Luna Tech

Everyone is racing to build an AI tutor. Paradigm built something stranger. An AI that watches you learn and jumps in before you even ask. Meet Clover the first ambient learning agent. Most AI sits there waiting for the perfect prompt. Clover watches your canvas. You stall on a slide, cross out a line, freeze mid-problem, and the hint is already there. I dropped in a lecture slide I was stuck on. It zoomed into the exact step that was breaking, traced it one move at a time, and circled where it fell apart. No prompt. It just noticed. This is the part nobody is ready for. If a school like this handed you a certificate one day, would you recognize it? What would you need to verify before you did? That question is coming faster than people think. paradigm.study @ParadigmStudy

You know that folder of raw footage you have been avoiding? Hours of clips, no labels, and the one shot you need is buried somewhere in the middle. filmora.go.link/26rUE Filmora 16 goes straight after that problem. The feature that stands out to me is AI Media Analysis. It auto-tags and describes your footage, so instead of scrubbing through hours of clips, you search for the shot you need and you can find the clip faster. For anyone sitting on a messy folder of raw footage, that alone saves real time and you stay fully in control of the edit. That is the whole pitch of Filmora 16: level up your editing without the added complexity. If you edit video, it is worth a look. #Filmora16 #Filmora #Wondershare

The Nvidia board just opened an independent investigation into its own CEO.

Everyone is going to run with the spy and romance headline. For an AI audience, the part underneath it matters more. A US intelligence memo reportedly alleged a Megaspeed executive was recruited by China to get close to Nvidia’s CEO. Nvidia says the interactions were nothing unusual and that it does not share classified information with customers. None of the memo’s claims are proven. The more concrete thread is the chip trail. Reporters visited the listed address of New Matrix, a firm cleared by US Commerce to buy advanced Nvidia chips, and found “a hollow gray shell” raising the question of whether it fronts for a blacklisted buyer. Strip the drama and one thing stands out: the most important company in AI now sits in the middle of a US China intelligence fight. That pressure is not going away.

Everyone is about to post “OpenAI hits $1.25 TRILLION.” Read the second paragraph. OpenAI reportedly wanted $30 billion in this round. So far it has closed $11 billion, with the full round maybe reaching $20 billion. That is a flagship lab landing well under its own target. Same news cycle, Bloomberg reported Anthropic’s IPO valuation expectations sliding from around $2 trillion to as low as $1.1 trillion. Two of the biggest names in AI, both getting marked down in the same week. After a wave of backlash against data-center expansion, the money is starting to ask harder questions. The models keep getting better. The checkbook is getting more careful.

🚨BREAKING: A folder of engineering discipline for AI agents just crossed one million installs. Not a model. Not an app. A set of rules one developer made public. It is called Skills for Real Engineers. Here is the failure everyone with a coding agent knows cold. You describe what you want, the agent runs off confidently, and hands back something that misread you entirely. Now you are debugging its misunderstanding instead of shipping. The usual escapes are worse: vibe-code and pray, or bolt on a heavyweight framework that seizes your whole process and buries its own bugs. Matt Pocock, the TypeScript educator, took his actual daily-driver agent folder and opened it up. Then a million people installed it. Here is what it actually is: → His real .agents directory the sharp skills he uses to do real engineering, not vibe coding published as is → Each skill is one file aimed at a specific failure mode: agent misalignment, needless verbosity, architectural rot → Real engineering motions, not toys: TDD workflows, a bug-diagnosis skill, breaking work into vertical slices, reviewing a diff against both your standards and the original spec → Small and composable, so you assemble the few you need instead of surrendering your process to a framework like GSD, BMAD, or Spec-Kit → Model-agnostic: runs in Claude Code, Codex, Copilot, or whatever agent you use → Two ways in: install the managed plugin and subscribe to his updates, or copy editable files into your repo and make them your own Most skill packs add capabilities browser control, API access, more powers. This one adds discipline. It makes the agent follow your engineering standards instead of inventing its own. One honest flag: this is real infrastructure now, wired into production PR-review pipelines, and that scale exposed the usual rough edges teams building automated workflows on top of it have hit reliability dips when a step flakes. Treat the skills as sharp tools to compose, not a set-and-forget machine. The people whose agents actually build the right thing did not write longer prompts. They installed the guardrails a real engineer already battle-tested, the same ones a million other builders just did. MIT licensed. 314k stars. 1M+ installs. 100% Open Source. (Link in the comments)

NVIDIA just open-sourced a fix for the most expensive lie in enterprise AI: the confident, wrong answer. The catch is that they shipped it without a single number proving it works. It's called Auto Ontology. Here's the problem it targets, and it's real. A company asks its shiny new AI agent, "how many active customers do we have?" The agent writes SQL against a database with a thousand tables, cryptic column names, and no documentation of what "active" even means. It returns a number. The number is wrong. And nobody catches it, because a wrong answer delivered fast and formatted nicely looks exactly like a right one. NVIDIA's own README says it plainly: a revenue metric answered faster is worth nothing if the number is wrong. Auto Ontology sits between the agent and the chaos. It reads your actual databases and automatically builds a mapping layer, a translation between the words your business uses ("active customer," "revenue," "churn") and the real tables and columns underneath. Then humans certify that mapping. Then the agent answers questions against the certified definitions instead of guessing. Generate, certify, serve. It exposes the whole thing over MCP, a CLI, and an API, so any agent harness can plug in. That is a genuinely good idea, and a boring one, which is usually the sign it matters. The semantic layer is the unglamorous plumbing that decides whether enterprise AI is trustworthy or just fast. Now read the fine print, because this is where it gets honest. NVIDIA ships zero published accuracy scores. The companion evaluation repo contains a benchmark harness, the machinery to test it on datasets like BIRD, but no results. They built the scoreboard and left it blank. The headline figure everyone's quoting, a 90% reduction in ontology build time, is NVIDIA's own internal claim, not an independent measurement. The clearest summary came from an outside writeup: it's an open, MCP-ready context layer that any agent can query, but you must run the benchmarks yourself to trust it. For a tool whose entire pitch is "stop trusting unverified answers," shipping it with unverified performance is a strange look. And it is not a weekend download. An independent audit cloned it and found roughly 129,000 lines of Python, two FastAPI services, a Postgres database with vector search, an optional secrets vault, and a default dependence on your own NVIDIA API key feeding their hosted model endpoints. The repo isn't accepting outside contributions, and auditors flagged a plaintext-credential default worth checking before you deploy. This is enterprise infrastructure that happens to be open source, not a tidy library you bolt on in an afternoon. So here's the balanced read, both halves true. NVIDIA correctly identified the thing that actually breaks enterprise AI agents, that they hallucinate against schemas they don't understand, and open-sourced a serious, Apache-licensed attempt to fix it. They also shipped it as a heavy stack, wired to their own paid model endpoints, with no proof of accuracy and the benchmarks left as an exercise for you. The idea is right. The gap between the idea and the evidence is the whole story. A tool built to stop agents from returning numbers nobody verified arrived with a performance claim nobody can verify yet. If you deploy it, do the one thing the tool itself is preaching. Run the benchmark before you believe the number.