The Startup Ideas Podcast

Grok 4.5 is a bigger deal than Fable

In this episode I bring Nick Vasilescu, co-founder of Orgo, back on the show to unpack the buzz around Grok 4.5. Nick makes the case for treating Grok 4.5 as a genuine AI co-founder inside harnesses like Hermes and OpenClaw, and he proves it live: spinning up cloud computers, wiring in tools, and building a full startup from idea to landing page to outreach. We race Grok 4.5 against GPT 5.6 Sol, tour Nick's agent stack, and talk through the cost paradox of a model this fast and cheap. Listeners walk away with a concrete playbook for standing up their own always-on agent today.Get Nick’s Agent Template Stack: https://startup-ideas-pod.link/nicks-stackTimestamps00:00 – Intro01:40 – Why Grok 4.5 release matters03:39 – Automation versus a co-founder05:16 – Setting up Hermes and Grok 4.5 on Orgo09:01 – Why Orgo to manage Agents11:23 – Grok 4.5 Cost discussion14:02 – Grok 4.5 Fast Execution and Unlock16:20 – The Agent tool belt19:13 – X MCP for trends20:37 – vidIQ for outliers and thumbnails22:11 – Finding new startup ideas26:15 – Grok 4.5 versus GPT 5.6 Sol30:56 – Ranking and Reviewing the startup ideas34:06 – The AI agency opportunity38:46 – Thumbnails over Telegram40:10 – Reviewing AI Agency Landing Page41:58 – Vertical MCPs and agent startups43:36 – Skill graph and the offer45:25 – Reviewing the Thumbnail Generated46:42 – Email Outreach Campaign48:07 – Reviewing Market Insight 1-Pager50:45 – From a Camry to a Ferrari52:12 – Reviewing Cold Email Outreach Sequence53:22 – Closing thoughtsKey PointsGrok 4.5 delivers Opus 4.8-level intelligence at a fraction of the cost and roughly 10-15x the speed of Fable.I learn to treat the model as a co-founder by handing it email, a phone number, a debit card, memory, and every connector that matters.Nick runs agents on Orgo cloud computers so they stay online, textable, and ready around the clock.Live, Grok 4.5 builds a landing page in about 40 seconds and wins on design and copy for me over GPT 5.6 Sol.The stack ships from idea to website, offer, thumbnail, and cold-email sequence in a single session.Nick's take: costs keep dropping while speed and intelligence keep climbing, so building your agent now compounds overnight.The #1 tool to find startup ideas/trends - https://www.ideabrowser.comLCA helps Fortune 500s and fast-growing startups build their future - from Warner Music to Fortnite to Dropbox. We turn 'what if' into reality with AI, apps, and next-gen products https://latecheckout.agency/The Vibe Marketer - Resources for people into vibe marketing/marketing with AI: https://www.thevibemarketer.com/FIND ME ON SOCIALX/Twitter: https://twitter.com/gregisenbergInstagram: https://instagram.com/gregisenberg/LinkedIn: https://www.linkedin.com/in/gisenberg/FIND NICK ON SOCIALYoutube: https://www.youtube.com/@nickvasilesInstagram: https://www.instagram.com/nickvasilescu/Personal Website: https://www.nickvasilescu.com/

We Tested OpenAI's GPT 5.6 for a Month

In this episode I sit down with Dan Shipper to see how he runs his work and personal life on OpenAI's Codex Desktop with the 5.6 model. He walks through his card-based email setup, daily feeds for his company and Slack, and the in-app browser that lets his agent collaborate with him inside tools like Proof. We build a small SaaS app live, called Turnaround, and use it to explore why maintenance is the real product in the AI era and where Codex-native software heads next. Along the way Dan shares his pirates-versus-architects framing, his approach to fine-tuning a copy-editing model, and the patterns — pulses, Mailroom, and router threads — that hold his system together. The throughline: pick one simple win, let context do the heavy lifting, and manage the system instead of running every task by hand.Learn how to get customers with AI Agents: https://startup-ideas-pod.link/GTM-agents-IBTimestamps00:00 – Intro01:16 – Codex and GPT-5.6 Overview03:40 – Training your own model: the step after skills04:49 – Automating Email, Slack, Meeting Notes with GPT-5.608:53 – Why GPT-5.6 sharpens the results10:26 – The light bulb moment with Codex15:05 – Building Turnaround live: a maintenance badge18:00 – GPT-5.6 vs. Fable: A tier and S-plus tier19:34 – LFG and goal: looping toward a finished build24:28 – Huge Opportunity: Codex-native apps29:33 – The design checkpoint and the "warm paper" quirk31:32 – Local models34:04 – From 70% to 100%: pirates and architects37:22 – Mailroom: giving Codex its own email address40:58 – Getting started: download, grant access, explore43:07 – Record and Replay: turning tasks into skills44:37 – Closing Thoughts: Start small and build over timeKey PointsCodex Desktop plus the 5.6 model runs as a full operating system for knowledge work — email, research, and building software from one surface.Context is the multiplier: an agent wired into your computer and the web turns every inbox and feed into cards with a clear next action.Maintenance is the real product in the AI era, now that anyone can one-shot a first version.Codex-native SaaS — software you and your agent share inside the in-app browser — opens a fresh category with healthier margins.A live build of Turnaround, a maintenance-status badge, reaches about 70% in one pass; an architect carries it the rest of the way.Start with one simple win, grow the system over time, and let curiosity lead the way in.The #1 tool to find startup ideas/trends - https://www.ideabrowser.comLCA helps Fortune 500s and fast-growing startups build their future - from Warner Music to Fortnite to Dropbox. We turn 'what if' into reality with AI, apps, and next-gen products https://latecheckout.agency/The Vibe Marketer - Resources for people into vibe marketing/marketing with AI: https://www.thevibemarketer.com/FIND ME ON SOCIALX/Twitter: https://twitter.com/gregisenbergInstagram: https://instagram.com/gregisenberg/LinkedIn: https://www.linkedin.com/in/gisenberg/FIND DAN ON SOCIALX/Twitter: https://x.com/danshipperYoutube: https://www.youtube.com/@EveryInc/videosEvery: https://every.to/

AI Agents are the new SaaS

In this solo episode I lay out why I believe building agents is the new SaaS: software is shifting from helping you do the work to doing the work with you. I walk through a full playbook — find a niche, pick a workflow with a paycheck attached, shadow the human, spec the agent, build the minimum useful version, sell a pilot like labor, then productize the repeatable parts. I share live market examples like Slang AI for restaurants and Same Day for home services, plus pricing models and a distribution strategy built on workflow teardowns. I close with a 30-day, zero-to-100 plan for launching an agent-first business. This one is for anyone eager to build with AI or simply become more productive.Timestamps00:00 – Intro01:38 – Building Agents is the new SaaS04:11 – Pick a valuable workflow06:12 – Shadow the Human First09:34 – Build the Minimum Useful Agent12:50 – The wrapper makes it SaaS15:50 – Sell the Pilot Like Labor (and Pricing)18:37 – Own the workflow21:45 – The Zero-to-100 Plan in 30 Days24:14 – Closing ThoughtsKey PointsAgent SaaS sells work as a service; the product is the job itself, priced like labor.Start with a workflow that already carries a paycheck: high frequency, clear finish line, existing software, learnable edge cases, and felt pain.Shadow a human across 10–20 real jobs before you write a single prompt — the detail is the product.Ship the minimum useful agent — draft-and-approve, triage, coordinator, or bounded action — and earn autonomy over time.The wrapper (logs, approvals, evals, analytics) creates trust and turns automation into real SaaS.Win distribution with workflow teardowns: show the old way, show the agent way, sell the painkiller.The #1 tool to find startup ideas/trends - https://www.ideabrowser.comLCA helps Fortune 500s and fast-growing startups build their future - from Warner Music to Fortnite to Dropbox. We turn 'what if' into reality with AI, apps, and next-gen products https://latecheckout.agency/The Vibe Marketer - Resources for people into vibe marketing/marketing with AI: https://www.thevibemarketer.com/FIND ME ON SOCIALX/Twitter: https://twitter.com/gregisenbergInstagram: https://instagram.com/gregisenberg/LinkedIn: https://www.linkedin.com/in/gisenberg/

“Learn AI” Is Bad Advice. Learn These Instead

In this solo episode, I lay out the six skills I believe stay valuable as AI grows more capable. I chose these six because each one is open to anyone, each one starts this weekend, and each one rises in value as AI improves. I walk through agents and local models, distribution, robotics, curation, the builder distributor, and IRL community building, with one concrete first rep for every skill. My goal is to hand you one simple, clear map of where the world is heading and exactly how to begin.Timestamps00:00 – Intro00:57 – Skill 1: Running AI Agents and Local Models04:51 – Skill 2: Marketers Who Build Distribution09:03 – Skill 3: Robotics Engineers Who Build and Source Hardware14:29 – Skill 4: Curators Who Yap and Make Short-Form Video19:05 – Skill 5: The Builder Distributor23:11 – Skill 6: IRL Community Builders27:34 – Build Your Skill StackKey PointsI chose these six skills because each one rises in value as AI improves.Skill 1 is the grown-up version of prompt engineering: I design an AI worker with context, tools, memory, permissions, and a goal.Distribution beats posting, so I learn where attention already lives and turn it into trust before I sell.Hardware is the new frontier: cheap arms, open-source robot learning, and supplier sourcing put robotics within my reach.As the builder distributor, I ship the product and win the attention in one loop, which makes the one-person company real.Real rooms grow scarce and valuable, so I build belonging, trust, and context as my edge.Numbered Section SummariesThe Premise: What Stays Valuable as AI Improves I open by picturing a near future where AI builds and writes almost anything, then ask which skills hold their value. I narrow it to six skills that anyone can start this weekend, each one climbing in value as AI gets better.Skill 1 — Agents and Local Models I describe the move from typing prompts to designing a small AI employee with context, tools, permissions, memory, a goal, and a way to check its own work. I add local models with tools like Ollama and LM Studio so you learn which jobs want a giant brain and which jobs want a reliable worker, and I suggest building a daily briefing agent with three sources as your first rep.Skill 2 — Marketers Who Build Distribution I explain that distribution runs far deeper than posting: it means knowing where attention already lives and the exact words people use to describe their problem. The winning marketer becomes part researcher, storyteller, media operator, and community builder, and the first rep is a distribution map plus 20 hooks for a single idea.Skill 3 — Robotics Engineers Who Build and Source Hardware I share my big insight: the last decade rewarded moving pixels, and the next decade rewards moving atoms too. With cheap cameras, low-cost arms like the SO-100 / SO-101, open-source work like Hugging Face LeRobot, and small VLA models, I suggest assembling a low-cost arm, teaching it one boring task, documenting every failure, and learning supplier sourcing on Alibaba.Skill 4 — Curators Who Yap and Make Short-Form Video I cover the curator who watches the timeline and says "this matters because…," translating new models, launches, and news for a specific niche. The algorithms reward raw, authentic yapping that carries a real take, and my rep is a seven-day curation sprint paired with a taste file of hooks, analogies, and titles you love.Skill 5 — The Builder Distributor I make the case that AI compresses the old build-versus-sell split into one person who prototypes the product, writes the launch thread, records the demo, DMs the first users, and iterates. The loop is the whole game, and my rep is a 48-hour loop: build the smallest version of one problem, then create 10 pieces of distribution before you feel ready.Skill 6 — IRL Community Builders I close with the old-school skill that grows more valuable as work moves to agents and feeds: real rooms full of ambitious people. Scarcity moves toward belonging, trust, and context, so I suggest hosting six to eight people around one sharp question and sending a recap that turns the room into a networkThe #1 tool to find startup ideas/trends - https://www.ideabrowser.comLCA helps Fortune 500s and fast-growing startups build their future - from Warner Music to Fortnite to Dropbox. We turn 'what if' into reality with AI, apps, and next-gen products https://latecheckout.agency/The Vibe Marketer - Resources for people into vibe marketing/marketing with AI: https://www.thevibemarketer.com/FIND ME ON SOCIALX/Twitter: https://twitter.com/gregisenbergInstagram: https://instagram.com/gregisenberg/LinkedIn: https://www.linkedin.com/in/gisenberg/

GLM 5.2 Clearly Explained (and how to set it up)

In this episode I sit down with Amir to get tactical about running local AI models as part of a daily workflow. We center on GLM 5.2 from ZAI, how it stacks up against frontier models like Opus 4.8, and how a fusion approach lets you sequence a heavy thinking model with a lighter execution model for the best output at the lowest cost. Amir walks through setup in Cursor and Codex via OpenRouter, shares real token-cost math, and demos GLM 5.2 refining a live app. By the end you will know how to start today, where local models shine, and how model chaining keeps spend in check.Timestamps00:00 – Intro02:09 – GLM 5.2 and Z AI04:01 – Specs: 1M context and Terminal Bench 2.105:22 – Making sense of benchmark scores06:42 – Setup in Cursor or Codex with OpenRouter10:18 – Local model upside: buy a machine, run tasks11:42 – Token cost: 44 cents versus $2.3813:36 – Future-proofing with an upfront hardware bet & The Uber subsidy analogy16:49 – Model chaining and the vision workaround19:23 – Token maxing vs routing tasks to the right model20:54 – Answering the "cost is irrelevant" crowd21:59 – Closing thoughtsKey PointsGLM 5.2 ships with a 1M-token context window and scores 81 on Terminal Bench 2.1, landing about four points behind Opus 4.8.A fusion approach (a term OpenRouter coined) sequences models: plan with Opus, execute with GLM 5.2, review with Composer 2.5 or Codex 5.5.Running GLM 5.2 in the cloud through OpenRouter costs roughly 44 cents for a task that runs about $2.38 on Opus 4.8 — close to a 5X saving.You can start today with credit-based access: load $20 in OpenRouter and route tasks to the right model.For images, Amir uses Opus 4.8 to read screenshots and describe them, then hands the layout to GLM 5.2 to act on.Teams are shifting from token-maxing to output-maxing, making model governance and chaining the smart playThe #1 tool to find startup ideas/trends - https://www.ideabrowser.comLCA helps Fortune 500s and fast-growing startups build their future - from Warner Music to Fortnite to Dropbox. We turn 'what if' into reality with AI, apps, and next-gen products https://latecheckout.agency/The Vibe Marketer - Resources for people into vibe marketing/marketing with AI: https://www.thevibemarketer.com/FIND ME ON SOCIALX/Twitter: https://twitter.com/gregisenbergInstagram: https://instagram.com/gregisenberg/LinkedIn: https://www.linkedin.com/in/gisenberg/FIND AMIR ON SOCIALHumblytics: https://humblytics.com/?via=communityX/Twitter: https://x.com/amirmxtYoutube: https://www.youtube.com/@amirmxt