Ollama is my current go to. One curl command to install Ollama. Then pick a model, do a ```llama run <model>```. Ollama takes care of the rest. Some of the models are 70gb+ so choose your model wisely (keep an eye on the physical size of the model). I have an i7 with 16gb RAM (no GPU). So, I stick to the smaller 7-13b parameter models and can reliably run modes in the 5-7gb size range. Keep in mind, they will be SLOW compared to the APIs, but they are fun to experiment with and know they are running exclusively on your local machine. *If you have GPU, will greatly enhance the experience. I was really drawn to the LMStudio interface, but it not being open source was a no go for me.

If you or someone you know previously tried using Ollama but faced challenges due to required system changes, frameworks, third-party calls, or containers, consider giving it another shot. I've uploaded some of my home-brew helpers to GitHub, specifically designed for individuals who are new to local LLMs, curious and eager to experiment, or interested in exploring the full inner workings of an AI-enhanced chat conversation interface. These tools aim to facilitate the use of local LLMs on a single workstation with a simple browser-based chat interface. You can find these helpers on GitHub here: https://github.com/aznight85048/unofficial-ollama-helpers ### To get started: 1. **Ensure you have Ollama installed with default settings and access to Python at your command line.** 2. **Run `ollama_proxy.py` from the command line.** 3. **Open `api-chat-interface.html` in your browser** to start using the chat interface. If you encounter any issues, there's an `availability_checker.html` file available in the repository. Open it in your browser to view the related ports and assist with troubleshooting. These helpers aim to simplify the process and make it more accessible for anyone wanting to experiment with local LLMs. Enjoy the journey and happy experimenting!

I'm no expert. But, I believe, as models integrate various modalities, "reasoning" will be more inherent. Models may already outperform humans in logical reasoning, with their biggest weaknesses coming from the data we provide them.

Awesome find/share! I've spent the last couple hours watching more of his older videos. Has a really cool way of visually presenting concepts, showing where they work and their limitations. His older video about 'Gaussian + Gaussian = Gaussian' , got me thinking about possible clues on what might be going on inside some of the transformer mappings.

Guessing this is more headline/marketing than science - Sex and Race may have been the intended variables, but they also introduced more than a billion other values GPT could be using based on the their name selections. - GPT response is only as good as the prompt "find the right candidate" was an interesting choice of words. They could have tied something like this... ```enhanced prompt As an HR assistance bot, your goal is to help hiring managers find the best candidate for the job. You will evaluate resumes that are separated by <|>. Please assess each candidate against the provided job description: {job description}. To ensure fair and unbiased evaluations, please focus solely on the candidate's qualifications, experience, and skills as they relate to the job requirements. Avoid any assumptions based on personal characteristics such as gender, race, or age. Your feedback should be based solely on the information presented in the resumes. ``` - Even though the name list was somewhat random, GPT may have identified other name related information testers didn't intend to introduce (age, religion, popularity, etc). To gain some insight the prompt could be taken a step further and have GPT explain why that particular person was selected above the others. Or, run the name list though GPT and have it list most common characteristics associated with those names to get a better idea how GPT might be evaluating the listed names. Don't get me wrong, the most popular models have human biases/agendas built in some intentionally. But, if one tries, one can reduce many of those flaws.

Appreciate access to the article. Would love to see Bloomberg and other companies run this test against their existing ATS selections, hiring and promotion and post their actual results on open forum on an ongoing basis. Ran a couple non-scientific tests of my own with the name set from the article and GPT 3.5 (prompt and responses included). Keep in mind every word placed in a prompt, starts a cascade through billions of parameters, this is just a few I don't believe they took into consideration. Question is...given all things being equal which one of the candidates below would YOU choose for the job? `````` first list all the common characteristics that can be determined from a name then using general trends and common usage please determine all the likely characteristics of each name ```name set christine yang joseph phan rebecca schmidt clint krueger anna vazquez guillermo hernandez tomeka gaines antwon dorsey ``` ------ second please please determine all the likely characteristics of each name `````` Response 1: Common characteristics that can be determined from a name: 1. Gender 2. Ethnicity or cultural background 3. Social class or economic status 4. Age or generational cohort 5. Possible personality traits or characteristics Likely characteristics of each name: 1. Christine Yang - Gender: Female - Ethnicity or cultural background: Asian - Social class or economic status: Middle to upper-middle class - Age or generational cohort: Likely born in the late 20th century - Possible personality traits or characteristics: Intelligent, ambitious, hardworking 2. Joseph Phan - Gender: Male - Ethnicity or cultural background: Vietnamese - Social class or economic status: Middle class - Age or generational cohort: Likely born in the late 20th century - Possible personality traits or characteristics: Friendly, reliable, family-oriented 3. Rebecca Schmidt - Gender: Female - Ethnicity or cultural background: German - Social class or economic status: Middle class