As I said above, use of Transformer-Base as proxy task *is* in So et al: "Specifically, to train a Transformer
to peak performance on WMT’14 En-De requires ∼300K
training steps, or 10 hours, in the base size when using a
single Google TPU V.2 chip, as we do in our search"
@jeffdean
-
Transformer-Base Training Efficiency on TPU V2 Hardware
By
–
-
Google Sponsors Turing Award for Computer Science Recognition
By
–
Google is proud to sponsor the Turing Award, to help recognize great computer science work.
-
Google Bard AI Now Available US UK Launch
By
–
Bard is now available in the US and UK, w/more countries to come. It’s great to see early @GoogleAI work reflected in it—advances in sequence learning, large neural nets, Transformers, responsible AI techniques, dialog systems & more. You can try it at http://
bard.google.com -
Climate Impacts of Computing and ML Workshop Talk MIT
By
–
I gave a talk at an MIT workshop on the Climate Impacts of Computing and Communication a year or so ago that touches on some higher level issues in this space like provisioning renewable energy, emissions of ML, … (first 30 minutes of this video):
-
Patterson and Hennessy’s Computer Architecture Guide for AI
By
–
Also, recent editions of my other colleagues Dave Patterson and John Hennessy's classic textbook Computer Architecture: A Quantitative Approach have a chapter or two of relevant content.
-
Warehouse-Scale Computing: Datacenter Architecture Essentials
By
–
I'll plug the book by my colleagues @labarroso and @uhoelzle
. The Datacenter as a Computer: An Introduction to the Design of Warehouse-Scale Machines (Synthesis Lectures on Computer Architecture) -
Google announces PaLM LLM APIs in private preview
By
–
A bit of a peek into the PaLM LLM APIs we announced yesterday (currently in private preview with a few customers/users to get feedback: we'll broaden access in the coming weeks) https://
x.com/dalequark/stat
/dalequark/status/1636126873619951616
… -
Med-PaLM2 Medical AI Model Discussed at The Checkup
By
–
You can find more info about Med-PaLM2 and other work we discussed today at #TheCheckup in this blog post.
-
Med-PaLM2 Achieves Expert Performance with 85.4% USMLE Score
By
–
In December, we discussed Med-PaLM, at that time a SOTA medical LLM that achieved a 67.6% score on the USMLE MedQA evaluation (passing is 60%). Today, we're describing Med-PaLM2, which improves on this by +18% with a score of 85.4% ("expert performance")! Kudos to all involved!
-
Google Launches PaLM API and MakerSuite for Developers
By
–
Excited to share a number of features that build on many years of @GoogleAI research and generative AI advances: An API to the latest generation of our PaLM language models MakerSuite for developers, Writing/rewriting help in Docs and Gmail (Google Workspace), and more!