Codex iterated a pure NumPy + cv2 closed-loop heuristic policy for VizDoom D3 Battle. No neural network training, no map, no object coordinates, no seed-specific routes. Just screen pixels plus public game variables, roughly the same signals a human player gets. It works surprisingly well. Notes and videos are... now in the blog:show more

what does the resultant decision tree look like roughly? a concrete example of a codex-coded if-else block would help

Isn’t this just….. GEPA

Admittedly I'm biased toward fine-tuning everything, but this is brilliant. No maps, no seeds, just pixels reacting in real time. Almost like actual gameplay instinct.

After we fully understand the magic of heuristic learning, the next step is how to remember these experiences.

Pure cv2 + screen pixels surviving D3 Battle with no seed-specific routes is the surprising part — most pixel-only scripts brick the moment HUD offsets or texture variants shift.

@tszzl NumPy + HRR Shhhh

Maybe it makes more sense and convincing to challenge a more difficult benchmark like Programbench and Riemann bench？or OpenAI Proof QA?Otherwise, this specialized system does not have any generalization and can only function in a few unique and narrow areas

Let's do it step by step, simple task first and hard task later!

amazing

This is the most interesting Codex angle to me: it makes old-school heuristics cheap to search and maintain again. Not everything needs a trained model if the loop can keep testing itself. Curious how brittle it is across maps or noisy UI.

🥰🥰

.After all these years, returning to Doom brings back a flood of emotions.😭

Cool idea and demo! Maybe in the long term the maintenance of the verification process will grow dramatically, since the verification + AI have already redefine what software means to us.

my guyyy pure numpy doom goes hard i kinda wanna see it break on a new map

如果不是让codex去写代码，而是让他根据类似过程为自己迭代一个执行动作的skill是不是也能达到这种效果？

Collab? 🤝 DM me 📩

the maintenance cost was always the bottleneck, not the expressive power of heuristics Codex as the author + maintainer changes that equation entirely. you get interpretable, debuggable, auditable policies — and the part that used to kill you (updating them as the env changes) just becomes another prompt might be the most underrated unlock in the whole codegen wave

HUD or scores?

Jiayi more tweets，pls😭

Heuristics are coming back because they were never weak, just expensive to grow and maintain. Codex changes that cost curve. Pixel loops plus public game variables are local decisions with tight feedback; enough of those can look planned without a global planner.

Amazing, excited!

投稿の完成度が高くて普通にレベル高いと思う

纯NumPy+cv2写Doom策略，没有神经网络训练，这跟之前翁家翌的启发式学习是同一个路线——知识存在代码里不在参数里。Codex迭代出来的策略代码直接可读可审计，比神经网络权重透明得多。不过VizDoom环境相对简单，状态空间有限，纯启发式规则能hold住。更复杂的3D环境里，规则系统的维护成本会指数级增长。

Interesting, this is basically hand-designed policy competing with learned agents. Curious how it scales beyond VizDoom environments?

Don’t you think this is kinda similar to how value iteration is done.