Happy Thanksgiving Eve, Promptobello fam! 🙏
Mr. Mushroom is currently trying to explain to his family why an AI agent is cooking the turkey this year. (If it tastes like thermal paste, we know who to blame). 🦃🤖
While you were prepping for Thanksgiving, the AI world decided to go full chaos mode. Ilya Sutskever dropped his first major interview since leaving OpenAI, Warner Music decided to stop suing and start partnering, and OpenAI is hemorrhaging cash at a rate that would make a VC cry.
The Vibe: The "Scaling Law" era might be over. We're entering the "Age of Research." Also, lawyers are settling, and miners are finding lithium from space.

The quietest man in AI, Ilya Sutskever (co-founder of OpenAI, now SSI), just dropped a 90-minute philosophy bomb in an interview with Dwarkesh Patel. If you don't have time to listen to the whole thing while peeling potatoes, here's the Promptobello breakdown of what the Godfather of GPT actually thinks.
1. The "Age of Scaling" is Dead(ish). We're back to "Research."
→ The Old Way (2020-2025): "Scaling sucked the air out of the room." Everyone just added more data and compute to the same recipe (Pre-training).
→ The New Way: Pre-training data is finite. We can't just 100x the data anymore. We need new recipes. We're back to the 2012 era where clever ideas matter more than raw brute force.
2. The "Vibe Coding" Problem
Ilya compares current models to a student who studied competitive programming for 10,000 hours but has zero "taste."
→ They know the syntax, but they introduce bugs when trying to fix bugs.
→ His diagnosis: Models have "unreliable generalization." They memorized the textbook but don't understand the spirit of the task.
3. Emotions = Math?
This is the trippy part. Ilya thinks human emotions are actually just highly efficient "Value Functions."
→ Evolution hard-coded emotions (like social anxiety or hunger) to help us make decisions quickly without processing too much data.
→ The Future: AI needs its own "Value Functions" to stop hallucinations and make better decisions.
4. SSI's Master Plan
He isn't trying to build a chatbot. He's trying to build a "Human-like Learner."
→ Instead of a model that knows everything instantly, he envisions an AI that enters the economy like a brilliant intern—learns on the job and eventually becomes a superintelligence through continual learning.
Mr. Mushroom's Take: 🍄 "So my anxiety is just a 'Value Function'? I need to debug my source code."

The music industry's "David vs. Goliath" battle just ended with a handshake.

The News: Warner Music Group settled its copyright lawsuit against AI music generator Suno yesterday.

◾ The Twist: Instead of destroying Suno, Warner is effectively partnering with them. The deal allows Suno to keep developing while "respecting artists' rights" (read: paying royalties).

Why it matters: This is the template. For two years, we've wondered if AI would be sued into oblivion. This proves the endgame isn't destruction, it's licensing.

◾ Bullish for: AI startups with actual revenue (Suno, Midjourney).

◾ Bearish for: Lawyers hoping for a decade of litigation fees.

Forget benchmarks for a second. Ilya Sutskever just explained why your AI keeps introducing bugs when you ask it to fix bugs.
The Diagnosis: Current models have "unreliable generalization." They're like a student who studied competitive programming for 10,000 hours but has zero intuition.
The Symptom: You ask Claude or GPT to fix a bug. It says "Oh my God, you're so right!" and introduces a new bug. You point that out. It brings back the first bug. Repeat forever.
Why It Happens: RL training makes models "too single-minded." They optimize for benchmarks, not real-world robustness. Ilya's exact words: "If you trust the data all the time... sometimes you can be doing the correct thing but there's a bug."
The Builder's Takeaway:
◾ Don't trust any model to self-correct in a single pass
◾ Build verification loops into your agent architectures
◾ The "Age of Research" means the next breakthroughs will come from how we train, not how much we train
This is why Ilya left to build SSI. He thinks reliable generalization is the unsolved problem.

1. Niche Wins: Find Your Grind raised $5M yesterday.
◾ What they do: AI for student career discovery.
◾ The Lesson: While OpenAI and Perplexity fight over "Search," niche startups are quietly raising money by solving one specific problem really well.
2. The Burn Rate: A new report from HSBC estimates OpenAI needs to raise $207 BILLION by 2030 just to keep the lights on.
◾ The Scale: That is roughly the GDP of Greece. Just for server costs.

Trending: Hit Hacker News front page today
What is it? An open-source multi-agent system with native agent-to-agent communication.
The Key Feature: Praval Deep Research — lets you run local research workflows on ArXiv papers.
Why we like it: Agents can talk to each other without hallucinating context. Native ChromaDB integration for memory. Think of it as Lego blocks for building AI agent teams.
Link: GitHub - aiexplorations/praval

Yesterday, Fleet Space announced they discovered a MASSIVE lithium deposit in Quebec using their combo of AI and satellite tech. This isn't just a small find – we're talking potentially the largest lithium reserve ever found in the region. Fleet's AI systems analyzed satellite imagery to pinpoint geological formations likely to contain lithium, proving yet again that AI + space tech = resource discovery superpowers. Battery manufacturers are already lining up with dollar signs in their eyes.