Alex Dhillon

If this is true, using the best public estimates we have of LLM resource use, solving this Erdos problem took 0.6–6.3 kWh of electricity and about 3–31 liters of water. So that is less than three almonds worth of water and the electricity equivalent of 2-20 miles of EV driving.

You just patched last month’s Nginx vulnerability that was actively exploited in the wild? It’s already time for a fresh 0-day RCE. The whole world is basically “pwned-by-default”, patching vulnerabilities before they’re exploited feels like a Sisyphean task... 🫠

Most people’s model for security bugs is wrong. People have acted as though there is always an infinite number, because humans couldn’t find them faster than they could create them, but the number has always been finite. AI systems aren’t going to find every single one immediately, but they are draining the available supply fast, and new ones aren’t going to be created as quickly. Once we start using AI to improve the engineering, the supply is going to drop a great deal indeed, and operations that rely on a steady supply of them are going to have to find other ways to work.

Over the past year, AI agents have learned how to self-replicate. In our test environment, an agent hacks a remote computer and copies itself onto it. Each copy then hacks more computers, forming a chain.

AI is taking over Williamsburg. 42% of the companies in this new Wburg office complex are AI companies (22/53). It’s called The Refinery at Domino. They even host a monthly AI Demo Day at the building that draws hundreds of people and have an Equinox fitness club on-site.

One of the bigger meta-patterns I've noticed is that engineers and conscientious people tend to overweight the value of internal consistency and logical consistency Our audiences are barely paying attention, and it is more important to resonate in simple ways than to worry deeply about the precision and consistency of our systems and logic Reality is incredibly complex, and any illusion you have that you have figured out a "consistent logical formula" for your work is probably wrong and unimportant Vibes matter a lot more than people think in a hyperdimensional world

I think returns to intelligence are nonlinear because decisions are path-dependent early choices in code, experiments, or strategy can compound positively or negatively over time for example by avoiding dead ends or preserving optionality it's why I am a big fan of very long running tasks and massive benchmarking budgets GPT-5.5 and Mythos Preview are only marginally more intelligent than previous models and have pretty much the same performance up to 10M tokens, but after that they go absolutely ballistic

2027 will be the year of the data breach