←
Scott14,405 posts
Scott avatar

Scott

@scottstts

From the infinite potential of energy to the total actualization of entropy, intelligence charts a course for the pursuit of meaning, mission and love.

◷ Joined September 2023▧ 2,312 Photos and videos
Scott avatar
Scott@scottstts·Nov 13

My new favorite, Astral Realm, made with Fable 5.2, rendered in @threejs 2hrs, 2M tokens, a LOT of iterations, everything besides the character is procedural. Even I was surprised how well it turned out

2:00
Scott avatar
Scott@scottstts·34m

Classic😂 OpenAI models are just too RL fried, i think the capabilities are there, just really hard to get them out

Didn't want to judge GPT Astra 6.2 and Sol 6.2 too early but now I will after testing them: Both Astra 6.2 and Sol 6.2 are lazy models. They do not follow instructions properly and take the easiest route (against instructions) that no other model would take. I've been building 3D things with Astra 6, Fable 5.2 and Opus 5.5. They ALL created the result either using Blender or entirely using three.js. But the GPT 6.2 models SOMEHOW, even though I CLEARLY said "USE ONLY THREEJS and ONLY CODE no assets", fucking took my reference images and put a wavy fisheye effect on the part of the image where the water is to make it look like real water. And then zoom and pan around in the images as if it was showing what it really built. In one aspect it is NOT lazy: it puts random asf slop UI around the 3D thing. I told it to render a video and give that video to me in the end. Only gave me the animated images stitched together playable on a website, not a video. The model also LOVES to cheat when building 3D demos with image generation. It puts them everywhere. I'd stick to Claude for now. This shit's not usable. Maybe for some more engineering focused users cuz I haven't tested that area yet but for my work which is mostly involving UI, 3D and taste this is NOT IT. Maybe computer use is cool with it but for me it is unusable. Wanna repeat for ME it's not good. You might have better experiences. I'll run some tests again really trying to avoid letting it cheat and post the results later if I get any good results that are worth posting. (below are some of the cheating outputs)

Scott avatar
Scott@scottstts·52m

Just a little toe dipping on plain 3d modeling with astra 6.2. I'd say it's really good, hard to say if it's better than fable 5.2, but if it keeps iterating on this i think it's promising this took just over an hour using a full 5hr limit on a plus plan

0:25
Scott avatar
Scott@scottstts·1h

I can still smell the RL fried brain from Astra 6.2 frequently ie if you give it a metric of work quality it will warp everything else to satisfy that one metric and completely forget what the task was trying to do in the first place Just feels they have done too much RL to it that it’s collapsed to a single will to meet a reward function, and nothing else matters to the model

Post attachment
Scott avatar
Scott@scottstts·1h

Okay I’m seeing great results from Astra 6.2, I’m gonna give it a real shot

Scott avatar
Scott@scottstts·2h

Interesting, I always trust his test more than normal benchmarks, and it appears Astra 6.2 is actually performing really well

I tested GPT-6.2 Astra on real work. 2 repos, 105 replanted bugs. Find and fix what you can. It's the new #1. The results: - GPT-6.2 Astra (max): 62.7 for $28.35 - GPT-6.2 Astra (xhigh): 60 for $23.53 - Fable 5.2 (max): 58 for $139.65 - Gemini 4 Carbon (high): 57.3 for $31.32 - Sonnet 5.5 (max): 51.3 for $154.03 - GPT-6 Astra (max): 45 for $33.03 From GPT-6 Astra to GPT-6.2 Astra: +17.7 points at max, +17 at xhigh. Whatever people say about other use cases, it's a code review beast. GPT-6.2 Sol and more effort levels drop soon in the thread 🧵

Post attachment
Scott avatar
Scott@scottstts·2h

Nvidia board too stupid to realize Jensen is Nvidia’s moat

TSMC now wants the cash up front. With so much money moving in circles between chipmakers, cloud providers and AI labs, you can see why. "Regarding orders, orders exceeding the credit limit are paid in cash, but some customers require upfront payment to secure designated production capacity." news.mydrivers.com/1/1166/11…

Post attachment
Scott avatar
Scott@scottstts·2h

lol I didn’t think OpenAI models could be even more annoying than Astra 6, apparently they did it

Okay I give up. GPT-6.2 Astra is complete garbage. It simply does whatever the fuck it wants, regardless of what you explicitly tell it to do. The prompt asked it to BUILD a 3D T-Rex. It downloaded an existing model and called it a day. I’ve NEVER experienced this level of instruction-following failure with any other model.

Post attachment
0:10
Scott avatar
Scott@scottstts·15h

Wow this is pretty amazing, i found a random drawing app online, asked Astra 6.2 to use computer use to draw Jensen this is 30x speedup This is probably the best computer use model right now ngl

Post attachment
0:30
Scott avatar
Scott@scottstts·17h

I’m just gonna say it, Astra 6.2 is worse than Astra 6 on 3d

Scott avatar
Scott@scottstts·17h

From what I can tell it’s not that much better on 3d and threejs either, even bit of regression

OpenAI wanted to sabotage Anthropic's IPO, rushed the release of GPT-6.2 Astra, and ended up benefiting its biggest competitor even more. GPT-6.2 Astra is SOTA in Three.js, Blender, and 3D. You'll see some absolutely incredible demos of that, but it has serious alignment and safety issues, and starts speaking an alien language during long workflows. What had everything it needed to be a wonderful launch turned into a failure because of the impatience of some executives.

Scott avatar
Scott@scottstts·17h

This is Astra 6.2 on jelly test, exact same prompt haven't done this for a while now, but since the entire jelly fever started with this jelly demo Astra 6 made, feels only right to test it again gotta say i'm disappointed, it looks worse

0:26

Asked Astra to make a delicious and bouncy jelly in @threejs and webgpu

0:44
Scott avatar
Scott@scottstts·18h

OpenAI models since gpt 6 have been RL-fried, this is like the exact symptom, regression to the mean

I was saving this for my video but I'll spoil it now. Ben and I both made new levels in games using 6.2 Astra during the early testing. He did Majora's Mask, I did Outer Wilds. We both prompted ourselves, but it was along the lines of "make a creative new level for [game]" We both got water levels with music themes, specifically around bells. Both were named "Bellwether". This model regresses to the mean like nothing I've ever seen. Go ask it to make a new level for a game you love, lmk if it chooses "bellwether" as the name.

Scott avatar
Scott@scottstts·18h

So all things considered, opus 5.5 is still THE GOAT?