This solo builder runs 24/7 local AI on his own hardware | Alex Finn
Alex Finn, a solo builder, discusses his setup of multiple local AI machines (Mac Studios, DGX Spark, RTX 5090) running 24/7 to power autonomous workflows for software development, security scanning, and market research. He explains why local models unlock unlimited AI usage compared to cloud APIs, and demonstrates his 'software factory' with automated build and review loops using Claude Code.
Summary
Alex Finn runs a fleet of expensive AI hardware in his office, including three Mac Studio computers with 512GB RAM each, a DGX Spark, and a custom-built RTX 5090 machine. His motivation stemmed from discovering OpenClaw in January, which sparked a realization that he wanted personal AI assistants living locally on his machines rather than relying on cloud services. He explains the trade-offs between different hardware options: Mac Studios excel at running large models like GLM 5.2 (Opus-level intelligence) due to unified memory, but process slowly; AI computers like the DGX Spark hit a sweet spot with decent memory and speed; traditional NVIDIA chips like the 5090 offer cloud-like speeds locally but with limited VRAM. Setup is simplified through agents like OpenClaw and Hermes that automatically detect hardware capabilities and install appropriate models across a Tailscale-networked private network.
Alex counters the common objection that cloud subscriptions are cheaper by arguing that local models unlock fundamentally different use cases: 24/7 unlimited token burning that would be prohibitively expensive on cloud APIs. His fleet automation includes continuous security scans every 30 minutes, code reviews every half hour, and a dedicated agent monitoring Twitter, Reddit, Product Hunt, and Hacker News for market signals. He uses a tiered approach, assigning higher-intelligence models to critical tasks and faster models to simpler data-gathering work. His most recent innovation is a 'software factory' with two CloudCode loops: a build loop that automatically creates features based on morning prompts, and a review loop that validates the code. After autonomous overnight work, he reviews the results on Slack and approves merges with a rocket emoji, reducing his day from constant prompting to a single morning conversation and evening review session. He discusses the reliability differences between agents, noting OpenClaw has more impressive moments but stability issues, while Hermes is dependable but less inspiring. His preferred local model is Ornith 1.0, which he claims outperforms Quen 3.6 on coding tasks despite knowing little about its creators.
About this episode
<p><strong>Alex Finn</strong> is an AI builder, YouTuber, and the creator of Vibe Code Academy, a community for people learning to build with AI tools. He runs one of the most ambitious local AI setups I’ve come across: three Mac Studio 512 GB machines, a DGX Spark, and a custom RTX 5090 build, all coordinated through a fleet dashboard he built himself. He’s spent five months figuring out which local models belong on which machines, how to wire them to Claude Code loops, and how to get a software factory running without babysitting it.</p><p><br /><strong>What you’ll learn:</strong></p><ol><li>How Alex chose between a Mac Studio (512 GB unified memory), DGX Spark, and RTX 5090, and what each is actually good for</li><li>Why Tailscale is worth installing even on a single machine, and how it lets one agent manage your entire hardware fleet</li><li>How the build loop and review loop in Claude Code work</li><li>How to allocate tasks by machine and model</li><li>Why unlimited local inference changes the use-case math in a way a $20 cloud subscription never can</li><li>What OpenClaw and Hermes are each best suited for, and why Alex runs five agents total with failover baked in</li></ol><p>—</p><p><strong>Brought to you by:</strong></p><p><a href="https://runwayml.com/howIAI"><strong>Runway</strong></a>—The creative AI platform for images, video, and more</p><p><a href="https://atlassian.com/howiai"><strong>Jira Product Discovery</strong></a>—Prioritize with insights, build with confidence</p><p>—</p><p><strong>In this episode, we cover:</strong></p><p>(00:00) Intro</p><p>(02:58) Alex's hardware stack</p><p>(03:48) What "ambient AI" means</p><p>(04:15) Alex's red-pill moment with OpenClaw</p><p>(07:04) Mac Studio vs. DGX Spark vs. RTX 5090</p><p>(13:24) How to set up local models with no technical knowledge (Tailscale + OpenClaw/Hermes)</p><p>(17:16) Fleet control dashboard: assigning 24/7 tasks across machines</p><p>(20:42) Local models as security scanners feeding Claude Code</p><p>(22:25) How Alex allocates GLM 5.2, Qwen 3.6, and Ornith 1.0 by task</p><p>(24:28) OpenClaw vs. Hermes: the honest comparison</p><p>(26:55) The software factory: build loop, review loop, rocket emoji</p><p>(31:55) Lightning round: favorite hardware, favorite model, prompting style</p><p>(34:46) Where to find Alex</p><p>—</p><p><strong>Tools referenced:</strong></p><p>• Claude Code: <a href="https://claude.ai/code">https://claude.ai/code</a></p><p>• OpenClaw: <a href="https://openclaw.ai/">https://openclaw.ai/</a></p><p>• Hermes: <a href="https://hermes-agent.nousresearch.com/">https://hermes-agent.nousresearch.com/</a></p><p>• Tailscale: <a href="https://tailscale.com/">https://tailscale.com/</a></p><p>• Codex (OpenAI): <a href="https://openai.com/codex">https://openai.com/codex</a></p><p>• GLM 5.2 (z.ai): <a href="https://huggingface.co/zai-org/GLM-5.2">https://huggingface.co/zai-org/GLM-5.2</a></p><p>• Qwen 3.6 (Alibaba): <a href="https://huggingface.co/Qwen/Qwen3.6-35B-A3B">https://huggingface.co/Qwen/Qwen3.6-35B-A3B</a></p><p>• Ornith 1.0: <a href="https://github.com/deepreinforce-ai/Ornith-1">https://github.com/deepreinforce-ai/Ornith-1</a></p><p>• Gemma 4: <a href="https://huggingface.co/collections/google/gemma-4">https://huggingface.co/collections/google/gemma-4</a></p><p>• Playwright (browser testing): <a href="https://playwright.dev/">https://playwright.dev/</a></p><p>• Vercel (preview deploys): <a href="https://vercel.com/">https://vercel.com/</a></p><p>—</p><p><strong>Other references:</strong></p><p>• DGX Spark (Nvidia): <a href="https://www.nvidia.com/en-us/products/workstations/dgx-spark/">https://www.nvidia.com/en-us/products/workstations/dgx-spark/</a></p><p>• Mac Studio (Apple): <a href="https://www.apple.com/mac-studio/">https://www.apple.com/mac-studio/</a></p><p>• How to design AI agent loops: schedules, goals, and subagents in Claude Code and Codex: <a href="https://www.lennysnewsletter.com/p/how-to-design-ai-agent-loops-schedules">https://www.lennysnewsletter.com/p/how-to-design-ai-agent-loops-schedules</a></p><p>—</p><p><strong>Where to find Alex Finn:</strong></p><p>LinkedIn: <a href="https://www.linkedin.com/in/alex-finn-1848684a/">https://www.linkedin.com/in/alex-finn-1848684a</a></p><p>YouTube: <a href="https://www.youtube.com/@AlexFinnOfficial">https://www.youtube.com/@AlexFinnOfficial</a></p><p>X: <a href="https://x.com/AlexFinn">https://x.com/AlexFinn</a></p><p>—</p><p><strong>Where to find Claire Vo:</strong></p><p>ChatPRD: <a href="https://www.chatprd.ai/">https://www.chatprd.ai/</a></p><p>Website: <a href="https://clairevo.com/">https://clairevo.com/</a></p><p>LinkedIn: <a href="https://www.linkedin.com/in/clairevo/">https://www.linkedin.com/in/clairevo/</a></p><p>X: <a href="https://x.com/clairevo">https://x.com/clairevo</a></p><p>—</p><p>Production and marketing by <a href="https://penname.co/">https://penname.co/</a>. For inquiries about sponsoring the podcast, email [email protected].</p>
Key Insights
- Alex argues that the primary value of local AI is not pure ROI but rather the ability to run unlimited intelligence continuously 24/7, which would cost thousands of dollars monthly on cloud APIs like ChatGPT or Claude.
- He claims Mac Studios enable frontier-level intelligence (running GLM 5.2 at Opus 4.8 level) through unified memory architecture, but this comes at the cost of very slow inference speeds (5+ minute responses).
- Alex demonstrates a tiered intelligence strategy where he assigns task complexity to model capabilities: frontier models for critical security reviews, mid-tier models for coding, and fast models for simple signal detection across social media.
- He asserts that OpenClaw delivers more impressive 'wow moments' but suffered month-long periods of breaking with each update, while Hermes is less exciting but more reliable and dependable.
- Alex argues that automation loops represent the future of software development, shifting from constant prompting throughout the day to a single morning ideation conversation and evening review of autonomous work.
- He claims that OpenAI and Anthropic remain vague about their internal loop systems because sharing specific methodologies would allow competitors to replicate their high-productivity code generation capabilities, suggesting loop infrastructure is a key moat.
- Alex employs a failover strategy with multiple agents (three Hermes instances and two OpenClaw instances) where approximately three are perpetually broken, but the remaining two can repair the others, eliminating single points of failure.
- He contends that local model inference speed isn't the limiting factor for many tasks—a security scan taking 24 hours is acceptable if it's continuously scanning code in the background without resource constraints.
Topics
Transcript
What is stacked around your office right now? I have three Mac Studio 512 gigabytes. We got a DGX Spark, as well as a computer. I just built an RTX 5090. Basically at all times of the day, each one of these computers is just burning tokens. The number one pushback I get on all this is your computers are so expensive. Isn't cloud models cheap? Isn't it $20 for a Chad GBT subscription? Well, that's not the point. The point isn't pure ROI. The point is the use cases it unlocks. You now have, because you have AI models running locally, the ability to run unlimited intelligence around the clock 24-7. If you were to do that with a…
Full transcript available for MurmurCast members
Sign Up to AccessMore from How I AI
GPT-5.6 Sol vs. Claude Fable: Why OpenAI’s new model crushes my benchmark
A creator conducts a detailed benchmark comparing OpenAI's new GPT-5.6 models (Sol, Terra, Luna) against Claude Fable across real-world tasks like PRD writing, prototyping, and coding. GPT-5.6 Sol emerges as the superior choice for practical product development, offering better design uniqueness, clearer communication, and more effective collaboration despite Fable's theoretical intelligence.
What a harness is and how to build one with Claude Agent SDK
A harness is custom code built around an AI agent to make it more effective for specific workflows. The speaker demonstrates building a Sentry bug-fixing harness using Claude Agent SDK, explaining how structured constraints, opinionated tooling, and custom prompts enable AI agents to solve complex tasks more reliably than general-purpose coding tools.
How I run autonomous coding agents from my phone with OpenAI Symphony + Linear | Alessio Fanelli (Kernel Labs)
Alessio Fanelli demonstrates how to run autonomous coding agents from a phone using OpenAI's Symphony framework integrated with Linear, showing practical applications from Pokemon card arbitrage to small business automation. He emphasizes moving from agent prompting to agent management through cloud-based orchestration and shares how AI enables efficient scaling of traditionally manual, inefficient business processes.
Sonnet 5 review: I ran 64 generations to find out if it's worth it
The creator develops the 'How I AI Bench,' a custom evaluation framework combining human vibe checks with LLM judging to assess AI models on practical tasks like PRD writing, prototyping, and agentic coding. Testing five frontier models including Claude Sonnet 5, the results reveal significant disagreements between automated scoring and human preference, with Sonnet 4.6 and Gemini 3 Pro ranking highest despite Sonnet 5 being the newest release.