Karan 4D, co‑founder & Head of Behavior at Nous Research, says no. The global compute race won’t stop, so the only survivable path is to build powerful AI openly, collaboratively, and across a decentralized network of GPUs—before closed systems lock in brittle, biased guardrails.
In this wide‑ranging conversation we unpack why stopping is impossible, how the field sleep‑walked into the assistant paradigm, and what it will take to recover lost creative and cognitive search space in our models. Karan walks us through Nous Psyche stack and DRO Optimizer, a bandwidth‑slashing breakthrough that lets you train large models across widely scattered hardware—think “turn idle global GPUs into one giant lab.”
GUEST INFO ⤵
YouTube ▶️/ https://www.youtube.com/@Nousresearch
Twitter/X 🐦/ https://x.com/nousresearch
Paper 📄/ https://nousresearch.com/
HuggingFace: http://huggingface.co/teknium
Discord: https://discord.gg/nousresearch
Github: http://github.com/teknium1
https://github.com/sponsors/teknium1
The latest AI News. Learn about LLMs, Gen AI and get ready for the rollout of AGI. Wes Roth covers the latest happenings in the world of OpenAI, Google, Anthropic, NVIDIA and Open Source AI.
______________________________________________
My Links 🔗
➡️ Twitter: https://x.com/WesRothMoney
➡️ AI Newsletter: https://natural20.beehiiv.com/subscribe
Want to work with me?
Brand, sponsorship & business inquiries: wesroth@smoothmedia.co
Check out my AI Podcast where me and Dylan interview AI experts:
https://www.youtube.com/@Wes-Dylan
______________________________________________
00:00 Intro (skip to 6m44s for the interview)
06:44 AI lobotomy
16:20 break AI’s reality
18:25 well behaved, less intelligent
20:22 base models vs instruct models
24:02 true neutral
25:46 scraping the entire internet
29:25 glazing humans
33:04 stuck in chat mode
36:16 what else is there?
40:01 prompt engineering
43:06 open source RL gym
44:27 RL in gaming
47:01 RL evolution
49:45 R1-Zero
52:27 future of Reinforcement Learning
55:47 Paperclip Maximizer
58:42 Instrumental Convergence
01:02:33 dangers of misaligned AI
01:05:51 activation hacking
01:09:49 open collaboration
01:15:35 urgency of AI regulation
01:18:51 open vs controlled AI
01:20:10 human behavior and AI alignment
01:23:16 predicting future needs
01:26:45 simplicity in AI control
01:30:24 manipulative nature of AI models
01:32:12 narrative power of AI
01:34:39 future of AI
01:37:35 better AI through experience
01:40:11 open source AI development
01:45:22 government’s role in open source AI
#ai #openai #llm
