mistral on 8x 7B MoE Model just dropped
@soumithchintala
-
Standards and Benchmarks for AI Safety and Innovation
By
–
it connects the Safety & Responsibility crowd with the Model/Data Innovation crowd, and tries to establish standards and benchmarks that these two sets of crowds can agree are good.
Think of it as establishing a standard that you can subscribe to if it benefits your cause — -
Commercializing Superior LLM Models: Safety Certification Strategy
By
–
You've created a superior llama/mistral-derivative model (like @teknium often does).
How can you convince the world to use it (and pay you)? Step 1: You need a 3rd party to approve that this model is safe and responsible.
the Purple Llama project starts to bridge this gap! -
Google’s AI Model Competes with GPT-4, Cuts Infrastructure Costs
By
–
Seems to solidly compete with GPT-4 on benchmarks.
Google has existing customers and surfaces to start the feedback loop, without worrying about adoption.
And Google will use TPUs for inference, so doesn't have to pay NVIDIA their 70% margins (like @OpenAI and @Microsoft has to -
Intrinsic Motivation Drives Research Impact at FAIR
By
–
and there are countless people that I know at FAIR that literally only cared about their intrinsic motivated rewards and did quite well (including getting super senior), because as a side-effect of their great intrinsically motivated research was great extrinsic impact. they
-
Scientist Evaluation Methods in AI Research and Publication
By
–
i haven't kept up since i moved to infra ~1.5 years ago (to take care of pytorch), but Scientists were evaluated in one of three ways (people could do parts of each too):
1. finish a fully formed work, and publish. You'd only get rewarded once the published work is validated as -
Publication quotas constraints for top AI researchers questioned
By
–
that's not a thing lol.
you think researchers like ross and kaiming would be held over publication quotas? -
A Decade of AI Innovation: Meta’s Transformative Contributions
By
–
10 crazy years — pytorch, detectron, segment-anything, llama, faiss, wav2vec, biggraph, fasttext, the Cake below the cherry, and so much more.
Can't say we didn't change AI and to an extent the world. -
PyTorch LLM Inference Optimization Beats Industry Standards
By
–
using simple pytorch code and advanced features, you get LLM Inference speeds that beat most LLM Inference engines out there. Checkout this great post and work from @cHHillee
-
Verse-jumping Technology: Parallel Self Access and Memory Integration
By
–
Reminds me of verse-jumps.
"In the Alphaverse, the late Alpha-Evelyn developed "verse-jumping" technology, which enables people to access the skills, memories, and bodies of their parallel selves by performing bizarre actions that are statistically unlikely." https://
x.com/katherine1ee/s
/katherine1ee/status/1729690964942377076
…