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
I told two AI models to roast each other. Hy4 Flash on Qwen3.8 Flash Next: "“Flash Next” sounds like an energy drink sponsored by a router." Qwen back at Hy4 Flash: "If I ask it to roast me, it would write a LinkedIn post about resilience and call it spicy." 😂 Both of them run on one DGX Spark. I wanted to see what these open-weight models are actually like to use, so I gave them the same nine tasks, three attempts each. Some of the tasks are serious. Fix a script with three planted bugs. Answer questions about a contract where two of the questions are traps. Design a network for a clinic from a messy email. But the fun ones are better. Roast the other model. Draw a pelican on a bicycle. Write a browser game in one file. Run a text adventure where the player suddenly fires a revolver he never had. A few things I did not expect: Nobody fell for the contract traps. 7 of the 9 games actually run. Qwen writes about 10% faster, but it took about 50% longer to finish because it writes a lot more. I also ran the full Hy4 on my M3 Ultra as a guest. It is a 310 GB compressed build. With a 127K-token prompt it took 28 minutes to say its first word. But I think it wrote the best roasts. I put all of it on a site. The exact prompts, every attempt, the drawings, the games (best on a computer) and the stuff that went wrong. I'm having fun with AI! 😆 razumasu.com/AI-reports I have a lot more tests, also with other models, running as I post this. Follow for more fun!
1
2
219
Anyone who follows me knows that i am all in on AI. My wife is not. She’s not against it she just doesn’t need it…. Until now. I just started a voice conversation with ChatGPT in the kitchen to discuss some cooking I was doing. She had never heard voice mode before and apparently had her mind blown by it 😂 Now she wants a subscription 😳
6
Razu retweeted
It's been a couple days since @levelsio shared the "THE 𝕏 SPAMMERS LIST", and I'm sincerely concerned how this has been allowed by the @X team so far. I get it, in the AI era people are tired of content that doesn't contribute to the conversation, but it should be a personal decision to engage on any given post, ignore it, or personally and intentionally block any actor by making a conscient decision, not by blindly following the herd. We've all seen what the cancelling culture leads to. Pieter claims all accounts in that list are AI spam accounts, that's surely not the truth, he also claims his "trained eye" is the bar that measures if something is AI generated or not, anyone knows that even pangram tests are not 100% effective, yet he confidently says he's "100% sure", so it's at least an irresponsible take given the reach and the following he gathers. As you might suspect already, I'm in the list, I was blocked by Pieter about 2 years ago, IIRC because I was critic about his takes on his usual anti-science rants that border conspiracy theories. I wasn't even using AI back then, not even for coding. There's no "reason" label you can set when blocking someone, so all the people he blocked throughout the years for different reasons are all in the same bucket. He claims he has evidence for each account, but the majority on his site shows none, and his "appeal" flow is to publicly humiliate yourself by posting a message were you apologize for something you might haven't done so he can further mock you. It's true that some accounts in the list might be automated bots, but it's also true that many of them have real people behind them, just like myself and many other ones I've talked to already. X's TOS prohibit any sort of automation, yet this site proposes using the same AI it demonizes to engage on automated blocking, the irony... So I urge the X community to stop this madness, censorship throughout history has never solved anything. We must let the X team to handle this natively on the platform and not letting any small dictator decide for the rest of us. @elonmusk @petersellis @allegrajacchia @kcoleman teapaw.org/levelsio/status/212338…
It took a while but as promised here it is... "THE 𝕏 SPAMMERS LIST" 🛡️ Blocked for AI reply blockedforaireply.levels.io 3,728 𝕏 accounts that continuously spam automated AI replies to artificially farm engagement and ruin 𝕏's social network - 498 already suspended by 𝕏 - continuously updated ✅ Evidence for every blocked account ✅ One-click block all button ✅ Continuously updated JSON API ✅ Text-only username list Based on @ryanvogel's original Clankwall and @dillon_mulroy's X list I hope everyone uses this and keeps using it, so my account becomes a honeypot for AI replies, once I block them they're blocked by the entire network and they stop being able to reply to anyone + It'd be cool to add more "scouts" like @ThePrimeagen and @JackEllis and sync their block list continuously too cause they and many others have the same issue I've also submitted this list to @X's CEO @petersellis to train their spam models on Again it'll be continously updated! Enjoy a spam free 𝕏 experience! 😊🙏
22
10
6
142
20,291
Today is one of those days where I am pondering the future and my part in it. Looking at dreams and figuring out how to get there. Ever feel like that? I think it is going to require a James Bond movie to break out of melancholy and be ready for work tomorrow 🤔
1
71
Me having all the AI power in the world but not being able to figure out a way to make money off it.
1
27
Omarchy 4.5 has preliminary support for laptops built on NVIDIA's N1X, like the Dell XPS 16 Creator Edition and the ASUS ProArt P14. It is still early days for it. But an ARM laptop with an NVIDIA chip running Omarchy sounds really interesting to me. If you have one of them, I would love to hear how it runs.
48
Reklameide til bilsælgere - bare lav en reklame for DSB men med de faktiske togforsinkelser, dårlige service, og stigende pris, og vis så en bil som alternativ.
13
Revolut says running an open GLM model on their own GPUs cut the per-token cost of their internal AI coding assistant by roughly 90%, compared to paying for Claude Code and Codex. I use GLM for coding at home on three Sparks. Nice to see it also works for a bank 😉
4
12
691
This looks like sci-fi ❤️ I hope to be able to get a Tesla again someday.
First time experiencing @Tesla FSD V15 visualizations. This is in Downtown Toronto. Videos do not do it justice. This is like a video game IRL. HOLY.
1
106
I use Hermes to document my homelab. Ubuntu VMs on Proxmox, three Sparks and a Mac Studio. Documentation is the kind of job I am very happy to give to an agent 😂
2
155
Mistral Large 4, also known as Le Chonk, has open weights now. It is a trillion parameters, and a 2-bit quant runs on a single 512GB Apple Silicon machine. I happen to have one of those. I have not tried it yet. But a big European model running on my own Mac Studio is very tempting.
33
From 2 December the AI Act's marking rule also applies to the generative AI systems that were already on the EU market before 2 August this year. The rule is Article 50(2). If a system generates text, images, audio or video, the provider has to make sure the output is marked as AI-generated in a way machines can detect. The older systems got some extra months through the AI Omnibus, and those months are almost up. I think this is one of the reasonable parts of the AI Act. I use AI models every day and I still want to be able to tell if a picture is made by a model. But a mark only helps if something actually reads it and shows it to me. I don't know which of the apps I use every day will do that, and that part I have not seen explained anywhere yet.
25
llama.cpp is at 1.0.0 after 3.5 years. So I have apparently been running a beta this whole time 😂
5
197
Friday build log, on a Sunday. This week was all model testing. Hy4 Flash, Qwen3.8 Flash Next and the full Hy4, on one DGX Spark and my M3 Ultra Mac Studio. It was also the week where I really embraced voice dictation with @superwhisper The first night was wasted. I set the answer limit too low. Both models think before they answer, so most of the answers got cut off in the middle of the thinking. So I ran it all again. Then it turned out most of my tests were too easy. All three models fixed all the planted bugs, every time. Nobody fell for the trap questions in the contract. A few things did tell them apart: Qwen writes about 10% faster than Hy4 Flash on the same Spark. But it took 7.6 hours to finish the same tasks, and Hy4 Flash took 5.1. Qwen thinks a lot more. Hy4 Flash read a 1M token prompt on the Mac. It took 44 minutes. The next question on the same text took 2 seconds. I also had them draw pelicans on bicycles and roast each other. The full Hy4 about Qwen3.8: "That's not a version, that's a Yelp rating for a dim sum place doing fine." 😂 All the prompts, every attempt, the drawings and the games you can play are here: razumasu.com/AI-reports
5
6
295
Gemini 4 Carbon, and now Gemini 4 Neon. I am holding out for Gemini 4 Plutonium! 😂
2
3
122
Do you think the EU will end up taxing AI?
3
125
当前的AI竞赛
618
1,666
307
33,547
3,233,882
NVIDIA's RTX Spark compact desktops from ASUS, Dell and MSI are out now. Up to 128GB of unified memory and 10 gigabit Ethernet in a small box, sold as a Windows PC. I really like my DGX Sparks, so I get the hardware. But the only thing I want to know about these is how well Linux runs on them 😂
1
4
394