By Guy G

Guy’s Journal Currently Focused on AI 

Journal 15 5.14.24

Mahallo it's been a hot second. Meta LLama 3 Started releasing OpenAi Sora and GPT 4o. Deliveries of AI devices including Limitless Pendant, Llama 3 in Meta Raybans Frames, Humane Ai pin, Rabbit R1, and Brilliant Labs Frames just today Google Gemini 1.5 Flash, Gemini in Workspace, Google Veo, Google Gemini Gems, and Google Gemma 2. Openai opening the marketplace to everyone, and Much Much More. Going into the next year we will see much much more to come at The WISRD lab there has been a lack of activity due to poor data yields and creative differences with various other members of WISRD. This will likely be my last entry of the academic year and I hope we may produce vastly more in the coming year than this with the delivery of the Brilliant labs frames and new data collection methods we hope to gain invaluable data and knowledge on AI in the coming world. Good day, Godspeed, Good Luck, and I will see you next Year. -G

Journal 14 2.13.24

It’s been about 2 weeks and I am. That’s it I am still. I am still in existence. I got excited about the Brilliant Labs Frame. Today I downloaded the newly released Nvidia Chat with RTX (Beta). I have had no time to test it yet as of now but am excited about its potential in localised anthropological LLM testing. After peer feedback, I will not be writing an Inquirer article this academic year. More Projects are in the works given this newly obtained news I will be focusing on adapting my inquirer pitch to be either a self-standing research paper or a larger study for next year. I have a hardware project in the early stages and will be scaling back active testing for the next week and a half to focus on other academic endeavors.

Journal 13 1.29.24

Hi, ya‘ll It has been a while and nothing has changed but a lot has happened. I talked to the tech team at Wildwood about getting Copilot Studio Either late last month or early this month. Still no Change. I Submitted my Enquirer pitch for peer review. Still waiting for feedback. I had some insane Ideas about how to drill into volcanos for energy and then I kid you not the next day Icelandic scientists announced that’s what they were going to do: https://interestingengineering.com/energy/iceland-drill-magma-chamber-geothermal-energy. So I am doing great. The Enquirer pitch is thus:

Can Teachers and students recognize AI-generated Text vs how well they think they can?

An Advanced ~100-question questionnaire asking participants to discern whether a given passage is AI-generated or not assigning a percent certainty rating to to each. The Data will be gathered from the Hello-SimpleAI/HC3 Dataset. At the beginning, the participants will be asked how well they believe they can discern AI written content from organic content they will again be asked this question at the end of the test which will be digital. The same data will be run through the Berkeley Artificial Intelligence Research (BAIR) Ghostbuster program as well as GPTZero a popular AI-generated content testing program. The Data will be compared and released with an opinion by the authors. Note: Both Tools for detecting AI content have a High likelihood of being trained on this dataset thus the exact details may change given more research.

Since then the concept has changed slightly in that due to the similarity in training data we are going to have Generative LLMs to distinguish between Ai Generated content with an accuracy score. Also thanks to the feedback from my peers the amount of questions will be significantly reduced.

Journal 12 1.3.24

Hi and welcome to 2024. I just submitted my request to Megan for Microsoft Copilot Studio which will majorly advance my LLM tailoring capabilities with GPT-4. Also, apple researchers published a paper called “LLM in a Flash” which made me so happy I can not tell you it will show that LLMs will be a lot more efficient in the future. I will. link the paper below. Also, I will replace all wisrd members with multimodal AI. But not this year. Also, I am in the planning stages of an experiment I hope to publish as a paper showing the most efficient way to build LLMs and/or whether you should be based on consumer data TBD. See you next time. LLM in a flash: LLM in a flash

Journal 11 12.19.23

Mama Mia, Here we go again. It is midnight and I have been messing with AI and it has been messing right back. I ran the main Mistral 7b LLM using Ollama.ai to run it locally on Apple Silicon Mac. (Note this is the September release, not the newer 8x7b Due to processing load errors with my device) I asked it (the September release known as Mistral by Mistral (just for clarification) one question and it ran for 90 minutes and only stopped because I stopped it. I gave an AI 30 characters, and it gave me 64,000 characters in response. It just kept looping and looping and AI once again has made me crazy. I put 3 other inputs before the Taylor-related loop of madness it (the AI’s responsive) changed over time from having blank responses and just answering them before I could answer them Unlike LLama 2. It went from intro to facts to ends to jokes to intro again. It drove me crazy and messed up my sleep schedule here is all of the terminal data as put into Google Drive: AI terminal Output Mistral Guy This is not crazy AI’s are not unknown to do this but this is a cornerstone of a commercial product of a 2 billion dollar company. Of course, I was able to terminate this program at any time for those who are just using a customer service program which (in my experience with AI-based customer service programs) can not be shut off. Bellow I would like to put a Similar conversation I carried with LLama 2 via Ollama this is an example of an acceptable LLM “conversation”[Figure 1]. Please note the reason for the rather unscientific prompt engineering is simply a conversational benchmark I developed while researching the Inflection-1-based pi.ai a very warm conversation-based AI that I have only seen rivaled once on the release date of Google Gemini Pro via Bard which answered the question “ARE you Gemini” with “LOL. Got that wrong earlier today. Bard is on Gemini Pro in English across most of the world as of Dec 6, 2023”. [Fig 2]. More projects are on the way. Also, an update as of this day my project is still technically “in development” due to some clerical errors I was not made aware of the process of becoming a full-fledged lab nonetheless I showed up at the WISRD board meeting and Due to multiple inefficiencies in the WISRD board’s processes, I doubt my lab will not be made official until at least February 2024.

Fig 1

Fig 2

Journal 10 12.9-10.23

I want to explain the confusing date I am writing this about 11:30 pm on Dec 9 so I will mark it as both. I would like to make this clear only because the context of my exhausted mind is the only justification for what will be written here. First of all, I have Upgraded WISRD-AI-2 to “Claude Instant” and thus is version 2.1 with version one being without any given data. While working on my poster which I will post sometime next year on this site anyway Megan mentioned that it would be cool to have a bot that gave out Fortunes So I made https://poe.com/Zoltar-AI and https://poe.com/Fortune_Cookie_Bot. Before I had even started working on my poster as I previously I made a NEOM info-related bot. Well, I added more data, and is running on ChatGPT https://poe.com/NEOM_AI. Zoltar is the fortune teller from Big and is run by Claude and the fortune cookie is on ChatGPT. To be clear When I say Claude I mean Anthropic AI’s Claude-Instant which is free to make a bot with on poe.com. I am Excited about the new possibilities brought by the varied available AI models

Journal 9 11.30.23

It has been a month and I have come up with many ideas that I intend to pursue later. But as I keep returning to AI again and again I have realised that if I don’t use it for my inquirer article I can still get a good poster out of it. I know that feels a bit cheap but the application is for people not for education. I have also realized that Poe.com allows information input to their LLM services. So I have updated the WISRD LLama2-based AI it is now available at: https://poe.com/WISRD-AI-2 and is now Based on ChatGPT, not LLAma2. It now has updated data after I received feedback that the LLM did not Know who Megan Nole (Our COO) was. Now with one week left, I will completely change my data poster. I have also in the past day created NEOM_AI a ChatGPT-based dataset with data from NEOM’s website I have yet to make this Public I hope I will but I am undecided. That being said it is an interesting test of how LLMs accept data from websites. I will update you if I make the AI available.

Journal 8 10.30.23

That’s it I am done. My current plans as they are can no longer be executed. A month of my life has been wasted. Let me back up. Dark Matter Day (The Institute Fair) was just fine. It was great a lot of people came and a lot of kids were entertained. However, it was incredibly tiring and may have broken my faith in humanity, because I can not believe how many high schoolers and grown adults could not tell the difference between an AI image and a real one. Given I was comparing one against the other many assumed it was my goal to trick them even still it was not encouraging for the future. Previously I would have said we need some sort of standard of watermarks for AI and yet I would never ever think it should be forced because the use of generative AI is to be a standing not another option for the world. I know that does not cover all situations but overall that is my opinion. Today I woke up to the news that President Joseph R. Biden has issued an Executive Order effectively kneecapping all public releases of AI. All public AI releases must be Vetted in some capacity and the gov will establish watermarking standards for the industry. My Secret goal for this project was to help Educate the Wildwood Administration about kids using AI. However given this heavy-handed gesture from the government it seems that without considerable time and resources to change any government-guided decision, the school would make. It is possible but given the current uncertainty I have decided to move on partially to something else. I want to keep working on AI but I can’t keep working knowing that it. is likely my work will do nothing as soon as the government decides they want to change some law or standard. I will keep the POE AI that nobody even knows about yet up until I can’t. I will try to set up a page on this site an “AI Art Gallery” if you will. And of course, I will still keep learning about AI and keep my eye on it until the situation improves. So with a heavy heart and an exhausted mind, I am Officially ending my AI project. I know it makes me sound like a broken record but stay tuned I will find something cool to work on.

Journal 7 10.22.23

So I have some updates. I was in Germany which sparked a lot of innovative ideas but I must first get done with my current projects. My Dark Matter Day project is almost done. Any kid that can guess between two images which is AI three times gets a sticker. Next, We have the big one Today WISRD_Llama2 my new Poe-based AI on Meta AI LLama2 70B goes live at: https://poe.com/WISRD_Llama2 !!!!

Journal 6 10.10.23

So a lot happened. Dark Matter Day is coming up and I have been having lots of fun using Dalle-3 through Bing chat to come up with wacky AI Images hopefully I have some below. Now I also experimented with Poe.com which I have learned is another way WISRD could host its own AI environment. II should have more on Monday. I hope this will involve flash speed testing as if it’s fast enough we theoretically we should be able run LLMs directly on there as long as its close enough like with Apple Silicon.


Journal 5 10.2.23

So I have not done much but it has been 12 days so I was told to update my journal. Now for various reasons, I was not actually able to get to the fastest computer in WISRD until Wednesday the 27th of September. however, this gave me plenty of time to research what to do so this is my plan:

On Thursday the 28th I was finally able to download Stable Diffusion 1.5(Stable diffusion is an open-source generative ai focused on images) locally onto the WISRD computer also used for games(which is part of the reason I was not able to use it for a week). This is important because running a local AI quickly and persistently is very important when trying to support “many” people(300ish) using it quickly and at all times (hopefully not 1-4 a.m. but I want them to have the option.

I decided to run some tests to ascertain the higher specification of hardware needed to accomplish my ultimate goals for this project.

The Specs of the Hardware used in the test:

Intel core i7-1200k

64 Gb DDR4 Ram 

NVidia Geforce 3070

Toshiba DT01ACA 930 Hard Drive

So what were the results? Well, they were mixed:

1.2 minutes total on Joe Biden local 512 by 512 4 images per batch random seed

4.5 minutes Total on Joe Biden local 512 by 512 20 images per batch random seed

21.8 minutes Total on Joe Biden local 800 by 1440 4 iMages per batch random seed

Let’s Break this down together the image prompt I used was simply “Joe Biden” and all used a random seed and one batch.

I would like to make it very clear there is no political reason I chose the prompt “Joe Biden” It was chosen by a fellow WISRD Brian P

Journal 4 9.20.23

So I had an idea for an automotive thing that I have not given up on but my new path is to create a forked AI to gather data on student usage of AI. Not to catch kids cheating but just for info on what people are doing not what they think they are doing. More to come…

Journal 3 9.14.23

Sooooo my original project idea involving understanding automotive computers may be dead. Turns out any school project that involves using a tor-browser may be a bad idea. For various file encryption reasons, this stage of this project is over. Maybe I will try again but just for reference I tried using these files: (http://rnsm777cdsjrsdlbs4v5qoeppu3px6sb2igmh53jzrx7ipcrbjz5b2ad.onion/14902e9881109f26bfe4385ff6d68cfa9bad6769/)

If you can’t access them try using a tor connection.

I have decided to head in a new direction if any opportunity presents itself I may dive back in but for now, stay tuned for the new direction I will take my wisrd career.

Journal 2 9.6.23

Just a ton of Chrome tabs are killing my brain as my (legally not malicious intent) evil plans solidify. Emails will be sent to those who require it as soon as my plan of attack is (almost) finalized. My journal is simply here so you understand my project I intend to only write if something interesting happens I do not intend to bore you with 100+ entries that just say “Did research learned “thing” It was cool this is why” This is my letter of intent for this journal for this year(of courses its not actually but just so we are clear a daily update will probably not be a thing in this journal unless something really interesting happens.

First Journal 8.23.23

I am overwhelmed I am not sure what anything is anymore I am tired and energetic. I don't know what to do but I am excited and ambitious this will be amazing. I am so excited I want to learn a lot more about everything. I intend to have a good time if not. This will be a mess for me to look back on.

Statement of interest, Spring 2023: I intend to utilize the resources and brain power of WISRD to grow my knowledge and opportunities in my chosen fields of biology, technology, and automotive applications.I am not one who enjoys writing out there thoughts without evidence. Now as Megan knows I have used various parts of other's projects to kick start my Into The Wild project I believe with more time I can come out of it with a magnitude more knowledge than anything else I could with the time at least at my age.