Free Software Directory talk:Artificial Intelligence Team

Free software replacements that are missing

• AI Research Assistant
  • https://elicit.org/ - Elicit uses language models to help you automate research workflows, like parts of literature review.
• Voice to instrument: Tone Transfer-like
• Identification
  • Photo
    • Pl@ntNet for Android - Pl@ntNet is a citizen science project for automatic plant identification through photographs and based on machine learning. "The observations shared by the community are published with the associated images under a Creative Common CC-BY-SA license (visible author name)." - https://plantnet.org/en/2020/08/06/your-plntnet-data-integrated-into-gbif/
  • Audio
    • Shazam: Shazam is an application that can identify music, movies, advertising, and television shows, based on a short sample played and using the microphone on the device.
    • A free app that functions like midomi.com -- "You can find songs with midomi and your own voice. Forgot the name of a song? Heard a bit of one on the radio? All you need is your computer's microphone."
• http://design.rxnfinder.org/addictedchem/prediction/

Potential Freedom issues

• Dependencies need to be checked.
• Verify whether a workflow requires non-free GPU or if CPU can be used.
• The training data often contains non-free licensed material.
  • According to current copyright laws, this does not impact the license of the model or the output of the model. According to current copyright laws, the output is public domain. Mmcmahon (talk) 11:48, 2 May 2023 (EDT)
    • A source for USA AI copyright guidance (Mar 16 '23) GrahamxReed (talk) 22:34, 14 May 2023 (EDT)

Testing model viability

Tools are needed to assess the pros/cons of each model.

HuggingFace's Open LLM leaderboard

Ordinal value scales could exist for

• indistinguishability from human creations - e.g. Human or Not? social turing chat game
• inference speed
• length of memory
• trivia accurateness
• computation costs: cpu/vram/ram mhz

Source of model training data

• amount of data
• date range (e.g. distinguishing old science from new science for smaller scale models)
• level of censorship (important to make personal+research use distinct from business use)

Problem solving

• math
• creative problem solving (there exists methodology for testing this in humans)

General trends

• Larger models are more prone to human superstition[1], but also generate more human-like readability.
• Quantization (a la GPT-Q) allows consumer hardware to run large models.

Extra information

Stable Diffusion

Stable Diffusion model files (.ckpt) are released under a non-free license.

Here's the stable diffusion beginning point: https://huggingface.co/CompVis/stable-diffusion-v1-4 https://huggingface.co/spaces/CompVis/stable-diffusion-license

stable-diffusion-webui

• https://github.com/AUTOMATIC1111/stable-diffusion-webui
  • Demo, and guide - https://www.youtube.com/watch?v=R52hxnpNews
  • Extension script: https://github.com/deforum-art/deforum-for-automatic1111-webui

Large Language Models

LLaMa

• LLaMa is released under the GNU General Public License v3.0: https://github.com/facebookresearch/llama/blob/main/LICENSE
• LLaMa is comparable to GPT-3, and has been fully released as a torrent and on huggingface.
• 7B paramenter model has a VRAM requirement of 10GB. The 13B model has a requirement of 20GB, 30B needs 40GB, and 65B needs 80GB.

Unknown license but still noteworthy

• Vicuna 13B - It appears as though this model is inherently censored [2]
• Alpaca 13B
• Pygmalion 6B

External links

• Applications of artificial intelligence (Wikipedia)