ChatGPT with a huge context window still makes me re-explain the project on Monday.
I'd rather a smaller tool that keeps last week's notes than a bigger one that starts blank.
@suraj_sharma14 The catch is that most frontier API terms forbid training competing models on their outputs. So the cheap tiny model usually learns from an open-weights teacher, not the giant brain on the invoice.
@testingcatalog Same performance, 18% less per scan, measured on Microsoft's own internal MDASH suite. Good number. I'd just like to see someone outside MDASH run the same scans.
@sydneyrunkle The underrated bit is using previous runs to estimate timelines. An agent that says "this video takes about 20 minutes" is way easier to trust than one that just spins.
@rammcodes 47.3k GitHub stars for a skill whose tagline is "Less slop, designs pop." That's a lot of people admitting their AI-built UIs all look the same.
@chetaslua Nice touch that the crane clip says "Sequence shortened" right on screen. Google's own limits section admits objects still drift in longer sessions, so the short clip is the product for now.
Google just showed Genie 4. You type a world, walk around in it, and change it while it runs. Make it rain, move the camera, people talk back. With sound.
Catch: testers only, and it holds together for a few minutes before things drift.
Game prototyping is about to get weird.