AI for science

Alex first saw promise for AI in theoretical physics after he asked o3 for help on his research. In the podcast, Alex recalls asking GPT for help with a calculation that would have taken days, and getting a result in eleven minutes. He immediately recognized how impactful AI would be for his work — even though his physicist colleagues and the larger community gave it a lukewarm or skeptical reception.

Model companies going downstream into B2B consulting

  • Anthropic's unnamed JV with Blackstone, Hellman & Friedman, and Goldman Sachs — funded with $1.5B ($300m each from main participants). "A typical engagement starts with a small team working closely with the customer to understand where Claude can have the biggest impact. From there, the company's engineers — alongside Anthropic Applied AI staff — will develop Claude-powered systems tailored to each organization's operations."
  • OpenAI's The Deployment Company, backed by 19 investors including TPG, Brookfield Asset Management, Advent, and Bain Capital — has raised about $4B so far at a $10B pre-money valuation. "Microsoft-backed OpenAI last month said that its chief operating officer, Brad Lightcap, will shift into a new role and lead special projects while reporting directly to CEO Sam Altman. Lightcap would oversee OpenAI's push to sell software to businesses through a joint venture with a private equity firm."

ChatGPT personalization

GPT-5.5 Instant became the default in ChatGPT. OpenAI rolled GPT-5.5 Instant out to ChatGPT and the API, positioning it as a de facto across-the-board upgrade — base intelligence, image understanding, and tone. The release also bundles stronger personalization: ChatGPT can now use saved memories, past chats, files, and connected Gmail, while exposing "memory sources" so users can see which context influenced a response.

ChatGPT also improved Bing's recommendations.

RL environments

RL infra is shifting from "single generation + reward" to long-running action systems. A new guide compares RL environment frameworks for the LLM era, focusing on what scales to thousands of environments. A detailed survey contrasted traditional RLVR with agentic RL, pointing to systems like Forge, ROLL, Slime, and Seer, and recurring concerns like TITO consistency, rollout latency, prefix-tree merging, and global KV caches.

Proactive assistant surfaces are becoming a product category. A leak around Anthropic Orbit describes it as a proactive assistant that synthesizes data from Gmail, Slack, GitHub, Calendar, Drive, and Figma without explicit prompting. Manus also added recommended connectors that are suggested in context when needed.