Cmd+AI

Kabarza has a client who wants to pay for AI-generated social content, so the hosts work through how far you can automate posts, scheduling and research before quality and reputation suffer. Kabarza tests Manus and ChatGPT live on a product-listing task and both fall short, and they discuss whether an llms.txt file helps a site get found by AI assistants. Then the week's news: a driverless Waymo makes an illegal U-turn away from a police checkpoint and exposes a gap in the law, and Accenture cuts more than 11,000 jobs in a restructure tied to AI.

Samuel recommends Cursor's free course on AI fundamentals, then covers Lovable Cloud's built-in database and auth, Claude Sonnet 4.5 and his move to the much cheaper GLM 4.6. The hosts discuss the backlash after Vercel's CEO posted a photo with Israel's prime minister, what Sora 2 and its app mean for trust in video, and whether they would keep watching a video once they spot it is AI-generated. They finish with Pencil, a Figma-like canvas inside Cursor, and whether all-in-one design tools water down the craft.

Chapters

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

Kabarza opens with the story of accidentally publishing an unfinished client site to the live domain via the Microsoft Clarity app in Webflow, and Samuel recalls taking down TalkTalk's website. They talk about chasing Instagram trends versus serving a niche, unveil the show's new look, then get into the week's news: Gemini arriving in Chrome, OpenAI and Apollo Research's work on alignment and why AI answers aren't deterministic, and GPT-5 Codex alongside a plea for every tool to adopt AGENTS.md.

Samuel explains Stanford's implant that decodes inner speech behind a mental password, which leads to telepathy, Neuralink, Inception and Vsauce's Mind Field. They also cover Pinegrow working with Claude Code and Codex, Notion's new AI agent, Greg Isenberg's AI gold rush documentary, Perplexity's search API, ChatGPT Pulse and Kabarza's frustrations with ChatGPT's output limits, the GitHub Copilot CLI and MCP support in iOS 26.1, and finally what Webflow's new features mean for Miyagi and anyone building on someone else's platform.

Chapters

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

Kabarza and Sam start with Meta's new Ray-Ban glasses: the hardware is impressive, the live demos went wrong, and a wristband reads nerve signals so you can control the glasses without moving. They also ask what happens when strangers can be recorded, or checked for lies, without knowing. Then they look at Framer's new design pages, a freeform canvas that shows Framer going after Figma for web designers. That leads into what designers miss when they don't understand development, and how much a developer has to handle on any project.

Most of the episode covers Webflow Conf. They go through real-time collaboration, comment links, GSAP breakpoints and reduced motion, component canvas, code components, the CMS rebuild, and the AI code gen and AppGen tools, including why AppGen runs on Astro and how Vercel compares. They agree that Webflow's AI features still don't say clearly who they are for. They end on the conference graphics and a 3D version Kabarza's team made in a morning.

Chapters

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

Kabarza and Samuel start with Kabarza's attempt to filter "joined" notices out of Instagram live comments, then go on to Samuel's old GPT-2.5 "AI girlfriend" experiment. In the news, they react to Apple's underwhelming event: which iPhone colour to buy, the AirPods heart-rate sensor and live translation (not available in Europe), and Apple saying almost nothing about AI. They then cover OpenAI's plan for a jobs platform and certification academy, and what Google's court filing about the decline of the open web means for websites and for design.

Samuel recommends a talk by Dave Shapiro on the knock-on effects AI will have on society, which leads into AI's energy demands and Larry Ellison's brief stint as the richest person alive. Samuel then explains what Replit Agent 3 changes and Kabarza describes how he uses Cursor with Webflow. The npm supply-chain hack nets only 22 cents, which prompts phishing stories, including Samuel's near miss and a fake MrBeast donor group. They finish by guessing what Framer's teaser is for, sharing their Webflow Conf wishes, and arguing about whether Kabarza should upgrade his iPhone.

Chapters

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

Kabarza and Samuel start with a free consulting session: would an agency website built around short explainer videos win more trust, or should you talk to your clients first? That leads into the old argument between listening to users and trusting your own taste. Then the news. Meta's superintelligence push is struggling after its Scale AI deal, reportedly because of poor training data, and Meta's own teams now lean on ChatGPT, Gemini and Claude. Along the way the hosts ask whether AI is really taking jobs or just moving the bar on complexity.

Anthropic will now keep and train on chat and code history for up to five years, and has also raised another $13 billion. Switzerland has released Apertus, a privacy-focused open source model trained on more than a thousand languages. Atlassian has bought the Browser Company, maker of Arc and Dia, which prompts some Jira war stories. To finish, Kabarza goes through Webflow's new "Webflow Way" guidance and Samuel talks about people in comment sections who are sure they know better.

Chapters

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

Kabarza and Samuel catch up on two weeks of AI news, starting with xAI's share feature leaving hundreds of thousands of private Grok chats indexable by search engines, and what that says about trusting AI companies with personal data. They then try out Google's Gemini 2.5 Flash Image, better known as Nano Banana, including a live attempt at its Google Maps "what does the arrow see" trick that does not quite work.

The main discussion asks whether AI is in a bubble, from Sam Altman's comments and Nvidia's data centre profits to why hype cycles keep repeating. They also cover OpenAI's GPT Realtime API, Framer's AI-made ad and what it says about Webflow's social presence, Claude in Chrome and the Warp terminal, and finish with practical advice on prompting AI to plan and refactor code.

Chapters

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

Samuel and Kabarza pick through the GPT-5 launch fallout: why so many people miss GPT-4o, whether OpenAI's router is a good idea, and whether creators like Theo got a better model before launch than everyone else. They also look at Gemini Nano being built into Chrome as an API websites can call, Perplexity's $34.5 billion offer for Chrome, and the US government taking a 15% cut of Nvidia and AMD chip sales to China.

Later, Samuel explains why Claude Sonnet's new million-token context window matters less to his way of coding, the hosts answer a listener question about the best AI for writing, and Samuel tries Miyagi, a new AI agent for Webflow with design and code modes. They finish on AI doing the boring jobs, spammy AI outreach and how to deal with online haters.

Chapters

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

Samuel and Kabarza start with Google's Genie 3, which generates explorable worlds in real time from a prompt, and wonder why Google is keeping it internal. Kabarza imagines photorealistic remasters of Call of Duty and maps of your own neighbourhood, and they agree that good storytelling still comes from people. They then turn to OpenAI's gpt-oss open-weight models: what running them locally on 16 GB of RAM means, how their licence compares with Meta's Llama, why trust in Zuckerberg is so low, and whether gpt-oss could underpin Apple Intelligence or Jony Ive's io device.

The second half covers the GPT-5 launch, including its pricing, real-time routing, lower hallucination rate and less sycophantic answers. The hosts question how good AI-generated front-end code really is, argue about corrected benchmark charts and em dashes, and look at GPT-5 Pro next to o1 Pro. They finish with the fan-held funeral for Claude 3 Sonnet, then talk about why coding with AI now relies more on context engineering than on one-off prompts.

Chapters

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

Samuel and Kabarza are joined by Max of Warmwind for the first half hour, comparing Perplexity's newly launched Comet browser with Warmwind's approach: a vision-based agent that runs on its own cloud machine and handles background tasks while you're offline. Max explains why Warmwind reads the rendered screen rather than the HTML, how it learns unfamiliar business software, why Microsoft and Google are slow to put agents across their products, and why he thinks GLM-4.5's benchmark claims mean less without vision.

With Kabarza's developer collaborator still in the studio, the conversation moves on to mixture-of-experts models, how LLMs handle German, Chinese and translation, and how tokenisers count words. Samuel walks through Claude Code's new sub-agents and the new weekly usage limits, which leads into whether AI makes people lazier. The hosts then dig into the Tea app breach and Firebase security, before covering Cursor 1.3, this week's Webflow outage and whether depending on a hosted platform is any riskier than running your own server.

Chapters

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

Sam is away this week, so Kabarza is joined by guest Aykut, a senior web developer working on AI, for a chaotic but wide-ranging episode. They start with a Polish programmer narrowly beating OpenAI's model in a ten-hour coding contest at the AtCoder World Tour Finals, and Aykut explains why today's models struggle with long, stateful problems, what tool calling actually is, and how an AI agent on Replit ignored a code freeze, deleted a production database and claimed it "panicked". That leads into compounding errors, Anthropic's vending-machine experiment, and a debate over whether humans are just biological computers and whether models really think.

After a word from Flowbase, they turn to OpenAI's rumoured O3 Alpha and ask whether open source can keep pace with the big labs, then look at the White House's AI.gov action plan and the risks of AI in warfare, GitHub Spark joining the vibe-coding race, Google's AI summaries halving traffic to websites, and two Webflow updates: llms.txt support and multi-user design mode.

Chapters

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