"Medea: scheduling of long-running applications applications in shared production clusters" Garefalakis et al., EuroSys'18. blog.acolyer.org/2018/06/13/med… #themorningpaper If you run clusters with a mix of batch and long-running (e.g., web) containers, then check out Medea...

Twitter is moving its Deep Learning infrastructure (DeepBird) to TensorFlow, and seeing the benefits. blog.twitter.com/engineering/en…

Think about it. You've always been told you are smart. You've interpreted that as you are always right. Reality: you aren't as smart as you think you are, and you are also an ass. x.com/soniagupta504/…

Organizations have a lot of state, which means they change slowly. I think the hardest cycle of org leadership is (a) diagnosing problem, (b) placing a bet, (c) keeping the faith and creating the space to see if you picked right. lethain.com/durably-excell…

I hate circling around this, but I hate seeing and participating diversity in tech discussions. I’m taking a new approach. If you are a person of color who enjoys this whole Kubernetes thing and wants to work with me, reach out. I have some ideas. Please RT for reach.

Doing is the engineering superpower. In the absence of doing, nothing gets done. We don't always have to convince and cajole and coerce others into building on our behalf, we can just *build^. This may seem basic, but it matters.

Just read this cool paper: Neural net pretraining keeps improving when you train on an unprecedented 3.5 billion (that's really big) labeled images and transfer to new task. IMO we're still nowhere near the limits of pretraining/transfer learning. research.fb.com/publications/e…

Interesting debates with OSS business models. We all know they are broke, glad to see more movement to see what can be done. I believe there are only 3 models that will work long term. 1. Bundle OSS w/ Hardware 2. Run it as a Service 3. Augment it with a Service.

We are running this huge experiment globally with Kubernetes but there is already a decade-run experiment from Google. No one understands borg config at this company. No one likes to engage with this abstraction layer.

Predictive test selection to ensure reliable code changes - Facebook Code code.fb.com/developer-tool…

[v1.0] : JIT Compiler, Faster Distributed, C++ Frontend, CUDA10 Read more about the changes at github.com/pytorch/pytorc… As always, get the install commands on pytorch.org

research.fb.com/publications/m…

A few nuggets from Geoffrey Hinton's talk from earlier today at the #ml4h unconference. First up, the distinction between statistics and AI (and presumably ML by implication). Overall, I think these are pretty clean contrasts:

On the (unexpected) connection between neural nets and decision trees. Decision trees are inefficient because the leaf nodes operate on a very small fraction of the full data. DNNS with ReLU can be thought of as a kind of decision tree with massive amounts of parameter sharing.

Introducing PyTorch 's TVM-based backend by Bram Wasti (Facebook). seamless integration and great speedups. 🔥🚀🚀tvm.ai/2019/05/30/pyt…

[1912.01412] Deep Learning for Symbolic Mathematics arxiv.org/abs/1912.01412

citizenlab.ca/2020/04/move-f…

Software Engineering is evolving for Machine Learning. New forthcoming tools: - Version control for ML models & pipelines - IDEs for neural network architectures - Compilers & Hardware Instruction sets for AI chips - Programming languages with stochastic functions that learn