yes and it seems pretty clear that very little of their workload is people going from zero to a 1M token query with no intermediate steps
13.03.2026 21:15
👍 0
🔁 0
💬 1
📌 0
yes and it seems pretty clear that very little of their workload is people going from zero to a 1M token query with no intermediate steps
The models stay the same but the model inference serving systems are improved
Yes in the simplest view of transformer models, doing prefill on a 1M token prompt is quadratically 25x more expensive than a 200K token prompt but I suspect Anthropic is banking on people gradually building up queries so they can use smart query routing and KV caching to reduce the cost
If we're (fairly) dinging critics for not being rigorous then we should make sure to be rigorous ourselves
I think it would be fairer to give figures for inference these days, since it dwarfs training (although even for inference, water is pretty negligible compared to power and other impacts)
... and markets are hovering around flat
that's one of the most disorienting things to me
There's definitely something going on at Xai that I'd be curious to understand... a week ago Musk was beating his chest about how epic and based grok 4.20 is, now he seems to be acknowledging it kinda sucks? And also:
bsky.app/profile/rbd....
Sure but ... "everyone's dunking on my post when I posted something else about a different thing" ... ok cool I guess?
Idk, you said what you said. I don't see why people should care that you also said something narrowly hedged about some hypothetical and not defined application that no one has built to your approval
???
Huh two senior engineering people leaving cursor just as they're supposedly raising at a $50B valuation??
www.bloomberg.com/news/article...
BLVD Moving truck with the slogan PAIN IS REAL BUT SO IS HOPE
Heavy truck
If you're willing to build a global filesystem that ties into the source control for your monorepo and make a huge investment in hermetic builds, it's super nice
This is true but I feel there is a huge gap in quality of implementation between rust and bazel. Maybe it's just the fact that bazel will always be a not-quite-as-good imitation of blaze
Money is useful for
1. Buying illegal drugs
2. Hiring hitmen
3. A medium of exchange that facilitates modern technological civilization
Which one do you use it for?
Asahi hasn't even shipped M3 support yet so you'll probably be waiting a loooong time for A18 support 😞
There's a lesson about saying nice things to people here
Sleep well, brother. Your race is run.
Unironically love this.
Yooooo what the fuck
Sometimes I worry that I undercut myself by being too self conscious, but I'm confident that *some* self awareness is a useful habit of thought
I'm glad Adebayo did it, I'm just sad he didn't chuck up a couple more shots and also get Harden's record for missed 3s in a single game
Why doesn't anyone just buy the same shoes but in a size that fits and wear that pair
I don't know what's funnier, 7-22 on 3s, or 36-43 on free throws
Lol. Also lmao
Email screenshot: Hi Roland, Sailplane.ai, backed by AI kingmaker Khosla Ventures (OpenAI's first investor) and seed specialist True Ventures, is building a "self-driving cloud" - intelligent agents capable of autonomously managing the largest and most advanced AI infrastructure on the planet. Think "Waymo for cloud infrastructure".
I guess instead of “Uber for X” the cool kids are building “Waymo for X” now?
I tried one of those car vs. bike calculators and with some pretty conservative assumptions it seems like the car comes out comfortably ahead on costs?
I mean, Some of My Best Friends are VC Guys, but you gotta admit this is kinda funny
I do not think this is well supported by evidence. It would be better to have fewer cars overall, but every study I've seen suggests that even if we just replace ICE vehicles with EVs, there are substantial benefits to air quality, including PM2.5 pollution
keck.usc.edu/news/study-l...
I agree executive turnover should probably be much higher at most companies. But I still think having an *interim* CEO specifically is usually downstream of some earlier governance failure.
Nice handlebar tape too