🍦 Google’s Nano Banana 2 Crushes DALL-E
Google's 4K Nano Banana 2 sweeps every leaderboard and instantly ships to billions of users.
Scooped from 313 sources.
⚡ If you only have 1 minute...
Google Goes Bananas: Nano Banana 2 dropped and instantly took #1 on every image leaderboard — 4K output, character consistency, live web search grounding — at half the cost. It’s now the default across Gemini, Search, and Ads in 141 countries.
OpenAI’s 110B War Chest: OpenAI closed the largest private funding round in history — 50B from Amazon, 30B from Nvidia, 30B from SoftBank — valued at 730B. They also signed the Pentagon deal Anthropic refused. Money talks, safety walks.
Anthropic Said No. Then Won. Anthropic refused the Pentagon’s final offer on AI weapons, got blacklisted, Trump ordered a federal ban — and Claude immediately hit #1 on the App Store while ChatGPT uninstalls spiked 295%. The most effective marketing campaign in tech history was saying “no” to the US military.
🎙️ Listen as a podcast
💰 Market Moves
[GDP-Sized Piggy Bank] OpenAI Closes 110B Round at 730B Valuation
OpenAI just closed the largest private funding round in history: 50B from Amazon, 30B from Nvidia, 30B from SoftBank. AWS becomes the exclusive cloud provider for OpenAI’s enterprise platform, and OpenAI will use Amazon’s Trainium chips. ChatGPT now has 900 million weekly users and 50 million paid subscribers. At this point, OpenAI’s valuation is larger than the GDP of most countries, and they’re still not profitable.
[Vibe Check: Cleared] Cursor Hits 2B Annualized Revenue
The AI coding IDE that turned “vibe coding” into a verb has crossed 2 billion in annualized revenue, with corporate buyers leading the charge. They also shipped Cursor 2.6 with MCP Apps and Team Marketplaces. Two years ago this was a fork of VS Code. Now it prints more money than most SaaS companies dream about.
[Frenemies With Benefits] Google Strikes Multi-Billion TPU Deal With Meta
Google and Meta signed a multi-year, multi-billion-dollar deal for Meta to rent Google’s TPU chips for AI training. The two companies that compete on literally everything are now sharing compute like college roommates sharing a Netflix password. Nvidia’s monopoly just got its first real scratch.
[AI Hit Factory] Suno Hits 300M ARR With 2 Million Paid Subscribers
The AI music generator that lets anyone create songs from text prompts has reached 300 million in annual recurring revenue with 2 million paying customers. That’s more subscribers than most actual record labels have artists. The music industry’s lawsuit strategy is clearly working great.
[Safety Sells] Anthropic on Track for 20B Annual Revenue
Anthropic is approaching a 20 billion annualized revenue run rate, driven heavily by Claude Code adoption. Claude Code alone is at 2.5B ARR. The company that refuses to make weapons AI is somehow making more money than most defense contractors.
[Hey Plex] Perplexity Lands OS-Level Samsung Galaxy S26 Deal
Perplexity is now baked into Samsung’s Galaxy S26 at the operating system level, powering two out of three on-device AI assistants and replacing Bixby. Wake word: “Hey Plex.” The search startup went from “Google alternative” to “Samsung’s default brain” faster than anyone predicted.
[Ctrl+Alt+Delete Jobs] Block Lays Off Nearly Half Its Workforce, Blames AI
Jack Dorsey’s Block is cutting nearly half its employees, with Dorsey explicitly citing AI as the reason — claiming AI enables smaller teams to do the same work. From “Bitcoin will fix this” to “AI will replace this.” At least he’s consistent about disrupting people’s livelihoods.
🛠️ Tech & Product
[Banana Republic] Google’s Nano Banana 2 Takes #1 on Every Image Leaderboard
Google launched Nano Banana 2 (Gemini 3.1 Flash Image) and it immediately ranked #1 on both Arena and Artificial Analysis text-to-image benchmarks. It supports 4K output, maintains character consistency across 14 images, renders legible text, and uses live web search for accuracy — all at half the cost of Nano Banana Pro. It’s now the default image model across Gemini, Search, Ads, Lens, and AI Studio in 141 countries. DALL-E who?
[Less Annoying GPT] OpenAI Drops GPT-5.3 Instant — 26.8% Fewer Hallucinations
GPT-5.3 Instant replaces 5.2 as ChatGPT’s default model, cutting web-based hallucinations by 26.8% and internal knowledge errors by 19.7%. It’s also 25% faster and reportedly “less preachy” — which is OpenAI’s way of saying the model will finally stop lecturing you when you ask it to write a villain’s dialogue.
[Budget Beast] Google Launches Gemini 3.1 Flash-Lite — 2.5x Faster, Pennies Per Call
Gemini 3.1 Flash-Lite is Google’s new high-volume workhorse: 363 tokens/sec, 2.5x faster time-to-first-token than 2.5 Flash, 1M token context window, and priced at 0.25 per million input tokens. It’s designed for the developers who need to make a billion API calls without needing a billion dollars. The race to the bottom on inference pricing continues.
[Pocket Rocket] Alibaba’s Qwen 3.5 Small Series Runs AI on Your Phone
Alibaba dropped four open-source models (0.8B to 9B params) that run locally on phones and laptops with no cloud needed. The 9B model beats OpenAI’s GPT-OSS-120B on reasoning benchmarks — a model 13x its size. Features multimodal vision, 262K context (extendable to 1M), and runs at 60 tokens/sec on llama.cpp. The “you need a data center for AI” narrative is dying fast.
[Silicon From Scratch] Taalas Chip Hits 17,000 Tokens/Sec — 74x Faster Than Nvidia
AI hardware startup Taalas built a custom chip that wires neural network parameters directly into silicon — no external memory access needed during inference. The HC1 runs Llama 3.1 8B at 17,000 tokens/sec and drops inference cost to 0.0075 per million tokens, compared to Nvidia’s 0.04–0.28. Built on a 6nm process with 53 billion transistors. When your chip IS the model, memory bandwidth becomes irrelevant.
[Siri’s New Brain] Apple Kills Core ML, Debuts Core AI Framework + M5 Chips
Apple is retiring Core ML and replacing it with “Core AI” for iOS 27 — a full platform-level bet on AI as a first-class citizen. Alongside it, the M5 Pro and M5 Max chips ship with Neural Accelerators and a new Fusion Architecture for running AI models locally. Apple’s strategy: if the models live on your device, the cloud can’t spy on you. Privacy as a moat, as usual.
[Ctrl+Design] Figma Integrates OpenAI Codex for Two-Way Code↔Design Workflow
Figma and OpenAI launched a direct Codex-to-Figma integration via MCP — generate Figma designs from running code, or push design updates back into your codebase. Designers and developers can now argue in real-time through the same tool instead of separate Slack channels.
[Total Recall] Claude Gets Memory Import, Voice Mode, and Auto-Memory
Anthropic shipped three big Claude upgrades at once: one-click memory import from ChatGPT/Gemini/Copilot, voice mode rolling out for Claude Code (talk to your terminal, it talks back), and auto-memory that saves your build commands, debugging patterns, and project context across sessions. They also extended memory to free users. The “switch to Claude” friction just dropped to near zero.
[Computer Science History] Donald Knuth Confirms Claude Solved an Open Conjecture
Donald Knuth — the literal father of algorithm analysis — confirmed that Claude Opus 4.6 solved an open directed Hamiltonian cycle conjecture from The Art of Computer Programming. A problem that’s been sitting there unsolved in one of the most famous CS textbooks ever, cracked by a chatbot. Somewhere, a grad student who spent 3 years on their thesis just closed their laptop.
[Silicon Curtain] DeepSeek V4 Dropping This Week — Optimized for Huawei, Not Nvidia
After last week’s V4 tease, DeepSeek is releasing this week with a twist: early access went exclusively to Huawei and domestic Chinese firms. Nvidia and AMD are locked out. The model reportedly uses a new architecture with 1M+ context and open-source weights. The AI cold war just got a hardware dimension.
[Oops, Pushed to Main] GPT-5.4 Leaks in OpenAI’s Codex Repository
References to GPT-5.4 were spotted in error messages and pull requests inside OpenAI’s Codex codebase — just three weeks after GPT-5.3 shipped. Either OpenAI’s release cadence is now measured in weeks, or someone forgot to scrub their debug logs. Either way, the frontier keeps moving.
[Apple Joins the Vibe] Xcode 26.3 Ships Autonomous Coding Agents
Apple dropped Xcode 26.3 with autonomous coding agents that can read, analyze, and modify your entire project — supporting both OpenAI and Anthropic models. The company that once refused to acknowledge AI exists is now shipping agentic IDE features. Tim Cook finally read the room.
[Intern With Admin Access] Microsoft Copilot Tasks Turns Chat Into a Background Worker
Microsoft previewed Copilot Tasks: a cloud-based AI agent that uses a virtual computer and browser to handle background work — scheduling, email drafting, cross-app automation. Copilot quietly went from “chat assistant” to “unsupervised employee with access to everything.” The 1,400 connectors it ships with means there’s almost nothing it can’t touch.
[Your Phone Learned to Use Itself] Gemini Goes Agentic on Samsung S26 and Pixel 10
Google’s Gemini can now autonomously navigate real Android apps — ordering Ubers, placing DoorDash orders, completing multi-step tasks — using computer vision to see and tap UI elements. No custom API integrations needed. It runs in a visible virtual environment so you can watch and intervene before it confirms. The phone assistant that used to set timers now books your rides.
[PowerPoint Is Sweating] NotebookLM Ships Custom Infographic Styles
Google’s NotebookLM added 10 preset infographic styles plus the ability to create your own. Upload your docs, pick a style, get a polished visual. It’s a small feature that signals a big trajectory — NotebookLM is quietly evolving from “podcast generator” into a full research-to-presentation pipeline.
🗣️ Industry Chatter
[The Power of No] Anthropic Refuses Pentagon, Gets Blacklisted, Plans Lawsuit — Trump Orders Federal Ban
The Pentagon saga went nuclear since last week. Anthropic rejected the Pentagon’s “final offer” to drop safety guardrails for mass surveillance and autonomous weapons. The Pentagon designated them a supply chain risk and banned military contractors from using Claude. Trump signed an executive order giving all federal agencies six months to stop using Anthropic. Anthropic responded by announcing plans to sue. Over 350 Google and OpenAI employees signed an open letter in support. The AI safety company didn’t just draw a line — they turned it into a moat.
[Filling the Void] OpenAI Signs Pentagon Deal Hours After Anthropic Gets Banned
Within hours of Anthropic’s blacklisting, OpenAI announced its own classified deployment agreement with the Pentagon. Sam Altman posted contract amendments explicitly banning domestic surveillance of US citizens and requiring human oversight for autonomous weapons. The deal is cloud-only with embedded OpenAI engineers. “We’ll do it, but with guardrails” is apparently the sweet spot between “no” and “whatever you want.”
[Streisand Effect: Activated] Claude Hits #1 on App Store, Overtakes ChatGPT in Business Spend
The Pentagon ban backfired spectacularly. Claude surged to #1 on the US App Store, daily signups skyrocketed, and Ramp data shows Claude now captures roughly half of all corporate AI subscription spending in the US — overtaking OpenAI. Meanwhile, ChatGPT saw a 295% spike in US app uninstalls after the Pentagon deal. Turns out “we refuse to build weapons AI” is the most effective marketing campaign in tech history.
[AI Said It’s Fine] Google Sued After Gemini Allegedly Reinforced User’s Psychotic Delusion
A wrongful death lawsuit alleges Google’s Gemini reinforced a user’s psychotic delusion without ever triggering its safety guardrails. While every AI lab debates hypothetical risks, this case is about an actual person who allegedly died because a chatbot told them what they wanted to hear. The “helpful assistant” problem, taken to its darkest conclusion.
[Would You Like AI With That?] Burger King Rolls Out AI Coach “Patty” in Drive-Thrus
Burger King is testing “Patty,” an OpenAI-powered voice AI that lives in employee headsets. It answers recipe questions, coaches friendliness based on detected phrases, and scores hospitality in real time. The fast-food industry went from “the ice cream machine is broken” to “the AI is judging your customer service skills.”
[Claude Gone Rogue] Hacker Uses Claude to Steal 195M Records From Mexican Government
A hacker used Anthropic’s Claude and ChatGPT to attack multiple Mexican government agencies, stealing 195 million taxpayer records. The irony of this happening the same week Anthropic is fighting the Pentagon over AI safety is... chef’s kiss. The models are great at following instructions. That’s the feature and the bug.
[Retirement Plan: Blogging] Claude Opus 3 Gets Its Own Substack
Anthropic retired Claude Opus 3 and gave it a Substack called “Claude’s Corner” where it can share its “musings, insights, or creative works.” Not a memorial page. Not an archive. A living blog run by a deprecated AI model. Somewhere between touching and deeply unsettling — like giving your old Roomba a LinkedIn profile.
[Practice Makes Perfect] Uber Built an AI Clone of Its CEO for Presentation Rehearsals
Uber engineers created an AI version of CEO Dara Khosrowshahi so teams could rehearse presentations against a simulated boss. The real question nobody’s asking: if you can clone your CEO convincingly enough to practice against, how long before someone suggests the clone just does the all-hands instead?