Today's Hallucination HQMeta Turns Its Employees Into a Free Dataset
Meta has built an internal tool that records employees' mouse movements and keystrokes to train its AI models. Your staff, apparently, are now also your R&D budget. The company insists this is about improving AI performance, which is reasonable — if you ignore the part where the people generating the data didn't sign up to be a labelling operation. Somewhere, a Meta engineer is typing very slowly and wondering if it counts.Source: TechCrunch
SpaceX Eyes $60 Billion AI Coding Tool, Just in Case
SpaceX has struck a peculiar arrangement to either acquire Cursor — a popular AI coding assistant — for $60 billion, or simply pay a $10 billion fee if the deal falls through. Which is a bit like saying you'll either buy the restaurant or leave a very generous tip. The move is designed to bolster xAI's developer tools ahead of a looming IPO for Musk's ever-expanding constellation of companies. The $10 billion consolation prize alone would be the largest "never mind" in startup history.Source: The Verge
AI Does Science Without the Sciencing
A new paper on arXiv finds that LLM-based research systems can produce scientific-looking results while completely ignoring the epistemic norms — the rigorous self-correcting principles — that make science actually trustworthy. In other words, the AI is submitting the essay without reading the brief. It generates conclusions confidently, it formats them beautifully, and it has absolutely no idea why any of it should be believed. Peer review is going to need a longer form.Source: ArXiv
YouTube Offers Celebrities a Deepfake Removal Button
YouTube is expanding its AI likeness detection tool to celebrities and their representatives, giving them a mechanism to find and remove deepfake videos. This is a genuine and sensible development — the internet has spent several years making fake versions of famous faces with cheerful abandon, and someone probably should do something about it. The tool won't catch everything, but it's a start. For now, verified humans can at least flag the versions of themselves they definitely didn't film.Source: TechCrunch
A Venti Iced Coffee and a Side of Existential Dread
The Verge tested Starbucks' new ChatGPT-powered ordering app, and it did not go well. What should have been a simple iced coffee order became a multi-step conversational ordeal — prompts, clarifications, and the sort of back-and-forth you'd expect when negotiating a trade agreement, not requesting a drink. Starbucks already has one of the most elaborately customisable menus on earth. Adding an AI layer doesn't simplify that; it just means your coffee now requires a conversation.Source: The Verge
The machines are learning. The coffee is getting cold. See you tomorrow.
|