engineering

Your Docker containers are slow, bloated, and vulnerable.  And you’re probably making the same mistakes 90% of engineers make in production. 
 
Take my advice and please
• Don’t use “latest” tags, instead use specific versions like node:18.17-alpine  

• Don’t build

open_clip/docs/openclip_retrieval_results.csv at main · mlfoundations/open_clip

Do the simplest thing that could possibly work

Cursor - The AI-first Code Editor

I finally found how to have the nice #SwiftUI Text+timer style animation in an app (the default behaviour in a widget). I just missed to specify a transaction for the Text. It's working like a charm now!

.transaction { t in
        t.animation = .default
}

cc @frboulais 🎁

GitHub - RahulSChand/gpu_poor: Calculate token/s & GPU memory requirement for any LLM.  Supports llama.cpp/ggml/bnb/QLoRA quantization

Want to build better charts?

Here's the full breakdown of how I built this one 👇

Improving Retrieval on Ramp with Transaction Embeddings

A nice @huggingface space to estimate VRAM for running an LLM

mf trained a simple model to translate ada-002 embeddings back to text and found something interesting:

sentence embeddings have arithmetic properties

the classic "king - man + woman = queen" still works just like w/ token embeddings — but can go a few steps beyond that as well

Traversing through CLIP Space, PCA and Latent Directions | Notion

Comparing full text search algorithms: BM25, TF-IDF, and Postgres

Hot tip - using gpt-4-mini as a reranker gives you better results, and now with strict mode it's just as reliable as any other reranker model.

Annotate | Craft walkthroughs that clarify code

I'm super impressed with Qwen-VL and CogVLM!

I've done a few (probably very naive) tests to compare LLaVA, BakLLaVA, Qwen-VL, CogVLM, and GPT-4V.

Tests include VQA, OCR, and zero-shot detection.

Any ideas on what else I should test?

Danny Lin on X

GitHub - GrantCuster/gemini-spatial-example: How to use bounding boxes with the Gemini API

Cognitive load is what matters

I couldn’t agree more, this is the secret

SwiftUI Aurora Background Animation — 🐙 Cephalopod Studio

Iconoir | Free Icons

One of the most useful tricks I learnt from @giffmana: automatically starting pdb debugger when program crashes.

It's just 5 lines of code that has probably saved me countless hours in trying to debug tricky bugs.

LLMs are slow. 
Caching can help, but developers should have control over how caches operate at different scales

In next week's blog, we will show you how to write a @pydantic compatible cache mechanism for use with Instructor. 

1. For a single session: (lru_cache)
2. For…

A Beginner’s Guide to Fine-Tuning Mistral 7B Instruct Model

Cabana - Design System for Figma

GitHub - markst/hotreloading-vscode-ios: Demonstrating iOS development environment in VSCode using XcodeGen + HotReloading.

Open AI released a great Prompt Engineering guide (link in comments). 

Here is a summary of their 6 strategies for getting better results when prompting GPT-4

GitHub - stas00/ml-engineering: Machine Learning Engineering Guides and Tools

i find it so insane alibaba can drop 3 new embedding models outscoring ada-002 on almost all benchmarks (with lower dims)

embeddings are great for similarity/search but also can improve how text is represented at train time

and you can run them on your own cheaper than ada2

OrbStack · Fast, light, simple Docker & Linux on macOS

This was much needed! Select a LLM, your vendor and your GPU and see if you can run it locally 🔥

