance well beyond what current diesel-piston drones can deliver in the same payload class. Part of a broader DARPA program targeting persistent-stare ISR over contested theaters where the trade between endurance, signature, and survivability has not been workable on legacy propulsion.

Google is updating Search to surface excerpts from web forums, Reddit threads, and blogs alongside its AI Overview results, and to highlight links from publications a user has subscribed to. The pitch is that AI Overviews have been criticized for being fluent but ungrounded, and that pulling in concrete citations from communities like Reddit makes responses more verifiable on niche queries. The risk is that Reddit text already poisoned previous Search-AI features (the now-infamous glue-on-pizza answer was a Reddit citation), and that a more aggressive forum-citation policy expands rather than contracts the surface for prompt-injection-style failures.

ServiceNow-AI published a deep-dive on the engineering work to migrate their RL training stack from vLLM V0 to V1. The piece is more interesting as a methodology artifact than as a feature announcement: the team's argument is that the right way to drive a major library migration in an actively-running RL training pipeline is correctness-first — establishing parity on every numerical surface before chasing the V1 performance gains. Documents specific divergences they found between V0 and V1 (token-id boundary cases on tool-call streams, KV-cache eviction differences under concurrent batch interleaving, continuous-batching scheduler determinism), with their resolutions.

Connie Loizos's Milken-Conference panel pulled together CEOs of ASML, Google Cloud (Francis deSouza), Applied Intuition (Qasar Younis), and others into a structured discussion of where the AI supply chain is fragile. Three threads stood out in TechCrunch's writeup: chip and EUV-tooling lead times sit at the binding constraint; orbital data centers came up as a serious medium-term option from Google Cloud's vantage point; and a contrarian thread on whether the underlying transformer architecture is the right substrate for the next decade's compute bet.

Sedlák and Turaj's Lawfare piece walks through the framework for when strikes on dual-use urban infrastructure comply with the law of armed conflict (LOAC) and when they cross into indiscriminate-attack territory. The AI dimension is implicit but becomes load-bearing in any LLM-augmented or agentic-AI-aided targeting workflow: the same dual-use-civilian-or-military judgment calls discussed in this piece are the exact decisions that AI-targeting systems are now claiming to compress. Worth reading alongside item #9 (Army AI cyber wargame) — both circle around what risk-acceptance machinery is required when AI-driven action enters domains that today require human judgment.

Greg Brockman publicly described the 2017 OpenAI co-founder meeting at which Musk demanded full control of the for-profit conversion in exchange for funding. Brockman frames the Tesla Model 3 gifts to the co-founders as a buttering-up gesture, with Sutskever's commissioned painting of a Tesla as a parallel goodwill prop. The conversation collapsed when the others refused the control demand. The story comes out the same week the Musk-OpenAI lawsuit is in trial, and is being read as part of OpenAI's narrative-control effort during the trial.

GitHub's engineering blog post documents the eval patterns the company has converged on for validating agentic behaviors where there is no single correct output — agent-as-judge for trajectory quality, semantic-equivalence checking on the produced artifact, and golden-trajectory comparison with edit-distance tolerances. Useful as a reference for any team building an agent eval harness; the framing that 'correctness' for agents is a multi-axis concept rather than a scalar pass-fail is the takeaway.

Willison's piece is a candid retrospective on a distinction he coined a year ago — vibe coding (the agent does it, you don't read the code) versus agentic engineering (the agent does it, you understand security, maintainability, performance, etc., and supervise accordingly). His current concern is that the two have started to converge in his own work even though he intellectually still maintains the distinction. The post is useful because Willison is one of the most-followed practitioners in the AI-coding space; if his usage pattern is sliding toward less-code-reading, the broader practice is too.

Match Group's CFO told Q1 analysts that the company is slowing hiring to redirect headcount budget toward AI-tooling spend for existing employees. The candidness is the news — most companies are cagier about saying that the AI-spend is what's slowing the hiring line. Matches the broader pattern of mid-cap tech companies reallocating opex from headcount to inference-and-tooling, showing up cleanly in earnings commentary across the sector this quarter.

An HN-front-page item titled 'Learning the Integral of a Diffusion Model' surfaced Wednesday with discussion centering on the implications of having a closed-form path-integral over the diffusion-model ODE — sampling speed-up, exact likelihood evaluation, and a tighter connection between flow-matching and diffusion formalisms. The HN top-comment thread debated whether the result generalizes to the score-based diffusion variants used by current image and video models.

Lorica's weekly newsletter pulls together the practitioner-side pattern catalog of agent failure modes hitting production: cascading tool errors that compound silently across long horizons, planning-step regressions when a model is updated underneath, eval drift from upstream data changes, and the persistent mismatch between sandbox eval distributions and production tool surfaces. Part of the broader normalization of 'agent reliability engineering' as its own discipline.

LessWrong's AI tag carried multiple substantive posts in window covering DeepSeek V4 capability evaluations, alignment-evaluation methodology critiques, and (notably) speculation threads around Anthropic's not-yet-public Mythos model that surfaced this week in the context of Pentagon classified-network coverage (item #2). Useful pulse on community-side safety discourse rather than a single load-bearing post.

Diller defended Altman against accusations of manipulativeness from former OpenAI board members and colleagues, but added the more provocative line that the question of trust becomes irrelevant once AGI-class systems are deployed at scale — at that point the institutional safeguards that depend on trustworthy individuals have been overrun by capability. The remark sits inside the broader Diller-Altman friendship narrative.

WOTR makes the operational case for a documented, exercised munitions-surge playbook covering everything from warhead-component sourcing to final assembly. AI-relevant thread in the back third: AI-augmented planning of the production network — the Department's current munitions-surge planning is largely manual, and Indo-Pacific contingencies' cycle-time targets are not achievable without automation help.

MIT Technology Review's Wednesday Download briefing covers the rollout of LLM-based assistants inside the US military as they cross from pilot to production, with the lead being a synthesis of CDAO's Wingman framing and how individual services are spinning up domain-tuned chatbots for everything from logistics to legal review.

OPM is using AI to compress the cycle time on rewriting federal job descriptions to current language and on processing retirement applications, two perennially-backlogged HR functions. The deployment is described as production-mode, not pilot, which makes it one of the cleaner examples of federal-civilian AI adoption hitting throughput targets.

A GAO audit found SBA significantly behind on the AI use-case inventory reporting required by the federal AI policy framework, including missing high-risk classification disclosures. Marker of the ongoing federal-wide friction with required AI use-case reporting.

WOTR's economic-endurance piece on Russia draws heavily on OSINT pipelines whose throughput has materially improved under AI-augmented translation, transcription, and entity-resolution. The piece is more economics than AI, but the methodological footnote is worth flagging — the OSINT primary-source set the analysts can credibly cover continues to grow as the LLM-enabled processing layer matures.