The only reason I became interested in this domain is everything is analogous to message handling systems and eventing which I had a ton of experience. I can clearly say that AI systems are behind of the systems we designed in early 2000s.

Just to keep things honest, the longetivity of a system in this field isn’t that long. Due to continuous capability growth, there are new major challenges every day other than the foundations.

The lack of systems thinking in AI production systems is destroying my soul. It’s a major cultural challenge and most of the best engineers I know stay away from this field due to past burnout.

There isn’t a single good example of a foundational model access API that provides composibility layers, bidirectional multimodal streaming, and optionally state management. People often get two of them correctly, not all three.

IMHO, a large portion of this failure has been caused by the collapse of Congress over the last decade. Congress, in essence, no longer functions as a co-equal branch of government.

can the entire world government be cool for like 5 minutes i'm kind of going through a lot rn

The org boundaries are arbitrary!

Almost all major projects I worked on started by 2-3 people sitting in the same meeting room and coding while debating for hours every day without any other distraction.

I think we will!

We are close partners. This only lets me think broadly rather than overfitting systems onto our own ;)

Never bet against the wish of people who are trying to build things by composition and open source. Building a few critical foundational blocks in this space will be more important than ever in the next few years.

I decided to cancel my transfer to DeepMind to be able to focus on distributed AI systems and open models. So much to do here, and last week's DeepSeek moment completely helped me revisit some long term goals.

Never bet against a few passionate people who want to unblock tens of thousands of companies.

When you zoom in, you see the same stages underneath a single company too.

- Early stage: Elitism among teams
- Advanced stage: Monopoly on decision making and execution
- Terminal stage: Blaming successful teams

Culture matters. It can save you or accelerate your decline to the terminal stage.

Decentralize everything.
Open things where possible.
Everyone wins.

The biggest winner of the AI race will be distributed systems people. Everything is converging onto a distributed network of stuff and it is only accelerating in the last two years.

I'm very very very abstracted away from the accelerator.

Everyone is in this for different reasons, but this is my personal reason. I'm looking for systems paradigms that make sense for this domain. Even though all the churn and burden, I wake up every morning because I'm passionately curious about the answer 😉

People are wondering why I'm enjoying the LLM period:
- Thick client vs thin client fight
- Actor model vs RPC fight

As an engineer, I cannot think of a better way to challenge our conventional thinking. It entertains me intellectually but is also making me a better engineer.

For someone who is programming for over 30 years, it's hard for me to believe every system, every model, every chip, every program will be fungible and regeneratable. There will be new systems/patterns that will end up becoming the new infra overlords, we just don't know which.

When you have a highly ambiguous systems problem, try to solve it for the distributed case first. When you solve it for the distributed case, it's easy to pack things together in a monolith. The opposite is almost always impossible.

Good software engineering is about encapsulating complexity while avoiding hiding away failure modes.

This month is my last at Google. In a month, I'm joining Google DeepMind.

Our job will be making DeepMind's work publicly available in ways that wasn't possible before. While this is one of the biggest challenges in my life, it's potentially going to be the most rewarding.

Building is very easy but it also lowers the bar to build stupid things.

I like to be used as OCD as a service.

The most fascinating reality to me is you have to build a whole new person around languages you speak, especially if they have distinct communication styles or are from different language families.

In my experience, people will express themselves most effectively when allowed to converse in the language they feel best suits the topic, regardless of their native language(s). Consistency within that language is the key for the best.

Perhaps most importantly, many people feel they have to change their personality to fit the communication style and culture of a particular language. Switching between these personas can be detrimental, especially if the conversation is about a difficult, nuanced topic.

Trying to have a deep, meaningful conversation in three languages is an exercise in futility. The mental juggling and context switch inevitably distracts even native speakers. Speaking multiple languages myself, I wouldn't put anyone through that, not even my worst enemy.

I am debating how close I became to retirement and people are in denial because they knew that I always wanted to work until 80s.

My justification is simple: I don't want to stop working. I want to start working ;)