Eric Meier

Give me back o3

DON'T EVER FUCKING FORCE CLOSE MY TEXT EDITOR FOR AN AUTO UPDATE AGAIN, CURSOR, YOU PIECE OF SHIT

AI products fail constantly—in ways both hilarious and terrifying. Regular software throws exceptions. But AI products fail silently. Meet @raindrop_ai : the first Sentry-like monitoring platform for AI products.

@benhylak Nice they finally caught up with cloudflare workers 🙌

..yo this is insane

This week has had the worst takes in AI I’ve ever seen.

Get on the COS private Beta

10 daily routine life hacks that will take your physique to the next level ⬇️

@benhylak You should take a look at AlloyDBs ScaNN integration

pgvector is insane. i don't understand why vector databases exist.

@hanchchch Wait, what? This is already a thing apparently

@hanchchch Native tracing inside ray would be dope indeed

anyone tried tracing data preprocessing with better methods than just printing timestamps, for example like otel and jaeger? I mean if you're dealing with millions of batches it may be worth putting more efforts in optimizing

@cloneofsimo Ray is amazing for this stuff at scale, makes sharding across all resources in the cluster super easy

I wrote a custom c++ dataloader and im getting 2000 images/sec with 4 threads Im very frustrated to see my reference pytorch implementation so slow, to a point i feel like something is straight up wrong with my implementation which ive been doing for past 5 years is there hidden config for torch.data.DataLoader that im missing that can fill this gap?

@zacharynado AI studio has no batch jobs though. And Gemini via vertex ai is buggy af

@dataghees jax.readthedocs.io/en/latest/pers…

@dataghees Yeah, without jit it’s quite slow. JIT takes ages on larger models though, but at least there is caching that can help you with that

@_xjdr has me learning Jax

@isidentical Nope, worked fine for us

was R2 down an hour ago (~8:30-9:30 am PST)?

Training models makes you realize how ridiculous AWS egress charges truly are. Thank you, R2 🙏

@hyhieu226 @_xjdr Torch xla is amazing for this. Do small tests / dev on GPUs / switch over to XLA for the large run

@_xjdr Most people also don't know that XLA takes 1 hour to "compile" any reasonably large Transformer on any reasonably large fleet of TPUs 😒

If google ever started selling TPU hardware and released internal tooling, they'd MOG nvidia so bad. Just a trillion dollar company waiting to be built. most people don't realize how good JAX + TPUs + (other stuff) really is.

@Christophepas What’s launching? 👀

@anthonyriera @usebreeew pinggy.io is super nice as well. No install required and dirt cheap

I love localcan! I use it daily to code @usebreeew > Allows me to code with Stripe webhooks simply > Consistent local https URL I can use with other services > Simple UX > Beautiful UI

@qtnx_ @Laz4rz Went from jax to torch xla, they switched to pjrt recently