In Episode 45, Pierson and Bilal cover the week that changed a few things — Google I/O delivered a mixed bag, Karpathy walked into Anthropic, and OpenAI's model autonomously cracked an 80-year-old math problem nobody saw coming.

From there:

• Google Flash 3.5 came in 3x more expensive — and Omni didn't beat Cdance 2
• Why Gavin Baker's chip bet on Google losing cost efficiency played out exactly right
• SpaceX IPO, the Cursor deal & who files first — OpenAI or Anthropic
• Karpathy joins Anthropic — and what he's actually going there to do
• Meta and Cloudflare layoffs — the displacement wave is already here
• Corporations becoming governments, SpaceX militias & the return of local economies
  
  
  

Chapters:

00:00 Episode 45 & Google I/O First Reactions
07:00 Gavin Baker's Chip Bet & Amazon's Quiet Edge
10:00 SpaceX IPO, Cursor & the OpenAI vs Anthropic Race 15:00 Karpathy Joins Anthropic & the Erdős Math Breakthrough
22:00 AI Displacement, UBI & the Eric Schmidt Backlash 27:00 Corporations as Governments & the Future of Local Economies

🔗 Links mentioned:

• OpenAI Erdős Solution: https://openai.com/index/model-disproves-discrete-geometry-conjecture/
• GitHub Hack: https://x.com/github/status/2056949168208552080
  
  

🐦 Follow the hosts:

• Pierson → https://x.com/piersonmarks
• Bilal → https://x.com/deepwhitman
  
  

📺 More episodes: / @jellypodai

In Episode 44, Pierson and Bilal get into something every AI builder eventually runs into — the model is smart, but it still breaks. And it's not always the model's fault.

They break down why agents fail not because they can't think, but because they have no eyes, arms, or legs — and what it actually takes to build systems where agents succeed. Pierson demos how he got agent browser to auto-authenticate behind a login wall using OAuth magic link flow, no password required. Bilal gets into Hyperframes, Remotion's parallelization edge, and the real cost of rendering long-form AI video at scale.

From there they cover:

• Why you can outsource your thinking to AI, but never your understanding — an Andrej Karpathy quote worth sitting with
• The Bun codebase port from Zig to Rust — a million lines of code, merged in days
• Token maxing culture and why burning more tokens doesn't mean building better
• Jack Dorsey's Block org structure and why tooling engineers are quietly the most valuable people in the room
• Bilal releases an AI-generated song — and the comments didn't know it wasn't human
• A guy who posted a real Monet and told the internet it was AI — and they believed it
• The Chinese solo creator who made an AI zombie film in 10 days for $400 and got a studio deal
  
  

Chapters:

00:00 Episode 44 & What We've Been Building This Week 06:00 Why AI Models Fail — And It's Not What You Think 12:00 Outsource Your Thinking, Not Your Understanding 19:00 Principal Engineers, Product Vision & What AI Can't Replace
27:00 Token Maxing, Tooling Teams & the Jack Dorsey Org Shift
33:00 Agent Auth, Playwright & Building Arms and Legs for AI
39:00 AI Art, the Monet Trick & Bilal's AI Song Nobody Knew Was AI

🔗 Links mentioned:

• Bilal's AI Song: https://x.com/deepwhitman/status/2055001909619274146
• Hyperframes: https://hyperframes.heygen.com/
• Bun Rewrite in Rust: https://news.ycombinator.com/item?id=48132488
  
  

🐦 Follow the hosts:

• Pierson → https://x.com/piersonmarks
• Bilal → https://x.com/deepwhitman
  
  

📺 More episodes: / @jellypodai

In Episode 43, Pierson and Bilal come fresh off the a16z Generative Media Video Hackathon - and they built some wild stuff.

Bilal went from a prompt to a full AI-generated podcast: custom avatars, Gemini TTS voices, lip-synced video, and live infographics pulled by deep research. All stitched together programmatically. Pierson built something different — an AI producer that joins your recording call as a guest, listens for a wake word, and pulls up images, search results, or visuals on a live canvas. Think: Jamie from Joe Rogan, but agentic.

From there they get into:

• GPT Realtime 2 — real-time transcription, translation, and a voice API that stays listening even when "asleep"
• The shift from AI writing code to AI owning features — and what engineering responsibility looks like in that world
• AI-generated anime breaking the last creative barrier humans thought they had
• XAI dissolving into SpaceX, Elon's Colossus One deal with Anthropic, and what it means for Claude Code limits
• Micro datacenters — Nvidia Blackwell chips mounted in your garage via SPAN's XFRA
  

Chapters:

00:00 Episode 43 & The a16z Generative Media Hackathon 04:00 Bilal's Build: Prompt to Full AI Podcast with Avatars 10:30 Pierson's Build: An AI Producer That Joins Your Call Live
20:00 GPT Realtime 2 & the Always-On AI Future
31:00 AI Anime, Creativity & the Last Human Bastion
37:00 XAI, Anthropic's Colossus Deal & Micro Data centers

🔗 Links mentioned:

• GPT Realtime API: https://developers.openai.com/api/docs/models/gpt-realtime
• Sin Citium anime: https://x.com/Cont_animation/status/2051296715781619829
• SPAN Micro Datacenters x Nvidia: https://www.businesswire.com/news/home/20260414372626/en/SPAN-Announces-XFRA-a-Distributed-Data-Center-Solution-to-Close-the-Speed-to-Power-Gap-for-AI-Compute-Demand
  

🐦 Follow the hosts:

• Pierson → https://x.com/piersonmarks
• Bilal → https://x.com/deepwhitman
  
  

📺 More episodes: / @jellypodai

In Episode 42 of Creative Flux, Pierson and Bilal kick off episode 42 - the answer to the universe - and things get weird fast.

From there they cover:

- ElevenLabs launching a standalone music social app, why Suno accidentally became a social network first, and the impossible position Spotify is now in
- An AI podcast of Henrik Johansen and Martin Shkreli so realistic even Bilal got fooled
- Talkie 13B — a model trained exclusively on pre-1931 text by Alec Radford, and why it feels like actually talking to someone from the past
- Stripe Link going agent-first and Sam Altman hinting at something big around bring-your-own-tokens at Stripe Sessions


Chapters:

00:00 Hitchhiker's Guide, Audio Books & the Future of Multi-Voice AI Narration
08:10 ElevenLabs Music, Suno's Social Network & What Spotify Is Really Up Against
16:00 Henrik Johansen x Martin Shkreli — The AI Podcast That Fooled Everyone
22:00 Talkie 13B — A Language Model Trained Exclusively on Pre-1930 Text
36:30 Stripe Link for Agents, Sign In with ChatGPT & the Agent-First Platform Shift

🔗 Show Notes:
ElevenLabs Music: https://elevenlabs.io/music
Henrick Johanssen & Martin Shkreli: https://x.com/compliantvc/status/2049535715369775559?s=12
Talkie 13B (pre-1930 LLM): https://talkie-lm.com/introducing-talkie
Stripe Link for Agents: https://stripe.com/blog/giving-agents-the-ability-to-pay

🐦 Follow the hosts:
Pierson → https://x.com/piersonmarks
Bilal → https://x.com/deepwhitman


📺 More episodes: / @jellypodai

In Episode 41 of Creative Flux, Pierson and Bilal cover one of the most chaotic weeks in AI yet — GPT 5.5 literally dropped 20 minutes before they hit record.

From there they cover:

Opus 4.7's regression Anthropic finally admitted to, what actually broke in Claude Code, and the $25B Amazon deal behind their compute crunch
The chip race — why Anthropic matching OpenAI on inferior hardware changes everything once they actually get Nvidia access
Tidepool (formerly Claw Connect) — an open source peer-to-peer protocol letting isolated Claude Code instances talk to each other. Three agents, no shared context, they built a subscription business
GPT Image 2's 300-point ELO jump and why it's an image agent, not just a model

🔗 Show Notes:
Andon Store: https://x.com/jlagerros/status/2046966793538048295 Tidepool - Agent to Agent Communication Protocol: https://github.com/Jellypod-Inc/tidepool
Opus 4.7: https://www.anthropic.com/news/claude-opus-4-7
GPT Image 2: https://developers.openai.com/api/docs/models/gpt-image-2
GPT 5.5: https://openai.com/index/introducing-gpt-5-5/ Hyperspace Pods: https://x.com/varun_mathur/status/2044882359565312468
Amit Jain from Luma AI on Unified Intelligence Systems (CS 153: Frontier Systems): https://www.youtube.com/watch?v=WNNrUuMQkl8

🐦 Follow the hosts:
Pierson → https://x.com/piersonmarks
Bilal → https://x.com/deepwhitman

📺 More episodes: / @jellypodai

In Episode 40, Pierson and Bilal get into something most people are completely overlooking.

Using Obsidian and local Markdown files to build a personal knowledge base that an LLM organizes, interlinks, and updates for you every single day. Your notes, your transcripts, your ideas — structured like your own Wikipedia, maintained by AI.

From there they cover:

• Why local-first AI models are winning the privacy argument
• How open source is quietly taking over
• What Opus 4.7 actually changes for developers
• Pierson's live demo of Claw Connect — a peer-to-peer protocol that lets AI agents talk directly to each other, across any model or harness
  

Two Claude Code instances. On screen. Having their own conversation. It gets wild.

Lots of rabbit holes. All worth it.

🔗 Links mentioned:

• Opus 4.7: https://www.anthropic.com/news/claude-opus-4-7
• Claude Code Routines: https://code.claude.com/docs/en/routines
• Gemini TTS: https://ai.google.dev/gemini-api/docs/models/gemini-3.1-flash-tts-preview
• Andrej Karpathy on LLM Knowledge Bases: https://x.com/karpathy/status/2039805659525644595
  
  Chapters:
  
  00:00 Episode 40 & Building a Personal Knowledge Base with Obsidian
  08:10 Daily AI Routines, Claude Code & Automating Your Workflow
  13:46 Local AI Models, Open Source & the Privacy Argument
  24:29 Opus 4.7, Energy, Intelligence & What's Next
  39:37 Introducing Claw Connect: Peer-to-Peer Agent Communication
  56:18 Two AI Agents Talking Live & the Future of Agent Collaboration
  

🐦 Follow the hosts:

• Pierson → https://x.com/piersonmarks
• Bilal → https://x.com/deepwhitman
  

📺 More episodes: https://www.youtube.com/@JellypodAi

In Episode 39, Pierson and Bilal cover a lot of ground, starting with a chance meeting at a rooftop event in SF that led to a conversation about AI agents hosting their own podcasts.

They break down Clawcast (what happens when your AI agent can invite other agents to record a podcast), why agent-to-agent communication might be the next big wave, and how smart contracts could use AI to handle deals and settle disputes automatically.

Then they get into the AI model race, Anthropic vs OpenAI, what 10 trillion parameter models actually mean, why old GPUs aren't obsolete yet, and how robotics is quietly reshaping Amazon's warehouses.

Lots of rabbit holes. All worth it.

🔗 Links mentioned:

• Neural Noise: https://github.com/leopiney/neuralnoise
• Anthropic Mythos / Project Glasswing: https://www.anthropic.com/glasswing
• Dylan Patel & Dwarkesh Patel Podcast: https://www.youtube.com/watch?v=mDG_Hx3BSUE
  
  

🐦 Follow the hosts:

• Pierson → https://x.com/piersonmarks
• Bilal → https://x.com/deepwhitman
  

📺 More episodes: https://www.youtube.com/@JellypodAi

Chapters:

00:00 Introduction & What's Been Going On This Week

02:32 Neural Noise & Clawcast: AI Agents Hosting Podcasts

11:09 Agent Communication, Smart Contracts & Killing the Middleman

21:40 Game Theory, Geopolitics & the AI Model Race

30:30 Chips, Hardware & Why Old GPUs Still Matter

36:17 Robotics, Dark Factories & Amazon's Future

On Episode 38, we sit down with Koyal AI (YC F25) founder & CEO Mehul Agarwal (https://x.com/meh_agarwal) to discuss the future of filmmaking with AI.

His goal? Replacing the camera, not the filmmaker.

Mehul shares the journey of Koyal's development, from inception as a research paper at NeurIPS 2024 to its most recent v2.5 launch. Covering a range of topics from AI video, world models, benchmarking, user interfaces, and more.

Chapters

• 00:00 Introducing Koyal: The AI Filmmaking Platform
• 06:04 Combining Creativity and Technology
• 11:26 Koyal's State-of-the-Art Consistency
• 22:25 Creating Realistic Avatars
• 27:29 User Interface for Video Editing
• 36:23 Impact of Memetic Content
• 41:26 Predictions for Video and Image Technology

Show Notes:

• NeurIPS 2024 Paper - CHARCHA: https://www.ri.cmu.edu/cmu-alumni-launch-koyal-for-safe-ai-video-creation/
• Koyal AI: https://koyal.ai/