Razu (@Razumasu) header
Razu (@Razumasu) profile picture

I run IT for a 17,000-seat arena (Live Nation). Local AI, agents & networking, at home and in prod. Writing about what works. Ex-founder. Horror author.

Denmark
Joined October 2025
Photo by Razu (@Razumasu)Photo by Razu (@Razumasu)Photo by Razu (@Razumasu)Photo by Razu (@Razumasu)Photo by Razu (@Razumasu)Photo by Razu (@Razumasu)Photo by Razu (@Razumasu)Photo by Razu (@Razumasu)Photo by Razu (@Razumasu)Photo by Razu (@Razumasu)
Pinned Tweet
The 512GB step on the new M5 Ultra Mac Studio costs $15,500 on its own. TechRadar worked out that's about 3.9 times more per gigabyte than the same step on the M3 Ultra last year. My M3 Ultra with 512GB just got a lot harder to replace. Holding on to it a bit tighter now....
9
4
3
84
12,134
I hope this isn’t true. I don’t want anyone to know how often I stop at red lights in GTA6…. 😳
Rockstar Games 周三解雇了我。第二天 GTA 6 就发售了 我是 Rockstar North 的高级引擎工程师,从事 RAGE 性能优化。三个月前,我在遥测模块里发现了一条本不该存在的数据管道,而且我能证明它流向了哪里 此刻每一份正在运行的 GTA 6 都在收集玩家的行为数据。这部分不算新闻。每款游戏都会收集遥测数据。隐私政策里写的是"用于改进未来产品的匿名游戏数据分析"。我发现的不是这个 遥测模块的数据会汇入 Take-Two 内部的一个系统,名为 Lens。Lens 并不会把匿名使用统计数据发送到 Rockstar 开发人员的数据看板。它为每一位玩家建立一份个人行为画像。不是匿名的。与你的主机账号、你的平台 ID、你的 IP 绑定 我在8月14日向我的技术主管报告了这件事。他说这是"用于玩家体验研究的扩展分析",让我别管这事。我要求提供数据处理协议。他说这是在 Take-Two 集团层面处理的,不在 Rockstar 的职责范围之内 我调取了 Lens 数据管道的完整文档。我发现的内容如下: 遥测的内容远远超出游戏时长和崩溃报告。Lens 以毫秒级精度记录每一次输入。按键时间点、摇杆动作、反应速度、PS5 自适应扳机上的扳机压力模式。它实时运行行为分析,并生成他们称之为"玩家行为签名"的东西 - 一份对每位玩家都独一无二的画像,就像指纹对每个人都独一无二一样 Lens 会从你的游戏过程中推导出:冲动控制评分、风险承受度指数、攻击性反应模式、耐心阈值、压力下的决策速度、消费冲动评级。这些不是调试指标。这些是心理行为类别 直到我找到了对外传输的管道,我才明白它们是用来做什么的 这些数据不会留在 Rockstar。也不会留在 Take-Two。Lens 通过内部 API 将行为画像传送给 Take-Two 旗下一家名为 Vantage Analytics 的子公司。Vantage 没有公开网站,没有新闻页面,没有招聘信息。它于2023年在特拉华州注册成立 Vantage 是一家行为 AI 公司。它用玩家画像训练预测模型,并把这些模型授权给商业客户。管道文档中包含一份不完整的客户清单。两家保险集团,一家信用风险建模公司,一个人力资源筛选平台,以及一家广告数据经纪商 你的 GTA 6 游戏过程此刻正在生成一份行为评分。没人追你的时候,你开得有多快。你会不会在红灯前停下。NPC 无端攻击你时,你怎么反应。你行动之前要想多久。你花游戏内的钱有多冲动。每一个选择都在喂养一个模型,企业正在付费查询这个模型,用来对真实的人做决定
15
Omarchy 4.5 has a new app called Hype that makes a presentation from a single Markdown file and exports it to PowerPoint. I want to try it for my next presentation at work. If it means I never have to open PowerPoint again, I'm in 😂
10
llama.cpp has release builds for s390x. So if you have an IBM mainframe standing in the basement, you can do local AI too…. I guess 👍
2
82
Omarchy 4.5 had over 740 pull requests merged from more than 180 contributors in twelve weeks. That is almost 9 a day. In twelve weeks the EU can just about agree on a date for the next meeting.
19
Ripple Preview from Ricursive Intelligence is in Ollama now. It is a 65B dense model built for agentic coding. I have not tried it yet. But I really want to see how it does next to GLM5.3-Flash on my Sparks.
1
3
183
Tencent's full Hy4 is out, and AngelSlim put up compressed builds yesterday. I wanted to know if the big one would run on my Mac Studio M3 Ultra 512gb at all, so I tried. It does. The 310 GB IQ2 build loads on the M3 Ultra with 512 GB and runs entirely on the Apple GPU. It needs the llama.cpp branch from PR #32517, which built on macOS without changes. Loading takes about two minutes the first time, and it leaves about 139 GB of memory free. It writes at 12.2 tokens per second, the same on three runs. After a 32K prompt that drops to 10.9, and after a 127K prompt to 10.0. The catch is reading. It reads about 75 tokens per second, so a 32K prompt takes close to 7 minutes before the first word. A 127K prompt took 28 minutes. Two things to know about these numbers. This is the IQ2 build at about 3.5 bits per weight, so it is not the full-quality model. And thinking was on, which is the default. For a single hard question I would use it. For long documents I would stick with Hy4 Flash, which read the same 127K prompt on a DGX Spark in under 4 minutes. I'm running it through a set of real tasks now and will post how it does.
4
2
359
Would be nice if the X app on iOS had an iPhone Duo layout.
1
120
Mentioned Omarchy to a guy today (because I talk a lot about it). He argued that the only reason people want to use nice looking versions of Linux is because they can't afford a Mac 😱
3
6
449
Omarchy 4.5 speaks 50 languages. The thing I'd want from a Danish desktop is to be able to search the menus in English too. Half the time I've got a guide open and I know exactly what setting I need, just not what I'd have called it in Danish.
1
46
What's the best way to manage multiple monitors with Omarchy? I feel like the native one isn't very good for me.
1
51
The X algorithm feels really random. Some days it shows me AI and tech posts like the ones I make myself, other days it just shows me....
2
4
156
I wonder how/when my posts appear in other peoples feeds, and if they feel like its one or the other... 😂
1
2
13
A spec I'd put on the front of every USB-C cable packet: the time it takes to copy a 50GB file between two fast drives. Put the number on the cable too. The packet isn't going to be there when I'm looking through a drawer in two years.
1
2
52
There's a `.diskused` command in the SQLite 3.54 shell. It's the first thing I'd try on a database file that's getting suspiciously large. I'd like to know which table ate the space before spending an evening optimising the wrong one.
1
3
52
SQLite 3.54 shows the column headers in columnar CLI output even when a query finds no rows. Lovely little detail. An empty result called `overdue_invoices` is considerably more reassuring than a blank bit of terminal.
21
Two dates worth keeping apart in the EU Cyber Resilience Act: reporting duties started on 11 September 2026 for manufacturers. For open-source software stewards they start on 11 December 2027. A checklist that just says “CRA reporting: September” is missing a fairly important column.
1
129
When someone shares a quantized model, I'd love a couple of prompts where they could tell it apart from the original. Even one example would give me something useful to try at home beyond asking both versions to write the same Python script and squinting at them.
17
Before I let a local coding model rename anything across a Rails app, I give it one awkward case: two things with the same name, only one of which should change. It’s quite interesting to see what it leaves alone and what it doesn’t.
2
3
222
Can you squeeze an AI into a much smaller file and still have it notice the same things? I tried it on my M3 Ultra Mac Studio with two versions of Qwen2.5-Coder: an 8-bit download at about 35 GB and a 4-bit one at about 20 GB. Same trained model in both. The smaller file rounds off the numbers inside it a bit more. That’s what people mean by quantization. I gave both versions 12 little Ruby puzzles from a made-up shop. My favourite was a sticker giveaway. You get a sticker on days 8 through 14. You turn up on day 8. Do you get one? The shop said yes. Both versions of the AI wanted me to fix the shop. They’d read “after day 7” and decided that somehow ruled out day 8 as well. I think you get your sticker. Across all 12 puzzles, the 4-bit version spotted the same four bugs as the 8-bit one. It usually answered sooner, too. What I liked was that even when they went looking for trouble in code that worked, they picked the same two bits. So in this little shop, squeezing the model down didn’t cost it a single catch. Now I’m curious what kind of puzzle would finally get the two to disagree.
1
128