Yaogang Lian

Indie developer building Quiver, Dino ( and other apps.

Beautifully crafted. Love is in the details. • 用心构建精美的软件

[Yaogang Lian @yliansoft]

Just asked Dino to build Remotion from scratch then use it to generate a promo video of Dino itself. 😀#inception https://t.co/cokL5o3RNS

[Yaogang Lian @yliansoft]

an observation: the highest leverage of AI at the moment is from the best people in non-AI fields

the best ones are quietly using AI to speed up their own research by 10x or 100x

[Yaogang Lian @yliansoft]

so coding is solved,

math is also solved,

time to solve physics:

Astra, pls unify general relativity with quantum field theory, use 100000 subagents, make no mistakes

[OpenAI @OpenAI]

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.

The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.

The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.

[Yaogang Lian @yliansoft]

@White1637402 Seems that more companies are realizing that flash models are better for business: cheaper, faster, good enough for most tasks.
In that sense Google is actually ahead of the game. 😄

[Yaogang Lian @yliansoft]

figured out how to run any AI agent autonomously for a month straight:

while true:
prompt agent "keep going"

seems to work well with Claude Code, Codex etc

[Yaogang Lian @yliansoft]

when everyone is rushing,
slow down

when no one reads code any more,
read code

use AI,
but don't lose yourself

[Yaogang Lian @yliansoft]

@Steve_Yegge That applies to human engineers too.

[Yaogang Lian @yliansoft]

@DW33643 生活规律最重要，很容易陷入散漫，然后难以自拔。每天坚持定时做些事情。祝福。

[Yaogang Lian @yliansoft]

@kunchenguid Less is more.

[Yaogang Lian @yliansoft]

My suggestion for buying coding plans in general: buy the monthly plan and turn off "auto renew". Things change too fast so better to reassess every month.

Also don't put all eggs in one basket: buy the cheapest tier from two or three coding plans so if one server goes down, switch to the other.

[Yaogang Lian @yliansoft]

I can recommend two coding plans that I have been using for quite some time.

1. https://t.co/n1gzo7LSP1: I am an early adopter, started using last year. They are really reliable. No 429s, no quantized models, amazing speed (before speed becomes cool).

2. @ollama cloud: have been using for a couple months. Their servers are in us and europe. Server uptime is decent. Amazing speed esp. for flash model - use the flash models if you can since the weekly limit will last much longer.

[Yaogang Lian @yliansoft]

@YinsenW_ Agent 主对话可以分支，回退，重复运行，的确跟 git 的很多运作非常相似。

[Yaogang Lian @yliansoft]

@christomitov @Zai_org @ollama yeah it does like to babble quite a bit, even repeating the same things two or three times, but for the tasks I use it for (relatively simple compared to what I threw at glm-5.3), it is faster even in terms of wall clock time.

[Yaogang Lian @yliansoft]

Speed does matter.

I was using GLM 5.3 (through @Zai_org coding plan) on some hard coding tasks for the past few weeks.

Today I have some simpler coding tasks so I decide to seriously try DeepSeek v4 flash (through @ollama coding plan).

I am a bit surprised by my own reaction - GLM 5.3 was not slow, but deepseek-v4-flash from ollama cloud is just unbelievably fast. And the speed honestly affected me more than I expected - kind of hard to switch back to glm-5.3.

[Yaogang Lian @yliansoft]

At the current 50% off price: $0.075 per 1M input tokens, glm-5.3-flash is the first model that meets the bar I set in my article:

"What we really need is a decent coding model (GLM 5.2 level) for $0.1 (or less) per 1M input tokens."

https://t.co/CHByNVvqUV

I hope they will make the 50% discount permanent.

[Yaogang Lian @yliansoft]

https://t.co/oG8qOaKYAB

[Yaogang Lian @yliansoft]

AI industry likes to coin terms that are not new.

OPC (One-Person Company)

= Solo entrepreneur

= 独立开发者

= 个体户

= Einzelunternehmen

[Yaogang Lian @yliansoft]

很可惜快手 @KwaiAICoder 决定退出大模型赛道。KAT-Coder-Pro V2.5 还是不错的，后训练再强一点应该就可以跟 glm 5 差不多水准。

不过完全能理解。大模型卷的实在厉害。

[Yaogang Lian @yliansoft]

The thinking level might be a bug on my side - I think I was using auto which defaults to max for this model.

[Yaogang Lian @yliansoft]

Ox Alpha does feel like a glm model - just from the output style it feels quite familiar, like glm 5.2 / 5.3 I use daily.

Used for a few hours today in Dino. Impressions:
1. Quite similar to glm 5.2 or 5.3 in terms of capabilities: tool calling, code quality, coding skills, etc.
2. It's a bit slow in my tests - could just be too many people trying this model right now
3. It likes to think for a long time even though I set to low thinking. glm 5.2 and 5.3 already tend to think quite long; this thinks even longer.
4. At the moment the api drops quite often. The capacity claim is certainly overhyped.

You can use Ox Alpha for free in Dino right now via OpenRouter. https://t.co/rxO1PTIHDu

[Yaogang Lian @yliansoft]

BTW, I put up https://t.co/4wMM0Q5lUz for sale. Was trying to build some open agent plugin spec but Vercel beat me to it. Happy to transfer at a fair price.