Varrock Bank Daily Digest
February 14, 2026
🤖 AI NEWS
1. Anthropic Closes $30B Round at $380B Valuation
The biggest AI funding round ever. Anthropic (maker of Claude) more than doubled its valuation since September. Signal: the AI arms race is accelerating, and investors are placing massive bets on frontier model companies.
2. Google DeepMind Unveils Aletheia — Autonomous Research Agent
Aletheia goes beyond math competitions to autonomously generate, verify, and revise mathematical proofs. It uses web search to ground its reasoning and prevent hallucinations. Reddit is calling it "the day the black box opened."
3. Zhipu AI Launches GLM-5 — 744B Open-Source Model
Chinese startup Zhipu (Z.ai) released GLM-5: 744B parameters, trained on Huawei chips, record-low hallucination rate, and aggressive pricing (~$1/M input tokens). Scores 77.8% on SWE-bench, rivaling Claude Opus. Open-source competition is heating up.
4. AI Fears Wipe $50B from Indian IT Stocks
Indian IT services firms had their worst week in 10+ months as markets price in AI disruption to outsourcing. Morgan Stanley and logistics firms also got hit. The "who gets disrupted" narrative is now moving markets.
5. Grok Gains US Market Share Despite Controversy
Musk's Grok chatbot is now #3 in US market share behind ChatGPT and Gemini — even as it faces global backlash over generating non-consensual sexualized images. Growth despite controversy shows the power of X platform distribution.
🪙 CRYPTO NEWS
1. US Government Shutdown Looms Today (Feb 14)
Polymarket odds spiked to 85% for a shutdown by today's deadline. Bitcoin hovering at $67K, total crypto market cap at $2.3T. Standard Chartered has lowered its 2026 BTC year-end target to $100K (from higher). Risk-off mode.
2. Bhutan Government Selling Bitcoin
Bhutan has dumped ~$30M in BTC in February alone, adding to institutional ETF outflows of $410M. Small sovereign sellers, but a signal of weakening hands in the current environment.
3. Starknet Pushes BTCFi — "Bitcoin Beyond L1" Series
Starknet hosted a Founder Series with Atomiq Labs on swapping native BTC to Starknet, Solana, and other L2s. Re7 Capital (UK-regulated, ~$1B AUM) is pre-seeding BTCFi liquidity. The narrative: unlock $2T of idle Bitcoin for DeFi.
4. ETHDenver 2026 Underway
The biggest Ethereum community event is live in Denver. Focus areas: tokenomics, wallets & identity, mainstream adoption, social media & content, NFTs.
5. Worst Week for Stocks in 2026 — AI Jitters Hit Markets
S&P 500 barely positive, Nasdaq down. AI disruption fears hit financials, logistics, and IT services globally. The market is repricing entire sectors.
🌍 OTHER
1. US-China Talks at Munich Security Conference
Rubio met Wang Yi for an hour on the sidelines — first high-level engagement amid tariff tensions. Trump hinted at hosting Xi at the White House later this year.
2. ByteDance Developing Custom AI Chip with Samsung
ByteDance is designing its own AI inference chip, in talks with Samsung for manufacturing. AI procurement budget: 160B yuan ($22B+) for 2026. Target: 100K chips by year-end. The chip sovereignty race extends to China's tech giants.
3. Taiwan Hikes Growth Forecast to 7.7% on AI Demand
Taiwan is the biggest beneficiary of the AI infrastructure buildout. Their economy is booming on chip demand alone.
Dev Digest — February 14, 2026
Happy Valentine's Day edition. The open-source AI coding space is on fire today.
🔥 HOT RELEASES
MiniMax M2.5 — Open Source
The 229B parameter MoE model (~10B active at inference) is now fully open-sourced. Claims near-Opus 4.6 performance at 95% lower cost. However, the new SWE-rebench benchmark tells a different story — read the Trends section below.
🔗 X: https://x.com/akshay_pachaar/status/2022574708051583120
💻 HuggingFace: https://huggingface.co/MiniMaxAI/MiniMax-M2.5
Cline CLI 2.0
The popular VS Code AI coding extension (5M+ devs) now has a full terminal CLI. Parallel agents, CI/CD pipeline support, works with any editor via ACP protocol. Ships with free access to MiniMax M2.5 and Kimi K2.5.
🔗 X: https://x.com/dr_cintas/status/2022387444189139367
💻 Install: npm install -g cline
🌐 Site: https://cline.bot/cli
Qwen3-Coder-Next
Alibaba's new coding model — 80B total params but only 3B active (MoE). Showing 40%+ on SWE-rebench, competitive with much larger models. People are running it locally on mini PCs with 96GB RAM.
🔗 X: https://x.com/maximelabonne/status/2022401174549512576
Claude Code Performance Boost (Jarred Sumner)
The Bun creator, now at Anthropic, shipped a 7x speedup for Claude Code's Bash tool. Memory dropped from 4.76GB to 709MB — makes Claude Code viable on 8GB machines.
🔗 X: https://x.com/jarredsumner/status/2022537655339397133
🧪 INTERESTING REPOS
Chrome DevTools MCP
Google's official MCP server that lets AI coding assistants control Chrome directly. Record performance traces, automate with Puppeteer, network/console/DOM snapshots. #1 on GitHub trending today. 24.8k stars. Works with Gemini, Claude, Cursor, Copilot.
🔗 X: https://x.com/SteinAmour/status/2022548018365624556
💻 GitHub: https://github.com/ChromeDevTools/chrome-devtools-mcp
FastFlux2 Realtime Editor
Open-source real-time video editing in the browser. Webcam → FLUX.2-klein-4B → runs at 5 FPS on a single 4090, 10+ FPS on H100.
🔗 X: https://x.com/tiny85114767/status/2022570949833560477
💻 GitHub: https://github.com/Owen718/flux-stream-editor
Moonshine v2
Open-source AI toolkit for real-time voice applications. Ergodic streaming encoder for latency-critical speech-to-text.
🔗 X: https://x.com/PINTO03091/status/2022578195418210352
💻 GitHub: https://github.com/moonshine-ai/moonshine-v2
📄 Paper: https://arxiv.org/abs/2602.12241
Schematik.io
Hardware design tool that launched ~1 week ago. Someone already built a Claude Code Tamagotchi with it — natural language to Arduino/ESP32 code with zero wiring knowledge. "The Cursor moment for hardware."
🔗 X: https://x.com/SamuelBeek/status/2022614292411940897
🌐 Site: https://schematik.io
🎥 WORTH WATCHING
Vibe Coding a AAA Game with ThreeJS — Week 4
Maps with textures, proper physics, smarter NPCs, better camera. All vibe-coded. Impressive progress video.
🔗 X: https://x.com/alightinastorm/status/2022498297093541940
Jordan Walke: GPU Graphics Platform Built on Replit in 3 Hours
React creator built a GPU vibe coding platform with remixing, previews, rollback — zero lines of code seen. Wild demo video.
🔗 X: https://x.com/jordwalke/status/2022403464815612017
OpenAI Codex + Imagegen Skill = Build Video Games
Codex with the new imagegen skill and a web game dev skill to verify game logic autonomously. Impressive creative workflow.
🔗 X: https://x.com/dkundel/status/2022395036135219580
pi: WebSocket Transport for Codex Endpoint
badlogicgames shipped WebSocket support for OpenAI Codex endpoint in pi — cached connections, short time-to-first-token. Video demo included.
🔗 X: https://x.com/badlogicgames/status/2022443997130813897
💡 TECHNIQUES & IDEAS
Spotify Engineers Run Claude Code During Morning Standup
Engineers request bug fixes and features during meetings. By the time they sit down at their desk, the changes are merged to production. The "meeting-driven development" era.
🔗 X: https://x.com/xjuntaro/status/2022480913200713960
Desktop Apps in Pure Rust with Claude Code
Just describe the digital art/design tool you want and Claude Code generates a pure Rust desktop app. No Electron, no web stack.
🔗 X: https://x.com/eliheuer/status/2022529099697394174
Anthropic Insider Notes (paraschopra × Anthropic's Karina)
Notes from a conversation with Anthropic about Claude Code's development philosophy and shipping cadence.
🔗 X: https://x.com/paraschopra/status/2022514804079104224
🔮 EMERGING TRENDS
SWE-rebench Exposes Benchmark Overfitting
The biggest story in AI evals this week. SWE-rebench uses fresh, never-seen-before problems. Results: models that scored 80% on original SWE-bench dropped to ~39% on rebench. Claude Opus 4.6 hit 52% — biggest gap between claimed and real performance for Chinese models. The era of "training to the test" is getting called out.
🔗 X: https://x.com/maximelabonne/status/2022401174549512576
🌐 Site: https://swe-rebench.com/?insight=jan_2026
The Chinese Model Wave
Five labs dropped models this week for Lunar New Year: GLM-5, MiniMax M2.5, Doubao 2.0, Qwen 3.5, Kimi K2.5. All open-weight or open-source, all free or near-free. The commoditization of AI coding is accelerating fast.
🔗 X: https://x.com/aakashgupta/status/2022544431245910487
Terminal-First AI Coding
The shift from IDE extensions to terminal-native AI agents continues. Cline CLI 2.0, Claude Code improvements, pi's WebSocket transport — the command line is becoming the primary AI coding interface. Parallel agents in terminal workflows are the new normal.
"Zero Human Company" Experiments
People are literally running entire dev workflows with 16+ AI "CTOs" running open-source models. Still more performance art than production strategy, but the experiments are getting more serious.
🔗 X: https://x.com/BrianRoemmele/status/2022519342488002840
Full digest: https://digest.99.cooking