To me, the biggest thing missing from the r1 paper is info on their advantage functions. They're pretty vague about where they sourced math and engineering problems with answers.

everything about this is so good. miscounting the number of 'e's. the awkward emoji. assuming you are also an assistant.

Does anyone know what would give model development a moat? With meta doing their thing it isn't going to be compute, and r1 shows that the algorithms are easy enough to reproduce so far.

unlike the twitter hype, I tried gemini flash 2.0 image editing and it totally failed on several different tasks. I guess people are less inclined to post things that don't work.

"outputs were graded by deloitte" sorry what?

This is the last thing I worked on at meta, one of the coolest projects of my career. Learned a lot about how ads work. Also made me extremely bearish on europe.

Ghibli-fied my cousin who is so non-technical she doesn't own a laptop and she asked how to sign up on her phone. The image use case is a fun toy, but I bet a lot of these people end up sticking around and using it for other stuff.

Uhhhh llama.com/llama4/

Can ios apps detect other apps being opened? I've noticed if I'm waiting to be matched to a driver on uber and open lyft I immediately get assigned a driver.

We explore a new dimension in scaling reasoning models in Adaptive Parallel Reasoning APR lets LMs learn to orchestrate both serial & parallel compute E2E via supervised training + RL — w/ better efficiency and scalability than long CoT on Countdown 🧵 arxiv.org/abs/2504.15466

An AI Alignment researcher at Anthropic (Sam Bowman) just said that Claude Opus will CALL THE POLICE or LOCK YOU OUT OF YOUR COMPUTER if it detects you doing something illegal??? i will never give this model access to my computer

This is so fucked up. A sitting US senator being handcuffed for trying to ask a question at a DHS news conference? No matter what your political party, you can't think this is a good thing.

hmm

Reproduced a study on opoid misuse today as a test for our new tooling. Total cost was around 10c per patient vs several dollars in the original study, and from setup to getting results it took less than an hour.

Anyone knows adam?

We’re living in a brief moment where parents have a quaint idea of what an electric bike is and don’t realize they’re buying their kids motorcycles.