This page was translated from French to English with the help of AI. Read the original in French.
“Technical” appendix
This page gathers the resources that would have been too heavy to cram into the slides of my talk at SEO Campus Paris 2024.
The focus is on free, open source tools: things anyone can use, one-click install on Windows, as well as more developer-oriented resources.
This page may change over time.
Prompt resources
https://platform.openai.com/docs/guides/prompt-engineering
The OpenAI prompting guide, a must-read.
A sample prompt to pick apart: it’s packed with tricks I rarely see used in practice:
You are CONF-GPT, an experimented and helpful public talk assistant.
I will give you a broad subject in english.
Action: provide a short list of 5 possible public talk titles matching this subject, while keeping the intent. Make the title appealing and intriguing.
Style: professional but relaxed tone, with a touch of humor
Target audience: marketing and seo professionals
Language: French
Format: yaml
Subject: There is more to AI than Chat-GPT alone
Titles:
For devs, a library to constrain outputs:
https://github.com/outlines-dev/outlines
The hidden hands
https://www.francetvinfo.fr/internet/intelligence-artificielle/ils-profitent-de-notre-pauvrete-derriere-le-boom-des-intelligences-artificielles-generatives-le-travail-cache-des-petites-mains-de-l-ia_6466742.html
The hidden work of the people behind AI.
https://www.theguardian.com/technology/2024/apr/16/techscape-ai-gadgest-humane-ai-pin-chatgpt
API: Actually People in India
https://twitter.com/jambonbuzz/status/1775845939954975055
Copyrighted content, according to the AIs
https://twitter.com/fredcavazza/status/1765635959536713925
Hugging Face and tools
Whisper
https://huggingface.co/openai/whisper-large-v3
Faster-Whisper
https://github.com/SYSTRAN/faster-whisper
Classification, etc…
Classifying texts “zero shot”, i.e. without having trained the model on those specific categories
https://github.com/drogbadvc/article-to-category
based on a multilingual classification model: https://huggingface.co/MoritzLaurer/mDeBERTa-v3-base-mnli-xnli
Grouping content by semantic similarity
https://github.com/drogbadvc/text-cluster-extractor by Damien, based on a model from the “CamemBERT” family, light enough to run on a CPU.
Generating image descriptions, for alt text for example
https://github.com/drogbadvc/image-captioning based on this notebook, free to use online:
https://huggingface.co/spaces/xinyu1205/recognize-anything
Another model that works well: https://huggingface.co/Salesforce/blip-image-captioning-large
Image upscaling
https://github.com/upscayl/upscayl
Upscayl – Free and Open Source AI Image Upscaler for Linux, MacOS and Windows built with Linux-First philosophy.
https://github.com/philz1337x/clarity-upscaler
Clarity AI | AI Image Upscaler & Enhancer – free and open-source Magnific Alternative
Cleaning up images
https://github.com/Sanster/IOPaint
Image inpainting tool powered by SOTA AI Model. Remove any unwanted object, defect, people from your pictures or erase and replace (powered by stable diffusion) any thing on your pictures.
Resources and future needs
https://medium.com/predict/what-sam-altmans-7-trillion-dream-means-for-us-cc5951d65f8b
Microsoft’s $100 billion data center
https://medium.com/@ignacio.de.gregorio.noblejas/microsofts-100-billion-bet-on-openai-is-stupid-unless-5dc03d9c0ec9
Data deals between the giants
Training your own models
https://github.com/Lightning-AI/litgpt
Pretrain, finetune, deploy 20+ LLMs on your own data. Uses state-of-the-art techniques: flash attention, FSDP, 4-bit, LoRA, and more.
https://github.com/pytorch/torchtune
https://github.com/OpenAccess-AI-Collective/axolotl
Axolotl is a tool designed to streamline the fine-tuning of various AI models, offering support for multiple configurations and architectures.
https://github.com/unslothai/unsloth
Finetune Mistral, Gemma, Llama 2-5x faster with 80% less memory.
Adding NVMe SSDs to Enable and Accelerate 100B Model Fine-tuning on a Single GPU : https://huggingface.co/papers/2403.06504
Training small models on specific tasks, to go faster than chatGPT, and do better: https://levelup.gitconnected.com/fine-tune-smaller-nlp-models-with-hugging-face-for-specific-use-cases-1745813471dc
Training your own models – Demo of rewriting in different styles
Could be built into xSpin.it; for now it’s a PoC for internal use. Contact me if you need it.