aivs.tv

AI Coding

Pelican learns to ride a bike with RL — GPT-6 Astra vs Kimi K3.1 Pro

0:00 / 0:35
Posted byYifei Zhou@YifeiZhou · on X
View post on X

1 likes63 views

Scoreboard2

  1. GPT-6 AstraOpenAI
    Author’s pick
    • 98/100 success; small policy plus direct parameter search
  2. Kimi K3.1 ProMoonshot AI
    • 29/100 success; linear PPO with curriculum learning

Prompt

Topic · Pelican riding a bicycle (SVG)

Pelican learns to ride a bike with RLThe post doesn’t quote the prompt wording.

Author’s verdict
By the poster’s own eval, GPT-6 Astra scored 98/100 and Kimi K3.1 Pro 29/100.

@YifeiZhou’s view (paraphrased)

GPT-6 Astra and Kimi K3.1 Pro build a pelican and bike in a Rapier physics sim, then train a policy to ride 20 m from a standstill, scored on 100 held-out seeds.

The post, itemised

Kind
Side-by-side
Task
Pelican learns to ride a bike with RL
Topic
Pelican riding a bicycle (SVG)
Mode
Not stated
Posted
Video length
35.5 s
Likes
1
Views
63
Bookmarks
0
Sponsored
No

NotesThe constraints ban training wheels, invisible props and editing world coordinates.

Cost & time, as stated in the postAs stated: both used about 960k training steps; success 98/100 (Astra) vs 29/100 (Kimi).

Related

Topic: Pelican riding a bicycle (SVG)GPT-6 AstraKimi K3.1 ProGPT-6 Astra vs Kimi K3.1 Pro

Posted by @YifeiZhou on X. aivs.tv claims no ownership: it hosts a copy for fast playback and links to the post on X. Request removal

More with these models