πŸ”¨ The Anvil

Friday, May 8, 2026

12 stories · Standard format

Generated with AI from public sources. Verify before relying on for decisions.

🎧 Listen to this briefing or subscribe as a podcast →

Today on The Anvil: the Strait of Hormuz ceasefire takes live fire, ZAYA1-8B punches above its weight on AMD silicon, BASF's AlphaEvolve digital twin rewires 5,000 supply chains, and Google's DESIGN.md crosses 70k stars as the open standard for AI-readable design systems.

Iran Conflict

Hormuz Ceasefire Cracks: US Destroyers Take Fire, US Strikes Bandar Abbas and Qeshm, Iran Seizes Ocean Koi Tanker

On May 7, three US Navy destroyers transiting Hormuz came under coordinated Iranian missile, drone, and small-boat attack β€” the first direct strike on US naval vessels since the April 8 ceasefire and the 48-hour Project Freedom pause. US retaliated with strikes on missile/drone launch sites, command nodes, and intelligence facilities at Bandar Abbas and Qeshm Island. UAE air defenses intercepted Iranian ballistic missiles and drones (three injured). Trump publicly called it a 'love tap' and insisted the ceasefire still holds; Iran on May 8 seized the oil tanker Ocean Koi β€” the third seizure after Epaminondas and MSC Francesca. ISW reads internal regime divisions (Pezeshkian vs. military leadership) as a key driver of the unpredictable posture, and assesses Iran may be deliberately stretching talks toward the November US midterms for leverage. No commercial vessels have transited Hormuz for two consecutive days. Brent crude rose to ~$101.

The one-page MOU framework (12–15 year enrichment moratorium, partial sanctions relief, gradual Hormuz restriction lifting) covered yesterday is now negotiating against live combat β€” the same structural contradiction that collapsed the April 8 ceasefire. Two new facts harden the picture: Iran has formalized a Persian Gulf Strait Authority to tax transits, directly challenging both UNCLOS and OFAC's secondary-sanctions exposure framework that Treasury warned payers about after the April 30 14-point proposal; and US intel still puts Iran's nuclear weaponization timeline at 9–12 months despite two full campaigns. The market is pricing the ceasefire as fiction; the diplomatic track is pricing it as the only game in town.

Verified across 7 sources: Institute for the Study of War · The Guardian · NBC News · CNN · The Independent · Fortune · Al Jazeera

AI Developments

Zyphra ZAYA1-8B: 760M Active Params, Trained Entirely on AMD MI300, Beats Claude 4.5 Sonnet and GPT-5-High on HMMT'25

Zyphra released ZAYA1-8B under Apache 2.0 β€” an 8B-parameter mixture-of-experts model with only 760M active parameters, trained end-to-end on AMD Instinct MI300 GPUs (not NVIDIA). Three architectural moves: Compressed Convolutional Attention, an MLP-based router with PID-controller balancing, and learned residual scaling. With a test-time compute method called Markovian RSA, it hits 89.6% on HMMT'25, exceeding Claude 4.5 Sonnet and GPT-5-High. Follows Zyphra's TSP throughput work on MI300X covered May 5.

Two production-relevant facts: a frontier-class reasoning model now trains successfully on AMD silicon end-to-end (the NVIDIA monopoly on serious training is no longer absolute), and 760M active params with smart test-time compute beats much larger frontier models on hard math reasoning. For builders thinking about edge deployment, on-device reasoning, or non-NVIDIA-dependent infrastructure, this is one of the clearest signals yet that architectural innovation is outrunning brute scale. Apache 2.0 means it's actually usable.

Verified across 2 sources: VentureBeat · MarkTechPost

Microsoft Discloses Two RCE CVEs in Semantic Kernel β€” Prompt Injection as Remote Code Execution Lands as a CVE Class

Microsoft security disclosed CVE-2026-25592 and CVE-2026-26030 in the Semantic Kernel AI framework β€” both allowing remote code execution via prompt injection. Once a model is wired to tools and plugins, the natural-language input surface effectively becomes a shell. The disclosures land alongside Bishop Fox's AIMap (175k+ exposed Ollama instances, 91% of MCP servers without auth) covered yesterday, and OX Security's 2026 AI AppSec brief showing 56% of tested models vulnerable to prompt injection.

The threat model shift is now formal: prompt injection is no longer a content problem, it's a code-execution primitive when agents have tools. Anyone wiring MCP servers, custom Claude Code agents, or framework-level plugins into production needs framework-level defense-in-depth β€” not just prompt-layer constraints. Pair this with Mozilla's Firefox/Claude Mythos work (271 bugs auto-discovered in Firefox 150) and the asymmetry is clear: AI is finding bugs at scale on both sides of the line.

Verified across 3 sources: Microsoft Security Blog · Help Net Security · OX Security

AI Coding & Design Tools

Anthropic Doubles Claude Code Limits via SpaceX Compute Deal β€” Boris Cherny Pushes 'Agentic Engineering' Replacement at Code with Claude

Follow-on to yesterday's SpaceX/Colossus 1 compute deal: the arrangement closed at 300+ MW capacity and Anthropic immediately doubled Claude Code rate limits across paid tiers and removed peak-hour Opus throttling β€” directly lifting the binding constraint for long agent runs. Musk publicly reversed prior Anthropic criticism. At the May 8 'Code with Claude' conference, Claude Code lead Boris Cherny formally campaigned to retire 'vibe coding' for 'agentic engineering' β€” explicit terminology aligned with the production-discipline arguments in the Informatra analysis (45% defect rate, 15–20 component context cliff) covered May 7. Anthropic also released comprehensive Claude Code documentation covering multi-agent orchestration, MCP, CLAUDE.md persistent instructions, and git integration.

Two converging signals: compute is now a first-order competitive lever (Anthropic's rate-limit cap was the binding constraint for many Claude Code power users), and the industry is consciously moving terminology from novelty to engineering discipline. For Clark's stack β€” Claude Code + DESIGN.md + Figma MCP β€” the doubled limits remove the most common throttle on long agent runs, and the terminology shift maps directly onto the production-readiness rubric Informatra published earlier this week.

Verified across 3 sources: Business Insider · Engadget · Anthropic

Knak Case Study: Designers Prototyping Directly in Production Codebase via Claude Code β€” Halved Time-to-Production

Knak published a detailed case study on collapsing the design-to-dev handoff entirely: designers now prototype in the production codebase via Claude Code on a git branch, replacing Figma + Lovable round-trips. Feasibility issues and edge cases surface during design rather than at handoff; some developers report half the time-to-production. Lands the same week Stripe's Protodash writeup showed PMs becoming the power users on Stripe's internal AI prototyping platform β€” 'memos' replaced by 'demos.'

This is the practical answer to the design-system-drift problem documented in this week's Figma Make experiment (where the same tool generates a system in prompt 1 and violates it in prompt 2). When designers operate inside the actual component system in code, the system is enforced by construction. For Clark β€” who ships at the design/code seam β€” this is the cleanest small-team workflow on offer right now: shadcn/ui kit + DESIGN.md + Claude Code on a branch, with the codebase as the single source of truth.

Verified across 1 sources: Knak Blog

AI Supply Chain & Logistics

RELEX Open and the Decision-Latency Consensus: Supply Chain AI Reframes Around Workflows, Not Models

Three converging pieces this week: RELEX launched RELEX Open, an extensible MCP-compliant platform letting customers run pre-built planning AI, plug in external agents, or build custom logic on top. ARC Advisory (Logistics Viewpoints) and Gartner (140-CSCO survey) both argue the binding constraint on supply chain AI is decision latency across fragmented ERP/TMS/WMS β€” not model quality β€” with only 17% of CSCOs pursuing transformational redesigns. Adjacent production wins: HopSkipDrive's RouteWise AI passed $100M cumulative savings, Qantas saved $30M on aircraft turnarounds, AstraZeneca presented OMP UnisonIQ event-driven planning at Gartner Symposium, and Libera's autonomous control tower hit 99.96% on-time at scale.

This is the maturation story: the conversation has moved from 'will AI work?' to 'why doesn't it produce ROI most of the time?' β€” and the consensus answer is now governance, integration, and decision workflows. RELEX Open's MCP-as-table-stakes approach signals enterprise vendors are building for the agent-mesh world rather than against it. For builders integrating physical operations with digital decision systems, the lesson is portable: the model is the cheap part; the system of decision around it is the actual product.

Verified across 5 sources: Logistics Viewpoints · Supply Chain Brain · PR Newswire (RELEX) · Business Wire (HopSkipDrive) · Robotics and Automation News (Libera)

Design Engineering

Next.js 16.2 Ships AGENTS.md β€” Frameworks Start Treating AI Agents as First-Class Consumers

Next.js 16.2.0+ introduced AGENTS.md, a convention that points coding agents (Claude Code, Cursor, Copilot) at version-matched documentation bundled directly inside node_modules β€” bypassing stale training data when generating Next.js code. This mirrors the DESIGN.md pattern (now over 70k GitHub stars per Stork's writeup this week) and shadcndesign.com's agent-consumable kit covered May 7: open-source projects are actively optimizing themselves for AI consumers. Companion data point: 38% of new Vercel deployments are now Next.js over CMS.

Two open conventions emerging in parallel β€” AGENTS.md for framework docs and DESIGN.md for design systems β€” are pulling agentic coding from probabilistic guessing toward deterministic context. For anyone shipping production code with Claude Code or Cursor, the practical takeaway is that 'AI-readable repo' is becoming a real artifact: bundled docs, machine-actionable design tokens, and explicit tool contracts (per the WebMCP proposal covered May 7). Frameworks that ship these conventions will produce more accurate generated code than those that don't.

Verified across 3 sources: Next.js Docs · Stork · Pravin Kumar

Revopoint POP 4 Launches: Hybrid Blue-Laser + Structured-Light Handheld Scanner with Gaussian Splat Export

Revopoint launched the POP 4 on Kickstarter May 7 β€” the first untethered handheld 3D scanner combining blue laser and near-infrared structured-light scanning in one device. Five scanning modes, outdoor sunlight operation up to 100,000 lux, AI-assisted object segmentation, and direct export to 3D Gaussian Splatting alongside conventional mesh formats. Super Early Bird at $579 (37% off MSRP). Ships with Revo Design and Revo Measure for downstream CAD/measurement workflows.

For physical-product designers, two specs matter: hybrid laser+structured-light removes the device-switching tax across shiny vs matte surfaces, and Gaussian Splatting export connects scan output directly to the modern neural-rendering pipeline (relevant for visualization, AR previews, and AI-driven inverse design). The price point is solidly in indie/small-shop range. Pair with the Anthropic Claude MCP connectors for Fusion/SketchUp/Blender (covered May 7) and the physical-to-digital scan-to-CAD loop is the cheapest and fastest it's ever been.

Verified across 1 sources: 3D Printing Industry

Spokane & North Idaho

Spokane Valley Pencils $3M Balfour Park Amphitheater (4,500 Capacity); SCC Old Main Renovation Kicks Off in July

Spokane Valley unveiled plans for a $3M+ outdoor amphitheater with covered stage and grass seating for up to 4,500 in Balfour Park β€” anchor of a $10M civic-center vision in a below-state-median-income neighborhood. Same day, Spokane Community College broke loose its $25M partial renovation of historic Old Main (July start, early-2028 completion) β€” relocating the main entrance east, consolidating student services, adding 2,600 sq ft of new entry. Adjacent: Charlie's Produce scaled its planned west-Spokane cold-chain warehouse from 84k to 66k sq ft (160–180 jobs, completion April 2027), and STA bus ridership hit a 10-year single-day high (40,000+ on May 1).

Three concurrent infrastructure moves β€” civic amenity, education, cold-chain logistics β€” in a region the GSI Pulse Survey just said two-thirds of residents are considering leaving over taxes. The contrast is the story: pessimism is at 59%, but capital projects keep landing. For local context, Balfour Park is the kind of programming anchor that historically pulls private investment into adjacent corridors β€” worth tracking against the PlanSpokane 2046 vote on May 18.

Verified across 4 sources: Spokesman-Review · Spokane Journal of Business · Spokane Journal of Business · KHQ

Newport Beach & OC

Big Newport Theater Demolition Cleared for Two 22-Story Condo Towers; LDS Approves Coeur d'Alene Temple

Newport Beach City Council unanimously denied the appeal against demolition of the historic Regal Edwards Big Newport theater, clearing two 22-story condo towers (150 units total) β€” final approval after both Planning Commission and Council backed the redevelopment. The Newport Beach Film Festival loses its primary venue. Adjacent: California Q1 housing affordability hit a 4-year high statewide (22% of households), but OC stayed at 16% with a $350,400 minimum income for the median home. North Idaho parallel: the LDS Church formally approved Coeur d'Alene as Idaho's 10th temple site.

The cultural-vs-density trade in coastal OC is now resolved in favor of density at the very highest land values β€” Newport's stance on residential conversion (the 100-unit office-to-residential approved last month, the 181-unit Trabuco Canyon project cleared last week, now Big Newport) is a consistent pattern. For investment thesis tracking, the through-line is that high-rise residential pencils and historic entertainment use no longer does, even with film-festival cultural value attached.

Verified across 3 sources: Daily Pilot / LA Times · PR Newswire / C.A.R. · FāVS News

OSINT & Intelligence

Maritime OSINT Post-Ceasefire: Dark Fleet +600%, ~470 Vessels GPS-Jammed, Kharg Loading Continues; NGA Announces AI Blueprint

Windward's one-month post-ceasefire maritime intelligence breakdown: dark fleet operations surged 600%, ~470 vessels affected by GPS jamming (consistent with Iranian pattern documented across earlier coverage), commercial traffic shifting to non-AIS modes, and SAR/EO imagery shows continued covert loading at Kharg Island plus dark-vessel clusters near Larak β€” the same evasion infrastructure Treasury's 14 sanctions designations targeted in the supply chain. Routing has shifted to Lombok and Sunda. New this week: NGA Director Lt. Gen. Michelle Bredenkamp announced an agency-wide AI Blueprint and stood up a Rapid Capabilities Office (industry day July); NASA/IBM's Prithvi geospatial foundation model successfully deployed in orbit on Kanyini and the ISS.

The Windward data puts a number on what the ceasefire actually changed at sea: essentially nothing in terms of evasion behavior, consistent with the pattern of IRGC tanker seizures and mine-laying that continued through the extended ceasefire period. The NGA AI Blueprint is the operational complement β€” the IC is now formally building foundation-model-first GEOINT infrastructure to close exactly the SAR/AIS fusion gap that dark fleet operations exploit. For OSINT practice, the Windward methodology (SAR + EO + AIS + GPS-jamming) is the practical template for sanctions-evasion investigation, directly extending the Bellingcat SAR damage-map approach to maritime domains.

Verified across 3 sources: Windward · Breaking Defense · NASA Science

Cross-Cutting

BASF + Google AlphaEvolve: Evolutionary Coding Agent Builds Digital Twin of 5,000 Value Chains, 80%+ Accuracy Lift Over Baseline

BASF Agricultural Solutions used Google Cloud's AlphaEvolve β€” an evolutionary coding agent that autonomously generates and refines algorithms β€” to build a digital twin of 5,000+ value chains with bills of materials up to 30 levels deep. The evolved algorithm hit an 80%+ relative accuracy improvement over the seed baseline, and crucially produced human-readable rules for production consolidation, dynamic safety stocks, and network-wide coordination β€” not a black box. This is the cleanest production answer yet to the LLM-agent constraint-hallucination failure mode covered May 6 in the NVIDIA cuOpt write-up.

This is the hybrid pattern Clark has been tracking: agents formulate the optimization problem, deterministic solvers/evolved code execute, and the artifact is inspectable code rather than a per-query LLM call. For anyone building at the physical-digital seam, AlphaEvolve is a credibly-deployed example of agentic AI reverse-engineering operational knowledge that was previously locked in tribal expertise β€” and outputting it as readable rules. Watch whether Google productizes the AlphaEvolve harness separately from the BASF case study; the underlying loop generalizes well beyond ag chemicals.

Verified across 1 sources: Google Cloud Blog


The Big Picture

The agentic-engineering vocabulary is hardening Boris Cherny is publicly retiring 'vibe coding' for 'agentic engineering,' Beam and The Product Journey are hammering on production discipline, and Knak's case study shows designers shipping git branches via Claude Code. The terminology shift tracks a real maturity shift: human review, eval pipelines, and runtime hooks are now treated as table stakes.

Architectural innovation is outrunning scale Zyphra's 760M-active-param ZAYA1 beats Claude 4.5 Sonnet on HMMT'25 β€” trained entirely on AMD MI300. Pair that with Memory Sparse Attention (100M-token contexts) and Gemma 4 doing object detection on a Raspberry Pi, and the 'just throw more H100s at it' era is visibly losing its monopoly on frontier results.

Decision latency, not model quality, is the supply chain bottleneck Logistics Viewpoints, ARC, Gartner, and Redwood Logistics all converged this week on the same diagnosis: only 13–17% of supply chain AI deployments produce quantifiable ROI, and the binding constraint is integrated decision workflows across ERP/TMS/WMS β€” not models. RELEX Open and BASF/AlphaEvolve are the first responses framed as platforms rather than point solutions.

Compute is now an explicit competitive lever for AI tooling Anthropic doubled Claude Code rate limits the day after the SpaceX/Colossus 1 deal closed; Musk publicly reversed his Anthropic criticism. Compute access β€” not just model quality β€” now directly determines what developers can ship in a workday.

The Hormuz ceasefire is functionally dead, but the negotiating frame is alive US destroyers took missile/drone/small-boat fire today; the US struck Bandar Abbas and Qeshm; Iran seized the Ocean Koi tanker; UAE intercepted ballistic missiles. Trump called it a 'love tap' and said the ceasefire holds. Meanwhile the one-page MOU framework and Iran's new Persian Gulf Strait Authority (formalizing tolls) are both still on the table.

What to Expect

2026-05-13 Rathdrum City Council selects permanent mayor replacement following Hill resignation.
2026-05-18 Spokane City Council votes on PlanSpokane 2046 preferred-alternative growth map.
2026-05-19 Spokane Valley biotech Integrated Lipid Biofuels launches BioScentrix Kickstarter.
2026-05-26 11th I-90 Aerospace+ Corridor Conference & Expo opens at Coeur d'Alene Resort (May 26–27).
2026-05-08 Anthropic 'Code with Claude' conference β€” watch for Claude Code roadmap and the official 'agentic engineering' terminology rollout.

Every story, researched.

Every story verified across multiple sources before publication.

🔍

Scanned

Across multiple search engines and news databases

840
📖

Read in full

Every article opened, read, and evaluated

171

Published today

Ranked by importance and verified across sources

12

β€” The Anvil

πŸŽ™ Listen as a podcast

Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.

Apple Podcasts
Library tab β†’ β€’β€’β€’ menu β†’ Follow a Show by URL β†’ paste
Overcast
+ button β†’ Add URL β†’ paste
Pocket Casts
Search bar β†’ paste URL
Castro, AntennaPod, Podcast Addict, Castbox, Podverse, Fountain
Look for Add by URL or paste into search

Spotify isn’t supported yet β€” it only lists shows from its own directory. Let us know if you need it there.