← All episodes

Why Your Company Should Own Its AI Model | E2278

| 13 products mentioned
Watch on YouTube enterprise ai small language models reinforcement learning autonomous delivery drones regulatory frameworks ai in filmmaking protein optimization

Companies shouldn't rely on frontier AI models for every task—instead, they should build and own small language models (SLMs) trained on their proprietary data for dramatic cost and speed improvements. Aragon (featured via CEO Josh Sirota) demonstrates this vision with a "company world model" that uses reinforcement learning to continuously update an SLM overnight, enabling autonomous agents that know your company's context without expensive token-heavy API calls. The episode also covers autonomous drone delivery infrastructure (Ikiru) and explores how AI is reshaping creative industries, from filmmaking to recruiting.

Key takeaways
  • Building a private SLM trained on your company's data and stored in model weights can reduce inference costs by 50-70% and speed up task execution 10x compared to frontier models like Claude or GPT-4.
  • Reinforcement learning for continuous learning allows SLMs to update themselves overnight based on user interactions, creating a self-improving "company world model" that adapts to your business without expensive retraining cycles.
  • Part 108 FAA regulations for beyond-visual-line-of-sight (BVLOS) drone operations unlock massive logistics efficiency by allowing one remote operator (or system) to supervise multiple autonomous drones, analogous to jumping from Tesla Autopilot (Level 2) to Full Self-Driving (Level 4).
  • Practical SLM implementation starts by identifying high-volume, repetitive workflows (e.g., email triage, quote generation) and pulling company data into vector databases using tools like Chroma and Ollama, then training the model nightly on that fresh context.
  • AI-generated backgrounds and lighting can reduce feature film budgets from $200–300M to ~$70M by shooting actors entirely on green screen with AI handling all production design, enabling more diverse storytelling without sacrificing creative control.
  • Tinned fish—especially mackerel and anchovies—offers 28g protein for ~160 calories, making it an efficient high-protein option for people managing weight or on GLP-1 medications who need to preserve muscle mass.

Recommendations (8)

Kimmy uses

"As far as the model that we've found that has like, you know, very good kind of base is Kimmy. So, we're also very excited for all the announcements that um they're going to be making."

Josh Sirota · ▶ 10:22

Ikiru recommends

"Great film about a guy who gets cancer. This will be my off duty for today... he decides, I'm going to live a little."

This Week in Startups · ▶ 1:00:20

Living recommends

"2022 living. You should check it out. It is a British remake of the same story of Icar."

This Week in Startups · ▶ 1:02:14

"This is the one right now that I'm I'm really feeling. It's called uh Wild Planet... 160 calories for 28 grams of protein."

This Week in Startups · ▶ 1:13:06

Chrome
Chrome uses

"We use Chroma, we use Olama and it'll put that inside of this Chroma DB"

Josh Sirota · ▶ 12:19

Ollama
Ollama uses

"We use Chroma, we use Olama and it'll put that inside of this Chroma DB"

Josh Sirota · ▶ 12:19

OpenClaw
OpenClaw uses

"We've with OpenClaw with perplexity computer with co-works confused at times"

This Week in Startups · ▶ 12:35

Perplexity
Perplexity uses

"We've with OpenClaw with perplexity computer with co-works confused at times"

This Week in Startups · ▶ 12:35

Mentioned (5)

DeepMind
DeepMind "Most of my team is uh you know ex deep mind and and you know PhDs from Berkeley research lab" ▶ 8:05
Zipline
Zipline "We know quite well MANA for instance and we know quite well Google wing and uh and even some peop..." ▶ 51:01
Anthropic
Anthropic "You can comp that with what anthropic charges which is like $15 per million tokens for opus" ▶ 25:17
Cursor
Cursor "Cursor's got a new model coming out. I saw they were raising at 50 million" ▶ 37:19
Claude
Claude "You got Codeex uh you got Claude Code. Uh Gro's doing great work there" ▶ 37:26