toxicity VS Lean and Mean Docker containers

I'm a PM at a human data company (https://www.surgehq.ai) that helps the large language model companies ensure their models are safe (we're the “clever prompt engineers” who helped Redwood assess their model performance).

We actually just published a blog today that includes our perspective on building “AI red teams” and best practices for AI alignment/safety: https://www.surgehq.ai/blog/ai-red-teams-for-adversarial-tra...

I'd love to chat. Want to reach out to the email in my profile? I'm the founder of a much higher-quality data startup (https://www.surgehq.ai), and previously built the human computation platforms at a couple FAANGs.

We work with a lot of the top AI/NLP companies and research labs, and do both the "typical" data labeling work (sentiment analysis, text categorization, etc), but also a lot more advanced stuff (e.g., training coding assistants, evaluating the new wave of large language models, adversarial labeling, etc -- so not just distinguishing cats and dogs, but rather making full use of the power of the human mind!).

> Rather than operating under a strict definition of toxicity, we asked our team to identify comments that they personally found toxic.

[0]: https://github.com/surge-ai/toxicity

Love language? So do we, and our mission is to infuse AI with that same love. At Surge, we're building the human infrastructure to power NLP — from detecting hate speech, to parsing complex documents, to injecting human values into the next wave of language models. Our first product is a platform that helps ML teams create amazing, human-powered datasets to train AI in the richness of language. We're a team of former Google, Facebook, and Airbnb engineering leads, and we work with top companies at the forefront of machine learning. Our tech stack is Ruby on Rails, React, and Python. We’re rapidly growing, and we're looking for full-stack engineers to join the team and develop our product. To apply, please email [email protected] with a resume and 2-3 sentences describing your interest in Surge. We love personal projects and writings too!

More information: https://www.surgehq.ai/about#careers

A blog post explaining the problems we are working to solve: https://www.surgehq.ai/blog/the-ai-bottleneck-high-quality-h...

And if you want to make the container quickly secure without bloats, maybe give this a try https://github.com/slimtoolkit/slim

Slim.ai presents the data in a more user friendly way than many of the other tools in this post. On top of its open source SlimToolkit for identifying the contents of an image, Slim.ai uses Trivy for vulnerability scanning.

What about https://github.com/slimtoolkit/slim?

A last practice that I do not use at all and which may interest you is to use slim toolkit to keep only the useful elements in your final image.

Anyone tried using https://github.com/docker-slim/docker-slim To minify an image?..

Distroless images are better left for people with serious need for lightweight images and good Linux knowledge because they require lot of planning with the build so that they stay light and work. If you need lighter images but docker isn't your main tool and you can't afford to take hours and hours of practicing different build strategies you can check docker-slim (https://dockersl.im/). With this tool you can easily size down the images.

Maybe this would help in that regard: https://github.com/docker-slim/docker-slim