Viet NT

code, caffeine and an unstoppable beard

[Viet NT @viet2nt]

@bd7349 @birdabo I'm burning through 2.2 Claude Max accounts per day (Fable), so I have to switch more work to Opus 5, and it seems to be doing better than Fable on small/medium tasks - especially Opus 5 low.

Fable too lazy and made too many small mistakes recently.

[Viet NT @viet2nt]

@thsottiaux switch:
- model depth
- the team's transparent/communication

improve:
- better inference of intent / EQ / handling+coordinate many tasks
- let users pay to reset the weekly limit (incoming?)

[Xidulu @xidulu]

1/ Sharing a new, interesting project I did during my internship at Microsoft AI Frontiers w/ @JohnCLangford

TL;DR: At decoding time, we feed **previous hidden state** into the input together with token embedding, and it boosts performance for free.

https://t.co/za3cZ4UwWV

[The Bugged Dev @thebuggeddev]

Your our history shouldn’t be taught in a boring way 🫠

I’ve always loved history, but I could never imagine what life actually looked like back then.

So I built Empire Atlas, a 3D interactive explorer of 8 historical empires using @threejs and @Kimi_Moonshot K3 🔥

It lets you explore how people lived, what their homes looked like, their maps, daily life, interiors, and more. Properly researched.

But the craziest part is that this was near one shot vibe coded with Kimi K3 🤯

When I previously built a 3D anatomy app with GPT 5.6 sol, I had to iterate on performance and optimization.

With Kimi, the moment I handed over the 3D assets (generated using @tripoai), prompt and design (by GPT Image 2.0), it created an 11 step engineering plan to build the entire thing.

The very first step it did was optimizing the assets.

It took nearly 500MB of 3D assets and brought them down to just 17.8MB using mesh simplification, Draco compression, and 1024px WebP textures.

Absolutely nuts.

It also generated 56 historical images across the 8 empires showing daily life, maps, interiors, and more using its image plugin with batch processing.

Those were converted to WebP too, bringing the total image size to around 10MB.

That’s a huge reason the experience loads so fast on website.

It's engineering workflow or intelligence has really impressed me so far. The only downside is that it took more than 5 hours, though 😅

Anyway, back to history.

In Empire Atlas, you can explore 8 different empires and see how people and our ancestors lived at that time. I really love those textures I was able to create using @tripoai.

You can explore their homes in 3D, and there’s so much more we could do with this.

We could extend these houses into fully explorable interiors and create increasingly realistic reconstructions of what life actually looked like. And maybe create fun education games too.

I genuinely think this can make history education so much more immersive. Much more than showing black and white images in boring textbooks.

Go explore your history now 👇

Live: https://t.co/Dbgxr13L0g
Code: https://t.co/sahbaZeq63

[Viet NT @viet2nt]

@sarjnarw @thdxr I can run 3-500 concurrent stream (real work) with deepseek api at 60-90 tps < $5 per run - how many spark is that?

[Viet NT @viet2nt]

cooking = )

[Viet NT @viet2nt]

@thdxr opencode $200/month got $800 credit when 🤑

[Viet NT @viet2nt]

@brianryhuang @_aidan_clark_ And gpt-2026?

[Viet NT @viet2nt]

@teortaxesTex 15-18x cost (80% cost is cache read) - 3-4x speed?

[Viet NT @viet2nt]

@iruletheworldmo so September or ?

[Viet NT @viet2nt]

@SeanCasGamer @johnlindquist In my test grok isnt very fast - will re-run

[Viet NT @viet2nt]

@SeanCasGamer @johnlindquist my executor fleet mostly consists of opus-low instances - extremely fast at decent quality (faster than sol-low and luna-medium)

[John Lindquist @johnlindquist]

Attributing Opus 5's poor performance to "user error" just sounds silly.

I've never heard this kind of language from OpenAI.

[Viet NT @viet2nt]

@VictorTaelin In other words, we are being trained by the model just as much as we are training on it. I don't think we yet understand what we're trading away in that process.

[Viet NT @viet2nt]

@VictorTaelin Some of the perceived decline is on our side rather than the model's.

Early on, novelty and expectation inflate our impression of a new model; over the following weeks we build an increasingly accurate map of its failure modes, so identical performance starts to feel worse.

[Viet NT @viet2nt]

Fable 5: I don’t give a shit about your code. All I care about is whether I’m over-apologizing.

[Viet NT @viet2nt]

@jahirsheikh8 It’s a by-product of VRAM malfunction: some bits flip from 0 to 1 or vice versa—like evolution, but at a much faster pace. /s

[Viet NT @viet2nt]

@teortaxesTex - luna in sub is 5-7x cheaper
- but flash in api can run 500-1,000 threads (real load) - not sure about luna in sub

[Viet NT @viet2nt]

@llllvvuu 10000 output token/ 40-50 tk/sec = 3.5-5 minutes