What does DeepSeek R1 & v3 mean for LLM data? Contrary to some lazy takes I’ve seen, DeepSeek R1 was trained on a shit ton of human-generated data—in fact, the DeepSeek models are setting records for the disclosed amount of post-training data for open-source models: – 600,000
RESEARCH
-

Stochastic Parrots Day 2023 Recording and Resources Available
By
–
Friends, now might be a good time to revisit our Stochastic Parrots Day held in March 2023. You can find the recording of the session and additional resources on our website: https://
dair-institute.org/stochastic-par
rots-day/
… -

Oil Energy Enabled Population Growth and Technological Acceleration
By
–
It’s no accident that human lifetimes and human populations both exploded in the late 1800s and early 1900s when oil discovery and usage ramped up. And it directly contributed to the rapid growth of technology, because it’s what allowed the vast majority of people to stop working
-
COT Benchmarking Training in Modern AI Models
By
–
these were easy to benchmark. COT would be easy as well but is trained into most models these days so i dont think it would change too much
-
Framework-Agnostic AI Optimizers vs DSPy Dependency
By
–
dspy more powerful, but requires using the dspy framework (in addition to their optimizers) i think framework-agnostic optimizers are more usable
-

AI as Creative Tool Beyond Labor Automation
By
–
The prevailing mindset that AI is only a labor-saving tool betrays a lack of understanding of why people create, says @StanfordHAI associate director @gewang in his latest @StanfordHAI piece. https://
stanford.io/3EhQxB1 -
Post-Training: Adapting AI Models for User Preferences and Safety
By
–
Post-training involves much less compute and faster iteration cycles than pre-training. With post-training, you can adapt the model to user preferences, craft the model personality, or introduce safety behavior, says @barret_zoph
. -

ChatGPT Development and Post-Training Techniques with OpenAI Leaders
By
–
At today’s @StanfordHAI seminar, @OpenAI co-founder @johnschulman2 and VP of research @barret_zoph described the early days of ChatGPT and post-training these popular models. https://
hai.stanford.edu/events/john-sc
hulman-and-barret-zoph-chatgpt-and-art-post-training
… -

BrightHeart Uses Meta’s DINOv2 for Fetal Heart Screening
By
–
How BrightHeart uses Meta’s DINOv2 to transform fetal heart screenings https://
go.fb.me/wx8nyt -

DINOv2 Detects Congenital Heart Defects in Children Faster
By
–
Using the open source DINOv2 model, a medtech company founded by two pediatric cardiologists developed tools to help clinicians identify or rule out signs of congenital heart defects in children faster and more accurately.
