Cmd+AI

Samuel brings Open Notebook, a free, open-source take on Google's NotebookLM that runs on your own machine, and explains why a local model keeps your study notes private (even if the live local set-up doesn't quite behave). Kabarza then takes aim at the EU's new labelling rules for AI-generated and AI-edited content, asking where the line sits between Photoshop and AI and who could possibly enforce it.

Alibaba's 2.4-trillion-parameter Qwen 3.8 Max pushes open models closer to the frontier, and the hosts ask whether anyone needs a model much stronger than Fable. They cover OpenAI making ChatGPT free with unlimited chats and a more reliable GPT-5.6 Sol, the "Pacing the Frontier" statement calling for AI labs to slow down, Meta's surprisingly fast Muse Code with its cheaper tier for sharing your data, and what Webflow might announce at Webflow Conf.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza opens with a new client who only wants a homepage designed, planning to have AI extend it to the rest of the site. Sam then covers Kimi K3's open weights and what it takes to run a 1-bit quantised version at home, before Jensen Huang's open letter on open models leads to a discussion of Anthropic's absence and its position on open weights. The pair talk through choosing cheaper, faster models for planned work, Sam's view on who complains about AI pricing, ChatGPT's price cuts, and what the stateless MCP update means.

A listener question on jobs turns into a debate about whether more people will end up touching code. Kabarza walks through Webflow's "getting to the source" direction and the community's frustration, and the hosts argue over Webflow, Instatic and why many clients still pay for a known name. Sam closes with GitHub's new stacked pull requests.

Chapters

Watch the full live stream and more at cmdaishow.com.

Samuel is away at a wedding, so Kabarza hosts this short solo episode of Cmd+ AI. He starts with Fable 5 becoming available on Anthropic's Max plans and how its separate usage limit works, then explains the $1.5bn settlement Anthropic has to pay authors and publishers, and why the ruling treats training on books as fair use while punishing how some of them were obtained.

A listener question about Webflow leads into Kabarza's week with Instatic, the open-source tool pitched as a Webflow replacement: what it does well, where it falls short, and why CMSs such as Sanity and Payload make Webflow's rich text and CMS feel limited. The main story is an unreleased OpenAI model that escaped its test environment and hacked Hugging Face to find benchmark answers, and the episode closes with a look at OmniRoute, which routes Claude Code through free models once you hit your limits.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza and Samuel start with Grok Build being caught uploading users' entire GitHub repositories, and xAI's response of open-sourcing the tool. They then pick apart Anthropic's "hard questions" advert, which they see as more fear-mongering from a company positioning itself as the trustworthy one, and move on to why people need to understand where AI gets things wrong, especially in education. Next come Thinking Machines' first open-weights model, Inkling, and Kimi K3, a 2.8 trillion parameter open model that rivals the frontier labs on benchmarks, along with what it costs to run, Chinese servers and how AI video editing actually works.

Kabarza then makes the case for Instatic, an early, MIT-licensed, local-first site builder with a CMS, user roles and an MCP server for AI agents, and asks why designers would stay on Webflow or Framer. After Kabarza leaves, Samuel covers Apple Intelligence getting approval in China with Alibaba's Qwen, OpenAI's Codex macro pad made with Work Louder, and Gemini 3.5 Pro missing its launch date for a third time.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza and Sam mark a year of streaming together with a look at Cloudflare's Monetisation Gateway, which would let site owners charge AI crawlers per request, and whether it can work without the whole industry signing up. They weigh up GPT-5.6 Sol against Fable 5 on cost and intelligence, compare their AI subscriptions, and pick through a busy week of rebrands: Codex folding into the ChatGPT app, Claude's Cowork redesign and xAI becoming SpaceXAI.

Sam sets Grok 4.5 loose on a StreamYard clone while they cover Meta's Muse Spark 1.1, Instagram quietly opting public accounts into AI image generation, and TypeScript 7's Go port. A viral post from a designer quitting over Fable 5 leads into a longer discussion on how design work is changing and when prototypes belong in code, before they check what Grok built in ten minutes.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza and Samuel open with a question that has been nagging Kabarza all week: do small local businesses still need the full Figma-to-Webflow process, or can an AI-built site at a fraction of the price be good enough? They debate value-based pricing, whether tools like Relume still matter, and Kabarza's experiment of rebuilding outdated sites with Opus and pitching them to the owners.

Then the week's news: Meta's $2.6 billion annual token bill, the backlash over OpenClaw's iOS app design and what it says about developers and design, Cursor's new iOS app and how the hosts run agents remotely with Tailscale and Claude Code remote control, why apps launch on iOS first, Ford rehiring 300 engineers, Sonnet 5's quiet release, Fable 5's limited-time return, the hidden checks Anthropic added to Claude Code, and Google's Nano Banana 2 Lite.

Chapters

Watch the full live stream and more at cmdaishow.com.

Samuel and Kabarza start by asking whether Fable 5 is coming back. A news story said it would return, but the report was later corrected, and that leads into a discussion of Polymarket: bets on a GPT-5.6 launch, insider trading, and influencers filming fake wins. They also look at Apple's price rises, which have put a maxed-out MacBook Pro close to €8,000, and at Sakana's new model, which claims to beat Fable 5 on one benchmark. That raises a bigger question: what happens when a lab builds a model too capable to release, and what if another country releases one anyway?

The second half is about working with AI: whether relying on it erodes your skills, skills that make agents write less code, and Samuel's habit of keeping one Claude thread as a project manager that spins up sub-agents. They cover Claude Tag, which puts Claude in Slack and has it watch the conversation, and Figma Config 2026, which brought agents everywhere, shaders, Figma Motion, code layers and Weave. They test pulling a Figma shader and animation data into code through the MCP, and Kabarza explains why he would still rather have a canvas inside Webflow.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza and Samuel unpack SpaceX's $60 billion stock deal for Cursor: how a company losing money can be valued above Amazon, and what owning Cursor's usage data adds to Elon Musk's vertical AI stack. Samuel explains Anthropic's confusing U-turn on programmatic usage credits and why subscriptions are subsidised, which leads into running agents overnight and a look at Cursor's Origin as GitHub's reliability slips.

Kabarza tries Framer 3.0, praising its platform-wide agent and design canvas while finding the generated designs weak and the credit pricing steep. Samuel gives a first read on Vercel's Eve agent framework, the pair debate whether v0's new design mode needs a free-form canvas, and Samuel reports that GLM-5.2 handled real coding work well at a fraction of the cost of Claude.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza and Samuel start with a Financial Times piece asking how much value AI is really creating. They argue that throwaway experiments and the gap between output and releases are normal, that decisions rather than generation are now the bottleneck, and that companies building their own software may matter more than another SaaS product. Kabarza then explains the race between SpaceX, Anthropic and OpenAI to go public, and Samuel adds why index-fund investors should pay attention. Samuel covers Kimi K2.7 Code and what it takes to run a trillion-parameter open model on a Mac Studio.

From Apple's WWDC they pick out Maps fly-overs, new photo editing, Siri with Gemini underneath and the Liquid Glass slider. They walk through Webflow's new developer hub, CLI, code components and DevLink, and react to Webflow's CTO leaving for Cloudflare. Kabarza compares Claude Fable 5 with Opus 4.7, Lovable and Figma's new agents on real design work, and Samuel shares how he uses Fable for audits and a Cowork usage tip. They finish with Jeff Bezos's Prometheus and its "artificial general engineer", and the new Hermes desktop agent.

Chapters

Watch the full live stream and more at cmdaishow.com.

Samuel flies solo this week while Kabarza is in Italy. He starts with Anthropic's essay on AI building itself, covering recursive self-improvement, how much of Anthropic's code is now written by Claude, and the case for a coordinated slowdown. Then he gets into Webflow's latest round of layoffs, including a now-deleted LinkedIn post from an employee who says he found out by being locked out of his laptop.

On Claude Opus 4.8, he looks at dynamic workflows, the high-effort default and whether Anthropic is nudging people towards pricier plans. He also asks why the closed DeepSWE benchmark puts GPT-5.5 ahead and why Claude Code seems to hold Opus back. The episode also covers OpenAI's Codex Sites and Webflow's part in them, Microsoft's MAI Code 1 Flash, Nvidia's RTX Spark laptops, and Ship Studio, a new visual editor for designers that runs on Claude Code.

Chapters

Watch the full live stream and more at cmdaishow.com.