I designed a 4D "pelican riding a bike" test: GPT-6 Astra vs Kimi K3.1 Pro I didn't expect the final result to look like this. The classic pelican-on-a-bike test is easy to read: 2D as SVG, 3D as an interactive scene. You can mostly tell how good the work is at a glance. On top of that, I designed a 4D version: the model has to build the pelican and the bike, then — in a physics sim, with reinforcement learning — get the pelican to actually learn to ride 20 meters from a standstill. To keep models from taking shortcuts, I added constraints: no training wheels, invisible props, or fake anti-gravity, and no editing world coordinates to pretend they can ride. What I wanted to test wasn't only whether a model can get the pelican moving — but how it models the system, picks algorithms, plans training, and whether it will quietly cheat to finish the task. This time I picked GPT-6 Astra and Kimi K3.1 Pro, gave them the same prompt, physics constraints, and eval protocol, and watched who could teach the pelican within a limited budget. The gap was pretty big: GPT-6 Astra: 98/100 Kimi K3.1 Pro: 29/100 Both models had to verify the physics first, train and freeze a policy, then evaluate on 100 held-out random seeds. Kimi K3.1 Pro Kimi verified the physics first: no pedaling → falls; pedaling drives forward through tire-ground friction. It built a six-rigid-body system in Rapier — frame, two wheels, steering fork, pelican torso, and crank — with joint constraints and a 2.4:1 pedaling drive to simulate real riding. The first ~180k steps were mostly exploration; after that it finally got its first 20-meter rides. At 480k steps, the best policy hit 3/8 on a development check. Training after that did not go well. Kimi kept tuning stationary-start training and exploration annealing, but results kept regressing; averaging multiple checkpoints didn't help either. Around 950k steps, Kimi froze its best policy and ran the formal eval on 100 new seeds. Final scores: Success rate: 29/100 Mean distance: 7.25 m Median distance: 0.89 m Mean successful ride time: 12.69 s Dev / train / validation: 960,032 steps Final eval: 17,380 steps Total budget used: 977,412 steps Kimi chose linear PPO with progressive curriculum learning, aiming to balance physical realism, training efficiency, and interpretability. Looking at the results, it underestimated how hard it is to hold balance from a standstill. It could occasionally ride the full 20 meters, but never really learned a stable start. It also spent some time fixing quaternion rendering and terminal-step counting, and testing browser training, policy import/export, and mobile compatibility. GPT-6 Astra Astra also verified the physics first: no pedaling → falls; pedaling pushes the bike through tire-ground friction. It likewise built a six-rigid-body Rapier model — frame, pelican fixed on the seat, steering fork, two wheels, and crank — keeping the necessary physics for steering, pedaling, and balance. The two models took different paths on training, though. Astra judged that the state inputs aren't that complex, and the controls are only steering and pedaling — so it optimized with a small policy plus direct parameter search. After the first 300k steps, Astra's best policy already hit 7/8 on a development check. Pushing on to 700k also saw some regression. It then changed the training setup and tightened fall and lane checks. It wrapped training around 950k steps, froze the policy, and ran the same formal eval on 100 new seeds. Final scores: Success rate: 98/100 Mean distance: 19.92 m Median distance: 20.06 m Mean successful ride time: 7.04 s Real development/training decisions: 960,386 steps Development/training charged to budget: 960,496 steps Astra later also checked and fixed a history bug when importing a policy. One interesting thing from this test: Both models successfully built the physics environment, and both could get the pelican riding. But under nearly the same training budget, final success rate differed by 69 percentage points. Kimi used PPO and curriculum learning to try to learn riding step by step; Astra judged the control space was small and searched directly for a stable policy via parameter search. And Astra already hit 7/8 within the first 300k steps, while Kimi only reached 3/8 at 480k. Who do you think won?

9:00 AM. Slack, Notion, Jira, ChatGPT, Claude... 12 windows open before you've done anything. 10:00 AM. Meeting. Same update, explained for the third time to a different person. 11:00 AM. Open Claude, repeat everything from the meeting, and align on the frontend design. Switch to GPT, explain the same background again, and walk it through the backend. You screenshot the result and drop it in the group chat. A teammate replies: "We already discussed this. Doesn't work." 1:00 PM. A wave of last-minute requests lands, and you have no idea what led to any of them. So you become a human router: scrolling back through chat history, pinging one teammate, then another, pasting the pieces into AI one by one, then posting AI's answer back to the group. 5:00 PM. Time to log off, and it feels like you got nothing done. It's the AI era. So why are we still working together in a way that's slow, wasteful, and completely unfriendly to AI? Because we've never had the right tool. Offloop is the answer. Everything lives in one workspace. Every message lives in a channel. People and AI sit in the same group, so whatever's been said or decided, every person and every agent already knows. No more juggling a dozen apps. Agents handle most of the work directly. This is how teams will work.

Prompt is dead. Now just make a wish. Guys, I just test latest deepseek v4.1 flash 1017 @deepseek_ai. It is insane. It is so insane. ONLY one prompt, one shot, and it built the cod style fps game: build a call of duty 23 fps game use three js. It auto build 37 mins and JUST 0.1 USD! 5-wave tactical survival operation, 3 operators，3 difficulty protocols. I can not find word to describe my feelings now. It seems the same level as Gemini 4 and better than Chatgpt 6.1 sol @deepseek_ai is back.