That is an excerpt of Sources by Alex Heath, a publication about AI and the tech trade, syndicated only for The Verge subscribers as soon as per week.
Reinforcement studying (RL) is the subsequent frontier, Google is surging, and the get together scene has gotten fully out of hand. These had been the by means of strains from this yr’s NeurIPS in San Diego.
NeurIPS, or the “Convention on Neural Data Processing Techniques,” began in 1987 as a purely tutorial affair. It has since ballooned alongside the hype round AI into a large trade occasion the place labs come to recruit and traders come to seek out the subsequent wave of AI startups.
I used to be regretfully unable to attend NeurIPS this yr, however I nonetheless needed to know what folks had been speaking about on the bottom in San Diego over the previous week. So I requested engineers, researchers, and founders for his or her takeaways. The checklist under of responses consists of Andy Konwinski, cofounder of Databricks and founding father of the Laude Institute; Thomas Wolf, cofounder of Hugging Face; OpenAI’s Roon; and attendees from Meta, Waymo, Google DeepMind, Amazon, and a handful of different locations.
I requested everybody the identical three questions: What’s the buzziest matter from the convention? Which labs really feel like they’re surging or struggling? Who had one of the best get together?
The consensus was clear. “RL RL RL RL is taking on the world,” Anastasios Angelopoulos, CEO of LMArena, instructed me. The trade is coalescing round the concept that tuning fashions for particular use circumstances, moderately than scaling the information used for pre-training, will drive the subsequent wave of AI progress. What’s clear from the lab momentum query is that Google is having a second. “Google DeepMind is feeling good,” Hugging Face’s Wolf instructed me.
The get together circuit was naturally relentless. Konwinski’s Laude Lounge emerged as one of many week’s hotspots — Jeff Dean, Yoshua Bengio, Ion Stoica, and a few dozen different prime researchers got here by means of. Mannequin Ship, an invite-only cruise with 200 researchers, featured “a dedication to the dance ground that’s unprecedented at a convention occasion,” one of many organizers of the cruise, Nathan Lambert, instructed me. Roon was dry about the entire scene: “you’ll be able to study extra from twitter than from actually being there … largely my on-the-ground feeling was ‘that is an excessive amount of.’”
Right here’s what attendees needed to say about NeurIPS this yr:
What was the buzziest matter amongst attendees that you simply assume extra folks might be speaking about in 2026?
Which labs really feel like they’re surging in momentum, and which of them really feel extra shaky?
What was one of the best get together you attended or had FOMO over?
Sure, some folks thought keynotes had been events. I assume academia lives on at NeurIPS in any case.
