Brucky (@danikkk_wqs)

The easiest way into a top project 4.3 million AI repositories on GitHub. 230 new ones show up every minute. PyG’s CONTRIBUTING.md says it outright: if a pull request isn’t merging because of size, complexity, or reviewer availability, there’s a contrib folder with looser review. that’s not a hack. that’s the official path. HKUDS/LightRAG (builds knowledge graphs from text) has 38,800 stars and hundreds of contributors fighting over every issue. their own RAG-Anything (same tech, but for images and tables) from the same lab has almost nobody. the pattern repeats across the whole ecosystem. semantic-router is the neighboring project inside vLLM with the same logic. torch_geometric.contrib is the same looser-review lane, just inside PyG. while everyone’s knocking on the front door, the one next to it is wide open

97% OF PERMISSION PROMPTS IN CLAUDE CODE GET APPROVED WITHOUT BEING READ a coworker sent me the number with no comment. that was enough to make it unsettling. anthropic admitted this on august 7, right when they announced it. but that’s not even the interesting part. they ran a controlled study with 1,053 paid testers. they planted a dangerous command inside a task. humans caught it 13.6% of the time. the auto mode classifier caught it 89% of the time. starting august 14, auto mode becomes the default for new sessions on pro, max, and team plans. it only blocks actions that are irreversible, destructive, or reach outside your environment. everything else goes through without stopping. not because it’s faster. because it turned out to be safer than a human clicking “approve” on autopilot. so you were clicking “yes” without thinking, and now a model does the thinking instead. and it’s better at it.

You're the bottleneck in your own system. People actually making money with AI in 2026 don't prompt manually anymore. They build loops - systems that plan, execute, and verify themselves. Example: One voice note in five finished posts out, for Instagram, Telegram, YouTube, and LinkedIn. No human involved between the start and the result. You write the goal once. The AI decides the next step, scores it, and fixes what's weak. Skip the verify step and the loopт "finishes" - but the output is mediocre, because nothing ever checked it. Skip the stop condition and it either runs forever, or you kill it by hand. Your first version will be rough. The point is to ship one that actually works. Fixing prompts, or building a system?

The best AI safety grade - a C+. Not from an underdog. From the leader. The Future of Life Institute rated major AI labs on risk management, transparency, and whether they keep their own promises. Anthropic - C+. OpenAI and Google DeepMind - C. Meta - D+. xAI, DeepSeek, and Mistral failed the assessment. Even the leader barely scraped a C+. Meanwhile, several labs have quietly walked back their own earlier safety commitments. And all of this is happening while AI gets trusted with cybersecurity, medical reviews, and autonomous agents. If you’re trusting AI with decisions, don’t ask: “Which model is smarter ?” “Who checked its safety ?”

Your prompt engineering course is trash now. Karpathy said it straight at Sequoia Ascent: stop treating LLMs as oracles. They're statistical "ghosts" they need to be guided, controlled, and constrained by rigid boundaries. Before: send text, get text back. The model forgot context on step five - and nobody knew why. Now the agent writes the code, runs its own tests, reads the error and rewrites the fix - on its own, until the tests pass at 100%. Behind this are three technologies almost nobody talks about. Transactions that roll back when the agent makes a mistake. Memory that deletes outdated facts on its own. Routing that decides which agent a given step actually needs. The human reviews it and hits "Deploy." No prompt engineering course will explain any of that.