← All digests

Varrock Bank Weekly — April 7–13, 2026

TL;DR

The US-Iran ceasefire lasted about 48 hours before collapsing — Trump announced a full Strait of Hormuz blockade Sunday, sending oil surging and BTC sliding from $73K back to $71K. Anthropic dropped Claude Mythos Preview, deemed "too powerful to release publicly" (cybersecurity reckoning or marketing genius — you decide). Intel joined Musk's Terafab chip complex, and the Trump crypto project WLFI imploded with Justin Sun calling it a "personal ATM."


AI & AGENTS

Anthropic's Mythos Preview: The Model You Can't Have — Anthropic released Claude Mythos Preview exclusively to Apple, Microsoft, Google, and select cybersecurity firms under "Project Glasswing." The model excels at finding and exploiting software vulnerabilities — reverse engineering stripped binaries, reconstructing source code, discovering zero-days. Anthropic says it's "too powerful" for public release. Vance and Bessent reportedly questioned tech execs on AI security before the announcement. Whether this is genuine safety concern or the most elaborate product launch in AI history, the result is the same: Anthropic just moved the Overton window on what "frontier capability" means. Every CISO should be budgeting for AI-powered red teams right now.

Gemini 3.1 Pro Quietly Dominates — Google's latest leads 13 of 16 major benchmarks and ties GPT-5.4 Pro on the Artificial Analysis Intelligence Index — at one-third the API cost. Gemini hit 750M users. The "Deep Think" mode is reserved for Ultra subscribers, but for enterprise workloads, the cost-performance ratio is making it harder to justify OpenAI's pricing. Google is winning on value, not hype.

Intel Joins Musk's Terafab Mega-Complex — Intel will manufacture chips for Elon's Terafab project alongside SpaceX and Tesla, powering robotics and data center ambitions. After years of foundry struggles, Intel is essentially becoming Musk's fab partner. Desperate move or strategic lifeline? Both — Intel gets guaranteed volume, Musk gets US-based chip production without TSMC dependency.

OpenAI Launches Safety Fellowship ($1B+ Commitment) — A 5-month program (Sep 2026–Feb 2027) offering $3,850/week stipends and up to $15K/month in compute. The OpenAI Foundation pledged $1B+ for life sciences, economic impact research, and AI resilience. Separately, OpenAI published an economic vision paper advocating public wealth funds, robot taxes, and 4-day workweeks. They're positioning as the "responsible AI company" while being sued by everyone — classic dual-track strategy.

Anthropic Acquires Coefficient Bio for $400M — The AI giant's biggest move into life sciences, acquiring a stealth NYC biotech startup in an all-stock deal. The team joins Anthropic's healthcare initiatives. This is Anthropic diversifying beyond pure model development — if AI-for-science is the next trillion-dollar market (it is), having domain-specific talent in-house matters more than another benchmark point.

$7 Trillion Infrastructure Reality Check — Reuters analysis: Meta, xAI, and others are targeting 110 GW of data center capacity. At Jensen Huang's estimate of $60B per GW, that's $7 trillion — nearly every dollar from hyperscaler cash flows and debt combined. The AI buildout is approaching absurdity. Watch for cancellations and delays in H2 2026 as capex meets reality.

Elorian: Ex-DeepMind Team Bets on Visual AI — Andrew Dai (former Google DeepMind) publicly launched Elorian, focused on improving how AI understands visual prompts. His thesis: current models have the visual understanding of a 3-year-old. If he's right, there's a massive capability gap waiting to be filled — and every multimodal application from autonomous driving to medical imaging stands to benefit.


CRYPTO & DEFI

Bitcoin Whipsawed: $69K → $73K → $71K — The week started with BTC at $68,860 amid Iran war fears. The ceasefire announcement Tuesday sent it surging to $72,700. It held around $73K through Saturday before Trump's blockade announcement Sunday dropped it back to $70,900. Currently defending $71K. The geopolitical correlation is undeniable — BTC is trading as a risk asset, not a safe haven.

Trump's WLFI Implodes: Justin Sun Goes Nuclear — World Liberty Financial used 5B of its own WLFI governance tokens as collateral on Dolomite (whose co-founder is a WLFI adviser) to borrow $75M in stablecoins. Justin Sun, once the project's largest backer, publicly called them a "personal ATM" and accused them of hiding a blacklist mechanism that can freeze user funds. WLFI responded: "See you in court." Gizmodo called it an "FTX-esque move." This is the ugliest crypto breakup of 2026 and a devastating look for politically-connected DeFi.

DeFi Yields Fall Below TradFi — Aave's USDC rate hit 2.61% APY, trailing Interactive Brokers' 3.14%. The era of easy DeFi money is officially over. This isn't temporary compression — it's structural. As rates stay elevated and on-chain activity normalizes, DeFi protocols need a new value proposition beyond yield. Real-world asset integration and institutional flows are the only path forward.

Starknet Enters Phase 4: Privacy + Bitcoin — StarkWare announced Starknet's fourth phase with three major developments: STRK20 framework for ERC20 privacy features, strkBTC (redeemable Bitcoin integration), and quantum-resistant cryptography for Bitcoin via STARK proofs. The v0.14.2 upgrade vote ran April 9–11, bundling protocol changes that prep for decentralized validation later this year. Major STRK token unlock coming April 15 — watch for sell pressure.

Blockchain Association vs. Citadel on Tokenized Equities — Filed a response with the SEC rebutting Citadel's arguments against tokenized US equity securities and DeFi trading protocols. The battle lines are drawn: TradFi incumbents vs. crypto natives over who controls the next-gen securities infrastructure. This will define whether DeFi can actually eat into equity markets or gets regulated into a sidecar.

Ripple's Acquisition Spree Continues — After $4B+ in 2025 acquisitions, Ripple added Solvexia (financial automation) and BC Payments (regulated license) in Q1 2026. They now control custody, prime brokerage, treasury management, stablecoin settlement, and regulatory licensing across 75+ jurisdictions. Ripple quietly built the full-stack institutional crypto infrastructure that everyone else is still pitching in decks.


MACRO & MARKETS

Iran Ceasefire Collapse → Hormuz Blockade — The two-week ceasefire announced Tuesday barely survived 48 hours. Iran's parliamentary speaker said 3 of 10 framework clauses were violated before negotiations even began. Peace talks in Pakistan collapsed Saturday, and Trump announced a full US naval blockade of the Strait of Hormuz on Sunday. CENTCOM says the blockade begins Monday (today). Oil had plunged on the ceasefire; expect a violent reversal. KKR projects WTI peaking at $130–150 this year. This is the single biggest macro variable right now — 20% of global oil and LNG flows through that strait.

Oil Crash → Oil Surge → Oil Chaos — WTI plunged 15%+ within hours of the ceasefire announcement (Phemex saw 300% surge in crude trading volume, daily volumes hit $85M ATH). Markets briefly rallied — S&P reclaimed key levels, Nasdaq hit its 200-day MA. Then it all reversed Friday/Saturday as talks collapsed. The dollar tumbled on ceasefire uncertainty and hasn't recovered. Energy volatility is now the defining market feature of 2026.

AI Capex: Undeterred by Geopolitics — Despite market turmoil, AI capital expenditure commitments from major tech companies have NOT reversed. Tech earnings are expected to significantly outpace the broader market. The paradox: companies are cutting costs everywhere except AI infrastructure. This bifurcation will define winners and losers through the rest of 2026.

UK Retail Investors Flee Stocks — ISA season (the UK's annual buying period) was a flop, with buyers staying away amid geopolitical pessimism. FT reported the weakest ISA flows in years. This is a leading indicator of retail risk aversion spreading globally.


LAUNCHES & MOVES

Elorian — Ex-DeepMind researchers launched a visual AI startup from Palo Alto, betting current models fundamentally misunderstand images. Backed by undisclosed funding.

Synchronicity Labs' Sync-3 — Most advanced lipsync model to date. Deepfakes just got better (again).

Infleqtion Quantum Guidance: $40M Revenue — Quantum sensing company posted $32.5M in 2025 revenue, guiding $40M for 2026. Deployed a 100-qubit system at UK's National Quantum Computing Centre. Quantum is moving from research to revenue.

Nebraska Passes Conversational AI Safety Act — First US state law specifically regulating AI chatbots that interact directly with individuals (companion AI, healthcare, commercial). Other states watching closely. The patchwork AI regulation era has begun.

Fintech Funding: More Money, Fewer Deals — Global fintech VC hit $12B across 751 deals in Q1 2026 (5% more dollars, 31% fewer deals than Q1 2025). Consolidation is accelerating — money is concentrating in fewer, larger bets.


NUMBERS THAT MATTER


LOOKING AHEAD

Strait of Hormuz Blockade (TODAY) — CENTCOM begins blockade operations Monday. Expect oil to surge, equities to sell off, and crypto to trade as a risk asset. The next 48 hours will set the tone for Q2 markets.

Starknet Token Unlock (April 15) — Large vesting event that traders have flagged as high-impact. Watch for sell pressure on STRK.

Nebraska AI Act Implications — Other states are considering similar conversational AI regulations. If you're building AI-powered customer/healthcare interactions, compliance planning starts now.

Anthropic Mythos General Availability? — No timeline for public release, but partners (Apple, Microsoft, Google) are testing now. Watch for capability leaks and benchmark comparisons. If Mythos is as good as claimed, the cybersecurity industry is about to be restructured.

Databricks IPO Watch — CEO hasn't ruled out 2026 but pushed timeline to H2. Toyota adopted their "vista" platform. If geopolitics calms down (big if), this could be the marquee tech IPO of the year.


GROWTH IDEAS

GUILD.XYZ

  1. AI Agent Gating — With Anthropic's Mythos setting the precedent for tiered access to powerful AI, Guild could offer "AI capability gating" — let communities gate access to AI tools/agents based on on-chain credentials. "Prove you're a security researcher to access this red team agent." This is a new gating primitive nobody else is building.

  2. Conversational AI Compliance Guilds — Nebraska just passed the first conversational AI law. Companies building chatbots will need compliance verification. Guild could position as the credentialing layer — "this AI was built by verified, compliant developers." Regulatory moats are the stickiest kind.

  3. Starknet Privacy Guild Integration — With Starknet Phase 4's STRK20 privacy framework, Guild could explore privacy-preserving role verification on Starknet — "prove you hold X tokens without revealing your wallet." Privacy + access control is a natural fit.

NOSTRA.FINANCE

  1. Real-World Asset Lending — DeFi yields are below TradFi for the first time. Nostra needs to diversify beyond pure crypto lending. Tokenized treasuries, real estate, or even tokenized oil futures (extremely timely given Hormuz) could attract capital that pure DeFi yield can't.

  2. strkBTC Integration — Starknet Phase 4 introduced strkBTC (redeemable Bitcoin). Nostra should be first to offer strkBTC as collateral for lending/borrowing. Being the go-to BTC lending market on Starknet captures the Bitcoin DeFi narrative before competitors.

  3. Institutional Privacy Vaults — Leverage Starknet's STRK20 privacy features to offer privacy-preserving lending pools for institutional users. "Borrow/lend without your counterparty seeing your portfolio." This is the feature that gets institutions off centralized platforms and into DeFi.


Developer Digest

Dev Digest — April 13, 2026

🔥 HOT RELEASES

Gemma 4 by Google Google dropped Gemma 4, a family of open models you can run fully on-device — private, free, no internet required. The community is already cooking: SuperGemma4-26B-Uncensored (0/100 refusals, 90% faster prompt processing, runs on 16GB VRAM), and Gemopus-4-31B — a Gemma 4 distilled on Claude Opus 4.6 reasoning. 🔗 X: https://x.com/outsource_/status/2043438707500368220 💻 HuggingFace: https://huggingface.co/Jackrong/Gemopus-4-31B-it-GGUF

GLM-5.1 by Zai China's most powerful open-source model — topping agentic benchmarks above GPT-5.4 and Claude Opus 4.6. MIT license, free to download and deploy. 🔗 X: https://x.com/dr_cintas/status/2043390476032901157

Hermes Agent by NousResearch — Workspace Update +7,454 stars in 24h (72.5K total). New release includes Task Board/Kanban, agent artifacts inspector, crew monitoring, multi-language support, configurable knowledge base, and Skills Hub. The agent ecosystem is exploding — people building full multi-agent orchestration loops with research/coder/QA agents. 🔗 X: https://x.com/outsource_/status/2043438707500368220 💻 GitHub: https://github.com/NousResearch/hermes-agent

Meta Muse Spark First model from Meta's new Superintelligence Labs. Natively multimodal, supports tool use, visual chain-of-thought, and multi-agent orchestration. Beating top major models on benchmarks. 🔗 X: https://x.com/dr_cintas/status/2043390476032901157


🧪 INTERESTING REPOS

Karpathy's LLM Wiki v2 → Second Brain Karpathy's Obsidian-based LLM wiki went viral (5K+ stars in 48h), now evolved into a "living memory system" with confidence scoring, memory tiers (working/episodic/semantic/procedural), knowledge graphs, forgetting curves, and contradiction resolution. Multiple implementations available. 💻 GitHub: https://github.com/NicholasSpisak/second-brain 💻 GitHub: https://github.com/safishamsi/graphify (70x token-efficient knowledge graph variant)

VoiceBox — Open-Source Voice Cloning Clone any voice from 3 seconds of audio. Runs 100% locally, no API key needed. 23 languages, 5 TTS engines, built with Tauri + Rust. 💻 GitHub: https://github.com/jamiepine/voicebox 🔗 X: https://x.com/midudev/status/2043689984939483218

OpenMontage — Agentic Video Production Open-source video production pipeline: 11 pipelines, 49 tools, produce a product ad for $0.69. Works with Claude Code/Cursor/Copilot as director. 12 video gen providers, 8 image providers, budget governance built in. Can run with zero API spend using local TTS + free stock. AGPL v3. 💻 GitHub: https://github.com/calesthio/OpenMontage 🔗 X: https://x.com/rryssf_/status/2043705860631998885

OpenScreen — Free Screen Studio Alternative Cinematic screen recordings with zoom effects, cursor highlights, background blur. No watermark, no subscription, free for commercial use. 27K+ stars. 💻 GitHub: https://github.com/siddharthvaddem/openscreen 🔗 X: https://x.com/sharbel/status/2042877211791818952

WeClone — Create Your AI Twin Export your chat history (Telegram, WeChat), fine-tune a model (Qwen2.5-VL-7B with LoRA), deploy as a chatbot that talks like you. All local + self-hosted. 16.5K+ stars. 💻 GitHub: https://github.com/xming521/WeClone 🔗 X: https://x.com/0xMerl99/status/2043642951532982594

Lingbot World — Interactive World Simulator Open-source interactive world simulator with long-term memory. 💻 GitHub: https://github.com/Robbyant/lingbot-world 🔗 X: https://x.com/tom_doerr/status/2043617543017951258

Karpathy-Inspired CLAUDE.md Someone turned Karpathy's AI coding agent rant into a single CLAUDE.md file: think before coding, write the simplest thing, only touch what the task requires, define success criteria. Cuts costs by being concise. 💻 GitHub: https://github.com/forrestchang/andrej-karpathy-skills 🔗 X: https://x.com/_vmlops/status/2043547904313614643


🎥 WORTH WATCHING

"Building AGI" Series — Multi-Agent Orchestration with Hermes Day 6 of building AGI: from single agent to coordinated crew (main/subconscious/research/coder/QA). Good walkthrough of multi-agent architecture patterns. 🔗 X: https://x.com/gkisokay/status/2043594876881969386

LLM Wiki Setup Tutorial Full YouTube setup walkthrough for Karpathy's second-brain system with Claude Code/Codex/Gemini integration. 🔗 X: https://x.com/NickSpisak_/status/2041747312737759532


💡 TECHNIQUES & IDEAS

Karpathy's AI Coding Agent Rules Karpathy called out every annoying thing AI coding agents do: wrong assumptions, bloated abstractions, 1000 lines when 100 would do, touching code it wasn't asked to touch. The community distilled it into 4 rules: think before coding, ask when confused, write the simplest thing, only touch what the task requires. 🔗 X: https://x.com/_vmlops/status/2043547904313614643

AutoReason — Iterative Refinement for Non-Testable Tasks Karpathy's AutoResearch concept applied to content/strategy/positioning — tasks where there's no test to pass. Key insight: asking a model to edit its own work repeatedly usually makes it worse (hallucinates flaws, scope creeps). Paper + code by SHL0MS, co-written with Hermes Agent. 🔗 X: https://x.com/shannholmberg/status/2043630276262432892

Marketing AI Knowledge Layer Single knowledge layer that ingests CRM, call transcripts, analytics, ad platforms, competitor data — refreshes every 30 min, all agents query the same brain. Data compounds over time: by month 6 the system knows more than any new hire. 🔗 X: https://x.com/shannholmberg/status/2043713370407514380


🔮 EMERGING TRENDS

Open-Source Model Distillation is the Meta Gemma 4 dropped days ago and there are already uncensored versions, Claude Opus distills, and merged abliterated variants. The gap between closed and open is measured in hours now.

Agent Orchestration Going Mainstream Hermes Agent (+7K stars/day), multi-agent crews, task boards, skills marketplaces. We're past "single agent does thing" — it's coordinated teams with specialized roles.

"Living Memory" Systems Karpathy's LLM Wiki sparked a wave: knowledge graphs with confidence scoring, forgetting curves, contradiction resolution. Moving from static RAG to dynamic, self-healing knowledge bases.

ui.sh Hype The AI design tool is generating serious buzz — devs redesigning entire sites in minutes. Still waitlist-gated but early users are shipping fast. 🔗 https://ui.sh


Full digest: https://digest.99.cooking