My MLSys keynote on AI writing systems code got more interest than I expected. The recording will take a while, so in the finest tradition of AI labs sharing blog posts, we’re starting the Core Automation Blog with this one https://t.co/h4uSOyrglf
Four years ago, average SF sentiment was “the smartest people I know are all joining web3.” These same people thought the economy would restructure around getting a tradable NFT each time you bought something.
AI is much more useful. But it's important to remember:
1) SF is a massive bubble, and
2) Every dominant technology and company throughout history has eventually been surpassed.
Every single one.
Just in AI, we won't be able to keep scaling compute. Chips will get 100x more efficient. New forms of energy are becoming economically viable (or socially acceptable). We'll invent new types of models that are useful in new ways.
These, plus 100x more things that already exist today, and 1,000x more that don't even exist yet, are all opportunities for new dominant products and companies to emerge.
Newspaper, telecom, and Cable companies were all monopolies at one point. New technologies came, they did not react fast enough, and new players replaced them.
Google used to have a monopoly in online ads. Until Facebook came. And then Amazon came. And then TikTok. Now AppLovin.
IBM used to dominate the computer market. Then Dell. Then Microsoft. Then Apple.
In all of these cases, new technologies, form factors, types of customers, and/or social dynamics emerged that allowed new massive companies to be formed.
Even in AI, OpenAI was once the dominant company. Anthropic came out of nowhere in the past two years. And in two more, there will be another, probably a lot more (ask a VC who plowed $1B into a new AI company that hasn't launched a product yet which one that could be).
SF is a great place. But it's a giant bubble that's insulated from the rest of the world. SF is usually right on the technology, but often wrong on the timing.
Patience is underrated. There will always be opportunities to escape the permanent underclass.
4. Agents need their own identity. Today every action is laundered through the human's IAM role, so the audit log reads "corey@duckbill did this" when the truth is "Claude's third retry at 2am did this."
First-class agent identities: scoped, attestable, time-limited, revocable.
I'm joining Carnegie Mellon's CS Department (and HCII by courtesy) as an assistant professor in Fall 2027!
I'll be recruiting PhD students next cycle. If you're interested in AI systems or human-AI collaboration, list me in your application. Stay tuned for more about my new lab!
Not sure why we didn't fix this a long time ago. kwargs support to custom autograd functions coming to a nightly near you soon: https://t.co/LKsvcSnvbs
Running on Apple Silicon will never be as fast as an H100. But for interactive workloads like computer use, wall-clock latency is dominated by the network, not the accelerator. Skipping a large image uploads buys you more than the H100 buys back. https://t.co/icZ5uXqsh0
Adopting Claude speak in my regular life, episode 1:
Partner: Did you do the dishes tonight?
Me: Yes they're done.
Partner: Why are they still dirty?
Me: You're right to push back. I didn't actually do them.
Python no longer has to be constrained by a single thread running at a time. Here’s a short tutorial to get you started with free threading:
https://t.co/U9YroZaEYI
After 5 amazing years, I’m leaving the PyTorch team at Meta. I did my best work there and got to work with some of the smartest, most OSS pilled engineers in the industry. More soon on what’s next: still systems, still OSS (but not everything), a smaller team with a lot of GPUs
Mistakes happen. As a team, the important thing is to recognize it’s never an individuals’s fault — it’s the process, the culture, or the infra.
In this case, there was a manual deploy step that should have been better automated. Our team has made a few improvements to the automation for next time, a couple more on the way.