Cmd+AI

Samuel surprises Kabarza with a cloned version of his voice, made in ElevenLabs from the show's transcripts, and plays an AI-generated three-minute roundup of last week's episode. That sparks a frank debate about whether AI-voiced content helps or harms a small show's brand, and where the line sits between a useful new format and something that takes the hosts out of the equation. They then pick apart the Twitter row over a $50,000 landing page, covering value-based pricing, what a "landing page" actually includes, and why so much of the cost is unseen.

In the news, GPT-5.2 lands to benchmark fatigue and a slow first impression in real coding work. A maximum-severity React Server Components exploit means everyone should patch now. Anthropic hands MCP to a new Linux Foundation group, Apple quietly releases an open-source image model, Figma adds pixel-editing tools that edge towards Photoshop, and Cursor's new visual editor raises the question of who it is really for. Along the way: Shopify Editions as site of the week, listener comments on Webflow's shrinking creator scene, Liquid Glass, and Hank Green's take on AI's water use.

Chapters

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

Kabarza and Samuel open with phones, watch batteries and why both have swapped Siri for ChatGPT or Gemini on the action button, which turns into a frank chat about why developers neglect front-end UI. Samuel shares what he has learnt running coding models locally and why one-shot benchmarks say little about working in a real codebase. Kabarza's site of the week is Shopify's Black Friday pinball machine and globe, built in three.js and streamed in 8K to the Las Vegas Sphere, which promptly crashes his browser on air.

The main news is OpenAI's reported "code red" after Gemini 3, followed by Gemini 3 Deep Think, DeepSeek 3.2, OpenAGI's computer-use model and Anthropic's acquisition of Bun, which prompts a discussion about why software engineering still matters. Samuel explains how he uses agentic browsers to build a sponsorship CRM and why Perplexity's BrowseSafe benchmark matters, before both hosts talk about tool fatigue and learning through the hard part. They finish by asking where all the Webflow content creators have gone, looking at Webflow's marketing, new component libraries from Timothy Ricks and Alicia Sanizarro, Bradford Huber's holiday challenge and Kabarza's client Christmas cards.

Chapters

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

Samuel flies solo this week while Kabarza is off sick. He starts with a trip to Anthropic's Claude pop-up in London and a morning of running several agents at once to build a sponsorship pipeline. Then comes Claude Opus 4.5: lower pricing, better token efficiency, a recent training cut-off, Claude Code in the desktop app and the new tool search tool for cutting MCP context. That leads into whether general-purpose models are now good enough and the real gains will come from specialists.

He tries Perplexity's virtual try-on live, with mixed results, and asks whether it could fix the pain of buying clothes online. From the comments, he looks at Meta's SAM 3 object-tracking model and answers a viewer who says Framer sites are slow, which turns into advice on testing with slow machines and Lighthouse. A comment about the Antigravity site leads to designing for your audience. He finishes with Cursor's physical Tab key reward and why it rewards hands-on coding.

Chapters

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

Google has finally rejoined the chat. Kabarza and Samuel unpack the Gemini 3 launch: the benchmark jumps, how it has been pushed into Replit, Stitch and AI Studio, and why its creative, less wireframe-y output stands out. Samuel argues it is a bigger leap than GPT-5.1 but still not a match for his Claude Code setup. They also give their verdicts on the Antigravity code editor, compare it with Trae's solo mode, and try putting themselves into Nano Banana 2 meme images, which leads to a worried chat about AI-faked passports and ID checks.

Before that, Kabarza's site of the week is Stripe Press, with its 3D bookshelf. Later, a listener comment asks whether AI turns developers into full-time reviewers, Samuel follows up on ditching MCPs in favour of plain code folders, and the hosts ask whether OpenAI can ever be profitable, what AGI really means, and whether Gemini 3 is already good enough for most people.

Chapters

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

Kabarza and Samuel start with a change of direction. Kabarza's agency is adding AI companies as a client vertical. Samuel is moving his own business towards building apps, and he talks about finding a niche by interviewing people he knows in one industry. They then look at the Mantis studio site, which leads into how a brand concept is built, and go through Webflow's new AppGen: what it builds on (Webflow Cloud, Astro, DevLink), how it compares with Lovable, and where it falls short.

Next come GPT-5.1's warmer tone and Samuel's switch to Trae and the new Codex models. A LinkedIn thread about PewDiePie learning to code leads into Affinity going free under Canva, then Anthropic's own write-up on why MCPs waste context. The longest discussion follows CJ's month off AI coding: why AI's non-deterministic output is hard to square with code, and why Samuel is putting his time into tooling and review rather than more features. They finish with Apple's practical-effects TV intro and an AI copy of it, and argue over whether the copy counts as art.

Chapters

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

Samuel and Kabarza open with Halloween costumes, a birthday trip to a 1940s bar and a crash course in Cockney rhyming slang, before a warning that LinkedIn trains AI on your posts by default. They look at a TypeGPU experiment and what TypeScript actually adds, Freepik's node-based Spaces canvas, OpenAI's Aardvark security agent, and a long debate over whether people will one day build their own software instead of paying for SaaS.

They also cover Ilya Sutskever's deposition and what it says about Sam Altman, PewDiePie's turn towards running and training his own AI, Amazon's legal threat to Perplexity over AI shopping agents, and Mr.doob vibe coding updates to Three.js. Webflow gets Finsuite's dev mode and new Cloud storage options, GitHub's report puts TypeScript at number one, Builder.io promises a design and dev tool for whole teams, and Kabarza reports back on spending eight euros trying to animate a thumbnail with Freepik.

Chapters

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

Samuel and Kabarza open with a story about a strategic workshop that went badly wrong, then pick a site of the week from Niccolò Miranda and get into Cursor 2: the new agent mode, the Composer model, running several models side by side, worktrees and the built-in browser. Kabarza explains why Cursor feels like Webflow did in 2019, and Samuel replies to listener comments on M1 Macs, unified memory and running LLMs locally, before both take aim at Framer's per-site editor pricing and explain why Webflow is better value than people think.

On the AI side they cover OpenAI's hire for a non-invasive, ultrasound-based brain interface and how it compares with Neuralink, the Neo home robot that turns out to be teleoperated, and OpenAI's livestream predictions on devices, AGI by 2028 and compute. The second half is design: Affinity going free under Canva, Adobe Max and Adobe's hold on the creative pipeline, Figma's Weave acquisition and Schema updates including slots, and Rivian's bike with its neo-brutalist interface.

Chapters

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

Kabarza and Samuel start with why Kabarza won't leave Arc, whether voice and text beat knobs and menus, and how a minimalist and a maximalist each want their tools to work. They then cover Claude Skills and why they save context compared with MCP servers, Claude Code on the web, and DeepSeek's OCR paper on packing text into images.

Most of the episode goes on ChatGPT Atlas: Kabarza's centre-a-div test in Webflow, the training setting that is on by default, and why agentic browsing suits slow background jobs rather than your main work. After that come Manus 1.5 and its trillion-token competition, a listener question on connecting your own database, Instagram's AI Restyle, Framer's AI-generated ad and what counts as slop, a man who sent an AI avatar to argue his case in court, Webflow's real-time collaboration and the website of the week.

Chapters

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

Kabarza and Samuel celebrate passing 200 YouTube subscribers and work through the week's news. Kabarza tries Google's Veo 3.1 and explains how a chat with a laser tag owner turned into an idea for AI-themed safety briefings, which leads Samuel into why bespoke software is now cheap enough for small businesses to build. They then ask whether vibe coding is really dying after a report on Lovable's falling traffic, and talk about the gap between a prototype and a shipped product.

Samuel explains how Claude Haiku 4.5 now matches what Sonnet 4 could do five months ago and how he uses it for sub-agents. The hosts weigh up NVIDIA's Spark mini computer, the new M5 chips and running models offline, before getting into which MacBook designers and developers actually need. They finish with a look at some unusual AI job adverts, including a CEO who wants an assistant to help clone him, and what that says about context, delegation and communicating clearly with AI.

Chapters

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

Samuel and Kabarza open with a late-night rabbit hole of AI soul covers of rock songs, and Kabarza's view that Sora has become a glorified meme generator. They then cover Chrome's new MCP server and how MCP servers fit into an AI coding workflow (and eat context), before working through OpenAI Dev Day: apps inside ChatGPT, whether chat is the right interface, Agent Kit, n8n's $180 million raise, Sora 2 and GPT-5 Pro API pricing, and the ChatGPT super-app idea compared with WeChat.

On the tooling side they discuss Cursor's new plan mode, Anthropic's open-source Petri alignment tool and the Claude Code plugin marketplace. Kabarza argues that OpenAI's circular chip deals point to an AI bubble, then takes apart Framer's new pricing against Webflow's. They finish with a viral Framer motion piece, how the new Webflow site was built with components and Mast, and Webflow's pitch as a headless CMS.

Chapters

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