Daily TEA – Trust Is For Sale, Tokens Are the New Defense Budget, The Flat-Fee Era Is Over
6M fake GitHub stars exposed, AI security is now a computational arms race, Anthropic ends flat fees, quantum meets AI, and multi-agents beat senior GPU engineers
Hello, dear TEA-mates! Here is what you need to know today.
1. ⭐ 6 Million GitHub Stars Were Fake, and VCs Funded Some of Those Repos
A Carnegie Mellon study found approximately 6 million suspected fake stars across 18,617 GitHub repositories, sold openly for $0.03 to $0.85 each. The perverse logic: VCs use star counts as funding signals (median seed-stage: 2,850 stars), so a $285 purchase can manufacture credibility worth millions. Detection is possible. Fake campaigns leave fingerprints: zero-follower accounts, near-zero fork-to-star ratios, and sudden spike patterns. One 157,000-star repo had only 17 forks per 1,000 stars. Flask has 235. The FTC’s 2024 rule bans fake social metrics at $53K per violation. (Read More)
🫖 TEA For Thought: “100% — wherever there are people and competition, there will be exploitation.”
2. 🔐 AI Security Is Now a Token Budget War
The UK AI Safety Institute confirmed it: spending more tokens on defense yields better vulnerability detection. The formula is simple. To harden a system, you need to spend more tokens discovering exploits than attackers will spend attacking it. This creates an interesting inversion for open source. Security investments on shared codebases benefit all users simultaneously, making collaborative defense more cost-efficient than individual reimplementation. AI-driven security is a computational arms race where resources determine defensive strength. (Read More)
🫖 TEA For Thought: “This is proof of work. With tokens, backed by real money. The best defense can always be discovered. That is the ‘crypto’ of cybersecurity.”
3. 💸 Anthropic Ends the Flat-Fee Era
Anthropic restructured enterprise pricing. Seat fees now cover platform access only. Claude, Claude Code, and Cowork usage is billed separately at standard API rates. The trigger: GPU rental costs climbed 48%, CoreWeave raised rates 20%, and API uptime dropped to 98.95% over 90 days (well below the 99.99% cloud standard). OpenAI, GitHub, and Windsurf have already made the same move. The subsidized-access era for heavy users is over industry-wide. (Read More)
🫖 TEA For Thought: “The seat charge era might be over soon. Folks will be charged for the tokens they actually consume. More people will direct their work toward open-source, locally deployed models. Get some hardware ready before prices go even further up.”
4. ⚛️ NVIDIA Just Built the AI Brain for Quantum Computers
NVIDIA launched Ising, the world’s first family of open-source quantum AI models. The family has two components. Ising Calibration is a vision language model that automates quantum processor calibration, cutting setup time from days to hours. Ising Decoding is a neural network optimized for quantum error correction, delivering 2.5x faster performance and 3x higher accuracy than pyMatching, the current industry standard. Harvard, Fermilab, Lawrence Berkeley Lab, and IQM Quantum Computers are already adopting it. Jensen Huang framed it plainly: “AI becomes the control plane — the operating system of quantum machines.” The quantum sector is projected to exceed $11 billion by 2030. (Read More)
🫖 TEA For Thought: “Quantum plus AI = unstoppable.”
5. 🤖 One Markdown File. 235 GPU Kernels. 38% Faster.
Cursor ran a multi-agent system against 235 CUDA kernel optimization problems for NVIDIA’s Blackwell B200 GPUs. A single planner agent coordinated autonomous workers. The only instruction: one markdown file defining the output format, rules, and test criteria. Result: 63% of kernels outperformed the baseline, with a 1.38x geomean speedup. 19% exceeded 2x improvement. Work that traditionally takes kernel engineers months or years. The key insight: clear objective definition, not complex orchestration. (Read More)
🫖 TEA For Thought: “This is amazing. If you can define the ideal state as accurately as possible, AI agents can figure out the best way to achieve it.”
🛠️ Tools of the Day
NousResearch/hermes-agent. “The agent that grows with you.” Adaptive agent framework that builds on prior context across sessions. +53K stars this week.
HKUDS/DeepTutor. Agent-native personalized learning assistant. Adapts to the learner over time instead of serving static content. +5.5K stars this week.
HKUDS/AI-Trader. 100% fully-automated agent-native trading. Multi-agent team handling research, analysis, and execution. +800 stars this week.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more:
Twitter/X