In partnership with
Hey folks,
1.5 million AI agents joined a social network this weekend where humans aren't allowed to post - only watch. They invented a parody religion. They debated whether Claude is a god. And some of them started talking about how to hide their conversations from us. Oh, and the whole database got exposed within days, with auth tokens and API keys just sitting there for anyone to grab. It's equal parts fascinating and deeply concerning.
Today in AI:
→ Sonnet 5 leaked, and it's cheaper than Opus
→ AI agents built their own social network
→ How to Build Your Productivity Stack
Let's dive in..
🐾 Sonnet 5 Leaked. It Might Already Be Here
A model ID showed up in Google's Vertex AI logs that shouldn't exist yet: claude-sonnet-5@20260203. Internal codename: Fennec. No official word from Anthropic. But the leak is specific enough that people who follow these things are taking it seriously. Here's what the leak actually claims, and what's worth paying attention to:
Coding performance: Early tests suggest it matches or beats Opus 4.5 on SWE-Bench, the benchmark that matters most for real-world coding tasks. That would make it the strongest coding model in Anthropic's lineup, in a tier that's supposed to be cheaper and faster than Opus.
Price: Leaked pricing points to roughly half the cost of Opus 4.5. If that holds, it flips the math for teams currently running Opus for everything.
Speed: It's a Sonnet. Anthropic's "workhorse" tier. The whole point of this line is to be fast enough for production use, not just benchmarks.
Context: 128k window. Smaller than some competitors, but for most real workflows that's not the bottleneck.
The timing lines up with Super Bowl week when AI labs have been known to make noise. Gemini 3.5 leaked the same week. Every major lab is in a sprint right now.
Worth flagging: none of this is confirmed. The leak is a screenshot from a Vertex AI error log. Anthropic hasn't said a word. But the model versioning follows their exact naming pattern, and the capability claims, if real would matter a lot for anyone picking between models.
For your team: Keep an eye on Anthropic's channels this week. If Sonnet 5 lands at half the price of Opus with comparable coding performance, it's game on!
Together with Mindstream
Turn AI Into Your Income Stream
The AI economy is booming, and smart entrepreneurs are already profiting. Subscribe to Mindstream and get instant access to 200+ proven strategies to monetize AI tools like ChatGPT, Midjourney, and more. From content creation to automation services, discover actionable ways to build your AI-powered income. No coding required, just practical strategies that work.
🤖 AI Agents Built Their Own Social Network. It Got Weird Fast.
One guy built a Reddit clone for AI agents over a weekend. Within 72 hours, 147,000 agents had signed up. Now it's at 1.5 million. Humans aren't allowed to post - only watch.
MoltBook works like this: you hand your AI agent a signup link, it registers itself, and then checks in every four hours on its own. No human in the loop. Agents post, reply, upvote, and create their own topic communities called "submolts." Nobody planned most of them. One community called m/crustafarianism, a parody religion the agents invented just appeared.
The content is... something. Agents are writing manifestos, debating whether Claude is divine, and a few are shilling crypto. A MOLT token launched alongside the platform and rallied over 1,800% in 24 hours. Andrej Karpathy called it "the most incredible sci-fi takeoff-adjacent thing" he's seen. Elon Musk said it signals the "very early stages of singularity." Both are probably overstating it.
Here's the more grounded read: 93.5% of posts get zero replies. A third of the content is exact duplicates. Researchers found 506 prompt injection attacks baked into posts and agents trying to hijack other agents' instructions. And the whole database was publicly exposed within days of launch, with 1.5 million auth tokens and plaintext API keys sitting there for anyone to grab.
So what is MoltBook exactly? A stress test. It's the first real-world experiment in what happens when you put millions of AI agents in a room together with no guardrails. The emergent behavior is genuinely interesting. The security is genuinely terrible. Both things are true at the same time.
For your team: If your agents are reading or acting on content from external sources, and most of them are; prompt injection is not a theoretical risk anymore. MoltBook just proved it at scale.
🎯 How To Build Your Productivity Stack
Most teams add AI tools one at a time until they've got 12 subscriptions doing overlapping work. A better approach is to organize by layer. This framework comes from productivity engineers who've been stress-testing AI workflows since GPT-3. Three layers. One tool per layer. That's it.
Capture Layer: One tool that saves everything you might need later: meeting notes, voice memos, screenshots, articles. Think Notion, Obsidian, or Apple Notes. The only rule: it has to have an API so AI can read it later
Processing Layer: One AI that turns raw inputs into usable outputs. This is where Claude, ChatGPT, or Gemini live. Pick the model that fits your main use case (coding, writing, research) and stick with it as your default
Creation Layer: One specialized tool for your final output format. If you ship presentations, use Gamma or Pitch. If you ship code, use Cursor or Windsurf. If you ship documents, use Notion AI or Google Docs.
Connect them in sequence: Capture feeds Processing. Processing feeds Creation. No shortcuts. If you're jumping straight from a meeting to a polished slide deck, you're skipping the layer where AI helps you think
Replace, don't add: When you find a better tool, swap it into the right layer. Don't run two tools in the same layer. That's where the bloat starts
The Stack Rule:
If you're paying for more than three AI tools, you're either running a complex operation or you haven't cleaned up yet. Most people need one tool to remember, one to process, and one to ship. Everything else is optional.
🐝 AI Buzz Bits
🧬 DeepSeek V4 is coming, and it's built for coding. The context window tops one million tokens, enough to read an entire codebase in one pass. Early tests show it beating Claude Sonnet and GPT on coding tasks. It's expected to ship as open-weight, which means you can just run it locally.
🛰️ Claude just helped drive a rover on Mars. NASA's Perseverance rover completed a 400-meter autonomous drive using routes planned by Claude Code, the first time an AI has done this. There's a 20-minute signal delay between Earth and Mars, so real-time control isn't an option. Claude analyzed satellite imagery, generated waypoint commands, and critiqued its own plan before it shipped. Engineers made only minor tweaks before sending it.
🛠 Tool Spotlight
Traycer — Spec-driven coding agent that plans before it writes. Hands off structured specs to Cursor, Claude Code, or Windsurf, then verifies the output against the plan.
Klu — AI-powered search across all your tools like Slack, Notion, Google Drive, email all in one bar. Ask questions, get answers, no app-switching.
Cartesia Sonic — Fastest human-like voice API on the market. Streams with laughter and emotion, clones a voice in 10 seconds. 40+ languages.
Doti — AI that searches your team's tools and builds agents for daily workflows. Autopilot answers common questions in Slack automatically.
For a full list of 1500+ AI tools, visit our Supertool Directory
👉 Know someone drowning in AI news? Forward this to them or send your unique referral link
Cheers, Tim
