Looking for some good recommendations to understand and use GenAI tools? There are many, but I hope my list here gives you additional ideas and interesting picks. If you are specifically looking for picture and video creation or coding tools, see the respective sections in this vault.
This paper introduced the Transformer architecture. Do I need to say more? The attention-based approach is pretty much the technical foundation of almost every modern language model.
The InstructGPT paper. Showed how Reinforcement Learning from Human Feedback (RLFH) was the puzzle piece to turn raw text-completion engines into conversational assistants as we know them today.
Sure, technically this is already outdated. But few books managed to provide such a concise documentation on how to build deep learning models from scratch, and there is more than LLMs (such as GANs, Diffusion, VAEs, and more), this is really „GenAI“ to the fullest.
While prompt engineering, based on advancements of LLMs in recent years, has lost its earlier importance now that models can handle fuzzy information much more robust, all GenAI users still should have, in my view, basic expertise on how to prompt efficiently. This is a great guide.
Covers both NLP and LLM basics and more advanced techniques. Completely free and without ads.
Since we have covered theory, technical foundations, and methodology in recommendations 1-5, let’s change to the current applications. ChatGPT still is the best know assistant of its kind, and it’s still a must-have for such a list.
Frontier in the best sense of the word. Excellent reasoning capabilities and very well-suited for long documents, complicated tasks and of course, coding as well. Don’t forget: This one is token-heavy.
Strong on websearch functionality (not surprising) and multimodality. Not the best in either token efficiency or frontier capabilities, but a solid contender.
Missing open-weight models on this list? Here is one. Ships broad range of model sizes. Apache 2.0 license with no commercial restrictions. If you want to run, finetune and self-host a model, open-source is a powerful choice and Qwen one of the strongest offers on the market.
Disclaimer: You need a Google AI Ultra plan to use this. The concept is pretty cool: The Genie 3 world model renders a navigable interactive 3D virtual world based on text and image prompts. And yeah, you can remix those worlds as well.