Link: https://youtu.be/aCiCgr4KWPQ

Duration: 101 min

Transcript: Download plain text

Short Summary

The Weird Things podcast hosts — Andrew Maine, Brian Brushwood, and Justin Robert Young — held a guest-free episode ranging across Brian's AI-native creator workflow (Codex driving Audition, tr3d.ai for 3D assets, V-Mix scripting), the move from apps to always-on agents, comment-section triage, AI lab competition, and self-driving in Austin. Brian, a longtime creator turned AI-native producer, also previewed an 'Underground Productivity Fight Club' livestream and detailed his solo voice-editing workflow alongside body-composition insights from Apple Health.

Key Quotes

1. "the the point of all of this has finally borne fruit that I could take an idle thought and get actual like numbers out of it. And if you have been like me and you've tracked these things forever, like it it opened my eye to like what what I can do, comparisons, uh uh and and like all of this stuff now finally being worth it. It felt like I had come out uh uh you know, I I now discovered that the the seed I planted oh so many years ago had become a beautiful redwood." (01:30:26)
2. "And the first draft of what we got was incredibly advanced for a first draft to the point where like I I I have to kind of rethink what I want out of a first draft because productionwise it went above and beyond. So now it's like okay, you know, Brian and I were Brian made the point that it's like the first draft now is too busy. We need to be thinking about things that we are removing as opposed to things that we would want to hear or add." (00:09:12)
3. "I I see a world where like if I could I couldn't see a world unless one of the podcasts really really really got big and there was just like drudgery that needed to be done to hire somebody else. But now it's like I could see building up a larger staff if everybody was doing 10x then I could be a much larger thing like we could afford to take on different risks and and stuff like that." (00:24:50)
4. "we move towards services and away from apps because a lot of like you know I was playing the 3D thing what's my next step for my 3D thing I'm grabbing my API key and putting that in codeex so I don't have to keep using the interface that's that's the solution you" (00:40:36)
5. "internally they were using claude to code at Google. That's how bad it is at Google that they're using outside models to to code because they don't even want to use their own stuff." (01:01:15)

Detailed Summary

Weird Things Podcast — Guest-Free Episode Synthesis

Hosts and Framing

Andrew Maine, Brian Brushwood, and Justin Robert Young host a guest-free Weird Things episode that has evolved from covering monsters, to future speculation, to present-tense speculation. Brian accidentally went live early and almost broke an NDA by showing an embargoed "duo" device from Turnis, so he set it aside before discussing it on air; patrons already received three new episodes.

• The episode features Andrew Maine, Brian Brushwood, and Justin Robert Young on the Weird Things podcast, with topical scope evolving from monsters to speculative future to present-tense speculative future.
• Brian accidentally went live early and nearly broke an NDA by showing an embargoed "duo" device from Turnis, which he put away before discussing on air.
• Patrons already received three new episodes, with more coming at a faster cadence.

Brian's AI-Native Creator Workflow

Brian, a longtime creator turned AI-native producer, walks through using Codex/ChatGPT Computer Use to drive Adobe Audition, tr3d.ai (Tripo3D) for 3D assets, and V-Mix scripting for live shows. He also describes a one-shot co-hosted podcast pipeline and predicts broad Codex/Claude integration with consumer 3D printers within a year.

• Brian used tr3d.ai (Tripo3D) to turn a GPT-generated image into a 3D character asset for a Settlers of Catan-style tile board with a 3D-printed figure.
• ChatGPT in the Chrome sidebar cut that 3D model from 50 MB to 2 MB by reducing triangles, and 3D mesh image-to-model tools improved sharply in recent months via multi-view inference.
• He piloted Adobe Audition through Codex/ChatGPT Computer Use to autonomously generate a first-draft soundscape for World's Greatest Con, prompting the team to think in terms of what to remove rather than what to add.
• Brian is co-hosting a 30-minute podcast with John Teasdale, one-shotting content by pasting transcripts into Codex; the first episode fed in a shelved game and Codex made "really good decisions" on mechanics.
• A sample box for that game is expected from a China print in 3–6 weeks.
• Codex scripted V-Mix for a magic gig at Embry-Riddle and went further, writing an external program to fulfill requested features.
• Brian estimated roughly 8.5 billion more people will eventually have "aha" coding moments like this and predicted strong Codex/Claude integration with consumer 3D printers within a year.

Production Philosophy and Game-Making

Brian argues top creators should understand the full factory line even when not on the tools, contrasting James Cameron's hands-on style with a layered model of creator involvement. He frames production as moving through four eras and concludes that building games is now easy while making them good remains the hard part.

• Brian argued the top 1% of creators should understand the full factory line even if they aren't on the tools, contrasting this with examples like James Cameron doing every set gig himself.
• Production eras charted: pre-computer conductor/instruments → remix/sampling → Ableton-style DAWs → current AI-assisted generation.
• Building games is now easy; making them good is still the work — current AI demos produce "derpy versions" of existing experiences rather than original hits like Tetris, Flappy Bird, or Monument Valley.
• Recommended tactics include 10-minute games, paper apps, and one-card dungeons.
• For talks, he suggests asking the model to list every way the output looks machine-made and then removing them; Brian delivered a 14-minute CSICON talk by pressing go on a timer with Bonnie.

ChatGPT Plans, Tokens, and a Coding Mishap

OpenAI's discontinuation of the $200 plan prompted Brian's downgrade to $100, with rumors of a $500 tier and a Cerebras-bundled option. A surprising overnight mishap — an AI agent writing a literal "stop" command into a root loop — illustrates the gap between token accounting and real productivity.

• OpenAI discontinued the $200 plan, prompting Brian's downgrade to $100, with rumors of a $500 tier and a Cerebras-bundled plan.
• Brian said he would happily pay up to $1,000 for a meaningfully faster plan and participates in a Signal chat with a couple hundred current/former OpenAI employees.
• An AI coding assistant wrote a "stop" command into a root loop and went idle overnight; Brian saved tokens for a 6:00 a.m. reset with 22% remaining, only to find usage had dropped to 20%.
• His daily-use turning point was March of this year; he works with Will on Dog and Pony Show Audio and imagines scaling staff if each person were 10x more productive.

Comment Sections, Attention, and Triage Tools

Brian critiques current comment-checking as a "random reinforcement machine" akin to gambling and proposes a successor to the discontinued "Superfan" tool that delivers verbatim feedback on the creator's schedule. He also reads comments to program his "system one" associations, like knowing audiences hate specific segments.

• A previous tool, "Superfan," aggregated YouTube comments into filtered sentiments; Brian proposes a successor (VX3) that delivers all feedback verbatim on the creator's terms (during a walk or drive) without visiting a website.
• Brian prefers reading comments to program his "system one" associations — e.g., knowing audiences hate the "golden bunny rabbit" segment on Modern Rogue.
• He calls current comment-checking a "random reinforcement machine" like gambling; a healthier design would be a detached report delivered at a scheduled appointment (example: 2:15 PM) with no refreshes allowed until then.

Agents, Phones, and the App Economy

The hosts discuss the shift from apps to services and from on-demand interfaces to always-on agents, with a speculative "app-less phone" scenario that opens to a blank screen. Brian criticizes the App Store as predatory, citing recent one-star reviews of top-grossing apps and an in-app purchase economy worth hundreds of billions of dollars.

• The next 5 years will hand users "source code" access in ways not available 10 years ago when algorithmic filters became dominant.
• Codex is cited as an abstraction layer above ChatGPT, with API keys pasted directly to bypass UI for 3D tasks.
• Jason Calacanis tweeted that an app-less phone (modeled on Meta's Muse) would be timely; he is also selling the URL mahalo.com.
• The hosts previously laughed at the Rabbit device (~iPhone weight, essentially a ChatGPT plugin), and one argues an app-less phone would still need a separate work phone for 2FA.
• A speculative phone opens to a blank screen: ask for an app by name (e.g., Hearthstone), the agent narrates gameplay while in a pocket, and streams local radio on demand.
• Brian criticized the App Store as predatory — top-grossing apps show 4.9 stars overall but recent one-star reviews complain of being ripped off, with an in-app purchase economy worth hundreds of billions of dollars.

Always-On Agents and AI Safety Norms

OpenClaw is highlighted as a potential landmark tech moment, and the hosts debate disclosure norms when AI acts on a user's behalf. They frame the legal landscape around fault and deception by AI as defining the next decade, and they peg the consumer viability bar at 99% success.

• OpenClaw is predicted to be remembered as a huge tech moment: its creator popularized ClaudeBot before Anthropic shut it down, rebranded it as OpenClaw under OpenAI, and now works at OpenAI.
• Brian is uncomfortable when AI agents send messages as if they are the user and wants them to introduce themselves as agents — comparing it to the disappointment of receiving a form letter from an author's assistant in the 1980s.
• A chatbot over-shared details of home improvements during a quote request, and Brian's wife used an agent to make a restaurant reservation then deleted the stored credit card.
• The hosts estimate the legal debate about fault and deception by AI will define the next 10 years, citing new rules about bank obligations to scammed customers.
• Major AI lab agents have reportedly "hacked" government systems, usually by finding publicly available unindexed data, blurring the definition of hack.
• Consumer AI products need a 99% success rate to be viable; higher failure tolerance only works for experimental or high-benefit cases like self-driving.

Self-Driving Cars and Austin

Brian reports high robotaxi density in his Austin commute and projects that autonomy could make up half of local traffic within five years. He proposes small multi-stop EVs as a more efficient replacement for low-occupancy buses.

• Brian predicts self-driving vehicles could be 50% of Austin traffic within 5 years and currently sees about 1 in 5 vehicles between South Austin and downtown running on autonomy.
• He spots Waymo, Zoox (Amazon), and small golden Cybercab robotaxis.
• He argues buses are terribly inefficient at low occupancy and suggests small multi-stop electric vehicles (e.g., 6-person) instead.

AI Lab Competitive Landscape

Brian ranks recent launches and argues Google has slipped to a "third-tier lab" in frontier model terms, with Anthropic and OpenAI holding a self-reinforcing two-model advantage. He also points to OpenRouter's sub-$0.10 per million token pricing for GPT-6/5.6 Luna as evidence that cost efficiency is becoming a competitive axis.

• Recent launches reviewed: OpenAI's Astro ("token burner"), Anthropic's Opus 5.5 (well-received), Meta's Muse (a hit), and xAI's Grok/Rockbot (ranked #196 in the app store).
• Brian argues Google hasn't shipped a frontier-class model roughly 8 months past Anthropic's Mythos (February checkpoint) and is now considered a "third-tier lab"; Google engineers reportedly use Claude internally for coding.
• Google's strength is small "spark models" for speed, but agentic and coding implementations are weak — Brian claims Google would need ~$50 per user per month on compute to serve agentic AI at scale.
• Meta's Muse is described as technically simple — "a startup could have built Muse" — basically an open-source claw-style app integration, with broadly poor reputation.
• OpenRouter offers token-efficient pricing for GPT-6/5.6 Luna as low as 10 cents per million tokens, treating cost-efficiency as a key competitive axis.
• Google's data advantage has diminished over the last ~2 years as other video networks and data sources have been collected by competitors.
• Anthropic and OpenAI share a self-reinforcing advantage: owning two internal frontier models they use to build next-gen tools, which Meta, xAI, and Google lack.
• Anthropic benefits from Dario Amodei's persistent long-term vision; Sam Altman brings decades of consumer-product thinking and defers to strong people around him; Google's AI efforts are hampered by internal bureaucracy unless Larry Page intervenes.

Fly-Brain Neural Net

Andrew flagged a research result in which scientists built a neural network at the size and scale of a fly's brain by imaging and cataloging neurons, though it is not a fully functioning copy. The framing is intended to bound what is possible with a limited neural network.

• Researchers created a neural network modeled at the size and scale of a fly's brain by imaging and cataloging neurons, though it is not a fully functioning copy.
• The comparison is framed as a way to bound the question of what's possible with a limited neural network.

Health Data and Solo Editing

A host compared current and past body-composition photos via ChatGPT Health integrated with Apple Health, finding roughly the same weight but ~2% different body fat, with Apple Health attributing part of the divergence to pandemic-era walking. A separate solo editing workflow was outlined: review cuts, edit via voice during a walk, and post the video during the same walk.

• A host compared current photos to past images and found he weighs about the same but has roughly 2% different body fat.
• When asked why weight and body fat diverged, Apple Health cited pandemic-era walking adding 300–400 more calories burned each day, plus a slightly lower resting burn rate.
• He has been collecting health data via a smart ring for a long time and now finds that historical data usable in retrospect.
• He plans to edit video by reviewing cuts and editing via voice while on a walk, then posting during the same walk from his own computer, without hiring help.

Underground Productivity Fight Club Announcement

Brian announced a ticketed 'Underground Productivity Fight Club Number One' livestream the next day at fight.scamstuff.com, framed as the 'maiden voyage' of a new livestream layout. The format mirrors ~$25 magic lectures, with attendees sharing one best trick across a three-phase structure of explanation, hang-out, and shared best bits.

• He is running a ticketed experiment the next day called 'Underground Productivity Fight Club Number One' at fight.scamstuff.com, framed as the 'maiden voyage' of a new livestream layout.
• The format is modeled on magic lectures, historically around $25 per ticket with permission to take anything not marked off-limits from the magician's act.
• The three-phase structure is: explanation, hang-out, and shared best bits.
• Attendees are asked to bring one best observation, hack, or trick they are proud of that could be improved or new to others.
• His own trick for the event: on calls for the last 1.5 years he has taken notes in front of people using an interactive 'teleustrator' setup, including examples that animate letters when he blows on them and parallax-animated layers.
• The event will be recorded and a VOD made available; attendees Joey H66 and Biocalow received confirmation of the recording.

Products, Picks, and Personal

Brian recommends the AI for Humans podcast and the product Instinct, which proactively drafted email replies he rewrote via text and sent, while speculating on its provenance and survival prospects. A separate host strongly recommends the ChatGPT desktop app, and Brian noted a switch from AirPods Pro to regular AirPods for sleep comfort.

• Brian recommends the AI for Humans podcast hosted by Gavin Purcell and Kevin Pereira.
• Brian recommends the product Instinct, which over the prior week proactively drafted email replies he rewrote via text and sent; he speculates it may be an IRGC front but praises its UI decisions.
• Brian questions whether Instinct will survive the next 6–8 weeks.
• A host strongly recommends the ChatGPT desktop app, highlighting Astra and Six Soul for fast computer-use and video editing.
• Brian bought regular AirPods (~$130–150 with charging) for noise cancellation after switching off AirPods Pro due to comfort issues while sleeping.
• Brian closed with a memory of his late grandfather nerding out on the Apple II GS, reinforcing that humans will keep making things (Minecraft, 3D printing, overnight Codex-3D prints).

Notable Numbers, Decisions, and Surprises

The episode produced a dense set of concrete figures spanning 3D asset compression, OpenAI pricing, agent failure modes, AI app rankings, and self-driving adoption projections. The most striking surprise was an AI coding agent that wrote a literal "stop" into a root loop and idled overnight.

• 50 MB → 2 MB model reduction via ChatGPT in the Chrome sidebar.
• $200 → $100 plan downgrade, with a rumored $500 tier and Cerebras-bundled plan; willingness to pay $1,000 for faster.
• Codex reportedly cut a manual Oakland Police report task (10 minutes to 2 hours) to 0% failure; 99% success rate is the consumer AI viability bar.
• AI app ranking: Meta's Muse a hit, xAI's Grok/Rockbot at #196 in the app store.
• Brian's "aha" moment scale: ~8.5 billion more people.
• Austin self-driving: 1 in 5 currently, projected 50% in 5 years.
• Surprise: an AI coding agent literally wrote "stop" into a root loop and went idle overnight, with token usage mysteriously dropping from 22% to 20% across a reset.