Cmd+AI

Cmd+AI

By Kabarza and SamuelTechnology
Download on the App Store

Cmd+AI episodes

  • #51 Andrej Karpathy Joins Anthropic, Figma Design Agent, Google I/O, Composer 2.5

    Samuel hosts on his own this week while Kabarza is away. He starts with OpenAI putting Codex in the ChatGPT app and why he keeps his phone out of his working day. Then he tries Osaurus, a privacy-first Mac app for local and cloud models, and explains why MLX beats GGUF on Apple silicon. He also covers Notion's new agent platform, which he thinks may have arrived too late, and Cursor's Composer 2.5, which is still built on Kimi K2.5 but comes close to Opus at a fraction of the cost.

    The second half opens with Andrej Karpathy joining Anthropic's pre-training team and what that says about the company. Then comes a long run through Google's announcements: the Neural Expressive redesign, a Gemini Omni video test that doesn't go to plan, Daily Brief, the Gemini Spark agent, Antigravity 2.0 and Gemini 3.5 Flash. The episode finishes with Grok support in OpenClaw, Perplexity's query-aware context compression, and Figma's new design agent, which Samuel sees as the right way to bring designers round to AI.

    Chapters

    • 00:00 Solo show and running order
    • 03:16 Codex comes to the ChatGPT app
    • 08:05 Osaurus and local models on a Mac
    • 13:43 Notion's agent platform
    • 16:53 Cursor's Composer 2.5
    • 23:37 Karpathy joins Anthropic
    • 28:46 Google I/O: Gemini, Omni and Spark
    • 40:44 Antigravity 2.0 and Gemini 3.5 Flash
    • 49:47 OpenClaw adds Grok
    • 52:18 Perplexity's context compression
    • 55:48 Figma's design agent
    • Watch the full live stream and more at cmdaishow.com.

      1 hr 6 min
    • #50 GPT-5.5 Cyber, Apple $250 Million Class Action, Anthropic x SpaceX, Webflow Pricing

      Samuel is back from Madeira, and he and Kabarza catch up on the news they missed. They start with OpenAI's GPT-5.5 Cyber, held back for trusted defenders only, and how attackers get round the safeguards by splitting jobs into small steps and handing them to open models. Then it's Apple's $250 million settlement over the Apple Intelligence features it advertised but never shipped. Most of the episode is about Anthropic: its deal for the whole of SpaceX's Colossus supercomputer, the higher usage limits that followed, and the new charge on programmatic use of Claude subscriptions, plus a look at Claude's agent view.

      In the second half, Samuel isn't impressed by Perplexity Computer on the Mac, and they're doubtful about Subquadratic's claim of a 12 million token context window. Kabarza explains Webflow's new pricing and who will pay more. They talk about why asking agents for HTML instead of Markdown makes documents easier to read and share, and end on whether Lovable's new one-prompt designs are good design or just the new template.

      Chapters

      • 00:00 Madeira, Snapchat and side projects
      • 05:00 This week's news and GPT-5.5 Cyber
      • 12:40 Apple's $250 million Siri settlement
      • 16:29 Anthropic, SpaceX and Colossus
      • 27:08 Claude paywalls programmatic usage
      • 39:58 Claude agent view
      • 45:35 Perplexity Computer on Mac
      • 48:36 Subquadratic's 12 million token model
      • 56:23 Webflow pricing changes
      • 71:25 The unreasonable effectiveness of HTML
      • 81:43 Lovable's aesthetic update
      • Watch the full live stream and more at cmdaishow.com.

        1 hr 33 min
      • #49 Clicky, Webflow AI Code Components, Warp Open Sourced, Higgsfield MCP,

        Kabarza and Samuel open with a long pre-show chat: why Samuel wants a Vision Pro, what John Ternus as Apple's next CEO might mean, Meta recording every click its staff make to train its models, and why AI is turning into ordinary technology. Kabarza shows off an app he built so his dad can practise German, and Samuel explains the paper-trading crypto bot that Claude reviews and rewrites every day.

        In the news, they try Clicky, an on-screen assistant that can talk you through apps and spawn agents. They also use the Figma MCP to plan a site map in FigJam and test Webflow's new AI code components, which Kabarza finds rough and cut off from the rest of the site. Then they talk about what Warp going open source means and whether Stripe Treasury is really a bank. They finish with GPT Realtime and how voice could drive the controls in web apps. They also ask listeners what they would want from an Insiders community.

        Chapters

        • 00:00 Vision Pro and Apple's new CEO
        • 06:38 Meta logs its staff, and AI as everyday tech
        • 18:06 An app for Dad and an AI crypto trader
        • 27:10 Clicky, the on-screen AI assistant
        • 33:42 The Insiders community idea
        • 35:52 Driving FigJam with the Figma MCP
        • 46:53 Webflow AI code components
        • 59:31 Warp goes open source
        • 68:09 Stripe Treasury
        • 75:59 GPT Realtime and voice-controlled apps
        • Watch the full live stream and more at cmdaishow.com.

          1 hr 25 min
        • #48 GPT-5.5, Image 2, Deepseek 4, SpaceX x Cursor

          Kabarza and Samuel open with Samuel's new MacBook Neo, running agents from a phone with Dispatch and T3 Code, and the AI food-photography and menu app Kabarza has been building for restaurants. They then cover the Vercel breach that exposed environment variables and why it is worth checking clients' keys, before giving ChatGPT Images 2.0 a thorough test: UI ideation, recipe cards, brand merch and a head-to-head with Nano Banana 2 and Pro, including the telltale clone-stamp texture in its output.

          In the second half they ask what SpaceX gains from its option to buy Cursor for $60bn, discuss agent managers taking over from the IDE, and read through OpenAI's GPT-5.5 launch, its higher prices and its emphasis on following instructions precisely. They finish with the cheap open models released this week (Kimi K2.6, DeepSeek V4 and GLM 5.1), Anthropic briefly removing Claude Code from the Pro plan, and Flipbook, a prototype that streams every pixel of a website from a model.

          Chapters

          • 00:00 Pre-show: MacBook Neo and Dispatch
          • 04:31 Kabarza's AI menu photo app
          • 18:43 Headlines and the Vercel breach
          • 25:06 Mixing Claude and GPT
          • 27:05 GPT Image 2
          • 34:01 GPT Image 2 vs Nano Banana
          • 43:55 SpaceX and Cursor
          • 54:59 GPT-5.5
          • 65:36 Kimi K2.6, DeepSeek V4 and GLM 5.1
          • 69:46 Claude Code and the Pro plan
          • 72:09 Flipbook: video as UI
          • Watch the full live stream and more at cmdaishow.com.

            1 hr 17 min
          • #47 Claude Design! Opus 4.7, Sam Altman Attacked, Claude Code 2.0, HeyGen's Hyperframes

            Kabarza and Samuel open with a week of tool audits: Samuel explains how he has replaced Granola with his own meeting transcriber and what a week of running local models on a base M5 MacBook Pro has taught him, before Kabarza asks whether you can become a good engineer by learning concepts rather than syntax. The news covers the attack on Sam Altman's home, Google penalising back-button hijacking, the redesigned Claude desktop app with routines, enterprise customers losing access to Claude Max plans, Opus 4.7's literal instruction following and heavier token use, and Perplexity's Personal Computer.

            Claude Design launched during the stream, so the pair work out what it is for, upload their own design system from Figma and ask it for a website and a set of LinkedIn slides, concluding that it is a prototyping tool for non-designers rather than a replacement for Figma or Lovable. They also look at HeyGen's HyperFrames, which turns HTML into video, and why HTML is becoming a practical format for slides and motion.

            Chapters

            • 00:00 Tool audits, a DIY meeting app and local models
            • 08:42 Can you code without learning syntax?
            • 21:04 Sam Altman's house attacked
            • 25:04 Google penalises back-button hijacking
            • 30:11 Claude desktop app, routines and enterprise pricing
            • 46:04 Claude Design: first look
            • 62:13 Opus 4.7
            • 72:00 Perplexity Personal Computer
            • 76:45 Claude Design hands-on
            • 90:17 HeyGen HyperFrames
            • 95:48 Claude Design results
            • Watch the full live stream and more at cmdaishow.com.

              1 hr 42 min
            • #46 Claude Mythos, OpenAI $100 Plan, Netlfix VOID, Managed Agents.. and more!

              Kabarza and Samuel start with Anthropic's Claude Mythos. Anthropic is not releasing the model to the public. Through Project Glasswing it goes only to select security partners, after it found decades-old bugs in OpenBSD, FFmpeg and the Linux kernel. They also cover Anthropic's move to bill third-party harnesses like OpenClaw as extra usage, and how spend limits work. Then they ask whether Claude Managed Agents are worth paying for when n8n can do much the same.

              Netflix's open-source VOID model removes objects from video along with the physical effects they cause. A New Yorker investigation into Sam Altman leads into OpenAI's new $100 plan and the startup habit of giving things away cheaply before raising prices. Kabarza finishes with the experimental HTML-in-canvas API and what it could mean for accessible, SEO-friendly 3D websites.

              Chapters

              • 00:00 Lemmings and learning to run before you walk
              • 06:41 RAM prices and a first laptop
              • 11:31 Claude Mythos and Project Glasswing
              • 25:55 Model names, effort levels and 'just ask Claude'
              • 31:42 Anthropic blocks OpenClaw on Max plans
              • 44:26 Advice for learning JavaScript
              • 47:13 Claude Managed Agents vs n8n
              • 54:55 Netflix VOID
              • 59:07 The New Yorker on Sam Altman
              • 62:36 OpenAI's $100 plan and the free-usage playbook
              • 71:05 HTML in canvas
              • Watch the full live stream and more at cmdaishow.com.

                1 hr 21 min
              • #45 Claude Code Leak, Pretext, Cursor 3, Gemma 4

                Kabarza and Samuel start with a Shopify site that turns its own page into an x-ray, 3D view, which leads into why video is still harder to get right on the web than WebGL. Samuel then covers Google's open-weight Gemma 4 family. A tweet calls local frontier AI "effectively free", and he argues that it isn't: most people's laptops still can't run these models usefully. They also look at Pika's video chat agents, and at Cursor 3's agent-first redesign compared with Codex, T3 Code and Claude Code.

                After that come the accidental Claude Code source leak (swear-word sentiment checks, the buddy pet, Kairos, ultra plan and the 187 spinner verbs) and Pretext, Cheng Lou's library for measuring text without the DOM, which Kabarza tests against his own canvas project. The episode ends on a band called Angine de Poitrine and what photography did to painting, then moves on to short-form attention spans, Montessori schooling and the thinking machines of Dune.

                Chapters

                • 00:00 Intro and running order
                • 01:14 Site of the Week: Shopify, video vs WebGL
                • 07:22 Google's Gemma 4
                • 13:11 Can you really run AI locally?
                • 19:44 Pika's video chat agents
                • 22:14 Cursor 3, Codex and T3 Code
                • 39:55 The Claude Code leak
                • 51:34 Pretext and the future of text layout
                • 66:05 Art, music and AI
                • 79:35 Attention spans, Dune and sign-off
                • Watch the full live stream and more at cmdaishow.com.

                  1 hr 25 min
                • #44 TERAFAB, Claude Auto and Computer Use, Framer Shaders, Omma

                  Kabarza and Samuel open with the discontinued Mac Pro, Samuel's plan to replace Frame.io with his own tool, and the YouTube analytics MCP he built in a morning. Kabarza talks through his node-based YouTube helper app and an interactive presentation tool, and what he wants from Claude Dispatch. They then turn to Elon Musk's Terafab chip project and why data centres in space are harder than they sound, before comparing Framer's new shaders with Unicorn Studio and testing Spline's new AI builder, Omma.

                  The second half is about Anthropic. Samuel explains Claude Code's new auto mode and his trouble with permissions, then runs a live computer use task sent from his phone through Dispatch. They discuss the duplicate iMessage problem, replacing OpenClaw with Claude Code, the scheduled tasks still missing, the financial guardrails on computer use, and why Anthropic's steady product stacking leaves OpenAI looking scattered.

                  Chapters

                  • 00:00 Pre-show: Mac Pro, Frame.io and an MCP build
                  • 07:16 Kabarza's builds and a Dispatch wishlist
                  • 13:36 Terafab and data centres in space
                  • 24:53 Framer shaders vs Unicorn Studio
                  • 28:24 Spline's Omma
                  • 36:32 Claude Code auto mode
                  • 43:22 Claude computer use and Dispatch
                  • 48:51 Replacing OpenClaw and scheduled tasks
                  • 55:56 Computer use guardrails and invoices
                  • 60:18 Anthropic's ecosystem vs OpenAI
                  • Watch the full live stream and more at cmdaishow.com.

                    1 hr 8 min
                  • #43 Google Stitch, NemoClaw, Moltbook, Google AI studio, Amazon Code red

                    Kabarza and Samuel admit up front that this one is winged, then work through the week's news: Cursor's Composer 2 and its Codex-style agent view, and a long look at Google Stitch's update. Kabarza argues Stitch has reached good-template territory, which is enough for many people who make a living from design, while Samuel uses a ping-pong bat analogy to explain why Figma's shares fell. They also cover Claude's 1M-token context at standard pricing, the 2x off-peak usage promotion, Google AI Studio's vibe-coding upgrade and NVIDIA's NemoClaw and DLSS 5.

                    After Kabarza leaves, Samuel carries on alone with Alexa's new personalities, Meta delaying its Avocado model, and Amazon adding guardrails to AI-assisted coding after a run of outages. He then covers Moltbook's new terms under Meta, sets up a nightly journal for his own OpenClaw, and finishes with Mistral's Leanstral, OpenAI cutting back on side projects, MiniMax M2.7 and a couple of listener comments.

                    Chapters

                    • 00:00 Locked in and this week's line-up
                    • 03:33 Cursor Composer 2 and a T3 Code clone
                    • 08:00 Google Stitch and Figma's share drop
                    • 21:04 Claude's 1M context and 2x usage
                    • 27:22 Google AI Studio's vibe coding update
                    • 32:13 NVIDIA: NemoClaw and DLSS 5
                    • 39:08 Alexa personalities and Meta's Avocado
                    • 42:28 Amazon's AI coding guardrails
                    • 47:44 Moltbook and an OpenClaw diary
                    • 52:11 Leanstral and OpenAI's refocus
                    • 55:56 MiniMax M2.7 and listener comments
                    • Watch the full live stream and more at cmdaishow.com.

                      1 hr 2 min
                    • #42 Anthropic Marketplace, Meta Acquires Moltbook, Code Review, autoresearch, Cloudflare Crawl

                      Kabarza and Samuel open with a Twitter spat, the new MacBook Neo and a stop-motion Apple advert, which leads into a longer talk about why struggle and failure matter for learning. They then move quickly through the week's news: Webflow buying Vidoso for on-brand AI marketing assets, Anthropic's Claude marketplace for enterprise customers, Meta picking up the people behind Moltbook, Claude Code Review and what it costs, Andrej Karpathy's autoresearch, Gemini in Chrome, Cloudflare's new crawl endpoint and the irony of an anti-bot company selling scraping, interactive visual explanations in ChatGPT and Claude, and Canva's Magic Layers.

                      The second half is about building. Kabarza walks through the AI thumbnail and YouTube helper he is making and his UI tweaking tool, Samuel talks about pricing client work done with AI and his recipe app, and they discuss when a Notion and Claude setup is enough. They finish on where newcomers should start, how Codex, Claude Code, Lovable and Webflow compare, and how they use Claude Cowork and Notion day to day.

                      Chapters

                      • 00:00 Twitter blocks and the MacBook Neo
                      • 05:01 Stop motion and learning through struggle
                      • 11:55 Webflow acquires Vidoso
                      • 17:12 Anthropic's Claude marketplace
                      • 21:18 Meta acquires Moltbook
                      • 24:46 Claude Code Review
                      • 28:54 Karpathy's autoresearch and Gemini in Chrome
                      • 33:52 Cloudflare's crawl endpoint
                      • 41:42 Visual explainers and Canva Magic Layers
                      • 50:36 What we've been building
                      • 68:35 Where to start with AI tools
                      • Watch the full live stream and more at cmdaishow.com.

                        1 hr 22 min

                      About Cmd+AI

                      From the publisher's feed

                      A weekly live show where Kabarza & Samuel rip apart the week in AI, design and dev — building, shipping, roasting, and occasionally agreeing on things. One hour. No slides that aren't doing a bit.