Censorship and bias in AI models

• A UK university study found ChatGPT has a liberal/left-leaning political bias
• The presenter believes AI should be neutral, logical, and not have political biases
• Some censorship is necessary to prevent offensive outputs, but too much can degrade performance
• Open source models offer a middle ground with less censorship

Running large language models locally

• Large language models (LLMs) are AI programs that recognize and generate text for various tasks
• LLMs are trained on huge datasets and use techniques like deep learning and fine-tuning
• Running LLMs locally requires significant hardware resources, especially GPUs and RAM
• Quantization reduces the computational and memory costs of running models locally
• Tools like LM Studio make it easier to run models locally without coding expertise

Mixtral 8x7B (Dolphin) - an uncensored alternative to ChatGPT

• Mixtral 8x7B, also known as Dolphin, is a recent compact version of GPT-4
• It uses a "mixture of experts" architecture with 8 expert models of 7B parameters each
• Dolphin is uncensored, allowing for diverse applications especially in research and programming
• The uncensored nature requires users to be ethical as it provides access to any information
• Dolphin can run locally on a high-end consumer PC with a recent NVIDIA GPU and ample RAM

Phi 3 - a performant small language model from Microsoft

• Phi 3 is a small language model (SLM) from Microsoft with 3.8B parameters
• Despite its smaller size, it rivals Mixtral and ChatGPT in certain tasks
• Phi 3 excels at general Q&A, direction, and role-playing but falls short in coding
• It is heavily censored but fast and resource-efficient, deployable on limited devices
• The high performance is attributed to filtered training data and synthetic data

Choosing the right model for your needs

• ChatGPT is still worthwhile for average non-technical users who can afford the subscription
• Open source models are appealing for professionals, researchers, programmers, and AI freedom advocates
• No single model fully replaces ChatGPT, but alternating between Mixtral and Phi covers most needs
• Mixtral excels for programming and research, while Phi is better for conversation and general use