OM DENNE EPISODE
The first ever live Built 2 Scale, streaming at 6am for the Australian drive to work and 3pm for the Austin drive home. Scotty and Matt dig into open source models like Kimi closing the gap on the frontier labs and what that does to cost per token for every business downstream, then get into the coming HR nightmare of token spend as a staff member four months into a European summer ranks number two in company token usage. They cover the shift from token maxing to LLM routing, why work is now completely mobile and completely voice, Granola landing on the Apple Watch and BuildPass following with Superite, and the 45 minute hackathon that produced Tinder for Talent, a LinkedIn scraping recruitment app so creepy Scotty's wife told him to delete it. Plus: live audience questions on AI glasses for site defect walks, whether founders will spend more time talking than typing, and how James Hird could use AI to fix Essendon.
Built 2 Scale | Season 2 Episode 5
TIMESTAMPS:
0:00 First Ever Live Built 2 Scale: Confirm or Deny the Bed in the Office
2:17 LeBron to the 76ers vs Jacinta Allan Leaving Victoria
4:00 Open Source Models Catch Up: Kimi, Fable Going Offline and Back
4:50 Good for Consumers and Businesses, Bad for the Frontier Labs
6:20 Token Maxing to Agent Architecture: Mapping Models to Your Org Chart
8:13 LLM Routing and the Uber Eats Milkrun Playbook for AI Pricing
11:46 The Token HR Problem: Staff Member in Italy Ranks Number Two on Spend
13:00 Do You Get Tokens on Annual Leave? Personal Use, Side Hustles and JD Perks
15:29 Tool of the Week: Scotty's Three Hour Voice Session Building the WOS
18:03 How to Frame It: Ask the AI to Interview You Until It Finds the Real Problem
19:35 Matt's 20 Minute Drive Home Debrief Built as a Claude Project
22:30 Whisper Flow and Why Typing Is the Dumbest Bandwidth You Own
23:30 Granola Hits the Apple Watch: The Passive Note Taker Moment
25:26 Breaking News: Superite Coming to the Apple Watch (Hit the Gong)
28:12 Site Diaries, Defect Walks and Pre Daily Briefings From Your Wrist
31:47 Building in Public: The 45 Minute Hackathon With an AI Engineer Per Team
36:03 Tinder for Talent: Swipe Right, Auto Message, Wife Says Delete It
39:11 Why Software Now Needs Tight Feedback Loops and Deep Verticals
40:34 Modular Software, MCP Servers and Why You Shouldn't Build Your Own CRM
43:44 Two Businesses Converging: The Builder Gets Technical, the Software Team Goes On Site
46:00 Outcomes as a Service: Coming for the Labour Bill, Not the Software Bill
47:31 Audience Q: How Far Away Are AI Glasses Capturing Defects on Site?
49:27 Scotty's Take: Instructions in the Glasses Beats Live Defect Detection
52:00 The World Model Layer: Live Site Data Rewriting the Schedule in the Background
53:12 Construction's Memory Problem: Everyone Goes Home and the IP Leaks Out
55:48 Audience Q: How Would James Hird Use AI to Fix Essendon?
56:54 Audience Q: Will Founders Talk More Than They Type? (Lovable Becomes Lava Bulb)
57:59 The VTT Hack: Flag Your Voice to Text So No One Judges Your Grammar
59:10 Tongue Controlled Mouse: The Weirdest Input Device Yet
This Episode Covers:
- Open source models like Kimi closing on the frontier and what that means for cost per token across every business
- Why hardware getting better and models getting smaller were always going to cross over in the middle
- Moving from token maxing to real agent architecture: mapping models to roles in your org chart by task type
- LLM routing and why not every request needs the most expensive model on earth
- The coming HR fight over tokens: annual leave, personal use, side hustles and tokens as an employment perk
- Voice as the new interface: Scotty's three hour session across a run, a gym, a sauna and a car producing a real artifact
- Framing voice sessions properly by asking the AI to interview you until it finds the actual bottleneck
- Matt's Claude project that turns the 20 minute drive home into a daily debrief and to do list
- Granola launching on Apple Watch and BuildPass bringing Superite to the wrist for site managers
- The 45 minute hackathon format: every team gets an AI engineer, lists ten pain points, ships a working prototype
- Tinder for Talent: the LinkedIn scraping recruitment prototype that replaced a $20K recruiter fee
- Why software companies need modular products, MCP access and deep customer feedback loops instead of point solutions
- AI glasses on construction sites: capture and processing are solved, the real time guidance layer is roughly six months out
- Construction's memory and communication problem as the biggest leak in the industry
KEY INSIGHTS:
- Open source is a consumer win and a frontier lab problem: Cheaper models plus better consumer hardware means prices have to come down. The only losers are the labs subsidising inference
- Token spend becomes a line item next to headcount: Finance teams already track salaries. Companies like RAMP are tracking AI spend. Expect role cost to mean salary plus agent cost
- Route the request, don't max the model: Sending a basic task to the most expensive model is the AI equivalent of Uber Eats VC subsidies. That era ends the moment margins matter
- Work is now mobile and voice first: A three hour session across a run, a gym and a car produced a foundational business artifact. The thumb to keyboard bandwidth is the bottleneck, not the thinking
- The Apple Watch finally has a real use case: Passive capture is less awkward than a phone on the table and frees your hands on site. Voice input is the killer feature, not health tracking
- Hackathons are 80/20 machines: Forty five minutes gets you a working prototype and a real idea. The last 20% takes 80% of the time, but you can't invest in an idea nobody has proven yet
- Don't rebuild the wheel, build what makes you different: A construction business shouldn't build its own CRM or payroll. Build the workflows unique to you on top of infrastructure that already exists
- Outcomes as a service is the next software frontier: Safety paperwork is hard to sell, finance is easier, winning tenders sells itself. The pitch is converting labour bills into tokens
- Instructions in the glasses beat defects in the glasses: Training vision models to spot a 5mm tolerance issue is expensive. Scrolling the SOP into a tradesman's eyeline scales quality across a workforce today
- The site as a world model: Live capture doesn't just flag defects, it feeds a system that reruns the schedule, spots downstream clashes and reschedules deliveries in the background
- Construction's leaky bucket: Everything seen and heard on site walks home at 5pm and never reaches tomorrow's crew. Any observation not shared to the project brain is lost IP
- Curious beats credentialed: The people getting the most out of these tools aren't waiting to be told what AI can do, they're assuming it can and finding the edges the hard way
About Built 2 Scale:
Weekly analysis of artificial intelligence business strategy, robotics adoption, and technology market dynamics. Hosted by Scott Wilcox (Australia) and Matt Perrott (Austin), providing actionable insights for founders, investors, and technology executives navigating the AI transformation.
Subscribe for AI business intelligence: /@built2scale
Available on Spotify | Apple Podcasts
#AI #OpenSource #Kimi #Claude #Fable #TokenSpend #LLMRouting #VoiceAI #Granola #AppleWatch #Hackathon #BuildPass #Superite #AIGlasses #Construction #Built2Scale #LiveStream #TechNews
