Machine Learning’s Post

The team at OpenAI just showcased the power of data-centric model development. They reduced inference costs by 50x while maintaining accuracy, with just 1,000 well-curated training data examples: https://lnkd.in/gHNv8fCn. Doing this programmatically is where Snorkel excels: https://snorkel.ai/

This guide of OpenAI o1 series models shows how prompting this new model is different and requires simpler prompts and a more structured input context. https://lnkd.in/eDw3Dx4w

A handy reference for CGPT from the folks at OpenAI: use clear instructions, provide reference text, break down complex tasks, prompt models to reason from first principles, offload tasks to other tools, and create comprehensive tests. https://bit.ly/3SjxLOg

OpenAI Whisper is one of the top open-source speech recognition systems out there today. Running a simple demo? Easy. Getting a performant system in production? Hard. Here are a few libraries for improving the inference latency and memory requirements of Whisper models: ✅ https://lnkd.in/epMkd4S5 (Whisper in C/C++, great for running on low-resource edge devices, highly memory efficient) ✅ Hugging Face transformers has native support for Whisper models with out-of-the-box batching provided ✅ https://lnkd.in/e-gFzUH6 (4x+ speedup over OpenAI default lib) ✅ https://lnkd.in/eeBWGbF8 (Whisper models in JAX, 70x speedup on TPU over OpenAI default lib) ✅ https://lnkd.in/eA7eBfcK (distilled models, 6x speedup, 50% smaller, within ~1% WER of base model) ✅ https://lnkd.in/efvSDECp (terminal tool transcribing 2.5 hours of audio in 98 sec) Any I should add? #whisper #openai #asr #aiinproduction

A handy reference for CGPT from the folks at OpenAI: use clear instructions, provide reference text, break down complex tasks, prompt models to reason from first principles, offload tasks to other tools, and create comprehensive tests. https://bit.ly/3SjxLOg

A handy reference for CGPT from the folks at OpenAI: use clear instructions, provide reference text, break down complex tasks, prompt models to reason from first principles, offload tasks to other tools, and create comprehensive tests. https://bit.ly/3SjxLOg

The OpenAI Assistants API now supports vision. You can create messages with image URLs or uploaded files, and your assistant will use the visuals as part of its context for the conversation.

Learn how to setup a local, private, quantised model with an OpenAI compatible API server that you can directly interact with via LMStudio and LangChain. The popularity of projects like LMStudio, PrivateGPT, llama.cpp, Ollama, GPT4All, llamafile, and others underscore the demand to run LLMs locally (on your own device). Let me introduce our latest course on Vexpower "Using LangChain + Llama3 Locally with LMStudio", where you can learn all about Local LLM Inference. Link: https://buff.ly/3wmkhJw

Gone are the days when text preprocessing involved manually crafting complex rules for stemming, lemmatization, and stop words removal, each requiring meticulous tuning for specific tasks. Enter the era of the Tokenizer algorithm. This groundbreaking tool streamlines the preprocessing stages, enabling us to channel our energies towards addressing the core challenges at hand. Exploring the OpenAI tokenizer tool has been an absolute delight https://lnkd.in/g63hMXfm

Context is still your biggest challenge in RAG - getting the right information from your documents to the LLM so it can answer questions. OpenAI just released is prompt guidelines for its new o1 models: “Limit additional context in retrieval-augmented generation (RAG): When providing additional context or documents, include only the most relevant information to prevent the model from overcomplicating its response.” from https://lnkd.in/eDDW8wan Despite repeated claims to the contrary, “context stuffing” still doesn’t work. You still need a product like Pinecone and an engineer who can optimize your LLM’s performance.