javidlakha

One week before the first lockdown - when most offices had sent their staff home - HelloFresh were knocking door to door on my street to flog subscriptions, not minding that they might have been spreading the virus.

How much control do I have over this process? I might not want this to be abstracted.

How does that work?

This looks awesome!

How does the distillation mechanism work?

Google have released the models and code for the Switch Transformer from Fedus et al. (2021) under the Apache 2.0 licence. [0]

There's also OpenMoE - an open-source effort to train a mixture of experts model. Currently they've released a model with 8 billion parameters. [1]

[0] https://github.com/google-research/t5x/blob/main/docs/models...

[1] https://github.com/XueFuzhao/OpenMoE

Me too - particularly when the story proof was more concise than the algebraic proof, or hinted at the deeper reason as to why something must be true. I remember them better too.

Congratulations on the launch! (I'm a beta user.) Audit trails, access controls and security certifications are big headaches when developing in regulated industries. Having these already set up has made it easier for us to experiment with and build on LLM APIs.

*Motivation*

I wanted to learn about serverless computing and so I built Unofunction, a Python Lambda that wraps LibreOffice’s headless mode and can convert any document format that LibreOffice can import to any document format that LibreOffice can export (e.g. DOCX to PDF).

*Implementation*

Calling LibreOffice in a Python Lambda required compiling LibreOffice for the Amazon Linux 2 base image (CentOS 7). I originally tried compiling LibreOffice locally (using Docker) on my 2021 M1 MacBook Pro (10 CPU cores) but gave up after a few hours. On an AWS EC2 c6i.8xlarge instance (32 vCPUs, $1.36 per hour in eu-west-2), compiling took ~30 minutes. In case someone else might need a prebuilt headless LibreOffice that has been compiled for Amazon Linux 2, I’ve uploaded the resulting image to https://hub.docker.com/repository/docker/unofunction/libreof...

A crucial implementation detail is that because AWS Lambda does not permit files to be written, except to the /tmp directory, LibreOffice needs to be called with the argument '-env:UserInstallation=file:///tmp/’. Otherwise, it will attempt user installation in .config/libreoffice and fail.

*Thoughts*

I enjoyed using the AWS CDK a lot. There is something magical about being able to deploy infrastructure by manipulating it as if it were a Python or a TypeScript object. I preferred it to Terraform’s declarative approach.[1] My only criticisms are that deployment can be extremely slow and that destroying S3 buckets can be clunky (they need to be emptied first).

Testing code deployed to AWS is hard. Neither AWS SAM nor LocalStack perfectly simulates a real deployment: code which works locally does not always work on Lambda. (An example of this is deploying files to the /tmp directory. SAM allows such files to be edited or overwritten; Lambda makes them read-only.)

[1] This article is a good comparison between Terraform and the AWS CDK: https://medium.com/swlh/cdk-or-terraform-88a464bedf9e