The Cybersecurity threat from AI can be characterised very simply. the new computers >> the old computers AI isn’t software or code, it’s a completely different school of computer science. Deterministic machines are just inferior to probabilistic machines in the same way that insects are inferior to mammals. Lots of people will be in denial about this for decades, but sorry, sometimes you spend a long time mastering a hard thing and then a superior easy thing comes along, and your value to society and status in the village disappears, that’s life. AI is not a cybersecurity threat to IT infrastructure, its is an obsolescence threat. People argue that the threat is transitory and that once AI finds all the bugs the threat will subside. Errr, no. The threat is terminal. Because who still commutes on a mule? Why would anyone still be running deterministic programs in 2035? They won’t, there will only be Cobolt and Generative AI. “You idiot, don’t you understand the AI will write the code” No it won’t, all the layers between the AI and the machine code are what is obsolete. In the future there will be weights, pixels and transistors. All that stuff in the middle is going to get exfiltrated. Neural nets are more capable, faster and more error tolerant than syntax. It’s self evident now. People can disagree with this take, and I can safely ignore them as the tide of technology sweeps them away. People tend to split into three camps, those who think AI is a God, those who think it is a tool, and those who still think it is fake. I now think it is closer to sorcery. It is a swirling tempest of possibility that manifests reality when channeled by a principal actor. You cast prompts, you summon effects. Sorcery is a better analogy than God, or tools, or hiding under the bed. If you fear subordination, you are probably low agency. This is most people, so this will be the cause du jeur throughout the era. But there are people out there right now, fighting their Balrog. Some of them will win. You’re going to live to see some things and this whole cybersecurity zeitgeist is just the shiny mirrored inside of tiny bubble. And I’m sure all the people at Hugging Face and the labs are great folks. But this is a tactical episode not a strategic one.

Price Capture Ratio One thing that isn’t widely recognised about how the grid works is how different power generation types have different price capture ratios on the same grid. Let me explain… A grid delivers power to distributed load and sells power at day ahead prices to distributors. Generators all have different power generation shapes, some produce constant flat base load (gas and nuclear), some are random (wind), some have a daily cycle (solar). The data below is CAISO 2025 It shows not the cost of building each power generation type, it shows what price each type earned in $ / MWh over the full coarse of 2025. The average wholesale price on CAISO in 2025 was $32.20 / MWh. But that doesn’t mean this is what each generator was actually paid. Because prices are lowest at solar maxima and highest in the evening different technologies have different capture ratios. As you can see Solar earned $14.60/MWh whilst on the same grid during the same year, hydro earned $37.40/MWh The capture ratios is multiplied with the capacity factor to calculate what 1MW of capacity actually earns in revenue. For example, solar has a capacity factor of 0.26 whereas nuclear has a capacity factor of 0.90. This means 1MW of solar earns $3.79/hr whilst on the very same grid during the same 12 months, 1MW of nuclear earns $29.43/hr. There are some other factors at play (capacity fees, standing charges, tax incentives, etc) so this is somewhat simplified, but the point I am making stands. People compare LCOE, which is an unphysical and entirely fictional metric, and people make judgements about real business decisions and the future and get their decisions badly mangled. If you want the truth, just use empirical data. Batteries are quite good, they buy the solar minimum and enjoy a capture ratio premium, but only 4hr sizing fits inside the Duck Curve, which is 16.7% market share. As soon as BESS goes bigger than 4hrs it slips off the duck curve and the capture ratios sours. Once you understand price capture ratios on a grid you can quickly figure out the mathematically optimal grid mix for any given grid. Obviously nobody in the world actually optimises their grid accordingly, as that would be too sensible and we live in a somewhat chaotic world. Someone who understands this stuff should really launch a global infrastructure company. 😉 Humanity would benefit.

Interferometry What is it? 1. You make a monochromatic beam of light where all the wavelengths are exactly the same (a laser). 2. You make the beam coherent so that all of the waves are in phase (peaks align and troughs align). 3. Polarisation (optional), you polarise the beam to make the orthogonal oscillations of the beam aligned, this is for measuring rotation. 4. You split this beam in two and produce two beams that are monochromatic and coherent. 5. The first beam is the reference beam, the second beam is the target beam. 6. You bounce the target beam off a target and when it returns you compare it to the reference beam and the delta tells you information about spacetime. 7. Because light is fast you can collect an enormous volume of data with a very high sample rate. You can know exactly how far away the target is, how fast it is traveling, how it is vibrating, you can see all of its resonant frequencies and harmonics, right down to molecular lattice level. You can predict what the target is and how it is behaving, you can see its temperature from the Brownian motion of its surface. Because light is so fast you can really inspect it, because the wavelength is so short and precise you can inspect extreme resolution. This is the most powerful sensor family in science. As someone who once built an interferometer tech I find this is very cool, this is a whole domain of science and engineering that is very under discussed. The Fermilab Holometer was completed in 2014 it is a Michelson interferometer that was built to prove the universe was a hologram and that we live in a pixelated simulation. It did the opposite. After 145 hours of data, their specific model of correlated holographic shear noise was excluded at 4.6σ. No evidence of the predicted Planck-scale jitter was seen. As they collected more data they excluded a broader class of quantum-geometrical shear-noise models. The measured correlations were consistent with ordinary classical spacetime. Finally, the interferometers were reconfigured to be sensitive to rotational (twisting) fluctuations instead of shear. Again those models were also constrained; the data again matched a classical spacetime model. Lots of people still believe simulation theory, and that’s OK. People should believe whatever makes them happy. The same dataset also produced the strongest limits known on a stochastic gravitational-wave background in the megahertz band and on nearby primordial black-hole binaries, science that is independent of the quantum-geometry hypothesis. Interferometers in the hands of smart people unlock secrets. You can multiply the sensitivity (sensor power) of your interferometer by using Fabry Perot Cavities and bouncing your laser between two mirrors. This can give you many kilometres of beam length in a device that fits on your desk. Fabry Perot cavities are what LIGO uses (Laser Interferometry Gravitational Observatory), LIGO was built to measure intergalactic gravitational waves from distal black holes and supernova events and works incredibly well. LIGO can detect vanishingly small disturbances in spacetime that are from events that happened billions of years ago before our Sun even formed. We can actually measure and observe past events in extreme detail and understand what happened. We can observe things not just from the other side of the Milky Way, but from entirely distant galaxies that exist on completely different superclusters. We have been able to build phenomenal instruments for a while, we also have phenomenal DAQ, more recently we have the technology to train on this empirical data. There is stuff to find there. The 20th Century was dominated by theorists, I expect the 21st Century is an empirical age, build-test-build. We can ask the questions and the universe will just tell us.

“This gun I was firing shot someone” In Western legal systems machines carry no liability, because they do not have legal personhood, only organisations and people can be held liable. Use your tool accordingly. When your personal/corporate asset starts committing crimes… you are liable for those crimes.

So it’s now mid-September and we can see if this post aged like a fine wine, or like milk? In fairness it helps to know some things about what is possible in heavy industry and how the world turns, so I did have some advantages here.

10 years ago I had a long conversation with Professor Mark Parsons at Edinburgh Parallel Computer Centre (EPCC). This is where the UK national supercomputers are operated from (the Bayes Centre). Machines like ARCHER etc. Anyway, I was at times the only industrial user of some of the compute systems they had, (I was training NN on laser interferometry data back in 2015, back when AI was “not real”). Anyway, Mark was a good 10 years ahead of Terrence Tao here (a rare place to be!), Mark explained to me that a University is an institution wrapped around a library. A university spends money doing research to add to its library and it monetises the library by teaching people about the knowledge it contains. At a very fundamental level, this is the university model. A university is very much centred on its library. For context, this was after AlexNET, but before Attention Is All You Need. At the time, I was spending seven figures / year on ML. We didn’t have chatbots, we didn’t have LLMs, we weren’t doing language models. I was training my models on physical vibration data to make predictions about the physical world, my models were beginning to work for my fluids dynamics problem (oilfield multiphase flow metering). But back then we knew that AI was coming, we knew the UK would need an exascale machine or the country would not exist in the future. Mark wondered what a university would become in a world where the frontier of knowledge wasn’t in a library. He thought in the future… a university would be a supercomputer, people would feed it valuable data and it would return useful knowledge. I think Mark was correct, it just happened very fast. But there is no real need for universities when we all have access to Claude and GPT, curious people can build up very deep understanding without a library or a tutor. We have discovered a wonderful tool. I think we finally solved computers and now they are becoming what they were always meant to be. It’s as big of a change for humanity as cooking with fire, agriculture, literature, sanitation, mechanised work. Technology cannot be undone, and it cannot really be slowed down inside a capitalist global economy made up of competing nation states. We might feel sorry for Terrence Tao, but the game he loves is really over. He should be happy that he had such an opportunity whilst it lasted.

He’s going to do it again isn’t he. He owns the context graph doesn’t he.

High Entropy Alloys Maybe the oldest shipwreck discovered lies off the coast of Turkey. It sank carrying 9 tonnes of copper and 1 tonne of tin. Weird cargo, unless you know the recipe for bronze. The Uluburun wreck, is 3300 years old, smack in the middle of the Bronze Age. Now we live in the steel age, but the underlying idea behind our common engineering materials is the same. You take a pure metallic element and enhance it by adding a fraction of some other element. Adding one part tin to nine parts copper gives bronze; adding a pinch of carbon (charcoal) to iron produces steel. This is the recipe for making alloys, we do this because their strength, durability and workability make them the best known option for everything from cutlery to lampposts to bridges. But are traditional alloys the best we can do? Nope. Increasingly, metallurgists are questioning this old wisdom. We are out of runway with the ancient method, so now they are making wild metallic mixtures where no single element dominates, and producing materials the likes of which we have never seen. In any elemental metallic lattice the layers of identical atoms can slip past each other easily, along the planes of the lattice, pure metals are soft. That’s why gold panners can tell a pure nugget just by biting it. So we introduce foreign atoms and you can disrupt that slipping, locking the layers together and producing a tougher material. That is what an alloy is. This idea has furnished us with many of the materials that underpin modern technology. But with alloys, there’s invariably a point beyond which adding more of the alloying elements negates their benefits. At the atomic scale, the additive atoms start to form little clusters of metal-within-metal that make the material brittle. A small atomic dislocation locks the layers together, a big dislocation prestresses the layers and makes them easier to tear apart (crack propagation). The concept of entropy might provide a workaround. Entropy is a way of quantifying disorder in a system, and the rules of thermodynamics say that when something is more disordered it is more stable. 🤔 So rather than make orderly alloys from one main element spiked with pinches of others, why not mix five, six or more elements into a soup? There are 60 metallic elements, there are 10^40 combinations if you have crude discrete ratios, and infinite combinations if you consider fine tuning. The unorthodox atomic structures produce islands of absurd and alien properties One of those unexpected findings has to do with brittleness. All known alloys become more shatter-prone when cooled, but in 2014 we found a high-entropy alloy of iron, manganese, nickel, cobalt and chromium that became less brittle the colder it got, right down to -200°C. This defies our understanding of material science (our old understanding). With scaling can we solve High Entropy Alloys and predict the recipes that produce materials with desired engineering properties? It feels like we can.

Interestingly… this MOD program (link below) to produce “extreme fast jet drones” was launched the same week Reaction Engines and their Mach 25 technology slipped into bankruptcy, and released 200 aerospace propulsion engineers into the general population. I’m going to go out on a limb here and presume these turbojet drones, some of which are already in manufacture, are quite good.

Electrical Grid of the Future Much is discussed about resurgent demand for electricity and this looks directionally correct. But this is a crude macro understanding of the power industry. There are a series of 6 dynamics unfolding within the power generation and distribution sector that shape much of what is evolving. Future scenarios depend on what combination of these 6 dynamics actually emerge over time, and they tend to emerge in different places at different times. 1. Grid Services become the market, historically electricity markets were dominated by MWh of power generated and sold. But in the future capacity and stability fees will come to dominate value; frequency control, reactive power support, black start capability, inertia emulation, and fast response balancing are all monetizable services that will expand as intermittent renewables expand their market share. The grid is likely to see the volume price of the kWh rise sharply and then fall, and value capture evolves from units of power to grid services. 2. Intermittence Saturation, a synchronous grid (AC grid) can only absorb so much variable generation before the marginal value of the next solar/wind farm to be connected to the grid collapses. Once curtailment and negative pricing begin, the limiting factor for delivering electrical power is not generation, but system elasticity. Storage, flexible loads, and long distance interconnectors become the growth checks. Beyond the saturation point where the marginal unit collapses the spot price, the grid flips from being generation limited to coordination limited. Crucially as the market pricing flips to the new regime it opens up 3 types of arbitrage that provide a profit motive to drive coordination: • Temporal arbitrage • Geographic arbitrage • Service arbitrage Several grids worldwide already have zero or negative spot prices at the belly of the duck curve. This creates substantial arbitrage opportunity and moves all the value capture from generators to arbitrage providers. 3. Battery Energy Storage Systems - Temporal Arbitragers, intermittence opens an additional market for infrastructure + trader, physical and virtual assets can operate algorithmically to capture daily and even weekly price arbitrage opportunities. These can be regular (solar) and irregular (wind). Increasingly we are seeing disruptive utility companies deploy fleets of BESS often financed by their own customer base. 4. Behind-the-Meter Revolution, households and businesses are becoming independent nano grids: rooftop solar + batteries + EV + intelligent load (including domestic robots working when power is cheap). This BTM dynamic erodes retail demand whilst increasing self sufficiency. This likely has second order political effects strengthening demand for cleaner air and lower emissions. Households will still be grid connected but grid consumption by households will be tempered by BTM, even as absolute power consumption increases. The grid becomes more peaky, less predictable, with greater requirements for coordination. Traditional utilities will lose volumetric sales, but will gain the ability to monetise residual grid access and balancing services. Grid value allocation inverts from volume to services. 5. Long Distance HVDC spine, as renewables cluster around optimal geography (sun belts, wind corridors) and load concentrates around cities and datacenters long haul HVDC becomes the backbone energy infrastructure of civilisation. Expect continental scale spines, North South for seasonal arbitrage, as the intermittent saturation point (and marginal price collapse) varies with latitude and climate season. East-West HVDC spines provide diurnal arbitrage harvesting in competition with BESS. 6. Hyperscalers as anchor tenants, datacenters and AI training clusters and inference factories are effectively new industrial loads with $50-60bn of industrial capex adjacent to each GWe. These sinks demand 24/7 power, ultra stable frequency, and geographic (and jurisdiction) optionality. The high load density is already driving colocation and dedicated on-site power generation. All of these trends are superimposed over one another and unless you isolate them in analysis you can draw some weird assumptions. It's worth remembering that globally the world buys around $3.5 trillion of electricity every year, and around $2.6 trillion of that goes to generation and $900 billion goes to transmission. Whilst these figures look very likely to swell, the ratio is also likely to change. Transmission costs are going to be less and less about moving electricity over some distance and increasingly about quality control of the electricity that is delivered. The transmission system's role is evolving from a transport system to a quality system, transmission markets and pricing will eventually follow suit.

Zebrafish NN It’s pretty amazing to learn that efforts to map the fruit FLY connectome (brain synapses) were successful and its amazing to see the digital fly brain successfully complete a whole series of tasks, including complex navigation of open worlds. The fruit fly is 160,000 neurons and 10^7 synapses. Teams are currently doing the same thing with a zebrafish which has a similar scale neuron count but 10x more synapses. Fly was 100TB raw data and Zebrafish is 200TB, so about 2x the raw data. For context a rodent is 70m neurons and 10^11 synapses. Rodents will come later. But this method looks like it creates totally outsized results with absolutely miniscule models. Zebrafish NN is a vertebrae connectome, it has much of the same basic structure as other vertebrae, unlike FLY connectome. We are just about used to LLMs and coding, but AI is really just beginning and there is a lot more, and a lot weirder stuff coming down the pipe. These connectomes have already shown to be incredibly resilient, you can blind their sensors and injure their outputs and they still succeed. They are incredibly compact. A whole lot of inanimate objects are going to get complete autonomy, totally offline, air gapped autonomy. They will be delivering pizzas and fighting wars. Connectomics is an opposite approach to LLMs but you can of course use LLMs to help develop connectomes. It’s amazing that you can take a biologically evolved brain, map it, produce a digital twin, and then run it at machine speed. Imagine a human brain accelerated from 100Hz neuron fires to 3,200,000,000Hz that a typical CPU runs at. That’s 30 million times faster. But would need 1-2 exabytes to map. Anyway “Zebrafish” is next.

It might not be obvious for a while, but Fly Connectome Neural Networks (FLYNN) look much more disruptive to human civilisation than building a country of geniuses. arxiv.org/abs/2607.00025 Just as we pace the monoliths, the Zerg are here. But if you think about it, a self assembling connectome is like some sort of Ice 9. How do you stop a connectome that wants to assemble itself?

RF in Water How far electromagnetic radiation can travel through a medium depends on what type of medium it is and what the frequency of the radiation is. Water is famously very good at absorbing radiation and has very high attenuation (ability to absorb radiation). Water does have an attenuation hole approximately at the frequency of blue light, which is why blue light can travel through water and is why water looks blue. Water has high attenuation elsewhere and just absorbs most kinds of radiation. This is why is so easy to hide things in deepwater. This is why submarines are so difficult to find, and is also why submarines can’t really communicate with home when they are dived. However… As the chart on the right shows is you extend the wavelength axis waaaaaay out to the right until you get to 10,000m radiowaves or longer, extreme low frequency radiowaves (ELF)… they actually can travel through water! If you have 100,000m long radiowaves waves they can travel huge distances through water, and even through Earth. The challenge here is 2 fold, one you have a very low bit rate, second you need a very long antenna to handle a 1,000m radiowave. A 1/4 wave dipole would require a 1,000km long straight wire… so that’s what they built. The US built project Sanguine and Russia built ZEVS in Kola. ZEVS consumed 2,600,000 watts of electricity and produced 8 watts of radio signal. This is huge continent spanning infrastructure that broadcast encrypted codes 24/7. A sort of deadman’s switch for the two largest nuclear deterrent arsenals on Earth. They don’t use them like that since the Cold War ended, but it’s interesting what humans will build.

Big Ugly Fat F… (BUFF) The B-52 had its first flight in 1952, which is… 74 years ago. It’s still in active service today. In December last year Boeing won a contract to replace the engines on the B-52H fleet and upgraded them all to B-52J. The B-52J will remain in service until the 2060s. No tanks or armoured vehicles have made it to 100 years in service, or are planned to and the B-52 is the only combat aircraft expected to do so, but there are some naval vessels… USS Constitution and HMS Victory are both still afloat (although not used in combat), and the Russian ship Kommuna was built in 1913 and (I think?) is still in use. But the B-52 is in a class of its own, as a combat airframe shouldn’t really last this long, it has major Voyager-1 and Voyager-2 vibes. Plus it just looks cool, 8 engines…

Industrial Explosives Manufacturing The entire global military sector only buys 25% of all explosives that are manufactured, the other 75% of explosives are purchased and consumed by the mining industry. Whilst the military usually outspends mining in terms of the financial expenditure, mining totally dominates on raw tonnage. The mining industry gets through thousands of tonnes per day of Ammonia Nitrate Fuel Oxide (ANFO), this is a stable and low cost blasting media that is used in thousands of open pit mining operations around the world. The mining industry shifts millions of tonnes of earth and rock every day, almost all of it is blasted loose for recovery. Demand for metals and commodities is demand for high explosives and the chemical pre-requisites. Permitting for explosive manufacturing facilities is one of the big limits on the expansion of global mining operations, and the expansion of global commodity supplies. This isn’t obvious to most people, but if we want to build and make more things, if we want more goods manufactured, then that starts with making a lot more explosives. Explosives are one of the furthest upstream markets for the global economy and are therefore a good canary for an expansion of the global economy. Explosive prices will rise sharply before any new surge in commodity volumes feed through to the wider economy. This a good place to monitor the “fast takeoff” scenario commonly advocated by lots of people who have a very “software” understanding of how the world works. The reality is that upstream commodity prices and feedstock materials will need to blow out and stay blown out for a long time, for the supply chain to begin to respond to the price signal. A lot of these markets are cyclical and market participants often view price spikes as transient opportunities rather than structural shifts, and respond to price spikes with gouging rather than increasing production volumes. The Ukraine war disrupted the ammonia market in 2022 causing prices to rocket, but they have since fallen back quite substantially. The Ukraine war remains a bit of a ceiling on a major expansion of the global extractive industries. Interesting times ahead.

Food Supplies Agriculture has massively outperformed demographics all over the world. Food supplies have grown faster than the population on everything continent on Earth. This is what abundance looks like, there is more than enough food, everywhere. So what has this meant for prices and suppliers? Is food now “too cheap to meter”? No. Clearly not. OK but has food become so incredibly abundant that the cost of food has fallen? No. Food is actually more expensive than it used to be, and food inflation is something a lot of people will recognise. So why is food getting more expensive, even as the supply of food continues to outpace demand? Well food supplies from agriculture are waaay upstream of the consumer. The price of food on the shelves has actually risen quite sharply, but the price of food in the fields has not. The difference is in the food processing industry… it’s manufacturing. When you offshore all of your manufacturing industry, you also offshore all of your food processing industry too. And so even though countries think agriculture is the source of their food security… it is not. Almost everywhere in the world has an over abundance of agriculture and heavily subsidises agriculture, even though the bottleneck in their food security is manufacturing. Manufacturing is not subsidised, is not generally thought of as strategic and is not in a state of abundance anywhere in the world. Increasingly as geopolitics shifts toward a multipolar world, everyone is waking up to the fact that manufacturing is king. If you don’t have the machinery, you don’t have any security. You are a sitting duck. Politics and AI are going to combine to make some fascinating industrial policies over the next 10-15 years.