Welp… Even Fable 5.2 itself hates its own work the most.
LLM drawings, same prompt:
1. Fable 5.2 xhigh (57m, $21.77)
2. GPT 6.2 Astra xhigh (18m, $2.88)
3. GPT 6.2 Astra max, slightly guided
Surprisingly, when asked to pick which one is better, Opus, Sol, Neon all consistently prefer 2 over 1. Personally i like 1 much more. Is there an obvious gap between LLM taste and human taste?
1
2
179
LLM drawings, same prompt:
1. Fable 5.2 xhigh (57m, $21.77)
2. GPT 6.2 Astra xhigh (18m, $2.88)
3. GPT 6.2 Astra max, slightly guided
Surprisingly, when asked to pick which one is better, Opus, Sol, Neon all consistently prefer 2 over 1. Personally i like 1 much more. Is there an obvious gap between LLM taste and human taste?
people shit so hard on 6.2 Astra naming "Bell", but honestly i got "Bell" from Fable 5.2 as well
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.
1
143
source code of my recent LLM-generated pixel arts
github.com/dkenw/arts/
1
27
1,546
Made this with Opus 5.5.
Not exactly a game, but more of a "visual and sound experience" of a peaceful night ride.
🔗 passing-light.modmod.tv/
*Best with headphone for spatial audio 🎧*
It's only 3 min. Hope you find it relaxing.
1
4
401
Made with @kitao's Pyxel, which is actually a retro game engine. In 1200 lines of python.
gist.github.com/dkenw/b0f4aa…
1
3
234
such a big gap between opus 5.5 xhigh and high
Introducing FrontierCode 2.0: our new coding benchmark that measures a model’s ability to go beyond the spec.
Most coding evals measure a model’s ability to write code that meets a spec. With FrontierCode 1.0, we asked a new question: would you actually merge the PR?
FrontierCode 2.0 is comprised of more diverse tasks, featuring larger scopes and higher mergeability standards. We are now asking: not only would you merge this PR, but should you?
One key feature of FrontierCode 2.0 is underspecified prompts. Rather than including every engineering decision, prompts are deliberately ambiguous, giving the models room to demonstrate their judgment.
The result is a benchmark that is far from saturated, with the top model scoring just 45%.


















