Sakana AI Wins ICFP with Self‑Improving Model

Beth explains how Sakana's Shinka Evolve used inference-time self-improvement to win the programming contest.

Play episode from 12:21

chevron_right

Transcript

chevron_right

Transcript

Episode notes

Beth, Andy, and Brian closed the week with a full slate of AI stories — new data on public trust in AI, Spotify’s latest AI DJ update, Meta’s billion-dollar data center project in El Paso, and Anthropic’s release of Claude Skills. The team discussed how these updates reflect both the creative and ethical tensions shaping AI’s next phase.

Key Points Discussed

Pew & BCG AI Reports showed that most companies are still “dabbling” in AI, while a small percentage gain massive advantages through structured strategy and training.

The Pew Research survey found public concern over AI now outweighs excitement, especially in the US, where workers fear job loss and lack of safety nets.

Spotify’s AI DJ update now lets users text the DJ to change moods or artists mid-session, adding more real-time interaction.

Spotify also announced plans with major record labels to create “artist-first AI tools,” which the hosts viewed skeptically, questioning whether it would really benefit small artists.

Sakana AI won Japan’s ICF programming contest using its self-improving model, Shinka Evolve, which can refine itself during inference — not just training.

Yale and Google DeepMind built a small AI model that generated a new, experimentally confirmed cancer hypothesis, marking a milestone for AI-driven scientific discovery.

University of Tokyo researchers developed a way to generate single photons inside optical fibers, a breakthrough that could make quantum communication more secure and accessible.

Brian shared a personal story about battling n8n’s strict security protocols, joking that even the rightful owner can’t get back in — a reminder of strong data governance practices.

Meta’s new El Paso data center will cost $10B and promises 1,800 jobs, renewable power matching, and 200% water restoration. The hosts debated whether the environmental promises are enforceable or just PR.

The team discussed OpenAI’s decision to allow adult-only romantic or sexual interactions starting in December, exploring its implications for attachment, privacy, and parental controls.

The final segment featured a live demo of Claude Skills, showing how users can create and run small, personalized automations inside Claude — from Slack GIF makers to branded presentation builders.

Timestamps & Topics

00:00:00 💡 Intro and news overview

00:01:30 📊 Pew and BCG reports on AI adoption

00:03:04 😟 Public concern about AI overtakes excitement

00:05:23 🎧 Spotify’s AI DJ texting feature

00:06:10 🎵 Artist-first AI tools and music rights

00:13:35 🧠 Sakana AI’s self-improving Shinka Evolve

00:14:25 🧬 DeepMind & Yale’s AI discovers new cancer link

00:17:24 ⚛️ Quantum communication breakthrough in Japan

00:20:28 🔐 Brian’s battle with n8n account recovery

00:26:01 🏗️ Meta’s $10B El Paso data center plans

00:30:26 💬 OpenAI’s adult content policy change