Real talk: No matter how good the AI models get at benchmarks, I still feel despite having months of context on my work, no model can draft me a good resume. I think for models in general, persuasive writing and crafting tasteful arguments is severely lacking.
There seems to be a disconnect between tests and reality.
@finkd I built tamserve for my startup. I am open sourcing the code now. @Meta SAM 2 model families are the best at tracking objects but they weren’t built for real-time and real-life use cases.
Tamserve through it’s serving engine runs them on live camera feeds 2x cheaper and up to 3.8x faster, same accuracy.
One copy of the model serves every camera, and memory stays flat, so it runs indefinitely. Meta's code uses 35x more memory after 30 seconds of video and crashes within minutes.
On one RTX 4090:
5 fps: 54 cameras vs 22
30 fps: 9 cameras vs 4
https://t.co/jLmPCmrAVF
AI website af but we wanted to diffuse the message…
We didn’t need fancy websites (a quick realization) because we weren’t selling to tech bros. We knew our language had to direct and try and touch the hearts of families and caregivers alike, working in the nursing space.
A cyber-minimalist ($500) website was money better deployed for other things.
Memoria was our attempt to notice.
A camera in the home or care facility, with AI running on the device itself, that learns a person's normal routine and alerts family or caregivers when it shifts.
Next: why we believed this could work.
Earlier this year, I co-founded a startup called Memoria. You've probably never heard of it.
I recently walked away from it, and I want to write down what I learned. First, what it was.
When an older person's health declines, the first signs are small. They walk slower. They skip meals. They sleep badly.
Nobody notices, because nobody is watching day to day. By the time a doctor sees it, it's often years late.
I think I’ve hit peak unemployment: I built a living, interactive version of Times Square out of ASCII characters.
Starting with a photograph, I built a scene where billboards animate, people walk and sit, traffic flows, and the sky changes with New York’s clock. Crowd density shifts throughout the day, and weather follows the seasons.
A few of the techniques behind it:
• 𝗣𝗲𝗿𝘀𝗽𝗲𝗰𝘁𝗶𝘃𝗲 𝗺𝗮𝗽𝗽𝗶𝗻𝗴 to fit animated ads onto the buildings’ billboards.
• 𝗣𝗿𝗼𝗰𝗲𝗱𝘂𝗿𝗮𝗹 𝗰𝗵𝗮𝗿𝗮𝗰𝘁𝗲𝗿 𝗮𝗻𝗶𝗺𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗰𝗿𝗼𝘄𝗱 𝘀𝗶𝗺𝘂𝗹𝗮𝘁𝗶𝗼𝗻 to give pedestrians articulated movement and personal space.
• 𝗧𝗶𝗺𝗲-𝗱𝗿𝗶𝘃𝗲𝗻 𝗹𝗶𝗴𝗵𝘁𝗶𝗻𝗴 𝗮𝗻𝗱 𝘄𝗲𝗮𝘁𝗵𝗲𝗿 to bring in sunsets, illuminated windows, rain, and snow.
• 𝗖𝗼𝗹𝗼𝗿 𝗾𝘂𝗮𝗻𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗱𝗶𝘁𝗵𝗲𝗿𝗶𝗻𝗴 to translate every frame into colored characters while preserving detail and atmosphere.
You can pause, rewind, change the date, or go full screen and watch it unfold.
This is No. 01 in 𝗔𝗦𝗖𝗜𝗜, a series I’m building. A glimpse in the video below; the interactive version is here:
Website: https://t.co/Keq7RXcmmO
Code: https://t.co/eDklkzDRKl
If you find it cool, give it a like on the site, share it, star the repo.
What scene should I build next?
Best viewed on a laptop + chrome browser
Two Claude Code skills I won't ship without:
1 Superpowers: planning, review, testing and debugging baked into every session https://t.co/TorfnGjomm
2 Vibe Security: one /vibe-security and it audits the whole app for the holes AI assistants love to leave behind. Leaked keys, client-side only validation, sloppy Supabase rules. https://t.co/uG6xfDq1qE
Vibe coding is fun. Vibe coding vulnerabilities into prod is not.
Hi, I am a fullstack engineer, AI enthusiast, and tinkerer who hopes to bring my followers along on my days in the tech industry, learning about papers, sharing new techniques, building cool stuff, and maybe dropping some advice from the experience I keep collecting day to day.
Excited to kickstart the journey into the pit of social media existence that is X 😅