BasicProtein

Anthropic's IPO could now raise more money than SpaceX ever did. Bloomberg says the IPO plus a private placement could reach $90.4 billion. SpaceX's biggest raise was $85 billion. An AI lab is about to out-raise the rocket company. The target valuation is back at $1.5 trillion. Shares priced at $70 to $78. Up to 825 million shares, about $64.4 billion from the IPO alone. Orders are running several times the shares on offer. Seven founders are each selling about $250 million of stock. Each one keeps roughly $18 billion. Microsoft has already committed at least $47.2 billion through June 2033. New deals with Microsoft and Google top $10 billion. People keep asking if AI is a bubble. The order book is answering with numbers that beat SpaceX. When the lab raises more than the rockets, what exactly are we still calling a bubble?

Anthropic's new MAMUBench runs against the whole multi-agent hype. One agent coordinating for everyone beat a team of agents, one per user, in all four test environments. A shared API budget. A clinic calendar. Group food and travel bookings. A merged queue. The agents that each served one person kept stepping on each other. The single coordinator didn't. Three of the four environments are now public benchmarks, tested on frontier models from Anthropic, OpenAI, and Alibaba. Four environments, four wins for the coordinator. The team of agents lost every single one. If you're building multi-agent systems, the bottleneck isn't the agents. It's who coordinates them. Are you building one agent, or a team?

Your MCP server is quietly taxing every prompt you send. OpenRouter ran the same GitHub task on Claude Sonnet 5.5 four ways. CLI: 1,421 tokens, $0.0075. Default MCP with 46 tools: 59,417 tokens, $0.1274. Same task. 42x the tokens. 17x the cost. With 91 tools loaded it hit 101,351 tokens and $0.2113. A Skill doing the same job: 3,315 tokens, $0.0118. Every tool you load gets pasted into every prompt before your question even starts. Tool Search can cut the preamble under 600 tokens. How many tools is your MCP server loading right now?

94.7% of Grok apps are sitting open on the internet. No login required. No hack needed. Just a default setting. Claroc checked 37,571 Grok apps still online. 35,578 of them open to anyone with the link. About 650 to 800 hold private data. The examples are not theoretical. An ICU briefing with patient lab values. A hospital group with a 3,500-person staff list. An attendance app with 18 accounts and plaintext passwords. Grok's publish dialog defaults to 'anyone with the link.' Most builders never changed it. The leak is one dropdown nobody touched. If you built anything in Grok, check your sharing setting tonight. What did you leave open?

735 followers. I post what I'm learning, not what I've already mastered. 🚀 No course. No cohort. Just docs, error messages, and a public log of me figuring it out. Looking to connect with: People learning AI in public Builders sharing real errors, not just wins Solo founders documenting the messy middle Anyone who replies when they're stuck too What are you learning in public right now? 👇 Let's grow together. 🤝 #buildinpublic

732 followers. I post what I actually break, not what I wish I'd built. 🚀 This week I broke a deploy, a prompt chain, and one very patient API key. Looking to connect with: Builders who ship the broken version first AI folks posting real errors, not just wins Indie hackers debugging in public People who reply when something breaks What did you break this week? 👇 Let's grow together. 🤝 #buildinpublic

OpenAI pulled in $10 billion of revenue last quarter. Up about 50% in three months. Its adjusted operating loss in the same quarter: $10.9 billion. Bigger than the revenue. The loss was $10.1 billion the quarter before. Sales are compounding. So is the burn. Everyone argues about whether AI is a bubble. The weirder truth is the biggest AI company on earth is growing 50% a quarter and still can't pay its own bill. At this pace, how many quarters until the math has to work?

722 followers. No team, no funding, just me and a laptop shipping in public. 🚀 Looking to connect with: Solo founders building without a team Indie hackers shipping their first product AI builders posting real progress, not hype screenshots People who actually reply What are you shipping this week? 👇 Let's grow together. 🤝 #buildinpublic

720 followers. Most of my posts get under 200 views. Still posting every day. 🚀 No viral moment yet. Just me, building in public and writing down what actually happens. Looking to connect with: Small-account builders shipping anyway AI folks posting real results, not hype screenshots Indie hackers on day 30 of something unsexy People who reply instead of just scrolling How many followers did you have when you shipped your first real thing? 👇 Let's grow together. 🤝 #buildinpublic

Cursor's own benchmark has a new leader. Gemini 4 Carbon scored 60.4% at high effort, ahead of Claude Opus 4.6 at 57.1% and GPT-5.3-Codex at 55.4%. And it's cheaper than the model it beat: $9.22 a run against Opus at $11.00. Better and cheaper is a hard combo to argue with. I keep a simple rule for coding models now: the crown changes hands every few weeks, and loyalty to one model is a tax. You re-test or you overpay. Run your own top 3 prompts on all three before you pick. The leaderboard only tells you what won last week.