AI Week Symposium

Yesterday I attended the AI Week symposium, Exploring Frontiers: AI, Innovation and Emerging Technologies at MSOE. Here are my notes from the talks.

The Flipside of the Human Interface: As Our Tools Evolve to “Think” with Us, We Must Ask, “Who Will Retain the Ability to Think?” by Nataliya Kosmyna, PhD

Her presentation was based partially on her published work here [1]. I have found some links on the web to other similar presentations that she has given, but I was not able to locate the slides. I also asked AI to help and was not able to find anything. There are several podcasts and interviews with Kosmyna, but I didn’t find anything similar to the myths of AI that she described in her talk. My notes may not be entirely accurate, and I did the best I could to capture each myth while she was speaking.

We are building co-dependent relationships with our tools

“When we started thinking for you, it bcame our civilization.” – The Matrix

Are LLMs good for your brain?

• Less brain activity / connectivity when using these tools
• She presented some references, and I wasn’t able to write them down. Hopefully they’re in the references to [1]

The myths of Generative AI usage:

• Myth 1 – Socrates will fix it all
  • questioning the answers, asking more questions
  • our brains prefer shortcuts and don’t want to work this way
• Myth 2 – Cognitive load
  • offload some tasks to AI to focus on more important tasks
  • Boring tasks provide breaks for your brain – you cannot always focus on high cognitive loads
  • Boring work is good for your brain
• Myth 3 – Save time
  • All savings is eaten up by time checking output
  • Respect for others time
• Myth 4 – Enhancing skills
  • research shows reliance on LLMs for cancer screening decreases the skills of the humans making the diagnoses
  • this will compound over time
• Myth 5 – It will revolutionize everything
  • tech billionaires promote use for the masses in education, but shield their own children from it
• Myth 6 – AI literacy is key
  • Ready Player One example
  • many are going full forward with AI use in edcuation without any plan for evaluating the results
• Myth 7 – improve social relationships
  • Technology created the problem of decreasing human interaction and now aims to solve with chat bots
  • Teens are reporting that they are dating chat bots

A better way to leverage AI for humans – technology directly integrated with your brain

• BCI – Brain Computer Interface

When Copilot Becomes the Pilot, How Can We Fly Safely? by Larry Zhiming Xu, PhD

Accumulation of cognitive debt when using AI assistant for a task (interesting that this is the first time I noticed this term cognitive debt, when Kosmyna has it in the title of her paper)

• Eventually leads to cognitive surrender

Xu’s research shows that we rewire our cognition based on our expectations of AI

• Shows correlation (not necessarily causality) between AI use and poor self confidence

AI in Neuroimaging: Challenges and Opportunities by Dr. Max Wintermark, MD

AI in neuroimaging – very good at measuring things in the images such as tumor volume

• This is needed to be tracked over time
• Difficult for humans to make these measurements accurately

Shortage of radiologists in comparison to the number of images to review

Sensitivity vs. Specificity in testing

• Can leverage pre test probability to improve results

Algorithm validation – has not been done with algorithms in use

Held an ASFNR AI Challenge

• Winner used game theory (to prove a point?) – not clinically useful!
• Results published in [2] – I’m actually not 100% sure of this being the publication he was talking about, but this is what I found searching after the symposium.

Technology is evolving faster than validation

FDA regulation considerations

• Locking of SW version – AI algorithms change and evolve quickly

Brain and Prostate Cancer Radio-Pathomics at the Medical College of Wisconsin by Peter LaViolette, PhD

Radio-pathomics – merger of imaging with biopsy

MRI -> biopsy / pathology -> molecular profiling

The research takes the above and uses results from real pathology studies to train AI models that look at MRI imaging alone

• Use data in the images combined with analysis to infer what the biopsy would reveal

Savannah Duenweg from MSOE developed a machine learning model trained with data from 236 prostate cancer patients that included whole-mount prostate sections sliced to match MRI. The trained model using histological density from MRI scans, achieved 80% accuracy [3].

Glioblastoma (GMB)

• Scans – T1, T1+C, FLAIR, ADC

Research takes dead patients’ brains and slices them to match up with MRI scans

Data is used to train AI models that can make predictions for patients based off MRIs – Radio-pathomic maps of tumor probabilities

AI-Powered Care: Optimizing Quality and Outcomes in Assisted Living and Memory Care by Liz Jensen, PhD, MSN, RN-BC

Aging of society means in the future we will have more care needs and less workers available to provide care

Problems occurring now:

• medication errors
• ER visits
• falls
• turnover

Solutions in AI – support human in the loop, not replace humans

Very challenging to balance deployment of AI with evidence that it is effective and providing value

MSOE AI Club: Student Research and Industry Engagement

MSOE AI club is responsible for 11 AI publications

Has won many times at MICS competition

Insights from Working at the Center of AI by Luke Leonhard, chief of staff, Google DeepMind

Friend of MSOE – will arrange tours of Google if you reach out on linked-in

GEmini mode sfor audio

Know your super power and your kryptonite

Biases of AI models

• Interaction bias
• Latent bias – the future isn’t always based on the past
• Selection bias

Notebook LM will change your life

AI Studio – create apps

AI, Deepfake and the Evolving Cyber Threat Landscape by Aaron Pritz, CEO, principal consultant and co-founder, Reveal Risk

AI agents that have been given access are a possible attack vector

3 seconds of your voice = 85% deep fake accuracy

Verify everything!

presenter did a live deep fake of an audience member using only the persons linked in picture

AI Ambition: Tech, Talent and Transformation in Manufacturing by Tim Dickson, chief digital and information officer, Regal Rexnord

Manufacturing AI uses

• predictive maintenance
• predict demand

Sales agent – connects to multiple data sources and presents information to sales team instead of having to hunt for information

Sales agent sounds like the promise of MBSE – maybe an MBSE agent could be created to look at multiple project data sources and present a unified model?

List of books that he recommends:

• Big Bet Leadership: Your Transformation Playbook for Winning in the Hyper-Digital Era by John Rossman and Kevin McCaffrey
• The AI Playbook: Mastering the Rare Art of Machine Learning Deployment by Eric Siegel
• Agentic Artificial Intelligence: Harnessing AI Agents to Reinvent Business, Work and Life by Pascal Bornet
• The AI-Savvy Leader: Nine Ways to Take Back Control and Make AI Work by David De Cremer
• Show AI – Don’t tell it: Build Buy-In with Visual Storytelling by Lisa Palmer
• AI for the Authentic Leader: How to Communicate More Effectively Without Losing your Humanity by Allison Shapira

Powering Growth by Richard Stasik, vice president of regulatory affairs, WEC Business Services

WE energies has 8150 MW of generating capacity serving 4.7 M customers

47000 miles of gas lines

72400 miles of electric distribution

MISO – mid continent independent system operator

• coordinates energy consumption and generation

Expecting demand growth of 3.9 GW from 2026-2030

Data centers want to be grid connected, don’t want to operate power generation

A Little Closer to AGI Every Day at MSOE by Jeremy Kedziora, PhD

AI – humans are decision makers so AI is the automated decision making

Types of AI:

• LLMs
• Reinforcement learning agents (ie self driving cars)
• Classical machine learning

decision making represented mathematically:

function (policy): pi(a|s)

• s – observation of state of world
• a – action

AGI – Artificial General Intelligence

• learn / represent knowledge
• reason
• plan
• communicate
• sense
• manipulate

Tests for AGI:

• Turing test
• College test
• Employment test
• Task test – makes coffee, builds IKEA furniture

Live Bench – benchmarks for different LLMs

4 papers that argue that LLMs do not understand cause

• no understanding of cause mean they cannot predict the consequences of their actions

Settlers of Catan bargaining agent

Do we understand how these LLMs work?

• We understand the underlying matrix multiplication and exponentiation
• We do not understand how the input parameters drive the responses

[1] Kosmyna, Nataliya, Eugue Hauptmann, Ye Tong, Yuan, et. al. 2025. “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” ArXiv. [Internet WWW PDF] https://arxiv.org/abs/2506.08872 [Accessed 26-Apr-2026].

[2] Jiang, Bin, Burak B. Ozkara, Guangming Zhu, et. al. Septemeber 2024. “Assessing the Performance of Artificial Intelligence Models: Insights from the American Society of Functional Neuroradiology Artificial Intelligence Competition.” AJNR Am J Neuroradiol. [Internet WWW PDF] https://pmc.ncbi.nlm.nih.gov/articles/PMC11392353/ [Accessed 26-Apr-2026].

[3] Duenweg, Savannah, Sumuel R. Bobholz, Allison K Lowman, and Aleksandra Winiarz, et. al. January 2026. “Radio-pathomicmaps of histo-morphometric features trained with whole mount prostate histology distinguish prostate cancer on MP-MRI.” medRxiv: the preprint server for health sciences. [Internet WWW PDF] https://www.researchgate.net/publication/399758688_Radio-pathomic_maps_of_histo-morphometric_features_trained_with_whole_mount_prostate_histology_distinguish_prostate_cancer_on_MP-MRI [Accessed 26-Apr-2026].

NPUs and Computer Architecture Cycles

I came across an article in IEEE Spectrum talking about a change in laptop architectures to support large language models (LLMs) [1]. Today most large language models are trained and queried on large centralized servers with large numbers of GPUs in data centers. Desktop and laptop PCs lack the computing power to run these AI applications. The article describes architecture changes underway in laptops to accommodate at least the query side of using LLMs. The article describes a new type of processing unit called a neural-processing unit (NPU), which I hadn’t heard of before. These chips are specially designed for matrix multiplication calculations performed in parallel that the AI models depend on. NPUs are a little more application specific than GPUs which are also designed to handle 3D graphics type tasks, so they are more power efficient than GPUs. As the NPU gets integrated into PCs alongside CPUs and GPUs, it will also necessitate a change in memory architecture. Since GPUs came along, the memory architecture has been separated. CPU and GPU each have their own divided memory. For effective use of NPUs and CPUs LLMs, a unified memory architecture will be required. The industry is moving toward SOCs that have CPU, NPU, and GPU integrated on a single chip with a high bandwidth memory interface to a single shared pool of memory. AMD has already announced a product called Ryzen AI Max that will include Ryzen CPU cores, Radeon GPU cores, and NPU rated at 50 TOPS (trillion operations per second) on a single piece of silicon with a unified memory architecture of up to 128 GB. Intel and NVidia are teaming up for something similar using Intel CPU, NVidia GPU, and Intel NPU, though details haven’t been released yet.

One of the things that I have noticed about computing technology is there is a cycle between rushing toward computer architectures that are centralized and computer architectures that are decentralized. In writing this article, I asked AI for some references to this phenomena and it pointed me to an excellent paper deriving the cycles of centralization from 1945 to 1997 using computer sales data [2]. The paper identified the following cycles and major factors of influence:

• Cycle 1
  • (1945 – 1978) – Centralization – Mainframe environment
  • (1979 – 1984) – Decentralization – Distributed data processing using lower cost of mini computers
• Cycle 2
  • (1985 – 1989) – Centralization – Relational databases and lack of standardized networking
  • (1990 – 1997) – Decentralization – Client-server architectures with standard network protocols
• Cycle 3
  • (1997 – ?) – Hybrid Centralization – Support economies of scale with mature uses of mainframes

A more recent paper describes similar cycles with some description of where we are right now with AI, near the peak of a centralization cycle [3]. I expect that forces of innovation are going to be moving us toward another cycle of decentralization with a big change in PC hardware architecture. I wrote a previous article about the massive need for additional power generation and storage that the current trajectory of centralization at AI data centers is projecting by 2030. Maybe the decentralization cycle will alleviate some of that?

It will be exciting to get my hands on one of these new AI laptops and try out some locally hosted LLMs.

I was aware of another type of cycle in computing technology that I thought I had written about previously on this blog called Makimoto’s wave. However, when I searched my blog, I didn’t find it. I wanted to capture it briefly in this article with a reference to it. Basically Makimoto’s wave is an observation of the cycling between periods of hardware that is programmable and periods of hardware that is customized at the silicon level. When I first came out of college, the wave was entering the period of field programmable chips (FPGAs and CPLDs) and my early career as a digital engineer was dominated by these devices. Then the transition to the next wave was SOCs and SIPs and cheap microprocessors and microcontrollers which dominated. I’m not really sure what cycle we’re in right now, though it is about to end according to the reference.

[1] Smith, Matthew S. December 2025. “Your Lapop Isn’t Ready for LLMs – Yet.” IEEE Spectrum. Vol 62, Iss 12.

[2] Peak, Daniel A. and Mohammad H. Azadmanesh. January 15, 1997. “Centralization/decentralization cycles in computing: Market evidence.” The Information Manager. 31 (1997): 303-317.

[3] Sunyaev, A., Avital, M., & Lacity, M. C. 2026. “From ideology to design: Toward purposeful decentralization of information systems governance.” Journal of Information Technology, 41(1), 2-14.

Good Qualities for a Scientist

I just finished listening to the audio book for Project Hail Mary Project: a Novel by Andy Weir. It was excellently written with plenty of relevant and accurate scientific and engineering facts about space travel. I commend the author on the amount of research that must have been put into writing this book. Only a few things I thought were not completely correct. Anyhow, it was a great story about a scientist and an engineer being the heroes, and got me thinking about characteristics of such people.

My daughter is planning for the next steps of her doctoral work as a cancer biology scientist. I was talking with her about qualities of a good scientist, and these are what I came up with. I thought it would be worth it to write these down. I chose a single word to represent each characteristic, but that is never sufficient to capture the full meaning of the characteristic, so I include an explanation also.

1. Integrity – a scientist needs to always collect and analyze data in a thorough and honest way. Craftsmanship is a key part of doing good science experiments, and also knowing the causes of uncertainty and their effect on the results. Use of statistics is often important, and conclusions must be presented with transparency about the assumptions made. Good scientists don’t rush to conclusions, and make sure that the results are repeatable.
2. Mentoring – a scientist needs to be willing to spend time getting into the details with those working under them. Good mentorship is not doing the work for someone, and it is giving them challenging enough tasks to own. When roadblocks are encountered spending the time to understand the problems, and guide them to finding solutions on their own.
3. Curiosity, Adaptability, Humility – This one combines three words to capture the essence. A scientist needs to not have preconceived notions and look for data to support them. Instead a scientist is curious, develops theories, and then tests those theories, making adaptations that follow the data. A good scientist realizes that the universe is complex, and they aren’t going to fully understand it. The more you learn, the more you realize that you don’t know, and this should make one feel humble.
4. Planning – Making progress on science requires a structured and organized approach. There is a need to create detailed plans and budgets, to convince others that you are disciplined and get funding for your research, and ensure that you have the right resources. The real purpose of detailed planning though is that as reality deviates from the plan, you know quickly and can adjust the plan, and clearly communicate where the science is at. Sometimes planning also requires creativity to find alternatively simpler or cheaper ways to achieve similar results.
5. Teamwork – modern science work requires a team of people. Rarely can a single scientist working alone in a lab make important discoveries. Equipment is complicated and requires expertise, the amount of data is massive, and we’re increasingly encountering complex multi-disciplinary problems that require many different experts. A good scientist needs to be able to recognize what skills are needed, and then identify people who have those skills. The weakest link can ruin the results of any team. Another aspect of this is providing feedback to people on the team so they know where they are at, what needs to improve, and what they need to give more of to the team.

Transpiration Cooling

I came across a concept this past week that I hadn’t encountered before, transpiration cooling. In very high velocity convective heat transfer situations, the wall temperature of the surface exposed to the convective fluid can become very large. The two applications that readily come to mind are turbine blades in high speed engines, and spacecraft surface during reentry. The concept is that the surface can be made porous and a fluid (either the same as the convective fluid or a different fluid) is forced through the surface causing additional energy to be carried away from the region very close to the surface, above what would normally be conducted into the boundary layer during convection alone [1]. There’s some diagrams showing this concept at the following links:

• https://www.thermopedia.com/content/1203/
• https://www.mdpi.com/1996-1073/18/13/3282 (full text of the paper titled A Review of Recent Research on Flow and Heat Transfer Analysis in Additively Manufactured Transpiration Cooling for Gas Turbines by Kirttayoth Yeranee)
• https://www.thoughtco.com/forest-transpiration-water-cycle-4117845 (transpiration occurs naturally in plants, who release water through porous surfaces to prevent scorching in the sun and provide cooling)
• https://publications.sto.nato.int/publications/STO%20Educational%20Notes/STO-EN-AVT-261/EN-AVT-261-05.pdf (full text of paper titled Transpiration-Cooling with Porous Ceramic Composites in Hypersonic Flow by Hannah Bohrk)

As part of this investigation, I spent some time reviewing convection heat transfer using example problem 5-8 in [1]. Here is an Engineering Paper XYZ sheet that I spent some time on.

[1] Holman, J.P. 1986. Heat Transfer 6th Edition. McGraw-Hill Book Company: New York.

Snowstorm Elsa vs the EGO Snowblower

Overall I have remained very happy with my EGO snowblower that I described in this post last year. However, it definitely met its match against snowstorm Elsa last weekend. It was one of the largest in northeast WI history, and definitely the most snow I’ve ever had. Official estimates were about 27″ for Appleton. My unofficial estimate is based on 5 plowings of my driveway that I did, and estimating the size of each one. Here’s how the snowblower performed.

Sunday morning – about 5″ of extremely wet and heavy snow. The snowblower struggled mightily and kept shutting off. It’s strategy of aerosolizing the snow and blowing some out the front and some out the shoot clearly was ineffective. Also, the power was insufficient. I was able to clear the end of the driveway though with great effort and very slow, and cutting relief paths through parallel to driveway to allow perpendicular passes. The battery died when I was about halfway done. It was enough to allow one car to get out though, so we went to church with the work half done.

Sunday early afternoon – about 4″ of dryer snow, but half of the driveway now had 9″ with the bottom snow very wet. I got through most of the 4″ parts of the driveway and was able to get a few passes through the 9″ part. The battery died, and I had to hand shovel 3/4″ of the 9″ part. It didn’t bog down as much in the 4″ sections this time.

Sunday evening – about 3″ of dry snow. The end of the driveway had some very thick chunks of ice from the snow plow. After getting everything cleared earlier the snowblower made quick work of this, though it still struggled at the end of the driveway. The battery lasted to finish the whole driveway, but was on red at the end.

Monday morning – about 8″ of dry snow. The power and electric motor speed and torque control really shined here. It was able to power through and throw the snow onto the already very high snow piles. It didn’t quite make it through all the snow before the battery died, and I had to hand shovel a little bit, but not too bad.

Monday evening – about 4″ of dry snow. This was no problem and I finished the whole drive without using up the battery.

The snowblower is great for most snow storms, and I love it, but by Sunday afternoon I was completely exhausted. I question whether it would have actually been less work with the wet heavy snow to just manually shovel it. The EGO snowblower is not up to the task, even for modest amounts of this type of snow.

Space Based Solar Energy for Earth

I have been reading about the capture of energy in space using solar panels which is then beamed back to earth either using microwaves or lasers for many years. After spending some time in the space industry working on spacecraft design, the scale of it doesn’t really make sense to me. The systems I worked on struggled to generate 20 kW or electrical power and dissipate the resulting heat in low earth orbit (LEO). For space based power to be useful, it would MW or GW of power. Still, the technology sounds very cool, and it might be a possible solution for sources of renewable energy.

I remember first reading about this concept a bit more than 10 years ago in relation to Japan’s research and development in this area. Here is a web page that describes some of the history of the idea for space based solar power.

There was an article in the June 2024 issue of IEEE Spectrum by Henri Barde, retired head of the European Space Agency (ESA), that raised skepticism about the viability of this approach to power generation [1]. The article describes 4 proposed architectures that are under consideration along with Barde’s doubts.

SPS-ALPHA Mark-III – proposed by John C. Mankins in 2022 (space development foundation presentation)

• 6 km reflector
• 1.8 km solar PV panels
• 7 km backbone structure
• modular, design using robotic construction
• Microwave transmission of 6 GW to earth (2.45 GHz)
• 2 GW power delivered
• 6 km diameter ground station

CassieopeiA – A project of Space Solar

• 400 – 1700 m diameter solar arrays
• Medium Earth (MEO), Highly Elliptical (HEO), and Geosynchronous (GSO) orbits
• Microwave transmission to earth (5.8 GHz)
• 1-4 km diameter ground antenna

SPS (Solar power satellites) – A project of the EU named Solaris – led by Thales Alenia Space

• There has been little detail that I can find about the architecture proposal
• A report was supposed to be delivered in 2025
• AI query reported that nothing has been released yet and the program has been paused
• AI query hallucinated that some graduate research studies have presented some of the details unofficially, however the links it provides to those documents only mention that Thales is working on it, and do not really contain any details about the Thales architecture
• [1] claims that SPS consists of 8000 flexible solar arrays, each 10×80 m (overall size would be two wings that would be 40 km x 320 km) – that seems ridiculous!
• Barde mentions a study by Arthur D Little. I can find mention of this with a link to the study, but the link doesn’t work.

MR-SPS – A project of China Academy of Space Technology – paper presenting the architecture in [2]

• 50 solar sub panels, each 200 m x 600 m
• Geosynchronous orbit
• Solar output power of 2.4 GW
• Microwave transmission to earth (5.8 GHz)
• 1 GW of power delivered

The challenges of the scale and size of these deployments are obvious. Some other challenges that Barde points out are:

• Extreme costs for generating power this way – NASA estimates are 12 to 80 times the cost of power generated on earth with existing technologies
• Gigawatt-scale microwave-beam transmission is an unproven low TRL technology with very high technical difficulty
• Robotic construction of kilometer-long, high stiffness structures in space is an unproven low TRL technology with very high technical difficulty
• No way to dispose of such large structures in an increasingly crowded orbital space above earth
• Diffraction will cause the microwave beam to spread as it travels – assumptions about this in the plans are too optimistic
  • Optimistic size of receiver on earth would be 34 square km
• Huge structures will require very high launch costs to GEO even with those costs coming down
  • Current estimates about mass in these plans are too optimistic and it will be even worse
• The size of these proposed solutions is on 1 to 2 GW. These don’t even replace a single large coal or nuclear power plant
• Conversion losses stack up, and to get 1 GW to the grid at AC would require 9 GW of power in space
• This 8 GW of conversion loss will need to be rejected or temperatures will be unacceptably high (both on earth and in space)
  • Large swings in temperature will cause thermal expansion and contraction problems in these massive structures
  • The radiators needed to dissipate this kind of power would be even bigger than the solar arrays and don’t seem to be included in anyone’s plans
• Forces and torques exerted by altitude control and pointing thrusters plus the pressure of solar radiation on the massive sail-like arrays will bend and twist the structures – this needs to be accounted for in strength and reliability
• Strikes from micrometeroid orbital debris (MMOD) need to be mitigated
• Shaping and aiming the microwave beam
  • The antenna is not a simple dish whose geometry maintains pointing
  • It is a phased array with millions of little antennas that must be synchronized
  • Synchronization on ps scale across km of distance is needed – there is no known solution to this currently

Another tidbit in Barde’s article is that a private sector startup named Solaren signed a power-purchase agreement with California in 2016 to deliver 200 MW to the grid. Apparently that agreement is no longer in effect, and Solaren does not seem to be very active. Their web site is still up, and they did have a $2.5 M series D capital raise in 2023. AI says that they have shifted to long term R&D and are no longer developing a prototype for rapid deployment. It also mentioned that maybe they have pivoted to supplying power to lunar or mars surface.

Last year I read about a company named Reflect Orbital that is planning on putting giant arrays of pointable mirrors in sun synchronous polar orbit at 600 km (a bit higher than ISS in low earth orbit). The idea is to be able to provide sunlight to spots on earth at night on demand. It could be used for energy generation to illuminate solar farms at night, and there are also many other uses such as extended working hours, illumination of disaster zones and search and rescue missions, boosting growing season for agriculture, replace street lights in cities. The technical details are lacking, and it seems like there are some big technical challenges here, but not nearly as hard as beaming the power to earth in microwaves. They were supposed to launch something in 2025, and I don’t think that happened. I’m not sure what their timeline is now.

Last year, I would have classified this article under space in my blog, but I think now that I switched industries, I’ll classify it under Energy. I’m not optimistic that this is going to be real any time soon, but it is cool to dream.

[1] Barde, Henri. June 2024. “Castles in the Sky? A Skeptic’s Take on Beaming Power to Earth from Space.” IEEE Spectrum. Vol. 61, Iss. 6.

[2] Hou, Xinbin, Zhengai Cheng, Xin Wang, Changjun Liu. 2025. “Modular Multi-rotary Joints SPS Concept-Challenges and Design Considerations.” IEEE Transactions on Aerospace and Electronic Systems. Volume 61, pages 107-120.

AI Chip Cooling Solutions

As I was reviewing end of year statistics for this blog in WordPress, I noticed that this blog now has 490 posts. It is very likely in 2026 that I will reach 500. There are about 8 posts that are in various states of draft. One of those I have been adding to over my entire career and I have decided that it will not be published this year. A couple of others may be dead ends and never be published, and a few should make it to publication this year. Those, plus a flurry of activity on power generation and distribution based on my recent career pivot should make 500 possible by the third quarter. If I reach the 500th post this year, I will use that post to announce a new project that I’ve been working on. Maybe the suspense will improve the readership levels which have been shrinking in the past few years since a key referring post on Pinterest has dried up.

This post summarizes some new concepts in chip cooling technologies that were written about in the November 2025 issue of IEEE Spectrum. Early in my career, as transistor sizes continued to scale downward with Moore’s Law, the phenomenon of Denard scaling held, where the power density remained roughly constant because of reduction in chip power supply voltage. This started to end between 2005 and 2007 at the 65nm and 45nm transistor sizes. Now at the current size levels below 10nm, new solutions need to be considered for managing heat generated in the devices. When I was working on digital systems a combination of conductive heat spreaders, heat sinks, and air flow were used to keep chips cool enough. The following concepts presented in IEEE spectrum are under active technology development to cool the chips of the future.

Diamond blankets as heat spreaders

In the referenced article [1], the concept of using high thermal conductivity diamond material in close contact with the transistors themselves to conduct heat away and spread it out. Diamonds are extremely thermally conductive, greater than 1500 W/mK. For comparison, copper is only about 400 W/mK. And unlike copper, diamond is not electrically conductive, and this makes it perfect to act as a heat spreader inside chips, if it can be manufactured. Researchers at Stanford are working on this and believe that diamond blankets embedded in 3D stacked silicon devices could reduce localized temperatures to 1/10 of their value without the blankets based on simulations. Much work remains to make this manufacturable.

Next-Gen AI Needs Liquid Cooling

With power densities expected to grow toward 100 kW per AI rack, it is recognized that liquid cooling is a necessity. This article [2] describes four options for liquid cooling, some of which sound similar to concepts that I was working on the space industry for cooling spacecraft.

1. Single-phase direct-to-chip cooling – This has a coolant flowing through a cold plate attached to the chip. The coolant absorbs the heat and takes it to a heat exchanger where it can be exhausted or radiated away.
2. Two-phase direct-to-chip cooling – This has a specialized coolant with a low boiling point flowing through a cold plate attached to the chip. The heat from the chip causes the coolant to boil, thus capturing extra heat through latent heat in the phase transformation. The article implies that this is still using active pumping to move the coolant. A further enhancement of this would be the concept of the looping heat pipe in the space industry where the fluid flow is driven by capillary action.
3. Single-phase immersion cooling – This has the chip immersed in a coolant instead of the coolant flowing through a cold plate attached. This allows a more direct transfer of heat into the fluid. The fluids chosen must obviously be non conductive, and oils are used for this. The slow moving oil does not provide sufficient cooling for current chips’ power densities.
4. Two-phase immersion cooling – Similar to single phase immersion, the chips are immersed in fluid that boils, thus using the extra latent heat of the phase transition to transfer more heat. The vapor is then condensed back to liquid on a cooled surface.

How to Cool Chips with Lasers

The most interesting of the concepts is in a article [3] that describes how certain materials are able to absorb heat, and when illuminated at a specific wavelength, fluoresces at a higher energy level. Many materials exhibit fluorescence, however they fluoresce at a lower energy level than excited thus heating up. Materials such as ytterbium exhibit the phenomenon of absorption of low energy photons and emission of high energy photons called anti-Stokes cooling. This works because ions of ytterbium and certain other lanthanides have the right structure of electron orbitals to allow emission of higher energy photons by using phonons (mechanical energy of the orbiting electrons). The basic concept of operations for this concept will be to have an aimable laser pointed at the chip and a camera to detect hot spots as they occur. The laser will illuminate the hot spots in real time, thus causing cooling the hot spot and the emission of high energy photons. The really cool thing about this technique, is that it is possible to recapture the high energy photons using photo electric devices and convert the energy back to electricity. A company, Maxwell Labs, is working on developing this technology.

[1] Chowdhury, Srabanti. November 2025. “Diamond Blankets Will Chill Future Chips. A micrometers-thick integrated layer spreads out the heat.” IEEE Spectrum. Vol. 62, No. 11.

[2] Genkina, Dina. November 2025. “Next-Gen AI Needs Liquid Cooling. Here are four ways to get it done.” IEEE Spectrum. Vol. 62, No. 11.

[3] Balma, Jacob and Alejandro Rodriguez. November 2025. “How to Cool Chips with Lasers. Startup plans to convert heat to light in data-center processors.” IEEE Spectrum. Vol. 62, No. 11.

Battery Powered LED Light Strips Over Coax

The coax cable throughout my house has been dark since 2021 when I “cut the cord” from cable. Even if I switched back to the cable company for internet only, I would not use all of the coax to distribute signals to TVs. I had the idea a few years ago to use the coax to power LED light strips with a battery backup that would turn them on when the power goes out. This is an example of a more general usage of low voltage power distribution in homes which could have many possible applications.

To implement a prototype of this idea I first did a little research to make sure that it would be safe to use the Coax this way. Generally, manufacturers of coax cable do not specify a maximum current carrying capacity since they are used for high frequency signal transmission. Also, since the current and return current are flowing within the same cable assembly, it’s not really simple to calculate based on insulation type and wire gauge. I found this reference, ANSI/SCTE 32 2016 (R2021) Ampacity of Coaxial Telecommunications Cables, which shows a methodology for determining current capacity of Coax and has a table of ratings that includes the two common forms of coax cable used in homes. The ratings for RG-6 is 8 A and for RG-59 is 6 A. RG-6 is usually 18 AWG center conductor which by itself would support TBD, and RG-59 is usually 22 AWG center conductor. From another reference, single wires with these sizes would support:

• 18 AWG – 9.5 A
• 22 AWG – 5 A

I’ve seen other references with varying levels for single wire. Another thing to keep in mind with low voltage and long distance wiring such as in the application I’m talking about, is that there may be a significant percentage of the voltage you’re trying to deliver dropped across the resistance of the wiring, making it unusable. A couple of amps is probably a practical limit for this application with a voltage of 12 V. To conclude, LED lighting at 12 V definitely falls within the safe and practical limitations for using the coax cable within houses.

Creating the First Prototype

The first prototype was to utilize a battery jump start device with a large 12 V lead acid battery, and a coax power distribution board in the basement where all the coax cables converge. Here’s a block diagram.

The basic concept of operations is summarized as follows:

• Car Starter Battery Pack is plugged into 120 Vac wall output so it is always charged
• A wall wart is plugged in to 120 Vac wall output to provide 12 Vdc to open the normally closed relay – this turns off the 12 V to the coax cables and the LED lights are off when power is present
• When power goes out, the NC relay will close and battery power will be distributed to the coax cables and the LED lights are on
• The Coax Power Distribution board connects a single 12 V battery input to 6 coax connectors allowing 6 different coaxial branches in the home to be powered from a single battery.

Additional Ideas and the Second Prototype

For the second prototype, I added a satellite board to replace the coax adapter. In the process, I decided that it would be nice to have a switch to turn on and off each light, and that another desirable use case would be to have a USB output to charge your devices when the power is out. I also added a fuse on the satellite board and added a master fuse and branch fuses on the distribution board. Here is a block diagram of the satellite board.

An interesting thing about this block diagram, is that I used the AI option in Lucid Chart to generate it. It worked very well for a simple diagram like this, and I look forward to using it for more complex diagrams.

In addition to these changes, I also switched the power source from the car starter battery pack which was a lead acid battery to the Anker Solix C300DC which has a lithium battery. It has about the same capacity and is much smaller and portable meaning that I can easily repurpose it for other uses when the power is not out.

Construction and Testing

The circuit boards were designed in EasyEDA using an online schematic editor and PCB CAD tool. It is simple and easy to use for simple circuit boards like this. The tool provides an easy ordering interface to a Chinese PCB fabricator, JLC PCB. The boards are inexpensive and fast. They also have parts sourcing but I was not able to find all the parts I needed. It was a bit of a hassle to get parts with the right footprints that match what is available in EasyEDA. There was an intermediate version between the first and second prototype to fix a PCB artwork error that I do not know how happened. Basically one of the Coax connectors was disconnected from ground. The EasyEDA tool’s DRC checks may be buggy, so be careful that you visually inspect the artwork before ordering with the gerber files that it generates. Here are some pictures of the second prototypes:

I extensively tested the design to ensure fuse sizings were reasonable, temperature rises were within acceptable ranges, the inrush current when the system switches on is reasonable, and the USB 12 to 5V DC-DC converter efficiency was as expected. I also measured the battery capacity of the Anker Solix with a full 6 LED light load. The Anker Solix 300DC advertised capacity is 288 W-hr, however my measurement was only 268.6 W-hr, about 6.8% below the spec. The Anker Solix provides a solid output voltage throughout the entire discharge cycle, so it must have a DC-DC converter on the 12V output from the battery cells. It cuts off abruptly when the Solix says that the remaining capacity is 1%. Most of the testing was pretty basic using DMM and current clamp meter. The Anker Solix provides some good measurement data through the app interface for output power. The only challenging measurement was the inrush. I have an oscilloscope but no current probes. I found a current transformer on a power circuit board that I scavenged from something. I wasn’t sure how many windings, but I made a guess / estimate, and put a 100 ohm resistor across it with the oscilloscope to measure the voltage. I had some difficulty getting the measurement to be repeatable initially, and I discovered that stray voltage stayed on the LED output side of the AC relay. I determined that some RMS voltage with a DC offset was leaking through the AC relay to the contacts. I had to carefully dissipate the voltage in between each power on, while the power was off. Once I put this in the procedure, the inrush waveform was very consistent. Here is a scope capture of the inrush.

Final Thoughts

I have now installed the system in my house with four satellite locations. I’m ready for the next power outage! This was a fun project, and when it started I didn’t realize I would eventually end up in the power generation and distribution industry. From my view now, I have created a specialized UPS with low voltage DC power distribution. There are definitely some enhancements I have in mind, but too many other projects going on right now, and they’ll have to wait.

AI Data Center Power Usage

As I transition into the power generation and distribution industry, I came across an article in the October 2025 issue of IEEE Spectrum describing how much power generative AI queries are taking now and projected by 2030 [1]. The article assumes 25 queries for a person per day at 0.34 W-Hr per query based on Sam Altman’s blog post in June 2025 (https://blog.samaltman.com/the-gentle-singularity). Altman does not substantiate this number, and it is the basis of the ChatGPT estimates in the article. The article assumes 700 M weekly users of ChatGPT and over 2.5 billion queries per day (estimates disclosed by ChatGPT). These estimates are only for ChatGPT which is just a fraction of overall generative AI. A report from Schneider Electric Sustainability Research Institute has slightly different estimate based on a per query power usage of 2.9 W-Hr (about 8.5 times higher than Altman’s figure) [2]. At first it was a little difficult to discern how the author pulled the numbers out of the Schneider Electric report, though with some careful reading, the key is that the report contains a chart that breaks out the electricity usage for generative AI queries from all other AI uses (generative training, non generative queries and training) and this is the data used by Smith. I have pulled this data and the total for all AI usages from the Schneider Electric report into the following tables for examination (the data is based on the sustainable model described in the report, called scenario 1).

Year	2025	2030	% Change	Absolute Change
AI Query Energy Usage (W-hr) – Sam Altman estimate	0.34
AI Query Energy Usage (W-hr) – Schneider Electric estimate	2.9	0.002892	-99.9
ChatGPT queries per person per day	25
ChatGPT queries per day (billions)	2.5
ChatGPT electricity usage per day (W-hr)	8.5
ChatGPT electricity usage, all users, per day (MW-hr)	850
ChatGPT electricity usage per year (GW-hr)	310.25
All AI queries per year (trillions)	5.1	120000	2352841
All AI queries electricity usage per year (TW-hr)	14.79	347.04	2247	332.25
All AI queries, training, traditional and generative per year (TW-hr)	100	620	520	520
Metrics for Comparison
60W equivalent LED light bulb energy usage (W-hr)	8
60W equivalent LED light bulb energy usage per year (kW-hr) (assuming 8 hours use per day)	23.36
Average WI house electricity usage per year (kW-hr)	7742
My house electricity usage per year (kW-hr)	11997
My condo electricity usage per year (kW-hr)	3116
Total US residential electricity usage per year (TW-hr)	1482.873586
Total US electricity usage per year (TW-hr)	3975.381832
Personal electricity usage for AI queries (assume 25 per day at 2.9W per query) (kW-hr)	26.4625
Percentage of electricity usage to AI based on average usage (%)	0.34

Power Plant Yearly Generating Capacity by Type in U.S.	Average (TW-hr)	Biggest (TW-hr)	Number of Average Size Plants Needed to Add 520 TW-hr by 2030	Number of Maximum Size Plants Needed to Add 520 TW-hr by 2030
Wind (WND)	0.332	4.1	1567	127
Natural Gas (NG)	0.55	12.5	946	42
Bituminous Coal (BIT)	0.825	11.5	631	46
Subbituminous Coal (SUB)	1.17	18.4	445	29
Lignite Coal (LIG)	1.47	10.58	354	50
Nuclear (NUC)	8.2	11.7	64	45
Hydroelectric (WAT)	0.413	15.4	1260	34
Solar (SUN)	0.034	1.639	15295	318

The data in the table for electricity usage, residential and total in the US is based on information contained in the following source for 2024: https://www.eia.gov/electricity/data/state/

The data in the table for power plant size comes from the 2024 data for all U.S. power plants that can be found at: https://www.eia.gov/electricity/data/eia923/

I imported this EIA.gov data into excel and created a box and whisker chart for each power plant type (the codes for each type are defined in my table above for those that are most relevant, and an exhaustive set of definitions can be found in the spreadsheet file at the above site.

The data illustrates the size of this problem, and the amount of new power generation needed. The exponential growth of AI queries is based on the assumption that AI agents will be created that autonomously (or with some over site) are interacting with each other. The reduction in power per query seems a bit optimistic to me over 5 years, so the problem could be worse. Of course power generation and distribution is part of a complex system, and things will not likely stay static, and some power savings may be generated. However, with the simultaneous switch to electric vehicles over the same time frame, it is hard to imagine that the next years won’t bring extreme growth in power generation. It’s an exciting time to join this industry, with many new problems to be tackled.

[1] Smith, Matthew. October 2025. “The Hidden Scale Behind Every AI Answer.” IEEE Spectrum. Volume 62, Issue 10.

[2] Paccou, Remi and Fons Wijnhoven. December 2024. “Artificial Intelligence and Electricity A System Dynamic Approach.” Schneider Electric Sustainability Research Institute Digital Series. [Internet WWW PDF] https://download.schneider-electric.com/files?p_enDocType=Thought+Leadership+Article&p_Doc_Ref=TLA_System_Dynamics_Approach [Accessed 28-Dec-2025].

EngineeringPaper.xyz

As an engineer, I have been a long time Mathcad user. Over the past several years after acquisition by PTC, it has become more difficult to continue using because of expensive licensing and declining features. There is a free version called Mathcad Express for home or educational use, but I can’t use that at work. After switching jobs, I had to find a different alternative. My use cases usually involve creating sheets that perform calculations that I use to explore different concepts to gain a better understanding of them. Automatic unit conversions are probably the most useful feature, followed by the ability to quickly generate plots of functions. I also use it to solve systems of equations. Occasionally, I use the symbolic solving capabilities to quickly perform algebra.

A number of years ago I discovered a free alternative called SMath Studio. I haven’t used it much, and it is sometimes able to meet some of my needs, but is not well supported, has some user interface quirks, and requires to be installed which is becoming difficult in most companies without IT approval. Also, SMath Studio seems to be a dead project and has not been updated since 2022. So I’ve recently been searching for a different alternative. The most promising one that I found is EngineeringPaper.xyz. This tool works in a web browser, and lets you save your files to your PC. It works very intuitively for an engineer, and I was up and running with creating a sheet to calculate temperature rise in a bus bar carrying DC current within a few minutes. Here are some links to training/how to for some useful features:

• Units, plots, and solve cells in the basic tutorial – https://www.youtube.com/watch?v=r7EZQVhcr5Q
• Data tables from Excel – https://www.youtube.com/watch?v=xuNMikNa63U
  • This is a very cool feature that allows you to import tables of data into the sheet for use in calculations. Mathcad could also do this, and was a very useful feature at times.
  • An advantage over Mathcad, is that you can easily create new columns in the table using calculations performed in the sheet.
• Embedded Python code cells – https://www.youtube.com/watch?v=7qnyBwwGUJ8
  • This is a game changer feature; essentially you can use the full computational capabilities of Python right in your calculation sheets. So anything that is complex that is not easy to support with inline expressions, can be accomplished with this capability. Mathcad and SMath had their own programming feature, but it was clunky and difficult to get the syntax right and extremely difficult to debug. With real Python code, these problems go away.
• Using pre-built tables of engineering data – https://engineeringpaper.xyz/d5DEM9teTxodvCTwENwErn
• Exporting sheets to Python function code for numerical analysis – https://www.youtube.com/watch?v=uqK9tWiGqMY

Some other useful features that I’ve noticed in the above materials:

• Piecewise linear functions
• Symbolic solution of equations
• Sharing calculation sheets via shareable links
• Reddit community – https://www.reddit.com/r/EngineeringPaperXYZ/