it's not just perception of growth

it is actual new chips deployed and actively used

they can just charge more for tokens than the cost of inference?

Thank goodness we’re going to have self-driving cars become commonplace in the next half decade

Have you ever had someone in your life that you wrote poetry or anything romantic for?

Was it like, generic romantic or did it capture you and your significant other in your particular quirks

Siiiiiiick

Why do you still work at Google?

Man you would not like this young hacker named Rob Pike

He had a portfolio of bets across the company and most have paid out, hence the higher revenue, profit, and stock price

They’ve added AI ad creation tools and AI ad targeting, both proving effective

Well no, because he still massively grew the core business even with the tens of billions of burned capital on the speculative stuff

approximately $11.37 billion in total cumulative revenue since they started breaking it out (vs $60b in cumulative losses)

So not NOTHING, but probably should have worked against healthier expectations (this is basically a game console, spend appropriately)

Right, they are spending on AI and reducing spending on less productive fronts

People who do not write code in any capacity are going to be clinging on to the METR study saying that developers only think they're more productive using AI for a long time past its relevance based on the actual models used in the study

very straightforwardly, you spend less time needing to review code that isn't working as the quality of the output improves

no it isn't and you do not have a grasp on the methodologies of the studies to defend them

is the effect of those interventions as strong as that from AVs?

and what are the public costs of doing those things vs AV deployment? I'm happy to let private companies spend that capital

They've made one of the fastest growing products of all time, Claude Code

you think that developers are going to stop using these tools?

The METR study uses earlier, less capable models

he was at Microsoft for four months

he's really a Google AI guy

seems to be a great model so far, probably my new default except for when I need really long context

Written pre-AI. Pure human slop

tbh I love Waffles

Because the point of the article is the new information about the world it conveys, not the exact wording used

diminishing returns to AI scale are an illusion

Economic value comes from completing long projects, not single questions. And accuracy drives how long a project AI does: small gains compound exponentially!

Reasoners are much more accurate, with big impacts

Paper: arxiv.org/pdf/2509.09677

Book 6

Lesson:

The basic paradigm of design is to coax effects from the world by identifying and manipulating its extant tendencies, rather than imposing form on it.

(They’re fighting over it because of the unprecedented adoption by normal people)

I don’t think the AI cycle is coming to a stop just yet

Ehhh I think you just want to murder people who make new things

Also I’ve been told I should share that Claude and I worked on this on the road yesterday

One can imagine a world where every iOS app developer has their own slightly different version of Apple's model