ThursdAI - The top AI news from the past week · From Weights & Biases, Join AI Evangelist Alex Volkov and a panel of experts to cover everything important that happened in the world of AI from the past week

ThursdAI - Jan 1 2026 - Will Brown Interview + Nvidia buys Groq, Meta buys Manus, Qwen Image 2412 & Alex New Year greetings

·30 min·1 clip

Will Brown explains why thinking longer can emerge from RL when the reward comes from correct answers.

1. ThursdAI’s January 1, 2026 episode covers AI news from December 31, 2025 and opens with Alex Volkov broadcasting from Los Angeles. 2. Alex Volkov is the host from Weights and Biases, and Will Brown from Prime Intellect joins as the interview guest because of his RL and meme-heavy public commentary. 3. The episode’s purpose is to combine a year-end news roundup with a recorded conversation about reinforcement learning, environments, and model behavior. 4. Alex says Meta notified people that it bought Manus.ai, which he describes as an agentic runner company covered on the show in March 2025. 5. Alex says NVIDIA bought Groq, spells it “Groq with a Q not with a K,” and says Jensen led a reported $20 billion deal. 6. Alex also mentions SoftBank’s final investment into OpenAI and says Qwen released “Qwen image 2512” with improved realism. 7. Alex lists Hunyuan translation service updates and says the week was quiet because many American foundational labs were on vacation. 8. Alex reflects on GPT-5, GPT-5.1, GPT-5.2, Opus 4.5, and “vibe coding” as examples of the year’s fast-moving AI cycle. 9. Will Brown explains RL as hill climbing around scores and says it can be viewed as synthetic data generated from the model itself. 10. He describes sampling many model answers, upweighting the better ones, and downweighting the worse ones after scoring. 11. Will Brown says verification functions became important because people realized the final answer in math can be checked directly. 12. He says the old focus on labeling every intermediate math step became less important once final-answer checking worked. 13. He explains that models can be rewarded for the right amount of thinking, not just for longer thinking. 14. He gives the example of a bonus for the shortest correct answer as one way to encourage efficient reasoning. 15. On environments, Will Brown says a math question plus a grader is an RL environment, and he compares environments to train sets and evals to test sets. 16. He says Prime Intellect’s Environments Hub is built to create, share, browse, and use RL environments with tools such as eval viewers and rollout viewers. 17. He gives a GPU textbook example where a community member turned Programming Massively Parallel Processors into an environment for training coding models. 18. Will Brown says creative writing is difficult to optimize directly, so the goal is to make it more verifiable by comparing good and bad examples. 19. The interview style is informal and technical, with Alex translating concepts into simple language and Will answering in compact explanations. 20. Listeners who follow RL, coding agents, and model training will likely get the most value, while people only wanting a pure news recap may skip the long technical segment.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from ThursdAI - The top AI news from the past week