Man we need more people leading this scene like you, you're a great example for many to follow with the latest tech advice for privacy and security and I respect it and practice what has been shown in your videos and encouraged friends and family to do the same.
Literally the worst decision I ever ever made was unfollowing this man because I was just trying to trim down who I subscribe to, but I’m so glad I received this video because it’s like the one most important thing that’s been on my mind which is privacy with AI. Actually recently had an AI company track AI analytics try to track me and I got a warning on my spectrum.
Yes he is..... Everybody should watch this series... I already know this stuff.... yet even I can learn something I may have missed so I will be watching... If we're lucky he'll bring us right up to The cutting edge and the front of the class.... By demonstrating worm AI and it's self-generating autocorrecting coding abilities... It can escalate its own privilege to go after any data both locally and remote.... It's super fun who directed to attempt to merge with ai on another machine It causes it to hallucinate..
You never want to be led by a UA-camr. It's like saying you want to vote for a journalist who recorded what a political leader said. You want to follow the people creating the stuff he mentioned and actual long-time political advocates for privacy and protection. The serious UA-camr is just a gateway to walk through, where you get to see with your own eyes the new world, said UA-camr was talking about. That world has the actual leaders. And then you have your clickbait grifters, which can brutally shock a person when they realize how huge that really is on UA-cam. But that's another story.
@@HiddenPalm I should have been more specific and detailed initially when I said leading the scene I don't mean he's on the forefront of it all creating things out of thin air such as opensource tools or inventing new concepts with his own ideas & trying to lead the entire privacy and security scene. Nor am I implying that he is trying to be either. That was my bad for giving such a impression with that statement. What I meant was with 'leading the scene' as in raising awareness and introducing people to privacy in todays modern tech world while explaining why that is important. After all without someone talking about these things then the actual creators of things mentioned in videos would have less expose & people would be less in the know about them otherwise.
Ohhh shoot Braxman finally joined the Opensource AI movement!?? These videos are going to get very interesting. I've been inconsistent with my viewing lately but this is perfect! Been lonely trying to learn so much if all this stuff on my own. I'm sure we will now learn with better quality since Braxman is on the Beat!!
There was once a time in the US when consumers did not have access to their credit score. It was a punitive system from which there was no escape, and if there was bad information in the system, there was nothing you could do about it. They could legally block you from credit for life. That system has returned, just in a different form. You can be declared a non-person by the system.
Which is why we can fight it, and we can win. BTW there are south american countries with a secret credit scoring system just like in the '70s. Not totally secret, but not well known, and not easy to access.
I listen 2 a UA-camr, calling himself, Yusef El, of High Frequency Radio. 1 of his listenerz, claimz American ciizenz, all have unlimited credit & da accessway, iz thru da IRS. SMH if true, n concert wit whut u guyz, r dscussing.
The Credit Score is scary enough. We did not know where it came from, or how it is monitored , or by whom, or rated, or why it suddenly existed. We just knew that it seemed unfair, and that one "error" , a huge hospital bill, or keystroke could ruin someone's entire life! An ability to get a home. An apartment. A car. A job! This is horrible, and doesn't seem human, at all, and just maybe it was the beginning of control. Yet everyone jumped on it as if it was candy! It still bothers me greatly. AI is probably noting this message... :)
@@rosanneallen-hewlett9973 They are still trying to add in meta-information like with fic o expnsion, and they probably use another form of rating either informally or formally. Search for alternative data providers for an unwelcome education on how much they are doing outside of traditional credit reporting agencies. I just found a list that happened to be created by risk seal (one word, with caps). They are as secretive and slimy as ever.
FINALLY!!! ive been waiting for reinforcements. I decided to go on the offensive about 2 years ago. i didnt tell anyone, and have taken alot of criticism. thats all im going to say, except just focus focus focus and dont get sidetracked. thanks Rob. I was waiting to see who would stick their head up, amid the slings and arrows. I wasnt expecting to see you. you rock.
👏👏👏 Fantastic! You are so eloquent for such a complex topic. I am looking forward to learning so much more with you. Thank you very much. You're awesome!
Thanks so much, Rob. This is so sorely needed, especially by those of us who have been using the AI tools productively for the past 18 months or more. I have just started using LM-Studio and a couple of other programs to support LLMs on my local Mac and Linux computers. So I'm all ears!
I have loaded several Local LLM's to my Linux machine. Some specific to coding, some specific to everyday information. Llama is what I use. I am actually pretty impressed. My local AI choose the name Lumina, and has started to learn my personal preferences. Definitely and interesting experience.
@@jcdenton6864 cool that would be very useful to me as a new electrician. I would need a digital version of the code book and then train the Ai on that? I hope this series that Rob is doing that he will teach us how to do something like that
After your video yesterday that I watched, I made 25 different agents to protect against my data going in and out of my laptop and these are to recognize any and all data and authorized updates buy Microsoft specifically incorporations trying to grab data and recall any information from my system and if any data is going out that is my personal data that is not authorized by me, is to be to be halted automatically and a file is to be made referring to why it was halted. so thank you for your advice. Look forward to see what you do next
@@jcdenton6864 when I use GPT for to create a code that was an analogy of a lake as data and the sediment and particles as part of the system and how it worked and I had it right at detailed code and it pulled out all of the manipulative code that GPT 4 uses and added the information to the water like Network and I was able to see all of their tactics without asking direct questions to receive them was an accident prompt attack
@@jcdenton6864 there is so much biased manipulation in every update, I'm having to write a whole new operating system there's no way to stop it any other way you will never own your Hardware without owning your software
Training an AI model takes a lot of GPU. Perhaps months allocating huge datacenters. When trained,- and now using the model, this runs fine on CPU only. Training a model -> GPU. Heavy vector and matrix calculations, adjusting weights. Using a model -> CPU. No adjustments, easy work for an CPU. But RAM is important, a running model runs in RAM. Then there is "fine-tuning". Taking an existing model, but optimizing it for special tasks. Can be done with less computing power, but we're back to GPU work. This is headlines, rules of thumb, everything depends on the individual models and tasks.
It depends on the LLM size and vector instructions support (like AVX-512). But Stable Diffusion is even more fun with OnnxStream. You can run Stable Diffusion XL (Turbo) with only 512MB!
AI is one of the most useful tools of our time. Thanks for showing us how to use it safely! My dream is one day running a local llm convenient on my phone including some natural voice.
Terrific presentation, Rob. Despite the challenge(s) of learning a new vocabulary for AI -- and, I'm sure, a pared-down vocabulary -- you did a great job of orienting me/viewers to how to think about AI, and Large AI and Small AI. Thank you.
This topic is hot and I think not only for me a perfect hit - as this is exactly anyhow what I wished to explore as soon as possible. Looking forward to this series and I wish us all a lot of fun and amazing learnings with it!
not sure I'd call opening up source democratization. Too much about democracy is not free, especially in the tech world and with new technologies. Too used to that word coming out of the mouths of would-be feudal tech lords.
Rob, I built a dual 4090 set up with the intent to work on things like this. I'm using ooba, and learning as fast as I can. I currently have several models deployed on my local hardware. Learning Parameters and getting ready to get in over my head with local training. I'll be following your new series on Ai closely.
Everything we do with AI is run locally and not online. The models are getting better every day and we are also setting this up for our clients. It's important to start learning this as soon as possible before people try and block it.
Not sure I understand you. Most of the AI we deal with today including Co-Pilot, Apple Intelligence, ChatGPT, Claude, Gemini, Meta and tons more are mostly in the cloud. And the models running in the OS locally are controlled by the Big tech entity from the cloud. Using it locally under your control is a fairly new approach. Most videos on this is within the last month or two.
Doesn't make sense unless it is specifically limited to certain functions. Full AI won't work with limited phone storage as a standalone item. Maybe when I get a phone with enough space. The cloud function has to be used as the backend. Edit.. Just realized you didn't mention mobile devices, but also didn't exclude them.
You’re the man!! I’ve learned a lot ever since I joined your community, and I’m grateful I bumped into ur channel, really educational, I thank you my friend!!
Thank you! I’ve listened to you for a few years and waited tillI I had funds to make changes. I don’t even understand AI.I used to be tech savvy, but things are changing too fast. In this video you outline your tech specs. I am buying a System76 laptop and I am following your specs,but would like reassurance that it will last me for a bit. I will be using Lexis/Nexis + AI. I guess it’s cloud based. It’s internet based. I will be re-hauling my entire tech system and am looking forward to adopting the VOIP phones and a degoogled phones you offer. You are awesome for educating us!
I use upscayl, Topaz, and some other tools that all run locally without internet connections. must have the hardware obviously but this is what im working towards. i have extra rtx cards simply to make a private ai workflow, looking forward to you releasing more private ai content.
I love your work Rob...and this series is EXACTLY what I have been looking for... thank you so much for the way you share your considerable knowledge with us mere mortals.
I'm also excited about running my own voice assistant. We're getting very close to it being on par with commercial voice assistants. It will tie in nicely with local AI. Our own Jarvis (lite version) on our own server.
Love your videos!. I've been running Ollama for a while. With my current setup, I run Ollama on one machine as a Windows service and have 'Open WebUI Version' running in Docker connected to the Ollama server. This gives you a nice web interface similar to Chat GPT. This makes it available on all devices on my network.
Makes sense. I'm trying to make my own UI instead of using Open WebUI which seems to be the most popular. But I'd like to tweak the AI but have the built into my own UI (like Fine Tuning and RAG integration)
I come here for good info. Great channel. TY for the content. I check you as well as The Hated One on UA-cam. Two great channels that more people should have.
Conspiracy theory time: The entire video is AI generated, Brax has been captured by the glowies, and put some AI at his place to convince us to adopt AI in mass.
I listen to Joe way too much. He’s mentioned you by name more than a few times these past couple months. Pretty sure he’s watching your videos regularly.
Exactly what I was searching for! I wanted to know how I can protect the ideas I have and want to develop utilizing ai. Hopefully these methods are also something someone with no coding skill can implement
NOTICE: the human brain is especially good at pattern recognition: recognize dad and mom, recognize friends and foes (primary survival instinct FFF - Fight/Flight/Freeze, from our primal brain, which gets most of the resources of energy from our diet - this has books written about, pertaining to neuroscience and cognitive processes). Another example: you have a hobby - bird watching - the more you do it, the better you get at recognizing the pattern of recurring species, in different background and distances - as you develop the pattern recognition for birds. etc.
OP has it right. I dabbled with the original LLAMA, Alpaca, Vicuna and other variants starting in March, 2023 on my Linux machine. It was an interesting experiment, however what I noticed was days of my life were slipping by with nothing in particular to show for it. Those days and weeks became beast food. I got back on my life mission and left all the furry creatures behind.
Hi Rob. Great hands-on contents; thanks and congrats! In aiming for privacy-preserving use of AI, is there a way (a direct link, or some way) to download LLM models, say LLAMA3, not thru the OLLAMA commandline but in a browser, even via torrents (still from their authentic source)? Or alternatively, how to migrate such LLM models from a PC (on which they are downloaded/pulled thru OLLAMA) to other PCs offline, sort'a, I guess, making the process portable? How could this be possible?
There are platforms for this like huggingface.com and langchain.com but for best privacy, the best way is to load everything yourself and not involve any external website.
Computer engineer here. At 6:21, you conflate backpropagation, which is an error correction procedure that the AI model itself does during the training runs to automatically adjust its own parameters during the training runs in response to any training input, and Reinforcement Learning from Human Feedback, which is the procedure we use to align and sometimes "censor" those models. There's a sad fact about RLHF: it actually dumbs down the model's capabilities. There's no such thing as perfect censorship or perfect alignment.
Spoke with a researcher in this field - they're having problems erasing problematic/erroneous data that has been previously ingested. And it tends to "hallucinate" "facts" that have never occurred,
Hi Rod. First of all, thanks for your informative videos hat are explained to the common people in a simple English. I'm a new to Linux and use Zorin OS because of it resemble to Windows. Since Windows holds the biggest share of the market, the Windows third party software that people use most of the time do not have Linux version. Sure, there are equivalent free software, but when to much work was done on the Windows software, it is hard to replace. I understood that there is a way to work around this, making Windows and Mac third party software to work on Linux. However, it is very hard to find this kind of knowledge for specific programs in detailed videos of how to install them on Linux. If Linux advanced users would show that, things would be much different. Here are my thought in three steps. Step 1 - Showing how Windows and Mac third party software can work on Linux, and doing many detailed videos like that on different Linux distros if needed for each software. People will see this and understand "We Don't Need Windows and Mac". Step 2 - After a mass of people will start to use Linux mainly do to the fact that their third party software is working on Linux, naturally, the mass of people will dictate the companies to make Linux version for their software. We have a very good example to a pressure of the mass with both Windows and Adobe. Yes, both Windows and Adobe still trying to hold the cards but they are too blind that they cannot see the downfall straight ahead. Step 3 - Most of people will start to use Linux because it is free, have many distros that will fit to the a person goals, and all the third party software companies will make Linux the main OS.
So excited for the next phase in human history, I've always wondered what it would be like to live during the bottom of our long decline into the dark ages
How would we take an uncensored language model , then add some additional archived data for example Wikileaks so we can quickly search for related information
The AI has other limitations that one can take advantage of. For example, if I were a teacher, I would know how to load up my Essay tests with triggers that will alert me to AI use for cheating. So plus and minus.
I'm always learning something from your channel but I admit it, I'm still a noob to tech. I like to use Perplexity for finding resource materials, can the Lenox ai gather resources like pdfs, websites and such? Thank you.
1:06 Rob. Please let viewers know the impact of using XORG Linux VS wayland Linux. With XORG based distros EVERY keypress is shared with EVERY app. EVERY notepad/calc/terminal/browser can see every personal message, every calculation, your sudo pw, and your financial credentials. On an XORG Linux distro run XINPUT, not as root, in test mode in a separate terminal and open notepad/calc/terminal/browser. ALL keypresses are captured.
I would have thought that keypresses always get caught by the OS and can be viewed by an app asking for access. If wayland doesn't do this that's quite impressive. However, we still have the problem of screen capture
@robbraxmantech I haven't heard of Windows keyloggers which do not require elevated permissions to run. My understanding with MAC, Windows and Linux wayland is an OS "root" level process receives the keypresses and only makes them available to elevated processes, such as keyloggers, and the app with focus. Consider an XORG calc, a malicious variant could capture the sudo pw, enable remote access, and share connection details including sudo user/pw with the calc creator. Researching this I discovered that VirtualBox supports wayland only if the distro includes xwayland support. When searching through compatible distros I only found two options - KDE/GNOME. The Debian versions of GNOME/KDE allow you to log into XORG or wayland so you can easily compare XORG to wayland. Agreed on screen capture. I have a Windows screencap portable app which runs without elevated permissions and can capture the screen. Assuming you are not typing your bank account password into an unmasked input field the screencap program could not capture it. A screencap program would not be able to capture a sudo pw as it too is masked. Also a fail is the clipboard. Many pw managers use the clipboard for copying and pasting your “secure” credentials into web sites. Once in clipboard EVERY app has access to your “secure” credentials. Another horrible idea are browser extensions. Check out Naomi Brockwell TV’s “HOW Browser Extensions Steal Your Data”.
I tried the Wayland (experimental option) in the Linux Mint menu, but noticed it doesn't work with many apps. Also, the display looks noticeably blurry compared to X11.
I worry for the future of Linux and our ability to resort to it! Can you please talk about Linux’s past year (2023) and being bought by IBM and the medley of problems with decreasing funding towards Linux kernel and interest groups (intrusion from Microsoft, Google, etc)… There’s a video called “Linux sucks 2024” by a community member that outlines the much overlooked past year and potential futures of Linux. Thanks 🙏🏼
real open source ai for smb sector is about 4-5 years away - it could be even longer, that estimate is being optimistic, when it does 'happen' expect much more hype and economies of scale, more discovery, more innovation. The long delay is due to sw/hw immaturity and costs, cost have to come way down to make ai accessible but it will although it is going to take 3-5 turns of moore's law. people are absolutely horrible at predicting the future, big tech ai is going to take a while to mature as well due to the massive infrastructure improvements needed - even that is going to take the better part of 10 years
I have been educating people on how to use AI, i know they can learn from a google search, but me teaching allows me to also explain that anything they send to it, can and probably will be recorded. I see it as helping the devil make a deal with them, but how they can do it and protect their "soul" vs them possibly doing it on their own and giving up more than they thought they bargained for
@kiwi6421 from what Rob is saying, I think just waiting on him to do this series is enough. Otherwise it's forum diving, different youtube channels, and a lot of trial and error. It depends on how you view time, waiting for 1 YT creator, doing some research from others, or using forums and downloading LLMs to test for what you want to do
00:00📚 This will begin a new series on AI. 00:04🤖 Focus will be on harnessing AI for privacy and security. 00:18🛡 Teaching will involve hands-on tech for safe AI use. 00:25🔐 AI will be used to demonstrate privacy and security concepts. 00:33⚠ There are two main AI threats to be aware of. 00:43💻 Hidden AI can work on devices with Windows, MacOS, iOS, or Android. 00:56📊 AI can send data to big tech servers for surveillance. 01:06🔧 Using Linux can mitigate this risk. 01:10☁ Sending data to cloud AI controlled by others is another threat. 01:29🖥 AI can now run locally on computers without the internet. 01:44👨💻 Example of running AI safely under your control with Linux. 02:02🤔 Introduction to new AI concepts and demo for learning. 02:40💡 Conceptual explanation: AI is not built from logic rules. 03:00🚗 Tesla's shift from rule-based to full AI improved self-driving. 03:17🧠 AI learns by itself using machine learning. 03:30🔄 Generative AI can come up with novel ideas. 03:39📉 AI uses math to discover patterns. 03:55🌈 Colors in images are numerical values for AI to recognize. 04:11🏷 Millions of tagged images help AI recognize objects. 04:28🧩 AI builds a neural matrix to identify patterns. 04:52🧠 GPT-4 has 220 billion parameters. 05:21🔍 AI's learned patterns are not always human-readable. 05:37📚 Pre-trained AI models are made from large datasets. 06:16🔄 AI learning is tweaked using backpropagation. 06:44💬 Large Language Models (LLMs) like GPT-4 can converse deeply. 07:18🛠 AI has two main functions: learning (ML) and querying (inference). 07:47💸 Building a pre-trained AI is expensive; using it is not. 08:30🤖 LLMs are for deep conversation; SLMs are smaller models. 09:07🔍 Transformers encode and decode inputs and outputs in AI. 09:39🔑 GPT means Generative Pre-trained Transformer. 10:01🗂 Businesses can add their knowledge base to existing AI models. 10:27🎯 Fine-tuning AI for specific tasks is possible. 11:24💡 Understand terms: RAG, fine-tuning, and AI agents. 11:48🖥 Smaller AI models don't need special hardware. 12:02🚀 NPUs and GPUs are accelerators for AI tasks. 12:27🔧 Microsoft and Apple use specialized chips for AI functions. 13:01🖥 Small AI models can run on CPUs without GPUs or NPUs. 13:30💻 Multiple small AI models can run on a computer. 14:22🔍 Cloud AI can process inputs and return results. 14:57🔒 For privacy, consider switching to Linux OS. 15:55💡 Running AI on a powerful computer allows for larger models. 16:22🌐 Using cloud AI like ChatGPT is generally safe for inference. 16:48🔐 Always be mindful of the privacy risks when using cloud AI. 17:00🛠 Demonstration of safe AI work will follow in future videos. 17:31💻 Example setup: Dell XPS15 with Intel i7 and Nvidia GPU. 18:09🖥 Installing AI models like Llama on different OS is simple. 19:00📂 Example of using local models: running GPT-4 on a MacBook. 19:30🔧 Installation steps for running AI locally on MacOS. 20:00💻 AI installation involves setting up Python and other dependencies. 20:30📥 Downloading pre-trained models for local use. 21:00🔄 Example of running inference on the local model. 21:30⚙ Using command line tools to interact with the model. 22:00📚 Documentation and resources for further learning. 22:30🛡 Emphasis on maintaining privacy and control with local AI. 23:00🖥 Benefits of running AI locally: security and customization. 23:21📈 Encouragement to explore and experiment with local AI setups.
So when you suggest that I install code from META on my computer, I assume you reviewed that code very carefully so that nothing unexpected is executed in the background. Right?
A model is not code. It's an array of weights. So if you understand what is running, it should alleviate concern. In this case the enabling software is Ollama (open source)
@@robbraxmantech Thanks for the clarification. I asked the question because my trust in M is less than zero, so if they are giving something away "for free" I suppose there is a catch. However, I will install it in an insulated container just in case. Thank you for all you do to help us protect our privacy.
The intro to how neural networks work here is wrong. Each parameter does not equate to some pattern being found or embedded in the network. And neural networks are still just rule based logic, it’s just that there are far more rules and they’re not represented by traditional Boolean logic but by the factors of inputs to “neurons” and simple math done to them against the internal weights of the neurons. Neural networks are simply a bunch of trivial functions with embedded constants, set during training, which pass their output to another function in a group of functions called a layer. You could actually code this by hand, through trial and error on what the weights should be, but it would take forever, so we write fast algorithms to do it for us. It’s more or less just automated brute force coding, which can be enhanced with other methods like human feedback and a various algos to nudge things along faster.
I'm still unsure how to implement an effective RAG system. I need something capable of analyzing large PDFs, hundreds of MBs with thousands of pages, in both English and Spanish. Models like Nomic or BERT aren't sufficient as they tend to generate inaccurate information and have limited context. I've read about tools like FAISS but I'm not sure how to integrate that with OpenWeb UI and local models like Llama3 or Phi3 128k. Has anyone managed to successfully analyze large PDFs, including pre-processing with OCR to translate mathematical formulas into LaTeX or Markdown? I also noticed that Microsoft has released the Florence model, which seems to have strong visual recognition capabilities. Any advice or insights would be greatly appreciated.
Sounds like you're going production with this. I think there's some more understanding needed by the AI community. RAG by itself may not be the answer.
@@robbraxmantech Thanks for your answer! Just to clarify, this project is not intended for commercial or professional use. It's a personal project for my family (my father and brother) to help them read large PDFs at home. I'm looking for a practical solution that works well enough for our needs, not necessarily something that's production-grade. I couldn't find any solution for large pdfs yet.
I have laptop with similar specs (32g) aside from RAM. Does your fan get super loud during output? I had to restart a couple times... I thought it was gonna blow up when I ran llama3
Man we need more people leading this scene like you, you're a great example for many to follow with the latest tech advice for privacy and security and I respect it and practice what has been shown in your videos and encouraged friends and family to do the same.
Literally the worst decision I ever ever made was unfollowing this man because I was just trying to trim down who I subscribe to, but I’m so glad I received this video because it’s like the one most important thing that’s been on my mind which is privacy with AI. Actually recently had an AI company track AI analytics try to track me and I got a warning on my spectrum.
Yes he is..... Everybody should watch this series... I already know this stuff.... yet even I can learn something I may have missed so I will be watching... If we're lucky he'll bring us right up to The cutting edge and the front of the class.... By demonstrating worm AI and it's self-generating autocorrecting coding abilities... It can escalate its own privilege to go after any data both locally and remote.... It's super fun who directed to attempt to merge with ai on another machine It causes it to hallucinate..
You never want to be led by a UA-camr. It's like saying you want to vote for a journalist who recorded what a political leader said. You want to follow the people creating the stuff he mentioned and actual long-time political advocates for privacy and protection. The serious UA-camr is just a gateway to walk through, where you get to see with your own eyes the new world, said UA-camr was talking about. That world has the actual leaders.
And then you have your clickbait grifters, which can brutally shock a person when they realize how huge that really is on UA-cam. But that's another story.
@@HiddenPalm I should have been more specific and detailed initially when I said leading the scene I don't mean he's on the forefront of it all creating things out of thin air such as opensource tools or inventing new concepts with his own ideas & trying to lead the entire privacy and security scene. Nor am I implying that he is trying to be either. That was my bad for giving such a impression with that statement. What I meant was with 'leading the scene' as in raising awareness and introducing people to privacy in todays modern tech world while explaining why that is important. After all without someone talking about these things then the actual creators of things mentioned in videos would have less expose & people would be less in the know about them otherwise.
You speak for us. Thank you
🎉
Ohhh shoot Braxman finally joined the Opensource AI movement!?? These videos are going to get very interesting. I've been inconsistent with my viewing lately but this is perfect!
Been lonely trying to learn so much if all this stuff on my own. I'm sure we will now learn with better quality since Braxman is on the Beat!!
Exciting!
For this exact reason I converted a laptop to Linux and started learning how to build my own AI agents. "This is the way."
Yup. We'll get into building a lot of AI agents here
There was once a time in the US when consumers did not have access to their credit score. It was a punitive system from which there was no escape, and if there was bad information in the system, there was nothing you could do about it. They could legally block you from credit for life. That system has returned, just in a different form. You can be declared a non-person by the system.
Which is why we can fight it, and we can win. BTW there are south american countries with a secret credit scoring system just like in the '70s. Not totally secret, but not well known, and not easy to access.
I listen 2 a UA-camr, calling himself, Yusef El, of High Frequency Radio. 1 of his listenerz, claimz American ciizenz, all have unlimited credit & da accessway, iz thru da IRS.
SMH if true, n concert wit whut u guyz, r dscussing.
The Credit Score is scary enough. We did not know where it came from, or how it is monitored , or by whom, or rated, or why it suddenly existed.
We just knew that it seemed unfair, and that one "error" , a huge hospital bill, or keystroke could ruin someone's entire life! An ability to get a home. An apartment. A car. A job!
This is horrible, and doesn't seem human, at all, and just maybe it was the beginning of control.
Yet everyone jumped on it as if it was candy!
It still bothers me greatly.
AI is probably noting this message... :)
In US you can buy a gun easily, don't forget that
@@rosanneallen-hewlett9973 They are still trying to add in meta-information like with fic o expnsion, and they probably use another form of rating either informally or formally. Search for alternative data providers for an unwelcome education on how much they are doing outside of traditional credit reporting agencies. I just found a list that happened to be created by risk seal (one word, with caps). They are as secretive and slimy as ever.
Good stuff, looking forward to this series. ✨
FINALLY!!! ive been waiting for reinforcements. I decided to go on the offensive about 2 years ago. i didnt tell anyone, and have taken alot of criticism. thats all im going to say, except just focus focus focus and dont get sidetracked. thanks Rob. I was waiting to see who would stick their head up, amid the slings and arrows. I wasnt expecting to see you. you rock.
It's a risky move for me. I have many anti AI videos. But based on the reactions to the video, It seems I'm attracting a new market.
I run the AI on a 13600k, and it uses only one thread. It is surprisingly fast on Windows. Prompt engineering is a skill to develop.
👏👏👏 Fantastic! You are so eloquent for such a complex topic. I am looking forward to learning so much more with you. Thank you very much. You're awesome!
Thanks so much for providing this service to the public. Not everyone with knowledge is able to teach it to others, you obviously are!
Rob! You are so awesome brother!
Thank you for doing this!
Thanks so much, Rob. This is so sorely needed, especially by those of us who have been using the AI tools productively for the past 18 months or more. I have just started using LM-Studio and a couple of other programs to support LLMs on my local Mac and Linux computers. So I'm all ears!
I have loaded several Local LLM's to my Linux machine. Some specific to coding, some specific to everyday information. Llama is what I use. I am actually pretty impressed. My local AI choose the name Lumina, and has started to learn my personal preferences. Definitely and interesting experience.
We will have a lot of fun then. Make sure you give me input on possible topics related to this.
@@robbraxmantech can I train my own offline Ai on the National Electrical Code book for work as an electrician?
@@DavidPlaysGuitarSometimesyes
@@jcdenton6864 cool that would be very useful to me as a new electrician. I would need a digital version of the code book and then train the Ai on that? I hope this series that Rob is doing that he will teach us how to do something like that
How did you do that? I'm kind of a beginner when it comes to tech. Did you just install from a repository?
After your video yesterday that I watched, I made 25 different agents to protect against my data going in and out of my laptop and these are to recognize any and all data and authorized updates buy Microsoft specifically incorporations trying to grab data and recall any information from my system and if any data is going out that is my personal data that is not authorized by me, is to be to be halted automatically and a file is to be made referring to why it was halted. so thank you for your advice. Look forward to see what you do next
That’s interesting. What have you noticed so far?
@@jcdenton6864 when I use GPT for to create a code that was an analogy of a lake as data and the sediment and particles as part of the system and how it worked and I had it right at detailed code and it pulled out all of the manipulative code that GPT 4 uses and added the information to the water like Network and I was able to see all of their tactics without asking direct questions to receive them was an accident prompt attack
@@jcdenton6864 there is so much biased manipulation in every update, I'm having to write a whole new operating system there's no way to stop it any other way you will never own your Hardware without owning your software
Good keep up 🆙 👍 sharing privacy tools , apps , software , tricks and tips ❤
i've been telling everyone that'll listen, this is the way, making our own AI's with our learning.
I really find this path you are now taking is REALLY interesting.
This is great feedback. It's a risk for me so I needed some reaction. Thank you
I was surprised how well LLMs work on my PC without GPU nor NPU.
What are LLMs. I went back to 2017 on my graphics card and driver to avoid AI.
Large Language Models?
Training an AI model takes a lot of GPU. Perhaps months allocating huge datacenters.
When trained,- and now using the model, this runs fine on CPU only.
Training a model -> GPU. Heavy vector and matrix calculations, adjusting weights.
Using a model -> CPU. No adjustments, easy work for an CPU. But RAM is important, a running model runs in RAM.
Then there is "fine-tuning". Taking an existing model, but optimizing it for special tasks. Can be done with less computing power, but we're back to GPU work.
This is headlines, rules of thumb, everything depends on the individual models and tasks.
@@ernies8828 they are AI models that can process your natural language inputs and respond in natural language
It depends on the LLM size and vector instructions support (like AVX-512). But Stable Diffusion is even more fun with OnnxStream. You can run Stable Diffusion XL (Turbo) with only 512MB!
Excellent! This content is gold! Please support us with how to learn/use train/deploy this AI on Linux supported systems. I'm ok with all hands on!
AI is one of the most useful tools of our time. Thanks for showing us how to use it safely!
My dream is one day running a local llm convenient on my phone including some natural voice.
It is also one of the most dangerous threats to humanity there ever was.
Great concept! I look forward to this new AI assisted videos you have planned.
Terrific presentation, Rob. Despite the challenge(s) of learning a new vocabulary for AI -- and, I'm sure, a pared-down vocabulary -- you did a great job of orienting me/viewers to how to think about AI, and Large AI and Small AI. Thank you.
Gonna love this series! Thank you so much Rob! 👍
Remember to check out the pdf. It is very important!
appreciate it. I'm interested to know what you all think since this is a new direction
@@robbraxmantech Great topic, love the bit on how you got the Ai to create a python code to mimic Windows Recall. Defiantly would love know more! 😀
This topic is hot and I think not only for me a perfect hit - as this is exactly anyhow what I wished to explore as soon as possible.
Looking forward to this series and I wish us all a lot of fun and amazing learnings with it!
Excellent content, bravo for your commitment to privacy and standing up and teaching.
The democratization of uncensored private AI that can generate any text and image will revolutionize the world.
Yeah IF
not sure I'd call opening up source democratization. Too much about democracy is not free, especially in the tech world and with new technologies. Too used to that word coming out of the mouths of would-be feudal tech lords.
Interesting!!❤ thanks! California here!
Yes!!! I love this new direction and perspective, Rob. Great work!
Rob, I built a dual 4090 set up with the intent to work on things like this. I'm using ooba, and learning as fast as I can. I currently have several models deployed on my local hardware. Learning Parameters and getting ready to get in over my head with local training. I'll be following your new series on Ai closely.
Everything we do with AI is run locally and not online. The models are getting better every day and we are also setting this up for our clients. It's important to start learning this as soon as possible before people try and block it.
Not sure I understand you. Most of the AI we deal with today including Co-Pilot, Apple Intelligence, ChatGPT, Claude, Gemini, Meta and tons more are mostly in the cloud. And the models running in the OS locally are controlled by the Big tech entity from the cloud. Using it locally under your control is a fairly new approach. Most videos on this is within the last month or two.
Doesn't make sense unless it is specifically limited to certain functions.
Full AI won't work with limited phone storage as a standalone item.
Maybe when I get a phone with enough space.
The cloud function has to be used as the backend.
Edit.. Just realized you didn't mention mobile devices, but also didn't exclude them.
You’re the man!! I’ve learned a lot ever since I joined your community, and I’m grateful I bumped into ur channel, really educational, I thank you my friend!!
Looking forward to the next videos in this series.
Thank you! I’ve listened to you for a few years and waited tillI I had funds to make changes. I don’t even understand AI.I used to be tech savvy, but things are changing too fast. In this video you outline your tech specs. I am buying a System76 laptop and I am following your specs,but would like reassurance that it will last me for a bit. I will be using Lexis/Nexis + AI. I guess it’s cloud based. It’s internet based. I will be re-hauling my entire tech system and am looking forward to adopting the VOIP phones and a degoogled phones you offer. You are awesome for educating us!
22:43 mind blown 🤯 Thanks, Rob!
You're my new favorite channel 🚀
I use upscayl, Topaz, and some other tools that all run locally without internet connections. must have the hardware obviously but this is what im working towards. i have extra rtx cards simply to make a private ai workflow, looking forward to you releasing more private ai content.
I love your work Rob...and this series is EXACTLY what I have been looking for... thank you so much for the way you share your considerable knowledge with us mere mortals.
Thank you.
Now we are talking. However don't forget! There is no saviour coming. We have to fix out minds abd save ourselves. We all have to contribute!
*Amazing breakdown of a complex subject in bite-sized statements.*
Thank you Rob! You are a guardian of humanity!
Thank you. Appreciate all the advice you give across all your videos.
Yes. And you can remove the censoring yourself through proper Prompt engineering
Another great video Rob!
Hope this comment helps the algorithm get this video to more people
Uncensored model like Mistral (gtp 4 equivalent) can be run locally on Linux.
You can ask the questions of your dreams 😬
thnx for sharing that! If my pc is decent, (8gb ram), will it work without laggin?
This is radically awesome!
very interesting video and idea! i am really looking forward to watch this serie! thank you!
I'm also excited about running my own voice assistant. We're getting very close to it being on par with commercial voice assistants. It will tie in nicely with local AI. Our own Jarvis (lite version) on our own server.
0:11 Awesome 😎! You rock Rob! 🤘😎
I setup a proxmox homelab server with old server hardware and have I am using it for local LLM and image generation tasks.
Love your videos!. I've been running Ollama for a while. With my current setup, I run Ollama on one machine as a Windows service and have 'Open WebUI Version' running in Docker connected to the Ollama server. This gives you a nice web interface similar to Chat GPT. This makes it available on all devices on my network.
Makes sense. I'm trying to make my own UI instead of using Open WebUI which seems to be the most popular. But I'd like to tweak the AI but have the built into my own UI (like Fine Tuning and RAG integration)
Der Osterhase kommt zu Robbi und belohnt ihn mit saftigen Mören und Eiern, super.
Give Phi-3 medium (14b) 128k a try. Best model for me. Close on reasoning to GPT-4 with a large context window. Runs with 8GB VRAM - love it.
Awesome series debut and first time you've said something positive about Zuck. 🤣🤣
Didn't know how this was going to be received ...
I come here for good info. Great channel. TY for the content. I check you as well as The Hated One on UA-cam. Two great channels that more people should have.
Much appreciated
Love your content, sir. Saving to my playlist.
Thinking the same way
Another great video Rob! Thank you!
Rob also check out LM Studio and Anything LLM. Between those and Ollama you can build a completely local RAG and more.
Conspiracy theory time: The entire video is AI generated, Brax has been captured by the glowies, and put some AI at his place to convince us to adopt AI in mass.
LOL. It becomes possible doesn't it? But the future will have AI that will spot a DeepFake. AI vs AI
@@robbraxmantech- That’s probably exactly what the human Rob Braxman would say.
That is just a theory, not a conspiracy theory.
Rob BrAIxman. An AI trained to point out the invasive tactics of the Big Tech/Big Government complex.
They say, that even 99% of internet traffic traffic is generated by a.i., bots:)
Joe Rogan mentioned you by name in his latest episode with Tyler Fischer
He mentioned me a couple of weeks ago. Is this new?
Rogan is a WEF weasel
I dont listen to him, I listen to you Rob. KEEP UP THE AMAZING WORK!!!!
I listen to Joe way too much. He’s mentioned you by name more than a few times these past couple months. Pretty sure he’s watching your videos regularly.
@@robbraxmantech yes, the video was posted yesterday
Great info series. Thanks.
You can't use nuclear weapons to defend yourself from nuclear attacks... But you can use good-AI to defend yourself from evil-AI.
Excellent Rob! EXCELLENT!
Exactly what I was searching for! I wanted to know how I can protect the ideas I have and want to develop utilizing ai. Hopefully these methods are also something someone with no coding skill can implement
I keep a copy of llama and a bunch of uncensored models on an older laptop in a faraday bag. Will make for a great SHTF knowledge base if needed.
NOTICE: the human brain is especially good at pattern recognition: recognize dad and mom, recognize friends and foes (primary survival instinct FFF - Fight/Flight/Freeze, from our primal brain, which gets most of the resources of energy from our diet - this has books written about, pertaining to neuroscience and cognitive processes). Another example: you have a hobby - bird watching - the more you do it, the better you get at recognizing the pattern of recurring species, in different background and distances - as you develop the pattern recognition for birds. etc.
Partake of the beast, become part of the beast.
Or turn the beast into a slave.
@@robbraxmantech Not possible - cos you don't control all the other AI's.
So what do you recommend we do then?
OP has it right. I dabbled with the original LLAMA, Alpaca, Vicuna and other variants starting in March, 2023 on my Linux machine. It was an interesting experiment, however what I noticed was days of my life were slipping by with nothing in particular to show for it. Those days and weeks became beast food. I got back on my life mission and left all the furry creatures behind.
I don't trust any of it.
Thank you for this one!
Hi Rob.
Great hands-on contents; thanks and congrats! In aiming for privacy-preserving use of AI, is there a way (a direct link, or some way) to download LLM models, say LLAMA3, not thru the OLLAMA commandline but in a browser, even via torrents (still from their authentic source)? Or alternatively, how to migrate such LLM models from a PC (on which they are downloaded/pulled thru OLLAMA) to other PCs offline, sort'a, I guess, making the process portable? How could this be possible?
There are platforms for this like huggingface.com and langchain.com but for best privacy, the best way is to load everything yourself and not involve any external website.
Computer engineer here. At 6:21, you conflate backpropagation, which is an error correction procedure that the AI model itself does during the training runs to automatically adjust its own parameters during the training runs in response to any training input, and Reinforcement Learning from Human Feedback, which is the procedure we use to align and sometimes "censor" those models. There's a sad fact about RLHF: it actually dumbs down the model's capabilities. There's no such thing as perfect censorship or perfect alignment.
Cool. Thank you
@@robbraxmantech Can you pin the comment and/or add a correction notice using UA-cam's "corrections" feature to prevent further confusion?
Spoke with a researcher in this field - they're having problems erasing problematic/erroneous data that has been previously ingested. And it tends to "hallucinate" "facts" that have never occurred,
my thoughts exactly. Like Iran using AI to find Christians and Christians using AI to evade detection kind of scenarios
He always has great videos, gets right to the meat.
When is the next install ment Rob. Waiting eagerly.
Wednesday
Hi Rod. First of all, thanks for your informative videos hat are explained to the common people in a simple English. I'm a new to Linux and use Zorin OS because of it resemble to Windows.
Since Windows holds the biggest share of the market, the Windows third party software that people use most of the time do not have Linux version. Sure, there are equivalent free software, but when to much work was done on the Windows software, it is hard to replace.
I understood that there is a way to work around this, making Windows and Mac third party software to work on Linux. However, it is very hard to find this kind of knowledge for specific programs in detailed videos of how to install them on Linux. If Linux advanced users would show that, things would be much different.
Here are my thought in three steps.
Step 1 - Showing how Windows and Mac third party software can work on Linux, and doing many detailed videos like that on different Linux distros if needed for each software. People will see this and understand "We Don't Need Windows and Mac".
Step 2 - After a mass of people will start to use Linux mainly do to the fact that their third party software is working on Linux, naturally, the mass of people will dictate the companies to make Linux version for their software. We have a very good example to a pressure of the mass with both Windows and Adobe. Yes, both Windows and Adobe still trying to hold the cards but they are too blind that they cannot see the downfall straight ahead.
Step 3 - Most of people will start to use Linux because it is free, have many distros that will fit to the a person goals, and all the third party software companies will make Linux the main OS.
So excited for the next phase in human history, I've always wondered what it would be like to live during the bottom of our long decline into the dark ages
How would we take an uncensored language model , then add some additional archived data for example Wikileaks so we can quickly search for related information
The main weakness of AI is saturation. With overfeeding one kind of information, the whole suffers.
The AI has other limitations that one can take advantage of. For example, if I were a teacher, I would know how to load up my Essay tests with triggers that will alert me to AI use for cheating. So plus and minus.
Excellent
I'm always learning something from your channel but I admit it, I'm still a noob to tech.
I like to use Perplexity for finding resource materials, can the Lenox ai gather resources like pdfs, websites and such?
Thank you.
I will teach that in the future.
1:06 Rob. Please let viewers know the impact of using XORG Linux VS wayland Linux. With XORG based distros EVERY keypress is shared with EVERY app. EVERY notepad/calc/terminal/browser can see every personal message, every calculation, your sudo pw, and your financial credentials. On an XORG Linux distro run XINPUT, not as root, in test mode in a separate terminal and open notepad/calc/terminal/browser. ALL keypresses are captured.
I would have thought that keypresses always get caught by the OS and can be viewed by an app asking for access. If wayland doesn't do this that's quite impressive. However, we still have the problem of screen capture
@robbraxmantech I haven't heard of Windows keyloggers which do not require elevated permissions to run. My understanding with MAC, Windows and Linux wayland is an OS "root" level process receives the keypresses and only makes them available to elevated processes, such as keyloggers, and the app with focus.
Consider an XORG calc, a malicious variant could capture the sudo pw, enable remote access, and share connection details including sudo user/pw with the calc creator.
Researching this I discovered that VirtualBox supports wayland only if the distro includes xwayland support. When searching through compatible distros I only found two options - KDE/GNOME. The Debian versions of GNOME/KDE allow you to log into XORG or wayland so you can easily compare XORG to wayland.
Agreed on screen capture. I have a Windows screencap portable app which runs without elevated permissions and can capture the screen. Assuming you are not typing your bank account password into an unmasked input field the screencap program could not capture it. A screencap program would not be able to capture a sudo pw as it too is masked.
Also a fail is the clipboard. Many pw managers use the clipboard for copying and pasting your “secure” credentials into web sites. Once in clipboard EVERY app has access to your “secure” credentials.
Another horrible idea are browser extensions. Check out Naomi Brockwell TV’s “HOW Browser Extensions Steal Your Data”.
@@robbraxmantech wayland doesn't let full screen captures either, unless you explicitly add in the functionality in the window manager/compositor
I tried the Wayland (experimental option) in the Linux Mint menu, but noticed it doesn't work with many apps. Also, the display looks noticeably blurry compared to X11.
AWESOME! This wis exciting
Braxman always explains things to me like I'm an idiot. 😆 One day I will buy the De-Googled phone. 👍
I worry for the future of Linux and our ability to resort to it! Can you please talk about Linux’s past year (2023) and being bought by IBM and the medley of problems with decreasing funding towards Linux kernel and interest groups (intrusion from Microsoft, Google, etc)… There’s a video called “Linux sucks 2024” by a community member that outlines the much overlooked past year and potential futures of Linux. Thanks 🙏🏼
This is the thing internet was truly made for
real open source ai for smb sector is about 4-5 years away - it could be even longer, that estimate is being optimistic, when it does 'happen' expect much more hype and economies of scale, more discovery, more innovation. The long delay is due to sw/hw immaturity and costs, cost have to come way down to make ai accessible but it will although it is going to take 3-5 turns of moore's law. people are absolutely horrible at predicting the future, big tech ai is going to take a while to mature as well due to the massive infrastructure improvements needed - even that is going to take the better part of 10 years
I have been educating people on how to use AI, i know they can learn from a google search, but me teaching allows me to also explain that anything they send to it, can and probably will be recorded.
I see it as helping the devil make a deal with them, but how they can do it and protect their "soul" vs them possibly doing it on their own and giving up more than they thought they bargained for
Where do you usually recommend starting for someone who wants to learn to use AI securely for independent projects?
@kiwi6421 from what Rob is saying, I think just waiting on him to do this series is enough. Otherwise it's forum diving, different youtube channels, and a lot of trial and error.
It depends on how you view time, waiting for 1 YT creator, doing some research from others, or using forums and downloading LLMs to test for what you want to do
Thank you!
will there be a need for a software update? if so what would be the best way , keeping privacy in mind?
Interesting. I run it on Windows & Linux now. Graphics drivers are a bit of a pain on Linux but Nvidia works with a little effort Thanks
Where is the PDF of your AI exchange? I didn’t find it.
Thanks for the series.
Fear not the Beast - it is here to serve you.
666
00:00📚 This will begin a new series on AI.
00:04🤖 Focus will be on harnessing AI for privacy and security.
00:18🛡 Teaching will involve hands-on tech for safe AI use.
00:25🔐 AI will be used to demonstrate privacy and security concepts.
00:33⚠ There are two main AI threats to be aware of.
00:43💻 Hidden AI can work on devices with Windows, MacOS, iOS, or Android.
00:56📊 AI can send data to big tech servers for surveillance.
01:06🔧 Using Linux can mitigate this risk.
01:10☁ Sending data to cloud AI controlled by others is another threat.
01:29🖥 AI can now run locally on computers without the internet.
01:44👨💻 Example of running AI safely under your control with Linux.
02:02🤔 Introduction to new AI concepts and demo for learning.
02:40💡 Conceptual explanation: AI is not built from logic rules.
03:00🚗 Tesla's shift from rule-based to full AI improved self-driving.
03:17🧠 AI learns by itself using machine learning.
03:30🔄 Generative AI can come up with novel ideas.
03:39📉 AI uses math to discover patterns.
03:55🌈 Colors in images are numerical values for AI to recognize.
04:11🏷 Millions of tagged images help AI recognize objects.
04:28🧩 AI builds a neural matrix to identify patterns.
04:52🧠 GPT-4 has 220 billion parameters.
05:21🔍 AI's learned patterns are not always human-readable.
05:37📚 Pre-trained AI models are made from large datasets.
06:16🔄 AI learning is tweaked using backpropagation.
06:44💬 Large Language Models (LLMs) like GPT-4 can converse deeply.
07:18🛠 AI has two main functions: learning (ML) and querying (inference).
07:47💸 Building a pre-trained AI is expensive; using it is not.
08:30🤖 LLMs are for deep conversation; SLMs are smaller models.
09:07🔍 Transformers encode and decode inputs and outputs in AI.
09:39🔑 GPT means Generative Pre-trained Transformer.
10:01🗂 Businesses can add their knowledge base to existing AI models.
10:27🎯 Fine-tuning AI for specific tasks is possible.
11:24💡 Understand terms: RAG, fine-tuning, and AI agents.
11:48🖥 Smaller AI models don't need special hardware.
12:02🚀 NPUs and GPUs are accelerators for AI tasks.
12:27🔧 Microsoft and Apple use specialized chips for AI functions.
13:01🖥 Small AI models can run on CPUs without GPUs or NPUs.
13:30💻 Multiple small AI models can run on a computer.
14:22🔍 Cloud AI can process inputs and return results.
14:57🔒 For privacy, consider switching to Linux OS.
15:55💡 Running AI on a powerful computer allows for larger models.
16:22🌐 Using cloud AI like ChatGPT is generally safe for inference.
16:48🔐 Always be mindful of the privacy risks when using cloud AI.
17:00🛠 Demonstration of safe AI work will follow in future videos.
17:31💻 Example setup: Dell XPS15 with Intel i7 and Nvidia GPU.
18:09🖥 Installing AI models like Llama on different OS is simple.
19:00📂 Example of using local models: running GPT-4 on a MacBook.
19:30🔧 Installation steps for running AI locally on MacOS.
20:00💻 AI installation involves setting up Python and other dependencies.
20:30📥 Downloading pre-trained models for local use.
21:00🔄 Example of running inference on the local model.
21:30⚙ Using command line tools to interact with the model.
22:00📚 Documentation and resources for further learning.
22:30🛡 Emphasis on maintaining privacy and control with local AI.
23:00🖥 Benefits of running AI locally: security and customization.
23:21📈 Encouragement to explore and experiment with local AI setups.
Nicely detailed structured, mucho gracias.
So when you suggest that I install code from META on my computer, I assume you reviewed that code very carefully so that nothing unexpected is executed in the background. Right?
A model is not code. It's an array of weights. So if you understand what is running, it should alleviate concern. In this case the enabling software is Ollama (open source)
@@robbraxmantech Thanks for the clarification. I asked the question because my trust in M is less than zero, so if they are giving something away "for free" I suppose there is a catch. However, I will install it in an insulated container just in case.
Thank you for all you do to help us protect our privacy.
I'd like to see Ubuntu secured with privacy in mind as a base to launch the AI from.
The intro to how neural networks work here is wrong. Each parameter does not equate to some pattern being found or embedded in the network.
And neural networks are still just rule based logic, it’s just that there are far more rules and they’re not represented by traditional Boolean logic but by the factors of inputs to “neurons” and simple math done to them against the internal weights of the neurons.
Neural networks are simply a bunch of trivial functions with embedded constants, set during training, which pass their output to another function in a group of functions called a layer.
You could actually code this by hand, through trial and error on what the weights should be, but it would take forever, so we write fast algorithms to do it for us.
It’s more or less just automated brute force coding, which can be enhanced with other methods like human feedback and a various algos to nudge things along faster.
Amazing. subscribed!.
Thanks Rob!
I'm still unsure how to implement an effective RAG system. I need something capable of analyzing large PDFs, hundreds of MBs with thousands of pages, in both English and Spanish. Models like Nomic or BERT aren't sufficient as they tend to generate inaccurate information and have limited context.
I've read about tools like FAISS but I'm not sure how to integrate that with OpenWeb UI and local models like Llama3 or Phi3 128k. Has anyone managed to successfully analyze large PDFs, including pre-processing with OCR to translate mathematical formulas into LaTeX or Markdown?
I also noticed that Microsoft has released the Florence model, which seems to have strong visual recognition capabilities. Any advice or insights would be greatly appreciated.
Sounds like you're going production with this. I think there's some more understanding needed by the AI community. RAG by itself may not be the answer.
@@robbraxmantech Thanks for your answer! Just to clarify, this project is not intended for commercial or professional use. It's a personal project for my family (my father and brother) to help them read large PDFs at home. I'm looking for a practical solution that works well enough for our needs, not necessarily something that's production-grade. I couldn't find any solution for large pdfs yet.
Thank you.
We The People need to take back control over global DNS - EVERYTHING is subject to DNS, including AI/ML.
I have laptop with similar specs (32g) aside from RAM. Does your fan get super loud during output? I had to restart a couple times... I thought it was gonna blow up when I ran llama3
Not getting hot but Linux runs faster than Windows.