Scott C. Lemon

I’m currently running multiple AI and IoT R&D projects in parallel and one of them relates to this post on cybersecurity. A friend and I were talking about experiments with AI Coding, and I haven’t been using my $20/mo Claude subscription since I’ve been focused on my Muse Gadgets projects on my Raspberry Pi. (More on that later!) Our conversation came up at one of our monthly Utah CTO Breakfast meetings. It was to use a coding agent to “take over” an old stagnant Open Source project, modernize and maintain it. On Monday at lunch time I started on this experiment. I won’t name the project yet, but it is a piece of software I still use, written in PHP in 2010, maintained until 2017, abandoned and archived. There are maybe 150 forks (including forks of forks) with people cobbling various fixes and minor enhancements. But it’s still stranded back at PHP v5.6 (!!) and with the recent security issues I realized I need to do something. I started a new project with Claude Code, forked the fork that I am running, and asked Claude to start by simply reviewing the codebase and the condition of the code. I explained that I wanted to eventually create a plan to fully modernize the project, add a slew of features, and also create a hosted SaaS version. It took maybe 15 minutes to think through and write the paragraph prompt. Within 10-15 minutes Claude came back and basically said: “Hey, I’m working on what you asked me to do, but at the same time I’ve already found 17 critical security issues with the code. I can immediately fix 12 of them so you can deploy them in your site. The other 5 are major rewrites of code and should just be addressed in the rewrite of the project. Want me to do that?” Amazing! And yes … of course. oh, and it also drafted emails for me to send to the fork owner to notify them also. This is how easy it is to have an expert AI review your code and find all kinds of things … including huge security holes in code that is almost 17 years old! —- A thousand fewer paths to root | OpenAI ow.ly/UvFl5104m4T

This is mind blowing. Weight-Space Programming. This is a massive advancement in understanding the internals of models, and not by us directly but through AI. If you compare this to the human brain, our brain contains a massive network of neurons that are interconnected and signals flow through those connections with varied signal strength. Some neurons might connect to thousands of others. Those are the same concepts of “weights” - the signal strength - and how models are interconnected “nodes” like neurons. (Yes, this is extremely simplified!) What they are finding is that AI can now examine the internals of a model to locate and discover “regions” of the model that are responsible for different aspects of output! This in itself is astonishing! But additionally, AI can now be used to make adjustments in the weights within the model to “reprogram” the model, or to shift its output! AI can understand the weights, and make changes at the weight level to improve AI models! This is like a level of brain surgery we have absolutely no concept of! This is making changes to the signal strengths of neurons in the brain to improve someone’s knowledge, behavior, or thinking skills. Again … completely next level. —- Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights ow.ly/27EG51046fz

I’ve been quite the last week or two as I’ve been deep in some Agentic AI R&D projects, and formulating some thoughts about our future trajectory. A few ways back I presented to a university class about the evolution of technology and where we arrived at just four years ago. We are all experiencing the Law or Accelerating Returns as described by Ray Kurzweil, and weekly news demonstrates that the rate of technological advancements is, itself, accelerating. In my presentation to the students I spoke about our advancing to “Intent-based Programming” which - IMHO - is where we are today. We no longer need to understand programming languages - AI knows far more and has been trained on more examples. This is just another virtual substrate transition, like moving up the ladder from machine code, to assembly code, to compiled code, to advanced frameworks, and now to Intent. Programming has now become the ability to carefully and precisely “describe” your intent - through requirements, user stories, use cases, security requirements, test cases, and more. If you know how to accurately describe the solution you want to build, AI can build it for you. Working with Muse on my Raspberry Pi has me realize that we are approaching a new horizon of “Intent-based Creation.” Of anything. Not just software. We are heading towards the ability to create anything in the physical world simply be describing it. The signs are all around us as we continue to connect physical “things” (The Internet of Things) to networks where AI Agents have access. A few recent examples might seem small or trivial, but a friend told me about his Grok Bot now designing 3D objects and directly controlling his 3D printer to print them. I have friends working on physical hardware design and development that can take ideas from concept to schematic to PCB to finished assembly - all via AI. I’ve now started to use Muse and my Raspberry Pi to perform sophisticated SIM/Cellular Module connectivity testing that is incredible. We are rapidly entering the age of “Intent-based creation” and it goes far beyond software. The skills to learn revolve around the ability to learn the lexicon of what you intend to create, and how to describe it accurately and in detail.

This is a really clever idea, and demonstrates how small hard is getting. I’d want to hear from people who actually used this - I don’t wear a watch - but this thin (2.8mm) fits to the back of your existing wristwatch to turn it into a health tracker. I’m curious how that additional thickness (thinness?) feels to people who wear watches, but this solves the problem for those who want the health tracking with the specific watch they choose to wear. —- Azume One - A 2.8mm sensor that turns any regular wristwatch into a health tracker (Crowdfunding) - CNX Software ow.ly/Mynr5102r9M

Security against AI driver cyberattacks has become a critical area now. Your own cybersecurity teams will need to be using AI to defend against AI. If your teams are not using AI already, they need to be. Just like coding, cybersecurity is shifting to become an intent-based “programming” task for AI Agents. —- VAPOR PANDA Uses AI-Driven ARTEX to Grow Global Espionage Campaign ow.ly/Iipn5102qvL

This is what occurs when security is an afterthought in #IoT, and when firmware updates are not possible or not performed. Security in #IoT does not have to be over the top, locked down systems. There are simple design considerations and features that would prevent these types of hacks. —- Pentagon suspects hackers behind freezer failures at US base stores ow.ly/gU6f5102304

Yet another measurement and impact of agentic software development. There was a recent article about Anthropic building its own source control solution as Git itself was slowing progress. Now even GitHub is having to grow to handle the massive influx of software that is being created by AI Agents. Software is now abundant! —- GitHub saw 7.4B commits in September, adding 4M CPU cores and 160 PB of storage to keep up - The New Stack ow.ly/M9If51022ow

Once again, model orchestration and routing becomes a critical feature of current systems. The decomposition of prompts - of Intent - and then allowing specialized and more cost effective models handle aspects where they can. —- GitHub debuts Copilot Fusion, a new model selection & orchestration system that beats Claude Code for half the API cost - The New Stack ow.ly/kR5Y51022i7

I recently presented to a masters degree computer science class in the Philippines on this subject. What will their careers look like? In my opinion, code is now cheap and abundant, and the focus has to move up the ladder - programming by “Intent” and then Verification of the results. If they are good, and accept the new reality they will never write code, and might never look at code! —- Hunger beats tenure: Advice for junior engineers ow.ly/MQRC51022aK

An artifact of content and training? An aspect that I never thought of. Using AI in a non-English language can impact your costs of using AI! —- The Tokenizer Tax: Why non-english AI users pay more for tokens - Rest of World ow.ly/oehH5102281

Yes, software has become abundant, but hardware is now coming quickly behind it. When I was designing and building #IoT hardware I learned a huge amount about the steps and process required. Even when I had learned many of the easier tedious steps, I hired experts to review and fine tune my schematics and PCB layouts. All of this is now rapidly moving to AI tasks that they will learn, evolve, and perfect. How long until hardware becomes abundant also? Not long! Hardware engineers need to learn Intent-based design and verification. —- Circuit Mind Raises $18.5M Series A ow.ly/i73L51022Br

This is truly impressive! They have created a full pipeline for video-to-sim-to-real for “programming” of robots. This demonstration taught a robot to perfect the playing of cornhole using an AI pipeline that analyzed a video of a human, created the physics simulation of the throwing, translated that into the robots joints and velocities, and wrote the robotics code for the robot to play. And of course it nailed the hole, seven times in a row! Accelerating the speed of watching to doing! —- Gemini 4 Argon Writes Physics Simulation Code From Human Video, Teaches Apollo Humanoid Cornhole ow.ly/jIjX5101Ggi

Insect cyborgs for military and civilian operations!! Yes, they are making remote control roaches! or some sort of large insects that can carry webcams and microphones, and that can be “steered” in different directions. And of course they are using AI to assist in the development and operation of these creatures. We live in the future! —- SWARM Biotactics ow.ly/vpHF5101joz

Culling the workforce? This is more than a quarter of Microsoft’s employees being offered the opportunity to leave under buyout conditions … before the layoffs begin. And this is not just due to the data center spending … it is the recognition that AI will be replacing many jobs by increasing the efficiency of the employees who embrace AI! —- Microsoft offers voluntary buyouts as AI leads to efficiency push ow.ly/KbFT5101iX4

The more that I interact with AI, and still experience a pretty lame Alexa, I keep wondering why Amazon doesn’t allow me to connect all of my AI subscriptions and access them through the Echo. Or do they and I just don’t know?

More serious science being solved by AI. And it’s still 2026. just over four years into the AI era! —- Working with academic communities to build scientific understanding | OpenAI ow.ly/4Fkg5101iUG

I’m very interested in upgrading my iPhone - finally - and the Duo has caught my eye. But I’m in no hurry, and I avoid early v1.0 products. This year down confirms my suspicions that Apple has a lot of work ahead of them to shore up the design of the Duo. I have no doubts they will get there, but I’ll see where they are post-CES and some time in late Q1 or Q2 of 2026! —- iPhone Duo Teardown: The Most Repairable Foldable Has a Huge Screen Problem - iFixit ow.ly/LUZP5100WIf

A friend of mine, Jim Chappell, forwarded this article and it’s a good articulate read in where we are with AI. As I responded to him, it’s a good report on where we are and the choices in front of us. And a few more items: - this is exactly what Ray Kurzweil forecast almost 30 years ago! - this is why people have been discussing “alignment” for so many years. For decades! - this is inevitable, although the specific timing is not know. Any attempts to “prevent” RSI will merely slow it down. We are in for a wild ride! —- What I saw inside Anthropic - by Kelsey Piper ow.ly/EoR85100WfN

HaLow is one of those standards that has simply stalled due to the short-sighted product marketing of the main chip vendors. I’m sure they have some some, and this is yet another product from Lilygo that is using a chipset from one of the Chinese vendors, but true interoperable IP based developer kits at low costs have been very hard to find. And there has not been enough catering to the popular developer ecosystems like Arduino and Micropython. I gave up on using HaLow quite a while ago as trying to get low cost solutions working was simply not supported by the chip vendors. Their developer efforts are lacking. I found one dev board vendor that is heading down the right path, and I bought some of them to explore at some point, but they have been on my backlog of projects to work on. —- LILYGO T-Halow P4 - An M.2 2280 board with ESP32-P4, WiFi HaLow, and ESP32-C6 - CNX Software ow.ly/omsO5100KC4

No pricing yet, but yet another humanoid robot for researchers and developers wanting to experiment with software development and applications. An impressive 1.35 meters tall, the video appears to show a variety of possible mobility features. I’ll be curious to see what kind of a price point they have got this to. —- RoboParty RP1 (ROBOTO 01) - A full-stack open-source humanoid robot with up to 30 degrees of freedom - CNX Software ow.ly/Q48t5100KBn