Wanaka Studio will build you a AAA-style 3D game out of a prompt, an idea worth considering the next time Rockstar delay a GTA launch. Describe what you want and an agent builds it while you work alongside it in a synced editor, asking for slower enemies or dragging the scene about yourself, and the characters and environments it generates come rigged, so they move rather than standing there like a shop dummy. Multiplayer and one-click publish are in the box. The team came out of ByteDance Seed and miHoYo, and their line is that LLMs won't replace professional software, so professional software has to be rebuilt around them.

Hindsight is a set of small models for the questions that aren't really writing, the yes or no, the pick one of these five, the how likely is this, all of which currently go to a trillion-parameter model because there was nowhere else to send them. hs-decide-1-flash hands back a typed answer in 167 milliseconds, quicker than the fastest frontier models manage a single token, and the reranking and embedding models run on a CPU with no GPU anywhere near them, from half a cent per million tokens, reached by pointing your existing OpenAI client at their endpoint and changing nothing else.

Eyecon watches you to check whether you're still watching, and pauses the moment you aren't. That covers the person who wanders in to ask you something, the thing you turned round to pick up, and the episode you slept through, all of which normally cost you four minutes of programme and another one dragging the scrubber about trying to find your place again. The only thing the front camera is looking for is whether your eyes are pointed at the screen, and it works that out on the Mac rather than anywhere else.