Tublat is now part of the @claudeai Startups program by @AnthropicAI. We already use Claude heavily across engineering and AI. Now we build faster. Thanks to the Anthropic team for having us. 🚀

Sometimes the messages that mean the most as a founder aren’t from customers, investors or metrics. They come from the people building the company with you. I asked one of our DevOps engineers what he genuinely thought about what we’re building at @tublatglobal and about the philosophy behind it. This was his answer. What makes me proud isn’t the compliment. It’s seeing the culture understood from the inside. Zero assumptions. Empirical proof. Performance measured, not claimed. Reliability engineered, not marketed. We’re still building. We still have a lot to improve. But this is exactly the kind of company I want Tublat to become. A company where infrastructure is treated as engineering, not a commodity. And where the people building it actually believe in what they’re building.

If @deepseek_ai ever launches token plans, a lot of AI companies are going to have a very bad day. @tianyi

Benchmarks are useful. But they’re not the product. I forked @OpenCode, connected models from different providers, and built a governance layer around them. The interesting result: you can build seriously good software with models that often look “weaker” on benchmarks, without relying on the most hyped models. The real unlock is orchestration. One model plans. Another builds. Independent models review implementation, architecture and security. A final model judges the evidence before the work is accepted. The model matters. The system around the model matters more. OpenCode Governance is completely open source and provider/model agnostic. github.com/jannotix/opencode… I think we’re going to see a lot more software built this way.

I recently tested several coding models on the same small SaaS MVP, covering architecture, backend, frontend, and an external API integration. My personal ranking: 1. GPT 5.6 Sol 2. Claude Fable 5 3. Grok 4.5 4. GLM 5.2 Last: Kimi K3 GLM 5.2 was the biggest surprise. It felt fast, capable, and genuinely useful for daily development. Kimi K3 was the opposite. It claimed that its PHP and JavaScript checks had passed, but when I verified the files myself, they had not. The code contained syntax errors, broken fragments, and inconsistent functions. Models can make mistakes. What I cannot accept is a model claiming that tests passed when they clearly did not. For that reason, Kimi finished last in my ranking, and I would not trust it in production without very strict supervision. I have also mostly stopped using Claude Code. Claude Fable 5 is excellent, but the usage limits make it frustrating for everyday work, even on a Max plan. In my experience, Claude is the most disappointing when it comes to the balance between quality and actual usable capacity. My current setup is simple: - GPT 5.6 Sol for architecture and code review. - GLM 5.2 for execution. GPT thinks. GLM builds. So far, this is the best daily development setup I have tested. @elonmusk @sama @thsottiaux @DarioAmodei @CarolGLMs @ppwwyyxx @zxytim #AI #CodingAI #SoftwareDevelopment #AIEngineering #CodeGeneration #CodingAgents #DeveloperTools #LLM #GPT #Codex #GLM #ZCode #ClaudeAI #Grok #Kimi #SaaS

Four products. One platform. flowdesk, Blead, Socialy and WABuzz are evolving into mainudge. One omnichannel team inbox with AI at the core. Messaging, social, automation, leads and collaboration: finally in one place. mainudge is coming.

This is the company I've spent years building. Not a startup. A structure. AXIOM - The core behind what scales. #Infrastructure #EnterpriseTechnology #CloudInfrastructure #AXIOM

@tublatglobal is stepping into AI with Torentia, an ambitious project to build a European intelligence. We don’t know where it will lead, but not trying isn’t an option. Starting internal, aiming to evolve into a SaaS. This is just the beginning. #Torentia #AI