Gianluca Iannotta @jannotix

@Nishant_N_25 Claude opus 5

@Zai_org is shipping again. GLM-5.3-FlashX is here: the same GLM-5.3 Flash intelligence, but pushed to up to 200 tokens/s with a 1M context window. And this is the part I find most interesting: Z.ai isn’t only improving the model anymore. It’s improving the infrastructure underneath it too. Faster inference, aggressive pricing, huge context, strong coding performance. GLM-5.3 was already one of the most underrated models I’ve used. Now they’re attacking its biggest weakness: speed. The Chinese AI race is getting seriously competitive.

If Gemini 4 leak is even close to real, @GoogleDeepMind is about to make a lot of noise. And it reinforces a bet I’ve been making for months. The companies I’m watching most closely for the next phase of AI: @Google → Gemini 4 @SpaceXAI → Grok 5 @AIatMeta → Muse Spark 2 @alibaba_cloud → Qwen 4 @XiaomiMiMo → MiMo 3 Not because I know which model will top the next benchmark. Because these companies have the ingredients that compound: compute, infrastructure, capital, distribution and massive feedback loops. @Meta’s Contributor pricing is especially smart. Give developers huge discounts in exchange for opt-in model-improvement data → more adoption → more real-world agentic data → better models → more adoption. That flywheel could matter more than winning one benchmark. The AI race won’t be won by whoever has the best model this month. It’ll be won by whoever learns the fastest for years.

@Zai_org ZCode is the best multi-model coding agent I’ve tried so far. Side chats are genuinely useful. The file editor is strong. The overall workflow feels much closer to how I actually want to work with multiple agents. What’s missing: browser login for agents that support it, like @OpenAI and @grok a native “follow edited files” mode so I can watch changes happen like in a real IDE more AI providers Add those three things and, in my opinion, @zcode_ai becomes very hard to compete with. Multi-model is the right direction. Now make the experience feel completely native. @ZixuanLi_

We’re changing how planning works in Cycle. Instead of trusting one model to decide the architecture, multiple models will independently reason about the same problem, challenge each other, surface blind spots and converge before execution starts. The goal isn’t “more agents.” It’s better decisions before a single line of code gets written. Planning is becoming a council. And this is only one of the pieces we’re adding to Cycle.

@chaosengineerr Codex

@WordPress 7.1.1 is out. 11 security fixes, 17 Core fixes and 19 Block Editor fixes. If you manage WordPress sites in production, this isn’t one of those “I’ll update it next week” releases. Patch first. Debug later. A huge part of keeping WordPress secure is still surprisingly simple: keep the stack updated.

WordPress @WordPress

2 days ago

WordPress 7.1.1 is now available with 11 security fixes, 17 Core bug fixes, and 19 Block Editor bug fixes. Update your sites immediately from the Dashboard or by downloading the release. wp.me/pZhYe-5Eh

15 64 239 20K 15

@ZixuanLi_ This is a serious game-changer, and it's definitely going to be my daily driver for coding now. It would be awesome if you could add a feature to show the agent editing files live, just like in @zeddotdev or VS Code.

Huge congratulations to @CompleteSkeptic and the @typesafeai team. Jev might be one of the most genuinely different AI launches I’ve seen in a while. Instead of building another model optimized to chat better, TypeSafe is asking a much more important question: what should AI look like when the user is software, not a human? Typed decisions. Calibrated confidence. Machine-native intelligence designed for automation. That’s potentially a much bigger shift than another benchmark win. We’ve spent years making models better at talking to humans. TypeSafe is trying to make them better at being part of software. This is worth watching.

Diogo Almeida @CompleteSkeptic

4 days ago

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

4K 8K 72K 37.2M 58K

@Meta is moving very fast. Muse Code is now native on Windows: no WSL, PowerShell-native, sandboxed by default, x64 + ARM64, zero-admin install. Another reminder that the AI race won’t be decided by who has the best benchmark this month. The four companies I’m watching for the real long-term race are: @AIatMeta @GoogleDeepMind @SpaceXAI @alibaba_cloud Compute. Infrastructure. Capital. Distribution. Models change every few weeks. Those advantages compound for years.

Meta for Developers @MetaforDevs

3 days ago

Muse Code from Meta is now available natively on Windows - no WSL required. It’s PowerShell-fluent, sandboxed by default, native x64 and ARM64, and supports zero-admin install. Coming soon: intersession messaging. Get started: irm dev.meta.ai/install.ps1 | iex Learn more:

86 141 1K 183K 529

@melissapekel @XiaomiMiMo @MiniMax_AI @GoogleDeepMind @antigravity @cursor_ai @Zai_org The $20 plans from Anthropic and OpenAI suffer from session limits, making it impossible to work seriously with them.

After months of heavy use across 20+ AI providers while testing Cycle, my subscription ranking is becoming pretty clear. Best around ~$20/month for sustained work: 1. @XiaomiMiMo 2. @MiniMax_AI Best overall value: 1. @GoogleDeepMind @antigravity 2. @cursor_ai 3. @Zai_org Best absolute model quality: 1. @AnthropicAI 2. @OpenAI But OpenAI and Anthropic also have the most frustrating limits for serious daily engineering work. That’s the part too many rankings ignore. The best model is not always the best subscription. Quality matters. But so do usable hours, predictable limits and how much real work you can actually get done before hitting a wall. Which AI subscription gives you the most productive work per dollar right now?

@charliermarsh Unless you bring in a router for third-party models, you guys will never take off. Look at zcode, it's growing like crazy and honestly outperforms Codex in several ways.

@bindureddy In my opinion, you're on the wrong way. I think it will be @GoogleDeepMind or @AIatMeta that ends up with the most powerful latest AI model.

Honestly, I can work all day for just $60 a month simply by assigning different models to specific roles. You can run Anthropic’s $20/month tier as the planner, pair it with MiniMax M3 at $20/month as the executor which is basically unlimited or swap that out for Antigravity's entry plan with Gemini 3.8 Flash, z.ai's base tier running GLM 5.3 Flash, or Xiaomi MiMo for $16/month. Then for reviewing, just use OpenAI’s base plan with GPT 5.6 Sol, escalating to Astra whenever needed.

I think GLM-5.3 is still massively underrated. And @Zai_org deserves more credit for something almost nobody talks about: during off-peak capacity, ZCode can give Coding Plan users free quota resets, while idle-time tasks can run without consuming plan quota. That’s smart capacity management. I’ve also been testing ZCode Computer Use heavily, and in my workflow it feels dramatically faster and burns far fewer tokens than the Computer Use workflows I’ve run with Codex and Claude Code. GLM-5.3 for serious coding. GLM-5.3 Flash for fast multimodal/computer work. Transparent usage stats. Off-peak capacity passed back to users. More AI companies should experiment with this model. Credit where it’s due, @Zai_org. @CarolGLMs 👏

@RyanLeeMiniMax @MiniMax_AI Code is already one of the most interesting coding products out there. Now give it real GUI Computer Use. Not just terminal, browser or file access. I mean full desktop interaction: mouse, keyboard, native apps, IDEs, dialogs, installers, settings, the same way a human actually uses a computer. That would make MiniMax Code dramatically more useful for real end-to-end engineering work. Who else wants this?