Berlin Brown and Software Development

 More information on GPU: https://jax-ml.github.io/scaling-book/gpus/

 Just learning about this, response from Open AI Chat GPT. On Lucy Guo Founding Scale AI In 2016, Lucy and Alexandr Wang founded Scale AI (via Y Combinator), originally conceptualized as an “API for humans.” They landed on solving a real pain point: data labeling for AI , especially for autonomous vehicles. One of their earliest customers was Cruise The Solution: “API for Humans” Scale AI positioned itself as an API layer between companies and a human workforce . Instead of building your own data-labeling team, you could make an API call like: { "task" : "label image" , "data" : "car.jpg" } And Scale would automatically route that task to humans (plus machine learning models for pre-processing), then return the annotated result . To the engineering team, it looked like calling any other API. But behind the scenes, it was a managed pipeline of people + ML models doing the work

  OneTrust Specializes in governance, risk, compliance (GRC) software with a strong focus on AI governance and data protection. Headquartered in Atlanta, serving over 14,000 customers. Website : https://www.onetrust.com/ Innova Solutions A global IT and business services firm based in Atlanta, delivering AI, generative AI, automation, data analytics, and cloud services. Founded in 1994, with over 50,000 employees and nearly $3.5 billion in annual revenue.  Website : https://innovasolutions.com/ Closeloop Technologies Inoxoft SDLC Corp Dualboot Partners (again) A42 Labs Novarata GoodF

    Sir Demis Hassabis is a British artificial intelligence researcher and entrepreneur. He is the chief executive officer and co-founder of Google DeepMind and Isomorphic Labs, and   Meta did not hire an engineer for $250 million, but rather an AI researcher named Matt Deitke was hired for a four-year compensation package worth an estimated $250 million. The massive deal was reported by The New York Times and other outlets in early August 2025.   Details of the agreement Initial offer: Deitke, a 24-year-old AI prodigy, first rejected a $125 million offer from Meta. Negotiation: After the initial rejection, Meta CEO Mark Zuckerberg personally intervened to double the offer to $250 million.   https://scholar.google.com/citations?user=k4VxCcYAAAAJ&hl=en And World Foundational Models https://www.nvidia.com/en-us/glossary/world-models/ https://worldmodels.github.io/ Vs Thousand Brains Project https://github.com/thousandbrainsproject

 Here are random AI LLM benchmarks off of Reddit. " This project benchmarks 41 open-source large language models across 19 evaluation tasks using the lm-evaluation-harness library. Benchmarks are grouped into three categories, with the corresponding tasks and metrics listed below." https://github.com/jayminban/41-llms-evaluated-on-19-benchmarks

Georgia and Atlanta Companies: https://www.salesloft.com -  a sales engagement platform that helps sales teams automate outreac Calendly is an automated scheduling software that allows individuals and teams to easily book meetings and appointments by sharing a link to their availability Technology Groups IEEE Atlanta  https://atlantatechvillage.com/ - Atlanta Tech Village (ATV): A popular startup hub known for networking opportunities.  https://www.civictechatlanta.org/ https://www.refactr.tech/ DevNexus - https://devnexus.com/ ATDC (Advanced Technology Development Center) Consulting Firms https://www.rippleit.com/ - small consulting firm in Atlanta

 This was too easy, easy CV capture. Based on openai chat gpt codce. And from:  https://myberlinaustin.com/ import cv2 # Load pre-trained face detector face_cascade = cv2 .CascadeClassifier( cv2 .data.haarcascades + "haarcascade_frontalface_default.xml" ) cap = cv2 .VideoCapture( 0 ) while True : ret , frame = cap .read() if not ret : break gray = cv2 .cvtColor( frame , cv2 .COLOR_BGR2GRAY) faces = face_cascade .detectMultiScale( gray , 1.3 , 5 ) for ( x , y , w , h ) in faces : cv2 .rectangle( frame , ( x , y ), ( x + w , y + h ), ( 0 , 255 , 0 ), 2 ) cv2 .imshow( "Face Detection" , frame ) if cv2 .waitKey( 1 ) & 0x FF == ord ( 'q' ): break cap .release() cv2 .destroyAllWindows()

 Here are more questions on that example from the LLM book. https://github.com/rasbt/LLMs-from-scratch/blob/main/ch05/01_main-chapter-code/gpt_train.py   From Chat Gpt - Open AI Response Yes, AI and Large Language Models (LLMs) typically start with randomly initialized weights, though more sophisticated methods have evolved from this simple technique. A random starting point is a crucial part of the training process that allows the model to learn and differentiate between features.   Why random initialization is necessary Breaks symmetry: If all the weights in a network were initialized to the same value, every neuron in a layer would learn the exact same features. This prevents the network from learning anything useful. By starting with random, different weights, each neuron is forced to compute something different from its neighbors, allowing for a more diverse and robust learning process. Enables learning from data: With random starting weights, the model's initial pre...

 Flag Burning and Trump Recent news.

 Businesses In Georgia https://calendly.com/about Prompt Owl AI Platform Company https://theaiinsider.tech/2025/04/29/worlds-first-agency-for-custom-enterprise-ai-workers-introduced-by-promptowl/ Micron Biomedical, Inc. https://www.micronbiomedical.com/ https://www.micronbiomedical.com/news/mi

 Here is my chat with ChatGPT on what is AI LLM? Break down the weights vs training data ELI5 Analogy: Learning Piano Training data = all the sheet music + recordings you practiced with Imagine you’re learning piano. You practice thousands of songs, scales, and drills. That’s the training data. Weights = your brain’s memory of “how to play” after all that practice After months of practice, you don’t carry around every sheet of music you ever saw. Instead, your brain has adjusted so you can play in certain styles, recognize patterns, and even improvise. That “muscle memory” is like the weights file. It’s just billions of little adjustments stored as numbers. Inference (running llama.cpp) = you playing piano on demand Now someone asks you to play something new. You don’t need the training books—you just use what you’ve already learned (weights) to generate the performance. ... Each token (word piece) is chosen based on probabilities learned in the weights If randomness (temperature, ...

 More Thousand Brains Videos: " The Thousand Brains Project is dedicated to building a new type of AI that works on the same principles as the neocortex. These principles are described in the Thousand Brains Theory and focus on sensorimotor learning with a repeatable computational unit modeled after cortical columns. "

 Review of Projects: Java Version: https://github.com/berlinbrown/htm.java https://github.com/htm-community htm core: https://github.com/htm-community/htm.core Thousand Brains Project https://www.youtube.com/watch?v=VqDVUWgJQPI https://github.com/thousandbrainsproject/feat.dynamic_resizing Jobs https://thousandbrains.org/company/careers/open-source-software-engineer/

Posed question to Chat GPT: response from chatgpt openai. ...  That is really interesting — you’re describing something like a "human-like AI operator" that interacts with the world only through the same physical channels a person would. Instead of APIs, direct memory access, or digital hooks, it would: See through a camera (screen scraping instead of data feeds). Hear through microphones (parsing audio instead of reading text). Act by pressing keys, moving a mouse, or even flipping physical switches. This would create an AI that is forced to deal with the world in the same messy way humans do , where perception and action are mediated by physical hardware. Some implications: Advantages Robustness — It can interact with any system humans use, regardless of APIs or locked digital environments. Security sandboxing — Because it only uses physical I/O, it’s less likely to bypass boundaries in unsafe ways. Human-aligned learning — Its training match...

 Good Meeting Atlanta Java Group: https://www.meetup.com/atlantajug/events/310453214/ "Pratik Patel is a Java Champion, developer advocate at Azul Systems, and has written three books on programming" "AI-Powered Data Exploration: Interacting with Apache Iceberg via Spark and LLMs"

 Have to read the book: Jeff Hawkins – A Thousand Brains: A New Theory of Intelligence Explains the "Thousand Brains Theory" in an accessible but deep way. Great for grounding AGI ideas in cortical theory.  Architectures, AGI Projects & Platforms OpenCog Hyperon – An open-source AGI framework using symbolic + probabilistic reasoning. SingularityNET / Cogito – AGI-focused decentralized AI project. Nengo (by Applied Brain Research) – A Python framework for large-scale brain modeling. Often used with Eliasmith’s work.

 Running the Chapter 5 Training Example,  This is an interesting example and without fully understanding it, asking Github Copilot, Chat GPT abotu the code. ... Here it is running, and also updates on the code.  The model has approximately 124 million parameters (hence the name) primarily from:  Position Embeddings: context_length × emb_dim = 256 × 768 parameters  Transformer Layers: 12 layers, each with attention heads and feed-forward networks  Even though "The Verdict" is only 3,600 words:  The model is designed to learn general language patterns, not just memorize the text   The vocabulary size (50,257 tokens) matches GPT-2's full vocabulary  The architecture follows smaller GPT-2 specifications (768-dim embeddings, 12 layers)  This is the total number of unique tokens the model can recognize and generate ''' ''' The full vocabulary (50,257 tokens) gives it the capacity to represent  many words it never sees in training It's bas...

 Bela Thank you for attention to this matter.

 Looking at the code from Sebastian Raschka on LLMs, here is a small snippet.  Training a model from scratch and basic test.  The key is understanding the basics. Here is the text to train, "I HAD always thought Jack Gisburn rather a cheap genius--though a good fellow enough--so it was no great surprise to me to hear that, in the height of his glory, he had dropped his painting, married a rich widow, and established himself in a villa on the Riviera. (Though I rather thought it would have been Rome or Florence.)" See: https://pytorch.org/ https://github.com/openai/tiktoken import matplotlib.pyplot as plt import os import torch import urllib.request import tiktoken   class GPTDatasetV1 (Dataset):     def __init__(self, txt, tokenizer, max_length, stride):         self.input_ids = []         self.target_ids = []         # Tokenize the entire text         token_ids = tokenizer.enco...

 So just random searches, Who is still using blogger, blogspot. https://returntoatl.blogspot.com/ https://diego-pacheco.blogspot.com/

  Darks Souls 1 and 2 are great games

 Yes, absolutely — the idea of building and training AI models using distributed computing , especially with tiny or resource-constrained devices like Raspberry Pis , is possible in concept and increasingly viable with the right technology stack. There are a few real-world technologies and approaches that relate to what you're describing. 🔹 What You’re Describing: You're imagining a world where: A small, local device (like a Raspberry Pi) can train or fine-tune a small model. These devices are networked , forming a distributed system . They collaboratively train a model or share knowledge (like a federated system). This is an intersection of: Distributed Computing Federated Learning TinyML Edge AI 🔧 Technologies and Concepts to Look Into 1. Federated Learning A decentralized approach to training models where each device trains on its own data locally, and only updates (not raw data) are sent to a central server or aggregator. Used by : G...

 https://github.com/thousandbrainsproject ... Reverse engineering the neocortex 🧠 to revolutionize AI 🤖. An open-source initiative.

 I asked this: 1. Better Search and Retrieval (RAG) Holy Grail: Turn every piece of company knowledge into something easily searchable, understandable, and actionable. LLMs are increasingly being integrated with retrieval-augmented generation (RAG) systems, which combine a company's internal documents (PDFs, wiki pages, logs) with conversational LLMs. Goal: Create “company copilots” — tools that answer employee/customer questions using internal knowledge. 🧠 2. Enterprise-Grade Agents and Automation Holy Grail: Use AI agents to autonomously take action on behalf of employees or customers. Example use cases: Automate Tier-1 and Tier-2 support tickets Handle complex workflows in finance, HR, legal, IT Companies want agents that are safe, traceable, and reliable , not just “smart.” 💬 3. Natural Language Interfaces to Software Holy Grail: Let employees use natural language to interact with business systems (SQL databases, CRMs, analytics, etc.). Think: ...

 Here are responses from Chat GPT.  Just random questions and responses from chatgpt.com What's Already in the Model? When you ask me a question, I first try to answer it based on what I already "know." I was trained on a huge amount of data up until mid-2024, which includes: Wikipedia Gaming wikis (like Fextralife and others) Walkthroughs Forums (like Reddit, GameFAQs) Guides and FAQs So for a lot of questions—especially about older games like Dark Souls 1 (2011)—the answer is already in the model. That means I can often recall details like: Boss names Soul drops Recommended level ranges Item and spell descriptions Lore connections If that had been enough, I would’ve just given you the info directly. 🌐 Step 2: Using the Web (Real-Time Search) In your case, the question was very detailed—you wanted: All bosses & mini-bosses Who give 4,000+ souls Exact soul amounts Recommended player level to fight them Even though I remembered some of that from training, it wasn’t en...