←
Justin.md8,291 posts
Justin.md avatar

Justin.md

@JustinPerea

Mathematical Statistician. Wesleyan MA Astronomy. BS Astrophysics. Building @mitamaton. Starting @goodmachinesai

◷ Joined February 2009▧ 674 Photos and videos
Justin.md avatar
Justin.md@JustinPerea·2h

Eclipse instead of just the Moon for a light and dark mode toggle.

0:14
Justin.md avatar
Justin.md@JustinPerea·Nov 17

Final pass. Going to see if using Astra 6.2 in Hermes improves the results.

GPT-6.2 Astra's Chongqing, pass four. We had it rebuild Raffles City alone, shape first in grey clay, then put it back in the city. The towers curve now. The city behind them is still too uniform. Four passes took 3 hours 43 minutes. @OpenAI

Post attachmentPost attachmentPost attachmentPost attachment
Justin.md avatar
Justin.md@JustinPerea·Nov 17

Okay after letting it run for 4 hours I have not been impressed. Looks like I will continue to stay on the $100 plan for now.

OpenAI@OpenAINov 17

Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…

1:59
Justin.md avatar
Justin.md@JustinPerea·Nov 17

Pass 2 is a slight improvement. Running pass 3 now. Astra 6.2

We asked GPT-6.2 Astra to critique its own Chongqing and fix what it found. First pass and second pass, same cameras. The second pass took 61 minutes. Hongya Cave changed the most. Raffles City got new facades, but the towers still don't curve like the real ones. @OpenAI

Post attachmentPost attachmentPost attachmentPost attachment
Justin.md avatar
Justin.md@JustinPerea·Nov 17

First test with Astra 6.2 using Blender. Going to ask it to do another pass because it's kind of lackluster...

We asked GPT-6.2 Astra to build Chongqing in Blender and make every choice itself. This is its first pass, untouched. It took 38 minutes. The bridge and Hongya Cave came out well, Raffles City didn't. Next we ask it to critique its own renders and fix them. @OpenAI

Post attachmentPost attachmentPost attachment
0:10
Good Machines avatar
⟳ Justin.md retweeted
Good Machines@goodmachinesai·Nov 17

We read a paper that counted screen-reader labels in data used to train phone-operating AI agents. In AndroidControl (2024), about 43% of the on-screen elements that need a label have none, by the authors' rule. Inputs are worst: 27% labeled. It tested no agent. Paper below.

1:43
Justin.md avatar
Justin.md@JustinPerea·Nov 17

Another banger by @PaulJun_

As promised, here is part two to "Brand as Software" — a practical look into how we're building it at @tryramp. I'm excited for this to become the future of brand and creative teams. Please enjoy!

Justin.md avatar
Justin.md@JustinPerea·Nov 17

Playing around with showing n-dimensions in a node graph.

0:14
Justin.md avatar
Justin.md@JustinPerea·Nov 17

Trying to figure out how to blow up a component into its atoms in 3D. I feel like I'm getting somewhere.

Post attachment
Justin.md avatar
Justin.md@JustinPerea·Nov 16

It's been driving me crazy how the chat bar is like 15 pixels wider the new nav bar in Claude mobile app

Post attachment