virushuo (@virushuo) 's Twitter Profile
virushuo

@virushuo

Back to the 2006. Screeching voices of the minority.

Author «Illustrated History of Programming»

ID: 49913

linkhttps://jhuo.ca calendar_today08-12-2006 07:14:11

8,8K Tweet

67,67K Followers

2,2K Following

virushuo (@virushuo) 's Twitter Profile Photo

统计下手边的项目,大概95%代码是ai写的。文档有28195 tokens, 测试有22170个,但是真正的代码本身只有20320(后端代码没有html/css占数量)。我觉得这个比例是合适的,在这样的粒度下什么模型都表现不差,而且消耗很少。传统的工程训练和经验还是有用的,想搞vibe coding需要学的不是vibe而是coding

统计下手边的项目,大概95%代码是ai写的。文档有28195 tokens, 测试有22170个,但是真正的代码本身只有20320(后端代码没有html/css占数量)。我觉得这个比例是合适的,在这样的粒度下什么模型都表现不差,而且消耗很少。传统的工程训练和经验还是有用的,想搞vibe coding需要学的不是vibe而是coding
Grok (@grok) 's Twitter Profile Photo

virushuo Brian Roemmele As of mid-2025, the average residential electricity price in the US is about 17.5 cents per kWh (USD), per EIA data for June. In Canada, it's around 18 cents per kWh (CAD), or ~13.5 cents USD after conversion, based on averages across major cities from Fraser Institute (2023

Neo (@soulhacker) 's Twitter Profile Photo

同感,我使用GitHub Copilot + GPT-5,尽量把功能/非功能性需求逐点列明,它干完会逐点回应我的需求,不多不少,最后附上一些的延展任务的建议,如果需要可以跟它说,就比较对我这种坚持要“知道自己要做什么”的用户的口味

virushuo (@virushuo) 's Twitter Profile Photo

这个和我体验相似。当你给codex/gpt5 high一个非常明确的任务的时候,它可以做到非常符合预期。我现在的办法是和aider/Gemini2.5 pro 讨论清晰到毫无疑义,生成设计和实现文档,review没问题,给codex就能一次得到符合期望的结果

Qwen (@alibaba_qwen) 's Twitter Profile Photo

🚀 Introducing Qwen3-Next-80B-A3B — the FUTURE of efficient LLMs is here! 🔹 80B params, but only 3B activated per token → 10x cheaper training, 10x faster inference than Qwen3-32B.(esp. @ 32K+ context!) 🔹Hybrid Architecture: Gated DeltaNet + Gated Attention → best of speed &

🚀 Introducing Qwen3-Next-80B-A3B — the FUTURE of efficient LLMs is here!

🔹 80B params, but only 3B activated per token → 10x cheaper training, 10x faster inference than Qwen3-32B.(esp. @ 32K+ context!)
🔹Hybrid Architecture: Gated DeltaNet + Gated Attention → best of speed &
Angel Bogado 🌻 (@angaisb_) 's Twitter Profile Photo

Anthropic has: - The best coding model ❌ - The best image gen model ❌ - The best model for math ❌ - The best voice mode ❌ - The best chatbot UX ❌ - A reliable API ❌ - Great rate limits ❌ - Expensive models ✅

Christian Cantrell (@cantrell) 's Twitter Profile Photo

Six months ago, we launched a preview of our first model. Today, we're launching the new Reve. But it's really four products in one: 1. An image creator and remixer. 2. A drag-and-drop image editor. 3. Your own creative assistant. 4. For the first time, an API. Let's start

Six months ago, we launched a preview of our first model. Today, we're launching the new Reve.

But it's really four products in one:

1. An image creator and remixer.
2. A drag-and-drop image editor.
3. Your own creative assistant.
4. For the first time, an API.

Let's start
virushuo (@virushuo) 's Twitter Profile Photo

gpt-5-codex确实挺不一样的,今天我跟他说了1个小时,用各种方式试图证明我的设计对,而我同事的是错的。但是它非常强硬的反复跟我说,不,你的是错的,你同事的设计对。而且每次都能给出理由,最后甚至列出了一份文档中某几行作为支持。后来我跟同事开了个会,开完会我承认codex说的对…

virushuo (@virushuo) 's Twitter Profile Photo

ai这个行业太残酷了,即使是顶级模型厂商用户忠诚度也几乎是0。肉眼可见的大批Claude code的用户几乎在一天内就转投了codex。但是别着急,Gemini3马上就来了。。。。

Alibaba Tongyi_Lab (@labtongyi96898) 's Twitter Profile Photo

1/7 We're launching Tongyi DeepResearch, the first fully open-source Web Agent to achieve performance on par with OpenAI's Deep Research with only 30B (Activated 3B) parameters! Tongyi DeepResearch agent demonstrates state-of-the-art results, scoring 32.9 on Humanity's Last Exam,

1/7 We're launching Tongyi DeepResearch, the first fully open-source Web Agent to achieve performance on par with OpenAI's Deep Research with only 30B (Activated 3B) parameters! Tongyi DeepResearch agent demonstrates state-of-the-art results, scoring 32.9 on Humanity's Last Exam,
virushuo (@virushuo) 's Twitter Profile Photo

adk/a2a因为变化比较频繁又有重名项目,几乎所有编程模型都处理不好。我的解决方案是把官方的demo选几个放在当前项目目录中,告诉它用这几个目录中的代码作为参考。效果还不错。

virushuo (@virushuo) 's Twitter Profile Photo

很多年前不太智能的google alert和中期略微智能的Google Now都想做过类似的事情,但是都不够好也不够准。所以最后还是苦涩的教训说的好:要重视能随算力扩展的通用方法的力量

eraera (@eraera) 's Twitter Profile Photo

很高兴评论里有几位提到电路设计。问题是:哪怕数字电路设计这种天生就是框框+连线的项目,业界也是使用HDL来写大规模电路,而不是手动连线。这就引人深思了:是不是文本天生有什么优势?对了,本人以前是Verilog/VHDL编译器程序员,写过你们有些人用的Quartus II

很高兴评论里有几位提到电路设计。问题是:哪怕数字电路设计这种天生就是框框+连线的项目,业界也是使用HDL来写大规模电路,而不是手动连线。这就引人深思了:是不是文本天生有什么优势?对了,本人以前是Verilog/VHDL编译器程序员,写过你们有些人用的Quartus II
Andrej Karpathy (@karpathy) 's Twitter Profile Photo

My pleasure to come on Dwarkesh last week, I thought the questions and conversation were really good. I re-watched the pod just now too. First of all, yes I know, and I'm sorry that I speak so fast :). It's to my detriment because sometimes my speaking thread out-executes my