loa

[loa @loaossin]

@Da7_Tech Yeahhh, provided there’s no “dumbing down,” it should remain usable until 2027.

[loa @loaossin]

@laopai1990 真实体感是，5.5依旧吊打一切。昨天用6.1sol跑了个任��修bug花了20h左右，还没有修复，5.5只用了30min，还优化了一些别的地方。而v4.1flash只是无限接近门槛，始终差一点。gemini4argon败在了价格，跑完了额度都没开始干活

[loa @loaossin]

@Yuchenj_UW It’s truly something to look forward to.

[loa @loaossin]

@xodud_rkd wow

[loa @loaossin]

@Denche2005 I don’t know. It’s true that you don’t necessarily have to go to a workplace anymore—you can get by just fine by tinkering with various little projects at home—but it ends up being far more exhausting than it used to be.

[loa @loaossin]

@leo114119 这次真的达成了“许愿式编程”了吗

[loa @loaossin]

@StefanoGPT 1. Internet censorship is too strict; many legitimate projects are blocked.
2. Accounts from China are constantly being banned (I've already had 20+ banned).

[loa @loaossin]

@math_fournity wow

[loa @loaossin]

@YxzRainy claude：只需略微出手

[loa @loaossin]

@yacineMTB Opus 5.5 is so cheap—it's simply impossible to say no.

[loa @loaossin]

Generally speaking, if the code the AI ​​generates differs from what I intended to write, then the AI ​​is just a "cyber jackass"—because I far surpass it when it comes to architectural design and the trade-offs involving key quality metrics. If the code matches exactly what I wanted to write, then the AI ​​is a "cyber beast"—or simply put, an "Input Method Pro Max" that solves the problem of my typing speed failing to keep up with my brain. If the code is close to what I intended but also spots paths or bugs I missed, then the AI ​​becomes "Lord AI"—my esteemed partner, my loyal familiar, or even the Kyubey with whom I’ve signed a soul contract to become a genius magical-girl coder. However, if I merely state requirements and the AI ​​handles the implementation without me having a clue what it’s actually doing, I’d call that "cyber possession," "Russian Roulette," or a "cyber game of *Three Kingdoms*": you never know if the AI ​​is a loyal subject or a traitor, or even whether you’re playing the AI ​​or the AI ​​is playing you. Development essentially becomes a power game where we scheme against and exploit each other—a relationship where both parties know the other is disposable and ready to stab them in the back the moment their value runs out. Of course, that’s not necessarily a bad thing; for simple prototypes or throwaway tools meant for one-time use, it works just fine.

[loa @loaossin]

Musk said on Oct 10 that real AI should handle situations it’s never seen, not just replay a demo. I think that matters even more for humanoid robots than for LLMs.

Robot videos are really good now. They dance, box, cook, fold laundry. Smooth enough to be a little unsettling.

But watch one put a cup in a kitchen cabinet and ask: what if the cup is six inches to the left? The cabinet door is half shut? There’s a cereal box in the way? Does it figure out what to do next, or run the same motions straight into the cereal box?

A robot that replays a rehearsed sequence and a robot that can actually help around your house are very far apart.

I don’t buy the “it’s all fake” take, though. Stable walking, joint control and fine manipulation are real engineering. Teleop is how a lot of training data gets collected. The problem is selling a scripted or teleoperated demo as autonomy. And Tesla isn’t innocent here either. The Optimus robots at its 2024 event were reportedly remote-controlled or human-assisted.

At least the testing is changing. Last month Figure said Helix 2.5 went into 30 homes it had no prior data from (tidying living rooms, folding towels, making beds) and hit 56% zero-shot. Their number, so grain of salt. But nobody picks 56% to look good, and “how does it do in a house it’s never seen” is the right question to ask.

The video I actually want:

One simple chore. First take, just do it. Second, I move the object. Third, I put something in the way. Fourth, I interrupt it halfway. One continuous shot, no cuts, nobody on the controls.

I don’t even care if it succeeds. I want to see what it does when it fails. Does it notice and try again, or freeze and wait for a human?

Every robot looks good when it works. Show me one failing.

[loa @loaossin]

Please give me a Muse invitation code

[loa @loaossin]

How exactly do you use the Rea project? I’ve integrated Codex, Claude, and Gemini, but none of them can reverse-engineer the output I’m looking for.

[loa @loaossin]

I wonder why—Grok Image processed two images in just 20 seconds, yet when I fed the same two images to the Grok CLI, it took half an hour. I have no idea what it was doing.

[loa @loaossin]

Opened a fresh Codex CLI window just to work on two images

Two images. That's it

And it keeps compacting the conversation over and over, like it's stuck in some stupid loop

What the hell is it even compacting? There's barely any context yet

[loa @loaossin]

Which AI (SI) company do you think will win at the end?

[loa @loaossin]

My favorite thing about OpenCode so far is that I can plug in pretty much any model API I want

I switch between models a lot, so this alone makes me want to keep using it

[loa @loaossin]

National Day. One of the few long breaks. A lot of people went home, back to places the new world doesn’t really reach.
Posts online said it: AI, GPT, Claude, FDE, vibe coding — like none of it exists. The world is still the old world.
Tonight, outside a restaurant, an old woman was tapping a sheet of tin. Sharp little sound. Asking people if they wanted to buy something.
I finished after eleven. She was still there, tapping the tin, asking.
I don’t know what I’m trying to say. Just that the world is moving fast, and some people are still sitting where they were.

[loa @loaossin]

Briefly in love with GPT again

Image limits are gone

I was complaining about it two days ago, but for now I’ll pretend none of that happened