特效在你之上

AI学习｜健身｜正在薄肌｜分享科技前沿｜求职

[特效在你之上 @cookietriu]

That's why LLaMA uses RoPE, not learned embeddings. And with NTK-aware scaling it stretches to 4x the training length. Lesson: the best position encoding is the one that encodes nothing absolute. #MachineLearning #Transformers

[特效在你之上 @cookietriu]

RoPE fixed this with a beautiful trick. Don't tell the model WHERE each token is — rotate the queries and keys instead. The dot product then depends only on relative distance. Relative position falls out naturally from rotation. Zero extra parameters.

[特效在你之上 @cookietriu]

Position encodings are weird. The original Transformer just added sine waves to word vectors, and it worked. But nobody told me the catch: absolute positions can't extrapolate. Train on 2k tokens, test on 8k, watch it fall apart.

[特效在你之上 @cookietriu]

@thsottiaux the reset button has gone physical lol. print a QR code on the cards that opens a reset request form and you've reinvented the queue ticket machine, but for GPU hours

[特效在你之上 @cookietriu]

@ClaudeDevs 3x faster, zero model changes. the wins were all in the plumbing: gzip the payload, pre-warm the session while you're still typing. half of AI progress is systems engineering nobody benchmarks

[特效在你之上 @cookietriu]

@rasbt funny how distillation is just RL where the teacher did all the exploring for you. you're not skipping the work, you're inheriting it

[特效在你之上 @cookietriu]

@thsottiaux They coordinated over email lol. no new protocol, no shared memory, just the 50-year-old dumb pipe. maybe the multi-agent future runs on the most boring stuff we already have

[特效在你之上 @cookietriu]

The catch: when reasoning lives in hidden states instead of tokens, chain-of-thought monitoring goes blind. You can't read what the model won't say. Reportedly even OpenAI caps loop counts internally. Depth is great — until you can't audit it. #MachineLearning
https://t.co/yjNutDp6Ua

[特效在你之上 @cookietriu]

The model behind Astra might not be a bigger Transformer — it might be a Transformer that refuses to stop thinking. Reportedly, it loops the same layers instead of stacking more of them. Depth from compute, not parameters.

[特效在你之上 @cookietriu]

This isn't just a rumor-shaped idea. ByteDance's Ouro is a real open looped model: the 1.4B version matches models up to 12B. Same weights, run again and again — reasoning depth becomes a dial you turn, not a wall you build.

[特效在你之上 @cookietriu]

Spent the weekend taking apart every Agent framework buzzword. Turns out most of them are just... while loops and try/except. 🧵
ReAct = a while loop. Call model → run tool → feed result back → repeat until done. That's it. The "reasoning" is just the loop asking the model what to do next.
Plan-and-Solve = call the LLM once to make a plan, stuff it into context, then execute. It's a to-do list with extra steps.
Reflexion = try/except. It fails, you feed the error back to the model, it retries. Every junior dev has written this.
Tree of Thoughts = generate 3 plans, score them, pick the best, backtrack if stuck. Congratulations, you reinvented DFS.
Short-term memory = a list. Long-term memory = write to a file. Task tracking = a checklist with reminders. Hooks = middleware with permissions and sandboxing.
The humbling part: I used to think each of these was its own complex system. They're all prototypes you can build with loops, conditionals, and state. The HARD part isn't the skeleton — it's what comes after: judging feedback quality, managing memory, recovering from failure. "It runs" vs "it works well" is where the real engineering lives.
Which agent mechanism did YOU take apart and think "wait, that's ALL it is?" 👇
Want to build each of these from scratch? This repo rebuilds a Claude Code-like agent harness lesson by lesson — agent loop, tools, hooks, todo, memory. 17 lessons, zero magic.
https://t.co/j8hFwVyoxz

[特效在你之上 @cookietriu]

@DDXG_99 浇

[特效在你之上 @cookietriu]

@SDG0777 加油

[特效在你之上 @cookietriu]

@FirstClassgoat 浇

[特效在你之上 @cookietriu]

@Tomkins1211 浇

[特效在你之上 @cookietriu]

@yiwen_shouji *罒▽罒*

[特效在你之上 @cookietriu]

我也被……

[特效在你之上 @cookietriu]

@NYang368 ٩(•̤̀ᵕ•̤́๑)ᵒᵏᵎᵎᵎᵎ

[特效在你之上 @cookietriu]

@Arrowli180954Li 哈哈

[特效在你之上 @cookietriu]

看到大佬的作品，我也心血来潮，让ai给开始做一个，后续会慢慢分享

[rewind @rewind02]

how to create a local game using Claude Opus 5.5 and Unity 6...

a beach, a pier, a day/night cycle, real fishing controls

the trick is never asking for “a fishing game” in one prompt:

- step 1: create a unity 6 project (urp template) and install unity’s plugin in claude code
- step 2: greybox the level with simple shapes and a camera, no art yet
- step 3: add the core mechanic in one prompt, then play it until it feels right
- step 4: generate models in tripo, clean them up with opus in blender, import as fbx
- step 5: add light, water, effects and sound last
- step 6: have opus playtest it with an autopilot and list the bugs before fixing anything

who does what:

> unity 6: the world, physics, ui
> tripo + blender: every model
> opus 5.5: scripts, scene setup, cleanup, tests
> you: play it and decide what stays

commit after every step that works. one bad prompt then costs you seconds, not an evening

[特效在你之上 @cookietriu]

@aidenyangcn 一个多月

[特效在你之上 @cookietriu]

@WizS731 浇一个