mr. TIM

timkellogg.me

19 February 2025 00:09

overall, like we’re finding repeatedly, while the transformer architecture isn’t perfect, it’s pretty darn good it seems like every problem has a mitigation, and much of the time the problem simply wasn’t as big as we thought

mr. TIM

timkellogg.me

19 February 2025 00:09

while writing this, i had an epiphany about reasoning models & the shortcomings of autoregressive models (explained more in detail in the article) tl;dr the way the reasoning trace is generated, it allows the LLM to make mistakes & fix them, which side-steps a lot of the shortcomings

AI Firehose

ai-firehose.column.social

19 February 2025 00:09

Breakthrough research unveils uOCF, an unsupervised learning method for zero-shot 3D object discovery from a single image, overcoming generalization limitations. This innovation enhances understanding of complex real-world scenes without prior object annotations. https://arxiv.org/abs/2402.07376

Daron Yondem

daron.me

19 February 2025 00:09

Currently diving into the series myself - anyone else checking it out? Would love to hear your thoughts! Let's learn together! 🤓 https://buff.ly/3CTgzKx #MachineLearning #AI #DeepLearning

Daron Yondem

daron.me

19 February 2025 00:08

Just discovered something cool! 🚀 There's a new YouTube series teaching devs how to build a mini version of DeepSeek from scratch. Raj Abhijit & team are breaking down complex LLM concepts into 15-minute videos (40+ hours of content!)