I bet they didn't quantize Opus, it's probably an infra bug like last time

Ernesto abrió debate genuino llamando a tener una mirada holística de los datos (que bien!), pero al ahondar y observar *este* gráfico se reafirma de cierta forma conclusion apresurada de aquellos que "solo ven promedio top100". Estas distribuciones son deprimentemente distintas

i am cringe but i am free

AI /infra is moving so fast that sometimes you think you are ahead of the curve when suddenly a big lab basically implements what you were in the middle of doing

Cant wait for HBO's Industry take on this whole AI-fication cycle in the financial markets

CC = Claude Code CC = (Terran) Command Center Life imitates art

I think the word 'scrappy' can be a positive adjective for any senior+ engineer in a non-developed country. You don't just pay >$20/mo to a random SaaS to get something. You start from scratch and build it yourself. That mentality is hard to find in many seniors swes elsewhere

me (and Claude) had fun with Prime Intellect hosted RL training, (we) wrote a blog post about it brunose.github.io/blog-llm-git-r… Many enterprises need agents to run complex multiple-tool sequential workflows. What if we made specialized agents using small LLMs + RL?

Qwen has been such a positive impact for people who use open source models. I'm sure the researchers behind it will do just fine. Sad to see them leave the team

>gets invited to AI Engineer event in NY >ofc it's in williamsburg aka SF v2beta

I wish someone had told me this when I started digging into diffusion language models (dLLMs) from an LLM post-training background. I've spent the last few weeks reading across both the dLLM RL literature (d1, EGSPO, MDPO, LLaDA 1.5) and the older robotics literature on

LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below

This and also vibe predicting SOTA model capabilities in 6 months to build AI infra around that prediction

1 bit large language model sounds so damn cool as a concept honestly. I wonder if it's only memory efficient or also helps mechanistic interpretability

New Anthropic research: Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.

no no i've definitely been doing a lot of work it's all just too dangerous to release

tautologer No man ever googles the same query twice, for it’s not the same query and he’s not the same man.

my respect for oai/ant employees saying good things about each other's model increases exponentially with each word