10 karma · joined December 11, 2025
What surprised me is that the public alpha18.1 model seems to have almost no visible limits at all compared with the usual chat products. That made me wonder what is actually different under the hood, because if this is already what alpha18.1 does, then what exactly is alpha18.5 Plus supposed to add on top of that? LOL
- Visual cards with titles and thumbnails
- Playlists keep research focused
- One private URL for editing, share links for viewing
- Fast import for friends and teams
The real open questions for a GPT 6 level system seem to be reliability, long term memory, tool autonomy, and knowing when it is wrong.
Curious what people here think is still fundamentally broken in current LLMs that a next generation model would need to address.
After I finally measured what I was shipping, I realized most of the client code was not adding value. It was just making hydration heavier, breaking accessibility, and hiding simple HTML solutions.
The post walks through the process I used to cut around 80 percent of the JS without going full “no JS”: listing real interactions in human language, leaning on native elements like details and dialog, using bundle analyzers, and deleting dependency creep. It also shows the small performance and accessibility checklist I use now.