What's pillowfort?

Tumblr clone.

Yeah they read the news this morning and whipped up a quick Torch->XLA translation layer and coordinated with four different inference engines (who also all wrote their own blog posts) and shoved this out the door quick to stay on top of the news cycle

Why be rude on purpose? OP clearly prefaced by saying it was a dumb question

they're releasing this today because the pytorch conf is this week...

Products can also be a moat.

Only to figure out that the vibration in his drone was due to vibration and it was easily fixed by swapping to rotors that he should've used in the first place. Oh well! Live and learn.

These things are not always as obvious as you make them out to be.

I was fighting a lot of similar instability when making my drone only to realize that a software regression many versions prior had introduced a bug that made flying the drone next to impossible. After fixing that bug a whole pile of issues went away. Debugging that on a drone that does not fly is super hard because you never get to the point where you can properly observe what is going on. So improving the observability by doing that camera trick was just what was needed here.

And that's why mathematicians are so pissed about the OpenAI math paper drop.

US dominance of the paraeto frontier is really striking. All over hackernews you hear about how open source models are "good enough but a lot cheaper" - it seems to me that they would be better served by using a frontier model on low thinking.

Whoa, your tinfoil idea blew my mind. It's totally possible too - all you'd need to do is perhaps select the nearest token in embedding space to the thinking token. Or perhaps train a model on those thinking vectors to map them approximately onto token embeddings.

Especially with OpenAI model's weird style of thinking.

Yup. We've been living in a world with ZERO obvious software mistakes and now that these LLM things are running wild, we've got more obvious bugs than ever!

Yes, they are building frontier models and should be responsible to handle basic text input validation.

An additional point here is that Anthropic is hiding behind bots so they don't have to fix it and support humans with real world problems.

It's frustrating to be forced into a financial burden just because they decided to cut corners on traditional software debugging protocols.

I'm normally more of an AI hater but this just seems objectively good so I'll ignore it and move on

It's because their datacenter businesses are growing so fast YoY. Why do you care about the difference in a few percentage points of interest if you're getting 37, 41, or even 82% returns YoY?

Correct hyperscalars are doing fine - as it stands.

The problem is OAI/Anthropic - but if OAI and Anthropic fail the whole thing comes apart. Goog, Meta etc will be fine as they had tremendous cash flow generating operations prior to datacenter offerings w.r.t LLMs.

If OpenAI and Anthropic somehow become insolvent to the point that they are unable to continue running inference on their existing trained models, their tens of billions of dollars of monthly revenue will need to go somewhere.