Hmmm, I have a feeling this plot might need an overhaul rather soon🤣.

I guess phi-2 was the lower left part of the triangle. I wonder what those guys have been up to in the last 6 months? 🤔

As part of our ongoing work on AI safety and security, we've discovered a powerful, yet simple LLM jailbreak that exploits an intrinsic LLM behavior we call 'crescendo' and have demonstrated it on dozens of tasks across major LLM models and services: …ndo-the-multiturn-jailbreak.github.io

At a time where 314B parameters models are trending, come join me at #NVIDIAGTC to see what you can do with 1 or 2B parameters :-) (and coming soon, what can you do with 3B?!?)

Elon Musk and Sam Altman may not agree on much of late, but do agree AI is built on strong math foundations, including algebra and calculus, applauding University of California for recent clarifications on math requirements for admission.

Many industry leaders signed:

mathmatters.ai

Interestingly the unicorn test is much harder today than back in 2022, because there are now a lot more crappy tikz unicorns on the web ... I guess this has some broader significance ...

Tune in on Tuesday for a really fun set of conversations on cutting edge AI research!

Introducing Mixtral, Phi2, Falcon, and Qwen support in #DeepSpeed -FastGen!

- Up to 2.5x faster LLM inference
- Optimized SplitFuse and token sampling
- Exciting new features like RESTful API and more!

For more details: github.com/microsoft/Deep…

#DeepSpeeed #AI

🚨Princeton ML Theory Summer School🚨

Aug 6 - 15, 2024

Speakers:

* F. Krzakala/L. Zdeborova Krzakala Florent Lenka Zdeborova
* S. Rakhlin
* B. Lourerio Bruno Loureiro
* M. Austern
* D. Krotov Dmitry Krotov
* J. Altschuler

Details: mlschool.princeton.edu

Sponsors:…

Feels good to have some external validation of our model ☺️. Thanks anton !

I did multiple fine tuning runs over tinyllama-3T and phi-2 using the same dataset (which I think is actually a good dataset of multiturn conversations)

Results? Tinyllama can't properly follow multiturns. The model gets confused and brings up completely unrelated completions

Starting the year with a small update, phi-2 is now under MIT license, enjoy everyone!
huggingface.co/microsoft/phi-2

Looking for a #postdoc and would like to work at Yale on emerging problems in the areas of foundations of AI/responsible AI?

Have your CV, statement, and 3 letters emailed by Jan 15, 2023
cs.yale.edu/homes/vishnoi/…

Strong math and empirical background required

Please fwd/rt!

Check out this short video for a brief discussion of the phi series with Microsoft CTO Kevin Scott , including why 'textbooks' in 'Textbooks Are All You Need' might not be exactly what you have in mind.

youtu.be/O-DjHgZt-Uk?si…

This video of a chat between Microsoft CTO Kevin Scott and MSR AI researcher Sebastien Bubeck is full of interesting insights about the phi family of large language models, and AI in general. Only 12 minutes, definitely worth watching. youtube.com/watch?v=O-DjHg…

2023 has been the most consequential year in Microsoft Research's 33-year history. In addition to countless internal (to Microsoft, OpenAI, etc) contributions, I bet that again our labs contributed more per researcher than any others to open research. Read more microsoft.com/en-us/research…

We're so pumped to see phi-2 at the top of trending models on Hugging Face ! It's sibling phi-1.5 has already half a million downloads. Can't wait to see the mechanistic interpretability works that will come out of this & their impact on all the important LLM research questions!

Phi-2 by Microsoft AI is now the #1 trending model on Hugging Face (hf.co/models). 2024 will be the year of smoll AI models!

Check out this short video for a brief discussion of the phi series with Microsoft CTO Kevin Scott , including why 'textbooks' in 'Textbooks Are All You Need' might not be exactly what you have in mind.

youtu.be/O-DjHgZt-Uk?si…

This is a really excellent video, including an insightful discussion of benchmarks like MMLU: youtube.com/watch?v=nPgs8T…

I'm immensely proud of the work that my colleagues in Microsoft Research showcased at this week’s NeurIPS Conference regarding foundation models of all sizes, including Phi-2, which matches or outperforms models up to 25x larger.