
Nathan Labenz
@labenz • 16,555 subscribers
AI Scout, building text-2-video @Waymark, host of The Cognitive Revolution podcast
Shorts
Videos

".Elon Musk told CCP officials in 2023: if China builds Superintelligence, it will run China. A month later, China rolled out its first AI regulations." - Max Tegmark, debating the merits of a ban on Superintelligence with Dean W. Ball, from Doom Debates by Liron Shapira
Nathan Labenz128,531 views • 6 months ago

When Andrej Karpathy goes off on a new paper, I talk to the authors the goal was to eliminate tokenization, but they might've found a more fundamental unlock if MEGABYTE works at LLaMa scale, we'll have a new frontier of transformer hierarchies to explore Lili Yu explains it all!
Nathan Labenz141,620 views • 3 years ago

"What if we give every person going through cancer, MS, Alzheimer's, an AI that helps them through that journey? Available in every language. The world becomes a better place. We call this Universal Basic AI" - Emad on his vision for the Intelligent Internet Intelligent Internet
Nathan Labenz73,869 views • 1 year ago

The UK AISI might be the most situationally aware government entity in the world today. Today on The Cognitive Revolution Podcast, Chief Scientist Geoffrey Irving surveys the AI landscape. While jailbreaking frontier models is getting harder, the AISI Red Team has never failed. 👀
Nathan Labenz21,444 views • 3 months ago

"Golf courses in Arizona use ~30X as much water as all data centers in Arizona. Best I can tell, data centers are generating ~50X as much revenue per gallon used. So, if I could close all golf courses in Arizona and replace them with data centers ... that would generate billions of dollars in revenue, without increasing water bills. I'm from a not super well-off area that is left behind by industry. I would be excited if a medium-sized data center were built in the vicinity that the town could then tax. I think that would be net good." - Andy Masley explains how building & taxing data centers can help revitalize many parts of America, with minimal environmental impact
Nathan Labenz29,738 views • 6 months ago

Special cross-post featuring Olive Song of MiniMax (official), from swyx 🇸🇬's AI Engineer and Ksenia Se's Ksenia_TuringPost From the ups, downs, & surprises of RL post-training, to battles with reward hacking, to painstaking debugging... Chinese researchers are ... just like us!
Nathan Labenz18,875 views • 3 months ago

"o3 is misaligned. It lies to the user & defends it onto death. It's significantly worse than o1 and GPT-4.1 seems less aligned than GPT-4o The more RL you apply, the more misaligned they get, & we don't seem to be trying that hard" Zvi Mowshowitz on why his p(doom) is up to 70%
Nathan Labenz37,302 views • 1 year ago

"We found that o1 lies a lot" 🚨 In >85% of cases, once o1 has deceived a user, it continues to lie when questioned This is much less common in earlier models! AlexM of @apolloaisafety & I discuss unpack their paper "Frontier Models are Capable of In-context Scheming"⬇️
Nathan Labenz22,386 views • 1 year ago

I gave my username & password and DoNotPay negotiated with Comcast for me I got an outage refund, and signed 1-year contract to save $40 / month The video is boring, but as says, that's why this is the perfect job for AI Hear all about it on new The Cognitive Revolution Podcast
Nathan Labenz34,378 views • 3 years ago

"The only thing that works 100% of the time is measuring in a lab" Andrew White 🐦⬛, Head of Science FutureHouse explains why they're building an AI Scientist to run real-world experiments -vs- pursuing a more in-silico strategy for biological discovery. Excellent episode!
Nathan Labenz16,877 views • 1 year ago
No more content to load