Been tinkering with a SaaS side project for a few months and the gap between what I can ship now versus a year ago is genuinely strange. Not in a purely good way either. The tools are good enough that I can move fast through parts of the stack I barely understand. Which works until it doesn't, and when it breaks I'm staring at code I didn't fully write trying to debug something I can't fully reason about. That's a new kind of stuck that feels different from the old kind. What keeps nagging a...
We spend a lot of time judging AI. So I asked five leading AI systems to judge us instead. “Humans operate an all-in-one general intelligence, high-definition computer vision system, and hyper-agile robotic chassis on the power of a ham sandwich.” ~Gemini
I spend way too much time looking at AI leaderboards, and this one caught me off guard. Maya-2-Native from Maya Research is currently ranked first for real-time Hindi on Voice Arena. Considering the leaderboard is built from blind listener preferences rather than curated demos it's an interesting result that seems to have gone largely unnoticed.
Maybe they have seen AI fail for so long due to poor performance that it makes AI safety experts seem like they are vastly overestimating what AI can do?
Comments URL: https://news.ycombinator.com/item?id=49111613
Points: 2
# Comments: 3
I spent about three weeks trying to get text-to-video to produce consistent characters across scenes for a short project. Different prompts, different models, different seed tricks. Every generation gave me a slightly different face, different body proportions, different everything. I kept thinking I was just prompting wrong. Turns out I was using the wrong tool entirely. Someone in a Discord server finally pointed out that text-to-video and image-to-video solve completely different problems. ...
