Nearly all good ML papers are at least implicitly position papers, providing substantive support for a meaningful position. Otherwise, what’s the point?

🚨 Nano Banana 2 completely smashed both the clock AND full wine glass tests in ONE IMAGE. "11:15 on the clock and a wine glass filled to the top" Holy. Shit.

Elon: we train in int Shareholders: confused what int is Elon: int operation is fundamentally better than float Shareholders: more confused, what is float Elon: we can do float but int is simpler with logic gates, int is power efficient, we do int Shareholders: still silent and confused, what the hell is a logic gate Elon: some of it is made in TSMC Texas Shareholders: everybody applauds Elon: man I love chips, it is a good chip sir. You know what, we gonna make a gigantic chip fab ourselves

How is memorized data stored in a model? We disentangle MLP weights in LMs and ViTs into rank-1 components based on their curvature in the loss, and find representational signatures of both generalizing structure and memorized training data

In 5 years from now, probably 95% of the tokens used by AI agents will be used on tasks that humans never did before. I just met with about 30 enterprises across 2 days and a dinner, and some of the most interesting use-cases that keep coming up for AI agents are on bringing automated work to areas that the companies would not have been able to apply labor to before. Most of the world hasn’t quite caught on to this point yet. We imagine AI as dropping into today’s workflows and just taking what we already do and making it more efficient by 20% or something. Yet most companies realize that most of the time they’re doing far less than they could because of the cost or limited capacity of talent. This shows up in different ways across every industry. In real estate it’s ideas like being able to read and analyze every lease agreement for every trend and business opportunity possible. In life sciences it’s being able to rapidly do drug discovery or improve quality by looking through errors in data. In financial services it’s being able to look through all past deals and figure out better future monetization. In legal it’s being able to execute on contracts or legal work for previously unprofitable segments or projects. And these are just the Box AI use cases that deal with documents and content. The same is going to be true in coding, where companies tackle software projects they wouldn’t have done before. Security of all systems and events they couldn’t get to. And so on. If you are working on AI Agents right now, the big opportunity is to bring enterprises “work” for problems that they couldn’t do before because it was nearly impossible to afford or scale. And if you’re deploying AI agents in an enterprise, consider what things you’d do more of (or differently) if the cost and speed of labor became 100X cheaper and faster. This is going to get you the real upside of automation.

Tesla has revealed that owners using FSD experience just one crash per 4.92 million miles vs one crash per 700k miles for the US average.

Cerebras beats Nvidia H100 but can it beat Blackwell? Blackwell inference endpoints are finally out and it’s fast. It runs GPT-OSS-120B at ~700 tokens/s, leapfrogging H100 and Groq. Cerebras clocked in at 3,000 TPS - still #1. Looking forward to Rubin!

Voice AI Meetup on Wednesday Nov 18th (in-person in SF, or live stream). Themes this time: - coordinating between multiple agents - incorporating thinking models into voice agents - using sandboxes for long-running sub-agents - dynamic user interfaces I'm always excited about the monthly meetups. But I'm *especially* excited about this one. I've been hacking on all these things lately both with enterprise partners and experimentally. We have a great group of people to talk about these "next next" emerging patterns. I'm also showing a demo of some stuff I've been working on for the past few weekends.