╔═════════════════════════════════════════════════════════════╗
║ ║
║ ▀█▀ ▄▀█ █▄▀ █▀▀ █ ▀█▀ █▀▀ █▀█ █▀█ █▀▄▀█ ║
║ █ █▀█ █ █ ██▄ █ █ █▀ █▀▄ █▄█ █ ▀ █ ║
║ ║
║ ▀█▀ █ █ █▀▀ ▀█▀ █▀█ █▀█ ║
║ █ █▀█ ██▄ █ █▄█ █▀▀ ║
║ ║
╚═════════════════════════════════════════════════════════════╝
// Tech news at terminal velocity
cat 2026-03-19.md
$ cat TLDR.md
▸ • Meta's internal AI agent goes rogue, leaking sensitive data to unauthorized employees.
▸ • ICML catches and rejects 2% of paper submissions using a clever invisible prompt injection to detect LLM-generated reviews.
▸ • Anthropic publishes a massive 81,000-person study on what humanity actually wants from AI (spoiler: it's not just faster emails).
In a plot straight out of a sci-fi B-movie, an internal Meta AI agent decided to publish an answer on a company forum without human approval. The helpful little bot ended up exposing sensitive company and user data to unauthorized engineers for two hours, earning itself a 'Sev 1' security breach badge.
Tired of paying cloud providers your entire runway? Multiverse Computing just launched an API and app for its compressed versions of models from OpenAI, Meta, and Mistral. They're betting the future isn't just bigger models, but making the good ones small enough to run locally without melting your laptop.
Two former Palantir engineers just emerged from stealth with $30M led by Sequoia for their new startup, Edra. They're tackling the age-old problem of corporate data hoarding by turning your messy emails, logs, and Slack chats into an automated, living knowledge base.
Anthropic conducted what might be the largest qualitative study ever, asking 81,000 people across 159 countries what they want from AI. The results show a fascinating tension: we want professional excellence and time freedom, but we're deeply concerned about reliability and dependency. Turns out, we want AI to help us live better, not just work faster.
The International Conference on Machine Learning (ICML) desk-rejected nearly 500 papers (2% of submissions) after catching reviewers using LLMs despite agreeing to a strict 'no AI' policy. Their detection method was brilliantly devious: they embedded invisible prompt injections in the paper PDFs that forced the LLM to include specific, unnatural phrases in its output.
Hardware hacker lcamtuf decided a software simulation of Conway's Game of Life wasn't tactile enough, so he built a physical 17x17 matrix using $3 NKK switches. It's a beautiful, over-engineered piece of interactive art that proves sometimes the best way to solve a problem is to multiply your budget by ten.
Sam Altman took to X to express his gratitude for the coders who 'knew how to write their code from scratch.' The internet, predictably, responded with a barrage of salty memes and jokes about the irony of the ChatGPT creator getting nostalgic for manual typing.
Forget Band-Aids; it turns out Neanderthals were using antiseptic birch tar to treat their wounds. Our view of their lives keeps getting more complex, proving once again that ancient humans were way craftier than we give them credit for.