Richard A Aragon
AI & ML interests
Articles
Organizations
TuringsSolutions's activity
Amazing!
I made a Github repository for my Hyperdimensional Computing Neural Network: https://github.com/RichardAragon/HyperDimensionalComputingNeuralNetwork
I made a YouTube video showcasing the model and some of my experiments with it: https://youtu.be/Eg51o519zVM
Sorry the audio quality sucks, I will buy a new microphone today. Why does some moron like me solve these things and not you? I know more about how computers work than you do, that's it. Swarm algorithms were big in the 90's and early 2000's. Computers were absolute dog doo doo then in one specific way, compared to now. That one way, which everyone overlooks, is the entire secret behind why swarm algorithms are so good.
Don't believe me? Ask Python-Chemistry-GPT yourself: https://chatgpt.com/g/g-dzjYhJp4U-python-chemistry-gpt
Want to train your own Python-GPT and prove this concept actually works? Here is the dataset: https://huggingface.co/.../TuringsSolu.../PythonChemistry400
https://www.youtube.com/watch?v=tuQI6A-EOqE
https://www.youtube.com/watch?v=yz0sd8ayenI
https://www.youtube.com/watch?v=I8jHRgahRfY
Model Link: https://platform.openai.com/playground/chat?preset=eCrKdaPe9cnMnyTETqWDCQAU
Knowledge Base Bots are internal facing as opposed to external facing LLM models, that are either fine tuned or RAG tuned, generally on systems and processes related data.
Learn more about Knowledge Base Bots at our website:
https://knowledgebasebots.com/
Geometric fractals do not allow you to sacrifice accuracy at all, that is not how geometry works. That happens to be how calculus works. It suddenly paid off to understand math theory. I didn't believe it when I did it either.
There is no such thing as a stupid question when trying to learn, that is how we learn. Here, this will help you more than anything else. You need to put in your own HuggingFace token, you need to change the model name, and you need to use a different dataset. I have the PFAF750 dataset in my profile, and like 90% of my datasets are a blend with P-FAF data.
Do not delete the RAM until you are done playing around with the model. When you upload the model to HuggingFace, it will be quantized. That model will perform worse than the model in your Colab notebook, it is how it is. That's the only way to keep it all free.
The training arguments in this notebook are for the Adam Optimizer and LORA fine tuning. That's 90% of what you need to know.
https://colab.research.google.com/drive/1KIRKGGB-LAqEhICQdtn_aJt8CzsWK6mH?usp=sharing
That formatting sucks lol. Use this instead: https://colab.research.google.com/drive/1YhKTZNGMIqniMqhjQnRgkQeXxb2rP2fv?usp=sharing
Just let me know if you want anymore help at all!
import torch
from torch import nn
from transformers import AutoTokenizer, AutoConfig, AutoModelForTokenClassification
Define fractal functions
def f1(x):
return x**2 + 0.1
def f2(x):
return 1 - (2 * x - 1)**4
Custom P-FAF Embedding Layer
class PFAFEmbedding(nn.Module):
def init(self, embed_size, fractal_funcs, num_fractals=None):
super().init()
self.fractal_funcs = fractal_funcs
self.num_fractals = num_fractals if num_fractals is not None else len(fractal_funcs)
self.p = nn.Parameter(torch.rand(self.num_fractals)) # Probabilistic weights
self.d = nn.Parameter(torch.rand(self.num_fractals) * 1.5 + 0.5) # Fractional dimensions
self.embed_size = embed_size
def forward(self, x):
# x: [batch_size, seq_length, embed_size]
batch_size, seq_length, _ = x.shape
x_expanded = x.unsqueeze(1).expand(-1, self.num_fractals, -1, -1) # Shape: [batch_size, num_fractals, seq_length, embed_size]
# Apply fractional dimensions and fractal functions
x_dim = torch.pow(x_expanded, 1 / self.d.view(1, -1, 1, 1)) # Apply fractional power across dimensions
# Apply fractal functions probabilistically
t = sum(p * f(x_dim[:, i, :, :]) for i, (p, f) in enumerate(zip(self.p, self.fractal_funcs)))
return t
Custom BERT Model with P-FAF Embedding
class AutoModelWithPFAF(AutoModelForTokenClassification):
def init(self, config, fractal_funcs):
super().init(config)
self.pfaf_embedding = PFAFEmbedding(config.hidden_size, fractal_funcs)
def forward(self, input_ids, attention_mask=None):
# Normal BERT inputs handling
inputs_embeds = self.embeddings.word_embeddings(input_ids)
inputs_embeds = self.pfaf_embedding(inputs_embeds) # Apply P-FAF transformation
# Rest of the BERT model
extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_ids.shape, input_ids.device)
head_mask = self.get_head_mask(None, self.config.num_hidden_layers)
encoder_outputs = self.encoder(
inputs_embeds,
attention_mask=extended_attention_mask,
head_mask=head_mask
)
sequence_output = encoder_outputs[0]
pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
outputs = (sequence_output, pooled_output) + encoder_outputs[1:]
return outputs # Return the base BERT outputs for compatibility
Define fractal functions to use
fractal_funcs = [f1, f2] # Additional fractal functions can be added here
Load pre-trained BERT and modify it
config = AutoConfig.from_pretrained("google-bert/bert-base-cased")
model = AutoModelWithPFAF.from_config(config, fractal_funcs)
tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-cased")
Yes, it will give BERT gsm8k scores that are on steroids lol.
A simple citation never hurt anybody lol. - Albert Einstein
I didn't think you could either. I don't think anyone should actually be able to. Thank you.
https://huggingface.co/blog/TuringsSolutions/pfafresearch
I think you are not wrong, it is the most plausible explanation. It is either that; for reasons that would be scientifically unexplainable in my head, that transfer learning does not work in this one instance and one instance only, or the paper is wrong. Given the evidence I know firsthand, I would say the paper is wrong. You rightly point out, it would not be the first time for one of the research institutions on that paper. It would not be the first time for any of them overall, let's keep it 100% real.
I don't know what the truth is regarding this situation but I do know one thing for sure. Our sources of truth are full of bullshit. And we wonder why that causes issues when we train models on the data.
This doesn't burst my bubble it makes me happy! I will look this up right away.
Alright, alright, alright. I am just an a--hole and I owe you an apology. This is good stuff, thank you.
OK, I believe you some now. Testing further.
Edit: I was wrong.
Edit: It works.
Make a copy of the colab and make this update you lying pos.
So, I followed all of your instructions. When you do that, it is very clear they are out of order, and YO APP DO NOT WORK, SON. Why are you trying to bs people?
https://colab.research.google.com/drive/1JzvIN8o0293LYy88MWl2WGDVfm4dLYZj?usp=sharing
Perhaps it works for you because those files are on your local computer and you never uploaded them....
It doesn't and I have the same error as before. Where are the files? This app does not work as presented. At all.
FileNotFoundError: [Errno 2] No such file or directory: '2013.parquet'
No such file or directory: 'vocab.json
Same for the parquet file. What are you using for the vocab? You know, what you claim that your app actually does....
This is absolute trash and doesn't actually work without any of the files you do not publicly provide anywhere. No way to verify this garbage app even works.
I'm going to take this and it is going to become a small part of a product of mine. I appreciate it!
https://arxiv.org/abs/2407.15211
Ah, I see why you are so interested in shitting on my work now. You are jealous! You could have just come out and said that in the first place.
You should research it and build something like this yourself.
The random cats and all the math is in it. I have also talked to UC Berkeley directly about it, they are one of the stars on the Github repository. You can open up the app.py, look at the math, and realize that AI is not coding, it is math.
Yes, exactly. My method is over 85% and it is cheaper. I don't understand what else there is to discuss? The random cat facts in the demo space is because that's the default API. Someone really likes cats and doesn't matter if you hit their API with 25 bots at once. Some don't like that. You can put in your own API, it is just the default.
You are incorrect in both the accuracy rate of the function calls and the fact that it is solved. I have been following this problem for 5 years now. I have multiple agent based frameworks I have developed myself:
https://github.com/RichardAragon/MultiAgentLLM
https://github.com/RichardAragon/MOBASwarmAgents
The people who deem themselves 'experts' in AI are quite the problem. It is not my job to educate them in any way. Learn math. It's all math. You don't know jack about the subject if you are mathematically illiterate and no amount of talking to ChatGPT can fix that.
It's the speed at which it happens. I cannot control the accuracy enough to solve it most optimally. But that is just a math problem. I don't know how to fix it, but it is 100% fixable. Need money to fix. Need someone to actually understand math when it comes to AI to get money. Why do so many people have an interest in AI but refuse to learn math?
Here is another one you will not understand, it is called HiveMind. The problem at the moment is that I cannot fully control any of this. If I could I would not be wasting my time here, I would be straight at Google HQ right now. Instead, they are stealing my shit because I cannot iterate on it fast enough. I'm positive I could control it fully, with money for more research. So, here it is publicly: https://colab.research.google.com/drive/1gXasjeZM_8u49go2Hn8cqA30Rm3JPc3V?usp=sharing
There are absolutely better solutions that exist to the problem than the simple demonstration I laid out here, that is correct. Thank you for looking. Do you have any actual questions? Yes, Swarm algorithms on their own do not have enough 'juice' to make them smart enough on their own. Give them a brain though, it's quite amazing.
Everything you laid out about the functionality is correct. Your understanding of where exactly LLM models are at when it comes to function calls is severely lacking and it is not my job to fix that understanding. Your understanding of this technology hollistically is very incorrect and I can ascertain that simply from the statements you have made. No, I have zero interest in debating them with you. I will more than gladly debate any actual prominent researcher or investor on these things.
I am indeed lacking social skills, yes. I only care to socialize over algorithms and math, very honestly. I cannot put that into any other terms.
I like how the first reaction from people who cannot do math is mania. That is what is severely broken about this world and makes the world quite insane. I pay because you are mathematically illiterate lol. Even happens on sites literally devoted to ML.
I also just proved it's theoretically possible to create text via diffusion using the same methods. Imagine being able to create Moby Dick in 3 seconds rather than a picture of a whale?
https://colab.research.google.com/drive/1VF4fQLKCCs8JVRXDOUOqZ9Kr6-LspAtU?usp=sharing
Maybe if I talk about this in non math terms more people will understand: There exists a 'Platonic Form' of the solution to the Traveling Salesman Problem. this Platonic Form is the most optimal solution to the problem that could ever exist. I take the Platonic Form and I make it a variable (x). Then I instruct a bunch of algorithms that I just placed this Platonic Form somewhere in the box but even I do not know where in the box it is, or even what it looks like. The agents all go in different directions and explore the box. When an agent finds a clue, they tell all the other agents. The agents all look for more clues until they find whatever I put in the box. Then they simply describe it to me.
No one has ever thought to Quantize Reverse Diffusion before? Really? What about the sausage one? Bravo and kudos to you either way!
If anyone is interested in investing in this technology, I just broke down in very simple terms for anyone who can actually understand it, exactly how it works. I would love to debate any part of the actual technology or math.
I also have novel implementations of it all over the place in multiple forms. I can showcase that through PSO and Gaussian Probability sampling, that I can solve literally any optimization problem that can be conceived of. Seeing as the solution space involves, literally any optimization problem that could ever be conceived of, I utilize AI to help generate things like a completely novel neural network from scratch with multiple matrix calculations and completely novel attention mechanism sometimes. AI models can't do math, right? It's all math, not code. Either AI models can't do math, or they can. I would be happy to debate that further with anyone but you if you would like.
Can I explain how the novel algorithms that I have been writing about for about a month now work? Yes, I can. Can you pay me at least $200M to do so?
Yes, your previous comment actually proved this further, I wasn't going to comment on it but I will now. The fact that you are focused on the AI generation of the code is quite hilarious. Even if I used AI to generate the code in its entirety and edited nothing within it, the implementation and the functions within the code are novel, as is the math. Since all of these go above your head, we are engaging in this frivolous discussion to reinflate your ego instead.
OK, I used AI generated code within my completely novel implementation of a Diffusion based SNN, I also wrote a paper on the subject in which I utilized AI. What else my guy?
It's obviously not the first time the guy has trolled me. I have no idea why but I attract trolls. Those who cannot do, troll. I handle them in the same way I handle everyone, that is why I am the CEO.
I can tell you like to harass people and seem to think I owe you an actual response. I don't know what you get out of this but I get increased viewcounts to my post either way when you comment which is all I care about here. I have reported you now, I hope you can get the message. Anywho, I am done chatting with the person with so much cred they have to use a fake name on a site devoted entirely to showcasing your cred.
Why are you still harassing me, Xander? Be well.
It is not hard to remember a unique name like Xander. Rather than criticizing other people's work, which you can't even comprehend in the first place, how about you learn how to build your own AI, Xander?
Wow you don't know how diffusion works or what is happening here so why are you commenting? Username does not check out.
Billionaires have been made for less than this. This is only one of the things it can it do. It can do API calls, function calls, optimize poker and blackjack odds, anything that is an optimization problem. It costs fractions of a penny and requires fractions of the compute of an LLM model. It can even communicate two ways with an LLM model.
@LeroyDyer Yes, it is just agent setup. The method is overall not that novel. PSO Swarm optimization was invented in 1995. That is the basis of my method. The breakthroughs are:
That you can simply use Gaussian Probability Distribution + Swarm Algorithms to solve ANY optimization problem (including function calls, API calls, image generation, etc.).
The power that SNN + LLM brings. SNN's have what we will call 'built in intelligence'. It's not that useful in practice. LLM models have decent enough logical reasoning capabilities. It literally requires a single function to setup two way communication between the SNN and the LLM, with the LLM model able to instruct and guide the SNN in every way.
So far, I have replicated: Diffusion models, API agents, and I can put 'AI' directly into a spreadsheet. All because of just these agents. I also slap a multi-head attention mechanism on top of them which is also important.
Here: https://colab.research.google.com/drive/1SeYnyovBEIqI-HC7MAqrdDW6ZWjImdWb?usp=sharing