Berlin Brown and Software Development

Numenta HTM (Hierarchical Temporal Memory) is a biologically constrained machine intelligence technology that models the structural and algorithmic properties of the mammalian neocortex https://github.com/berlinbrown/htm.java This is adding updates to the project. The visualizer simulates a room thermostat reporting 24 hourly readings each day: ```text 12 AM 64°F → 1 AM 63°F → ... → Noon 76°F → ... → 11 PM 65°F → next day ``` This example runs the complete learning path: ```text hour ─────────→ periodic ScalarEncoder ─┐                                        ├→ Spatial Pooler → active columns → Temporal Memory temperature ──→ ScalarEncoder ─────────┘ ``` The temperatures are simulated sensor readings, but they are ordinary real-world values processed by the actual encoder and Spatial Pooler. Numbers such as `[2, 3, 5]` are internal column addresses, not temperatures and not numbers the mo...

I had GPT rewrite this post. https://www.myberlinaustin-atl.com/2008/06/is-java-new-cobol-yes-what-does-that_14.html In 2008 there is a phrase floating around the software community: “Java is the new COBOL.” Usually it is followed by complaints about EJBs, XML configuration, enterprise frameworks, or somebody announcing that Ruby on Rails is going to replace everything. I think that misses the more interesting question. The important question isn't whether Java sucks. The question is what happened to COBOL, and whether Java developers should learn something from it. COBOL was created in 1959 and became enormously important for business computing. Banks, insurance companies, governments, and other large organizations built critical systems with it. Many of those systems still work. That is worth emphasizing because old technology is not necessarily bad technology. If a banking system has processed transactions correctly for decades, replacing it simply because the programming langu...

  Many years ago, in 2008, a programmer researched DNA and how the cell works and wrote a software-related analogy for the process. It was a great overview, especially because he translated a lot of the biological terminology into concepts that engineers could understand. I kept thinking about the article afterward and wanted to add my own spin to it. The author didn't really do anything wrong—the article is excellent—but I personally wished he had reduced the software analogy to a paragraph or two that an engineer could quickly visualize. As for the underlying biology, I am not a scientist, so I am certainly not trying to refute or replace his explanation. The article ends with an analogy that I particularly enjoyed: DNA is not a computer programming language, but there are some interesting similarities. We can loosely imagine a cell as a computer running its own system. Each cell contains essentially the same DNA, but different cells use different portions of that information. Wh...

 I wrote these old blog entries a while back.  Now rethinking them. https://www.myberlinaustin-atl.com/2008/07/simple-lisp-implementation-in-java-ode.html https://www.myberlinaustin-atl.com/2008/07/neophyte-schemelisp-interpreter-in.html Lisp can look strange if you come from Java, C, C++, or another procedural language. There are parentheses everywhere, functions seem to be mixed with data, and even a simple expression like (+ 1 2 3 4) does not look like the code most programmers are used to reading. Instead of starting with Lisp theory, I think it is easier to start with something every computer science student has probably implemented at some point: a linked list. A linked list has a node containing some data and a reference to the next node. You can walk through the nodes, print them, add their values, or perform some other operation on each element. Keep that basic data structure in mind because it gets us surprisingly close to understanding Lisp. Consider the Lisp expre...

2005 — Numenta Jeff Hawkins co-founds Numenta around the idea of reverse-engineering the neocortex. This eventually produces HTM and the Thousand Brains theory. 2010 — DeepMind DeepMind is founded by Demis Hassabis, Shane Legg and Mustafa Suleyman. Its path is heavily influenced by reinforcement learning, neuroscience and learning through environments/games. 2012 — AlexNet / deep-learning explosion 2013–15 — DQN / Atari DeepMind demonstrates deep reinforcement learning directly from game pixels. 2015 — OpenAI 2015–16 — AlphaGo DeepMind combines neural networks, search and reinforcement learning; AlphaGo defeats Lee Sedol in 2016. 2017 — Transformer Google researchers introduce the architecture that becomes the foundation of modern LLMs. Google DeepMind's history specifically credits the Google Brain team with the Transformer work. 2018 — GPT-1 2020 — GPT-3 / AlphaFold 2 era 2021 — Anthropic 2021 — A Thousand Brains Hawkins publishes the book laying out the Thousand ...

Random Picture of Toy AI Project There has been a lot of discussion lately about artificial intelligence ending humanity. Maybe it happens in one year. Maybe it happens in ten. Depending on whom you ask, there is some percentage chance that AI eventually wipes us out. I am not going to lean completely one way or the other. I am an engineer, not an AI safety researcher, but I do have some thoughts about the logic behind that proposition. A recent CNN article gave one of the more direct versions of the argument: https://www.cnn.com/2026/09/17/tech/how-will-ai-exterminate-humanity-cec One thing I continually hear, even when reporters interview leaders of major AI companies, is some version of this question: "Could AI destroy humanity?" And sometimes the answer is essentially: yes, it could, possibly within some number of years. What I don't hear discussed nearly enough is how . That seems like an important part of the argument. If someone says there is a meaningful probabili...

 One of the interesting things you can do with the new AI LLM technology is rebuild AI with AI.  Here is a project to take some of the python based technology and convert into Java, a small toy LLM in Java and Scala wrapper.  It is not practical due to performance issues but you can see the use of basic procedural and OOP concepts for the LLM.  Used AI for the documentation and generated code. https://github.com/berlinbrown/vanus-play-ai-java An educational decoder-only transformer implemented with Java numerical kernels, Java autograd, and a small Scala command-line trainer. It uses no PyTorch, pretrained weights, native BLAS, or external tokenizer. Architecturally, yes: causal self-attention, RoPE, RMSNorm, SwiGLU, backprop through an autograd tape, AdamW, and autoregressive next-byte prediction via cross-entropy are the same pieces production transformers use (see Architecture below). Nothing special-cases words or does string matching; every response comes out of...

 This is a repost from an earlier blog post.  More scientific sounding. Abstract Modern software systems present computation through increasingly high levels of abstraction. An expression such as 4 * 4 appears to the programmer as a primitive operation, while its physical realization ultimately depends on a hierarchy of compiler transformations, instruction execution, digital logic, and state transitions. Cellular automata provide an alternative model in which global computation arises explicitly from repeated applications of simple local rules. This paper examines cellular automata from the perspective of software computation rather than exclusively as mathematical dynamical systems. Beginning with Boolean state transitions and Conway's Game of Life, we develop the connection between elementary cellular automata and conventional digital computation. We then examine a multi-state cellular automaton capable of representing the computation of . Rather than invoking an explicit m...

https://openai.com/index/an-alien-mind/ The article from "Jakub Pachocki, Chief Scientist at OpenAI"

 I thought this was interesting, repost. Here is an example program that shows how to implement kinematic equations for projectile motion using Java and the Swing 2D graphics libraries. Equation used to plot projectile path along the X and Y axis The Java code for these equations are simple, here is the current implementation of the doSimulation routine. The routine calculates the X and Y positions of the projectile over time.  The project only consists of two classes, the class for rendering the simulation and initializing the application.  The other class contains logic for calculating the X and Y positions. Java code, DoSimulation routine, see s.i and s.k for the X/Y positions Java 2D Cannon Physics Simulation Java Source https://jvmnotebook.googlecode.com/svn/trunk/blog/java/SimpleCannonPhysicsJava So today we see man a highly evolved creature who not only acts but thinks and feels. All these thoughts, feelings and emotions are interrelated. The body and the mind of m...

 Made updates to :  https://www.myberlinaustin.com/ Berlin from Austin is a professional software engineer with more than fifteen years of professional application development experience and a proven track record of producing quality distributed software. My non-professional programming career began in Austin, Texas when I was 10 years old with the TRS-80 and early BASIC. And more: https://www.myberlinaustin.com/new-home.html