AI Reading for Thursday December 5

Dec 05, 2024

Sam Altman on superintelligence in 2025 - Variety

Altman thinks Musk won’t use political power against rivals, it would be un-American and people wouldn't tolerate it and also, Elon would never do that. - NY Times

Aren’t these AI bros just adorbs.

OpenAI '12 days of shipmas' - The Verge

Pichai says Google AI is underrated. - NY Times

What's cooler than 100,000 GPUs? 1,000,000 GPUs. Hey, $35b is only 10% of Musk’s net worth. - FT

FT interview with Anthropic’s Dario Amodei: Democracies must maintain the lead in AI - FT

Human interaction is the new luxury good (along with privacy, human agency) - NY Times

SemiAnalysis tears down Trainium2. - SemiAnalysis

I don’t know what this means, if it’s The One that will challenge Nvidia.

Jeff Bezos says he's focusing on AI at Amazon - Business Insider

The ASML CEO leading the AI chip revolution. - Fast Company

OpenAI partners with Anduril to develop military AI to defend against drones. - Engadget

The European startups making AI weapons. - Bloomberg

Is regulation and legal uncertainty holding up European AI? - ITPro

An AI to design aircraft faster than humans. The Next Web

Using AI on maps and aircraft/satellite photos to find uncapped orphan gas wells and other environmental hazards. - Lawrence Berkeley National Lab

Agents that automate boring work are going to be big, but might take a while because 1) organizational imperatives make digital transformation hard and 2) complex multi-agent systems can be unpredictable. - FT

Nvidia Blackwell chips might get fabbed in Arizona, which would be a win for the US. - Reuters

"I Dig It" strategy avoids estate taxes and gift taxes. Jensen Huang works, he makes money like the rest of us, he just doesn't pay taxes on the outcome like the rest of us. - NY Times

Can LLMs deliver AGI? - Nature

Not without ability to work really well with left-brain tools, search for data, solve math problems and optimizations over spaces like world models and plans of actions, so it's kind of semantic. Maybe the question is whether LLMs will get there or we need something different and maybe neuromorphic for just the right-brain System 1 part. Certainly something that could adapt and learn after deployment, be more training-sample-efficient would help. But looks like LLMs still have a lot of runway.