[DVYio Newsletter](https://newsletter.dvy.io/visit/6BcELfBB) It's Davey Barker again — in case this landed in your inbox and you're wondering where it came from. Lots of new faces this week, so welcome if you're just joining. And if you've been here a while, good to have you back. If you've got thoughts, suggestions, or just want to say hello, I'm on all the usual socials (links in the footer). [OpenAI's new model can adjust its thinking while it works](https://newsletter.dvy.io/visit/vwB5ER6d) [OpenAI's new model can adjust its thinking while it works](https://newsletter.dvy.io/visit/vwB5ER6d) OpenAI just released GPT-5.4, and the model can now show you its reasoning before it finishes, mid-response, so you can steer it while it's still thinking. Context window is up to 1 million tokens (roughly the length of several long novels), which matters for agents handling long multi-step tasks rather than one-shot queries. It's also the first general-purpose model OpenAI has shipped with native computer-use baked in, meaning it can actually operate software rather than just describing how to. Deep research on niche queries is reportedly improved too, and it holds context better over long thinking chains. [Apple's cheapest MacBook yet is actually good](https://newsletter.dvy.io/visit/EnVq6rbC) [Apple's cheapest MacBook yet is actually good](https://newsletter.dvy.io/visit/EnVq6rbC) Apple's cheapest portable Mac just got cheaper. The MacBook Neo starts at $599, which is almost half the price of the MacBook Air. - Apple's A18 Pro chip - 13-inch screen - aluminium body in four colours - "up to" 16 hours of battery - 1080p webcam The A-series chip is the one found in iPhones, not the M-series that is typically found in Macs. There's no MagSafe charging port (you're back to USB-C only). The display doesn't hit the same peak brightness. It's a real Mac, just not the top-shelf one. For students or anyone who's been put off by Mac prices, this is probably the most compelling entry point Apple has ever offered. [Wikipedia locked down after rogue JavaScript started deleting pages](https://newsletter.dvy.io/visit/ebDrQs6u) [Wikipedia locked down after rogue JavaScript started deleting pages](https://newsletter.dvy.io/visit/ebDrQs6u) Someone snuck a malicious JavaScript file into Russian Wikipedia. The script was built to hijack admin accounts and use those elevated permissions to mass-delete articles. It got found before spreading to main pages. But the Wikimedia Foundation didn't want to risk a delayed response, so they flipped Wikipedia into read-only mode while they dealt with it. No editing allowed site-wide until the threat was contained. The actual deletions appear to have been limited to Meta-Wiki (Wikimedia's internal software project site) rather than the main encyclopaedia. So the damage was minor. The intent, though, was clearly broader. What's interesting here is the attack vector: not a brute-force login, not a database exploit, but patient JavaScript sitting in a user subpage, waiting for the right conditions. It's a reminder that a platform built on open editing has an unusually large surface area to defend. This Week I Learnt [Japan has has two incompatible power grids for 130 years old.](https://newsletter.dvy.io/visit/RyoRmdu6) [Japan has has two incompatible power grids for 130 years old.](https://newsletter.dvy.io/visit/RyoRmdu6) [— Most countries pick one frequency for their power grid. Japan ended up with two, and they don't talk to each other. In the 1880s, Tokyo's early electricity suppliers bought generators from a German company. Those ran at 50 Hz (think of it as the rhythm of the electricity, how many times per second it reverses direction). Osaka's suppliers went to American manufacturers instead. Those ran at 60 Hz. Nobody thought to check whether the two systems would ever need to connect. Why would they? These were tiny, local operations hundreds of kilometres apart. They grew. The Tokyo grid spread across the north and east; the Osaka grid across the south and west. By the time anyone noticed the problem, both systems were enormous, deeply embedded, and essentially impossible to swap out. Japan ended up as the only developed country in the world with two incompatible national grids running in parallel. The 2011 earthquake made this very concrete, very fast. The tsunami knocked out power generation in the north-east. The south-west had spare capacity. But moving that electricity across the invisible frequency boundary is genuinely hard. You need specialist conversion equipment, and Japan only had a limited amount of it. So one half of the country sat in the dark while the other had power to spare. Converting everything to a single frequency has been discussed, and then quietly shelved, repeatedly. The cost is simply staggering. So the fix, for now, is appliances that accept both frequencies, and hoping the conversion stations are enough when the next crisis hits.](https://newsletter.dvy.io/visit/RyoRmdu6) When an AI is thinking through a problem, do you want to see its reasoning as it works? Yes, I want to steer it Yes, but just curiosity No, just give me the answer Depends on the task Last Week's Poll Results for: Should governments be implementing UBI right now? Yes, we're overdue 11% Yes, but slowly and carefully 57% Not yet, more evidence needed 24% No, fix existing systems first 4% No, it's the wrong approach entirely 4% [Image](https://newsletter.dvy.io/visit/4UJUVVRF) [Image](https://newsletter.dvy.io/visit/4UJUVVRF) [Image](https://newsletter.dvy.io/visit/4UJUVVRF) [Tiffany & Co. in Beijing, China](https://newsletter.dvy.io/visit/4UJUVVRF) [— Designed by MVRDV](https://newsletter.dvy.io/visit/4UJUVVRF) [The table for one is becoming the table to get](https://newsletter.dvy.io/visit/h2ZNwsR5) [The table for one is becoming the table to get](https://newsletter.dvy.io/visit/h2ZNwsR5) [— About one in five Broadway tickets this season was bought by a single person. That's double what it was just a few years ago. ATG Entertainment noticed and launched "Solo Seats" events specifically for solo theatregoers. The theatre industry, in other words, is starting to follow the money. It's not just Broadway, either. Google searches for "restaurant for one" hit a 20-year high in January. Solo travel is now a $95 billion market, with projections putting it at $190 billion by 2030. The pattern is consistent across dining, theatre, and travel. Part of this is demographic — more Americans live alone than at any previous point, and solo leisure is the natural extension of that. But there's probably something cultural happening too: the stigma around doing things "for groups" by yourself seems to be quietly dissolving. Going to the cinema alone has been normal for years. Theatre is catching up. Restaurants are next, if the search data is any guide.](https://newsletter.dvy.io/visit/h2ZNwsR5) [78 minutes a day, everywhere on earth. Your commute is basically universal.](https://newsletter.dvy.io/visit/J9peTnZ2) [78 minutes a day, everywhere on earth. Your commute is basically universal.](https://newsletter.dvy.io/visit/J9peTnZ2) No matter where you live or how much money you have, you probably spend roughly 78 minutes a day getting from place to place. That's the finding from a study across 43 countries covering more than half the world's population. Rich or poor, urban or rural, the number barely shifts — somewhere between 66 and 90 minutes, depending on who you ask. Nobody fully understands why. Part of it seems to be practical (there's only so much of your day you're willing to lose to travel), and part of it seems almost instinctive (humans appear wired to explore a certain radius around where they live). The mode of transport changes, the distances change, but the time budget stays roughly fixed. It has implications on energy use. Faster vehicles and more efficient engines haven't actually reduced how much energy we burn on transport globally because when travel gets easier, people just go further in the same amount of time. So if the 78-minute budget is basically locked in, the only real lever is how much energy gets burned during those 78 minutes. The researchers argue we should be measuring transport energy per hour of travel (not per kilometre), and designing cities around low-energy options (walking, cycling, public transit) so that the time people were going to spend travelling anyway costs the planet as little as possible. [CO₂ as the fluid for power turbines is now a commercial reality after China just switched one on](https://newsletter.dvy.io/visit/EsUr9o2D) [CO₂ as the fluid for power turbines is now a commercial reality after China just switched one on](https://newsletter.dvy.io/visit/EsUr9o2D) Most power stations boil water then use the steam to spin a turbine, but this one uses CO₂ instead. At high enough pressure and temperature, CO₂ enters what's called a supercritical state: it behaves like both a liquid and a gas at the same time. In that state, it's denser and more energy-rich than steam, which means you can extract more electricity from the same amount of heat, in a smaller physical footprint. China just switched on the world's first commercial-scale plant running on this principle. It's not a prototype or a test rig. It's generating real power on the grid. The reason to care: most of the world's electricity still comes from heat (burning gas, coal, or splitting atoms), and the efficiency of that conversion process is one of the biggest levers we have on emissions and cost. A technology that wrings more electricity out of the same heat source could slot into existing infrastructure rather than replacing it wholesale. [Can LLMs learn to update their beliefs?](https://newsletter.dvy.io/visit/JH6q3t88) [Can LLMs learn to update their beliefs?](https://newsletter.dvy.io/visit/JH6q3t88) Most AI systems are pretty bad at updating what they think they know. Ask a chatbot for recommendations across a conversation and it'll often just ignore what you told it earlier, or fall back on lazy defaults like "pick the cheapest option". Good reasoning under uncertainty requires something called Bayesian inference (the maths for how you should update your beliefs when new evidence arrives). LLMs don't naturally do this. They pattern-match on training data rather than maintaining a running probabilistic model of, say, your actual preferences. Google Research tried a different approach: instead of hoping the model picks this up implicitly, they trained an LLM to directly mimic the outputs of an optimal Bayesian model on a flight recommendation task. Essentially using the Bayesian model as a teacher. The results were promising on their own, but the more interesting finding was generalisation — the reasoning skills transferred to other tasks the model hadn't been trained on. Which suggests it wasn't just memorising flight-booking patterns. It was picking up something more like the underlying logic of belief updating. [300,000 living neurons on a chip were playing a first-person shooter within a week](https://newsletter.dvy.io/visit/wKr3ux6e) [300,000 living neurons on a chip were playing a first-person shooter within a week](https://newsletter.dvy.io/visit/wKr3ux6e) Living human brain cells, grown on a tiny chip, figured out how to play Doom. Not simulated brain cells. Actual biological neurons. The chip itself is human brain tissue (around 800,000 cells) sitting on a bed of electrodes that can send and receive electrical signals. The cells fire, the game responds, the cells adjust. It's a feedback loop, and somehow, within a week, the neurons got good enough at it to navigate Doom's corridors. Cortical Labs did something similar with Pong back in 2021, but Doom is a different beast: first-person perspective, spatial awareness, actual decision-making about which direction to move. The bigger news is that it's now programmable in Python, which means researchers don't need a specialist team to experiment with these systems. It's not beating human players yet. But it learns faster than conventional AI, and with far less energy. Whether that translates into something practical (controlling a robot arm is the example being floated) is still an open question. [Image](https://newsletter.dvy.io/visit/8yW2dARj) [Image](https://newsletter.dvy.io/visit/8yW2dARj) [Image](https://newsletter.dvy.io/visit/8yW2dARj) [Image](https://newsletter.dvy.io/visit/8yW2dARj) [A Norwegian village 200km from the sea that cruise ships can still reach](https://newsletter.dvy.io/visit/8yW2dARj) [— On a map, Skjolden (population 238) looks landlocked. It sits deep in inland Norway, ringed by mountains, 200km from open water. It's also a seaside village. Technically. And ocean-going cruise ships can dock there. The reason is the Sognefjord, a narrow channel of seawater that cuts roughly 200km into the Norwegian coast. It's one of the longest fjords in the world, and for most of that length it's over 1,200 metres deep (deeper than most of the North Sea). Even near Skjolden, at the very end of the fjord, it's still over 100 metres deep. Deep enough for a cruise liner to navigate the whole way without running aground. So you get this strange situation where a village with fewer people than a small office block sits 200km from the sea… and yet people sunbathe on its shores in summer, watching cruise ships go past.](https://newsletter.dvy.io/visit/8yW2dARj) [Why are some languages spoken faster than others?](https://newsletter.dvy.io/visit/4BGtebkP) [Why are some languages spoken faster than others?](https://newsletter.dvy.io/visit/4BGtebkP) [— Take any sentence you know in English and imagine translating it into Japanese. The written version gets longer. Switch to Thai and it shrinks back down. But read all three aloud and you'd finish at almost exactly the same moment. Christophe Coupé and colleagues studied 17 Eurasian languages and found that languages with dense, information-rich syllables (where each syllable is harder to predict from the one before it) tend to get spoken more slowly. Languages with simpler, more predictable syllables get spoken faster. The two effects cancel out, leaving information transmission roughly constant across languages. Think of it like compression. A zip file is smaller than the original, but it contains the same data. Languages are doing something similar: different encodings, same throughput.](https://newsletter.dvy.io/visit/4BGtebkP) [“Tired: This meeting could have been an email. Wired: This startup could have been a spreadsheet.”](https://newsletter.dvy.io/visit/MU6TQ37f) [@mhoye@mastodon.social](https://newsletter.dvy.io/visit/MU6TQ37f) [Web](https://newsletter.dvy.io/visit/aVzyyWVk) · [Twitter](https://newsletter.dvy.io/visit/qLUozFdy) · [LinkedIn](https://newsletter.dvy.io/visit/2A9d7myh)