Mostly Technical

148: AI Psychosis Day

· Summarized

Share this episode
Anyone with this link can open it.
Share Link
Start at 0:00

Key takeaways

  • The hosts are deeply engaged in AI experimentation and development, emphasizing local model training and practical applications for automation.
  • They see AI as a tool that can enhance productivity and creativity, especially in visual arts, storytelling, and infrastructure automation.
  • The competitive landscape of AI providers is heavily influenced by capacity constraints, with OpenAI currently leading in performance due to its larger infrastructure.
  • Personal hobbies like photography and media consumption are intertwined with their tech interests, showcasing a balance between work and leisure.
  • Social arrangements like writing retreats and mastermind meetups play a key role in professional camaraderie and creative inspiration.

Summary

Labor Day Reflections and Personal Projects

The hosts discuss how they're spending Labor Day differently from others, emphasizing their dedication to work on this holiday with some humor about drinking and family plans. Ian talks about his experiments building a fantasy world with AI for a physical book, leveraging agents for generating stories and 3D models using Blender, alongside fun ideas for combining art and storytelling. Aaron mentions his personal project of building AI tools, training small models for terminal-based agent management, and integrating AI into his workflow. Both highlight the creative and experimental nature of their work, blending fun with technical innovation.

Technical Deep Dive: AI Model Training and Optimization

Aaron describes training a tiny local AI model on his Mac Studio using label data from his terminal. The goal is to classify terminal states to reduce manual oversight and improve automation—especially for permission prompts and status checks. He details the process: capturing screen states, labeling data, training quickly, and testing the model’s performance. Ian shows enthusiasm for this approach, emphasizing the speed (11-minute training) and potential applications. They explore the challenges of model accuracy, platform support, and the importance of local, fast inference for practical use.

AI and Infrastructure: Models, Capacity, and Market Perspectives

The hosts discuss AI model ecosystems, particularly the performance and capacity differences among OpenAI, Anthropic, and others. They observe that OpenAI currently leads in capacity and perceived quality, despite some criticisms about model capability over time. Aaron and Ian touch on the importance of infrastructure and logistics in AI deployment, with observations on token usage, model pricing, and strategic shifts in AI services. They also reflect on the competitive landscape, capacity constraints, and which providers are gaining or losing ground, emphasizing that AI wars are often about resource management.

Personal Tech and Media — Photography, Cameras, and Media Consumption

Ian recounts his recent weekend taking thousands of horse photos, how he manages large photo libraries with Lightroom, and his camera gear choices, including high-end lenses and gear for specific effects like background blur. Aaron shares his casual approach to media, sitcoms like Seinfeld and Friends, and how he enjoys light entertainment. Both connect their hobbies back to their tech interests, such as optimizing workflows and exploring new gear for content creation.

Upcoming Events and Social Gatherings

The hosts talk about their plans to meet in Las Vegas for a social and mastermind gathering at a cabana. They plan to relax, network, and have fun, emphasizing the informal, laid-back nature of their get-together with friends and colleagues. They mention the allure of cabanas, good food, drinks, and the social atmosphere. The discussion extends to potential exclusive tickets and humorous ideas about funding access, showcasing their enthusiasm for community and shared experiences.

Sponsors

  • Svix · Svix has solved all the hard problems about scaling web hooks so you can avoid triaging failed deliveries and incrementally building features. Check them out at svix.com or link.svix.com/mostlytech.
  • Bento · Check out Bento, the email marketing and CRM platform built for the AI era. They have new transactional email pricing with rates as low as $0.9 per 1,000 emails. Visit bentonow.com.
  • DropInBlog · DropInBlog is a headless blog CMS with a first-class layer of SDK, offering domain, design, and stack integrations with Astro, React, TanStack, and Big Commerce. Find them at dropinblog.com.
  • Valorin Security · Valorin Security provides friendly hackers like Stephen Rees Carter to catch vulnerabilities before attackers do. Learn more at valorinsecurity.com/mt.
  • Typesense · Typesense is an open-source, lightning-fast search engine for your site or app, natively integrated with Laravel Scout. Check them out at typesense.org.

Show notes

Ian and Aaron discuss the release of GPT-6 "Astra", building Solo for Linux, Ian's camera setup, why Outro isn't out yet, and so much more.

Sponsored by Svix, Bento, DropInBlog, Valorin Security, and Typesense.

Interested in sponsoring Mostly Technical?  Head to https://mostlytechnical.com/sponsor to learn more.

  • (00:00) - Happy Labor Day
  • (03:02) - Rippin Astra Tokens
  • (17:04) - Solo On Linux?
  • (26:09) - 4000 Photos of Horses
  • (41:04) - What's New With Outro?
  • (49:27) - Aaron's Training a Model
  • (01:00:32) - No More Astra Left
  • (01:09:06) - Vegas, Baby. Vegas.

Links:

Follow this show on AudioBrevity and new episodes get summarized for you. Create a free account

Couldn't load the audio. A blocker may be stopping the show's tracking links. Open the file