Keeping Up with AI #1

MODEL OF THE WEEK: VICUNA

Vicuna is an open-source model for text generation (similar to ChatGPT). It was created through fine-tuning the LLama model (from Meta) with conversations available on ShareGPT.com. It achieves over 90% quality performance compared to GPT and Bard. You can chat with a free chatbot version of it online here: chat.lmsys.org.

TERM OF THE WEEK: FINE-TUNING

Fine-tuning (in Portuguese "afinação fina") is a technique used in Machine Learning where we take a pre-trained model and adapt it for a specific task. Instead of training the model from scratch, which would be costly and require large amounts of data, fine-tuning leverages the model's prior knowledge and adapts it for the desired task. A model has several layers, and the fine-tuning technique retrains the edge layers of the model.

Example: let's say we want to train a model to be specialized in the task of classifying documents for our company. We could take a pre-trained model like GPT or Vicuna and fine-tune them to learn this specific need.

PROMPT ENGINEERING TIP OF THE WEEK: The persona template

When we want to create a prompt that requires specific skills or knowledge, it's a good practice to use this template. It's applied like this, let's say I want to write a movie script, with the persona template I would start the prompt like this: "As an experienced movie scriptwriter...".

Why do this? What's the difference between the persona template and a normal prompt?

To make it easier, let’s imagine ChatGPT; in its natural state, it is a base of general knowledge, just like us with all the things we know superficially but not very deeply. So if we ask something from it in this state, it can give a generic and shallow answer or even not in the context we wanted.

But when we say that we want it to answer in a certain persona, it understands the context we want and digs deeper into its knowledge base to generate the answer.

Okay, but why use adjectives like experienced, good, famous, recognized, etc.?

Just like in our world, in GPT’s knowledge base there are examples of content from good scriptwriters and others not so much, adding these adjectives refines the AI's understanding of the expected result, making it discard those that do not fit what we requested.

LAUNCH OF THE WEEK: Balance

Use GPT-4 to create arguments against your anxiety scenarios.

Link: balance.dvy.io

SCIENTIFIC ARTICLE OF THE WEEK: Mind2Web

Dataset for developing and evaluating generalist models for the web capable of performing any complex task on websites following text instructions.

Link: osu-nlp-group.github.io/Mind2Web/

VIDEO OF THE WEEK: GPT TOOK OVER MY HOME

---

RECOMMENDED TOOLS:

SEARCH SITES:

* Phind: an internet search site with AI, it analyzes the main results, returning an answer and the links. Link: phind.com

* Perplexity: an internet search site with AI, same functionality as Phind, but with the distinction of being able to focus the search in certain contexts such as academic, Reddit, etc. Link: perplexity.ai

* GPT4: for those who subscribe to ChatGPT Plus, using the GPT4 model with the browsing tool usually gives good results in searches.

IMAGE GENERATION AND EDITING:

* Bing Image Creator: image generation. bing.com/create.

* Mage.space: image generation and editing, models refined for specific tasks. mage.space

* Playground IA: image generator and editor. playgroundai.com

* Adobe Photoshop Beta: Photoshop edits via text prompt.

* Adobe Firefly: image generation and editing. adobe.com/br/sensei/generative-ai/firefly.html

* MakePose: image generation with specific pose manipulation. makepose.com

* Clipdrop: various advanced image editing tools. clipdrop.co

* Runnway: various advanced image editing tools. runwayml.com

DEMOS ON HUGGINGFACE:

GMFTBY/PandaGPT: model capable of processing text, image, audio, video, and thermal image inputs.

jeffistyping/Youtube-Whisperer: transcription of YouTube videos.

facebook/MusicGen: music generation through text.

akhaliq/AnimeGANv2: transforms any photo into a drawing.

nttdataspain/Image-To-Text-Lora-ViT: describes any image.

Thank you!