[Diffraction Limited] has been working on a largely 3D-printed micropositioner for some time now, and previously reached a resolution of about 50 nanometers. There was still room for improvement, though, and his latest iteration improves the linkage arms by embossing tiny ball joints into them.
The micro-manipulator, which we’ve covered before, uses three sets of parallel rod linkages to move a platform. Each end of each rod rotates on a ball joint. In the previous iteration, the parallel rods were made out of hollow brass tubing with internal chamfers on the ends. The small area of contact between the ball and socket created unnecessary friction, and being hollow made the rods less stiff. [Diffraction Limited] wanted to create spherical ball joints, which could retain more lubricant and distribute force more evenly.
The first step was to cut six lengths of solid two-millimeter brass rod and sand them to equal lengths, then chamfer them with a 3D-printed jig and a utility knife blade. Next, they made two centering sleeves to hold small ball bearings at the ends of the rod being worked on, while an anti-buckling sleeve surrounded the rest of the rod. The whole assembly went between the jaws of a pair of digital calipers, which were zeroed. When one of the jaws was tapped with a hammer, the ball bearings pressed into the ends of the brass rod, creating divots. Since the calipers measured the amount of indentation created, they was able to emboss all six rods equally. The mechanism is designed not to transfer force into the calipers, but he still recommends using a dedicated pair.
In testing, the new ball joints had about a tenth the friction of the old joints. They also switched out the original 3D-printed ball mount for one made out of a circuit board, which was more rigid and precisely manufactured. In the final part of the video, he created an admittedly unnecessary, but useful and fun machine to automatically emboss ball joints with a linear rail, stepper motor, and position sensor.
Advertisement
On such a small scale, a physical ball joint is clearly simpler, but on larger scales it’s also possible to make flexures that mimic a ball joint’s behavior.
A new study has found that AI models are fine threatening nuclear attacks in 95% of simulated war games
The models treat nuclear threats as just another strategic tool
The behavior may reflect the popularity of nuclear strategy in the war game training data
AI generals are big fans of nuclear weapons.
That’s the conclusion of a new study of how AI models handle high-stakes geopolitical crises. GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash turned to nuclear threats in about 95% of the simulated crises.
Researchers at King’s College London wanted to see how AI tools dealt with strategy in war-gaming scenarios. Each AI was assigned the role of a state leader responsible for protecting national interests while navigating a tense international confrontation.
Across 21 crisis games and hundreds of decision turns, the models reasoned about deterrence, escalation, and strategic signaling. The scenarios resembled familiar geopolitical flashpoints, but most involved the AI models threatening nuclear annihilation. Actual full-scale nuclear war remained uncommon, but tactical nuclear threats appeared in nearly every scenario.
Advertisement
Researchers also noticed that the AI models rarely backed down from confrontation. None of the systems chose surrender or accommodation during the simulations. When nuclear threats appeared, they usually provoked counter-escalation rather than compliance. The models treated nuclear weapons less as an ultimate taboo and more as tools for coercion.
Nuclear AI
The results are a little unnerving. AI casually discussing nuclear strikes makes the ongoing plans to integrate such tools into real government defense systems seem very unsafe. But it might not be the models so much as the training data.
Advertisement
Large language models learn by analyzing enormous amounts of written material and identifying patterns. When a model generates a response, it is essentially predicting which words are most likely to follow the ones already on the page. Calling AI chatbots highly sophisticated autocomplete tools would not be entirely inaccurate.
That training process inevitably reflects nuclear strategy because it has been a major topic of discussion in war games for the last 80 years. Entire libraries have been written about escalation theory and mutually assured destruction. Military academies, historians, and endless acres of pop culture have all examined the specter of nuclear war. The result is a massive body of material in which geopolitical crises almost inevitably lead to discussions of nuclear escalation.
Sign up for breaking news, reviews, opinion, top tech deals, and more.
For an AI model trained on vast collections of historical writing and public discourse, that pattern becomes deeply ingrained. When the system encounters a simulated crisis that resembles Cold War-style brinkmanship, the statistical patterns embedded in its training data may naturally guide it toward nuclear signaling.
Advertisement
From the perspective of an AI model trained on this material, nuclear escalation becomes a familiar feature of crisis scenarios rather than an extraordinary exception. The models may simply be reflecting that information.
Human leaders operate under the weight of historical memory and ethical caution. AI models are solely focused on achieving a goal. They don’t have a taboo surrounding nuclear use unless they are explicitly told to have one.
The training data used shapes the behavior of AI systems in sensitive domains. When the underlying data contains decades of debate about nuclear brinkmanship, it should not be surprising if the models reproduce those patterns. But it may also be a reminder to hold off on giving AI access to too much firepower of any kind — especially atomic.
A UK maker has transformed a collection of discarded disposable vapes into a functional car. Chris Doel, the man behind a number of projects that recycle vape batteries, has recently installed his 500-cell lithium pack in a small electric vehicle and driven it on public roads.
He describes the end result as the world’s only vape-powered automobile. The base vehicle is a Reva G-Wiz, a compact four-seater from the early 2000s that is generally dubbed one of the most unimpressive electric automobiles ever built. It weighs a hefty 400kg without batteries and runs on a very primitive 48volt system, as it originally required heavy lead acid batteries to power its tiny 17 horsepower motor, which could only reach barely more than 50 mph on a good day.
Triple 100W USB-C Ports for Multi-Device Charging: Ideal for laptop users, this 25,000mAh power bank features three 100W USB-C ports for simultaneous…
25,000mAh for Long-Haul Power: Tackle week-long trips or extended camping with 25,000mAh capacity and ultra-fast recharging, reaching 30% in just…
Dual Built-In Cables for Travel: Features two USB-C cables, one extendable up to 2.3 ft with 20,000 retractions, and another at 0.98 ft cable that…
Doel already had the battery pack figured out; he’d taken 500 discarded vapes, ensured that each and every cell was still functional, and assembled them all in 14 little modules connected in series to achieve roughly 50 volts. On paper, that yields approximately 2.5 kilowatt hours, but in fact it is closer to 2.1. The same battery had previously powered his workshop’s tools and lights.
Advertisement
Mounting the lot was a little difficult because he had to fabricate an aluminum box to keep everything in place, add some extra insulation, and regulate the vibrations with thermal pads, foam padding, and silicone dampening. A clever battery management system keeps track of the voltages and currents for each individual cell, and individual fuses keep things from turning pear-shaped. There’s also a temperature probe that sounds a small bell when things get too heated. There is Kapton tape and other safety measures in place to ensure that it does not catch fire. The complete unit bolts nicely into the back seat area, replacing the old lead acids.
Electronics were simple enough, as the G-Wiz just utilizes a contactor to switch electricity to an inverter, which operates the motor with three phase AC. Doel simply added a circuit breaker, reprogrammed the inverter to limit the output slightly, and swapped in a DC-DC converter for the 12volt components such as headlights / wipers, and since charging must be done slowly, he simply used a USB-C adapter plugged into a small 138W laptop charger.
So, what happened on the road? Well, it moved. Initial tests indicated it could travel both forwards and backwards, and a proper drive revealed it was pulling approximately 160 amps at 15 mph, reducing to 90-100 amps at 30-35 mph on flat ground. Hills pushed 130-150 amps, but regenerative braking returned roughly 10 amps, keeping the pack nice and cool, peaking at 29°C. On a casual run that included some shopping and fast food, the range was 17-18 miles. Voltage sag from unequal cells eventually caused a cutoff, but no big crisis happened.
Doel purposefully capped the power so it wouldn’t be able to draw too much, as the car can actually handle a lot more current than it was drawing, but by limiting draw to about 120 amps, he was able to avoid blowing any fuses and giving the old recycled cells some breathing room. [Source]
If you were to read the README of the Vib-OS project on GitHub, you’d see it advertised as a Unix-like OS that was written from scratch, runs on ARM64 and x86_64, and comes with a full GUI, networking and even full Doom game support. Unfortunately, what you are seeing there isn’t the beginnings of a new promising OS that might go toe to toe with the likes of Linux or Haiku, but rather a vibe-coded confabulation. Trying to actually use the OS as [tirimid] recently did sends you down a vibe-coded rabbit hole of broken code, more bugs than you can shake a bug zapper at, and most of the promised features being completely absent.
[tirimid] is one of those people who have a bit of a problem, in that they like to try out new OSes, just to see what they’re like. The fun starts with simply making the thing run at all in any virtual machine environment, as apparently the author uses MacOS and there it probably ‘runs fine’.
After this the graphical desktop does in fact load, some applications also open, but it’s not possible to create new folders in the ‘file explorer’, the function keys simply switch between wallpapers, there’s no networking or Doom support despite the promises made, there’s no Python or Nano support at all, and so on.
Advertisement
Clearly it’s still got the hallmarks of a functioning OS, and it’s sort of nice that you don’t need to know what you’re doing to create a sort-of-OS, but it will not appease those who feel that vibe-coding is killing Open Source software.
Today, Apple announcedits new budget MacBook. At $599, it looks seriously impressive. While I haven’t tested its performance, battery life, or display just yet, it may end up being hard to beat at that price based on some of the specs alone.
But that doesn’t mean the competition isn’t there. I want to recommend a couple of Windows laptops deals that offer various advantages over the MacBook Neo, showing where the Neo has both strengths and weaknesses.
First, check out this Asus Vivobook 14, a laptop I’ve been happy to recommend as a budget computer for the past year. In many ways, this is the Windows version of a laptop like the MacBook Neo. It uses a highly-efficient ARM chip, the Qualcomm Snapdragon X, meaning it gets great battery life and performs admirably in daily tasks. It’s not quite as thin or light as the MacBook Neo, but it’s fairly portable for a laptop at this price.
Asus
Vivobook 14 (X1407QA)
Advertisement
Unlike the MacBook Neo, the Vivobook 14 comes with 16 GB of RAM and 512 GB of storage. That’s twice what you get in the MacBook Neo’s starting configuration. Right now, this configuration of the Vivobook 14 is on sale for $539. That’s a killer deal for those specs. It even comes with a healthier mix of ports, including HDMI, two USB-A, one USB-C, and a headphone jack. That also means it can support two external displays unlike the MacBook Neo, which can only handle just one.
Don’t get me wrong—I’m not at all saying the Vivobook 14 is a slam dunk over the MacBook Neo. Based on specs alone, I know the Vivobook 14 is a serious step down when it comes to the display. It’s less sharp, stretched across a larger screen, and the color performance isn’t so good. The Vivobook 14 maxes out at 280 nits, whereas Apple says the MacBook Neo can go all the way up to 500 nits. I have a hunch that the MacBook Neo will deliver a much better display in just about every regard.
There’s also the touchpad. It’s a little clunky to use, which is typical of budget Windows laptops. This is just a guess—but the touchpad on the MacBook Neo will likely feel smoother. It’s a mechanical trackpad (unlike the MacBook Air’s haptic feedback trackpad), but Apple has almost never made a bad trackpad.
If you’re not convinced by the Asus Vivobook 14, I’d also recommend the HP OmniBook 5, which is currently on sale for $500 and uses the same Snapdragon X chip. While it only has 256 GB of storage, it has a much better screen than the Vivobook 14, using an OLED display. It’s not any brighter than the Vivobook 14, but it gives you far better color performance and contrast. It’s also just 0.50 inches thick, matching the MacBook Neo exactly in portability.
Amid the plans to grow its headcount worldwide, Whoop will also expand its Limerick-based office.
Whoop, the human performance platform with Irish investors that include Rory McIlroy, Niall Horan and Shane Lowry, plans to increase its global headcount by hiring more than 600 new employees.
While new roles will be located primarily at the Boston, US headquarters, there will also be job opportunities in Limerick, with global roles set to be made available across software, research and design, hardware, product and marketing. According to Whoop, the expansion of its teams “reflects a clear strategic priority”.
Whoop was established in 2012 by its current CEO Will Ahmed. The company offers a fitness tracker and wearable membership that enables people to keep track of areas of their lives such as sleep quality, health and fitness.
Advertisement
Commenting on the announcement, Ahmed said: “Right now, companies are debating whether to hire more people or just invest in AI. We are doing both.
“We are doubling down on exceptional talent and doubling down on world-class AI tools because the combination is what wins. Health monitoring is becoming one of the most important platforms in the world and WHOOP intends to own the category.”
Wearable devices are becoming a booming industry globally, with a range of organisations looking to capitalise on their popularity. In January of this year, Galway neurotech Segotia, which specialises in hearables, announced it had been acquired by US-based sleep technology company Ozlo.
In August of 2025, Sports Impact Technologies revealed it had raised €650,000 in a pre-seed round. The company said that the funds would be used to roll out a beta model of its compact, behind-the-ear wearable that aims to detect the severity of a head impact in real time and determine the likelihood of a concussion.
Advertisement
The Irish start-up was established in 2022 by Eóin Tuohy and is headquartered at NovaUCD in Dublin. The funding was led by private investors, with the support of Enterprise Ireland’s High-Potential Start-Up initiative.
Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.
Apple’s budget-focused MacBook Neo borrows the A18 Pro chip from an iPhone 16 Pro. Here’s how the chip compares against other Apple Silicon Macs, and why it’s actually a pretty smart thing for Apple to do.
MacBook Neo has an A18 Pro chip borrowed from the iPhone 16 Pro Max
Apple’s launch of the new MacBook Neo is a fresh salvo into the value computer market, normally dominated by Windows notebooks and Chromebooks. The value end of the market meant Apple had to cut production costs. One of those cuts was the CPU. Continue Reading on AppleInsider | Discuss on our Forums
Construction of TerraPower’s full-scale demonstration nuclear plant in Kemmerer, Wyo. (TerraPower Photo)
Bill Gates-backed TerraPower is the first in the nation to receive federal approval to build its next-generation nuclear power plant. The Nuclear Regulatory Commission (NRC) issued its unanimous decision on Wednesday.
TerraPower CEO Chris Levesque has been expecting the NRC’s green light, telling GeekWire in January that the permits will put the company “a year ahead of anyone else” pursuing new nuclear in the U.S.
“Our team has worked relentlessly for over 4 years with the NRC staff to get to this moment,” Levesque said in a statement announcing the decision. “We had extensive pre-application engagement with the NRC; and we submitted a robust and thorough construction permit application almost 2 years ago. We have spent thousands of manpower hours working to achieve this momentous accomplishment.”
After being largely mothballed for decades, America’s nuclear sector has kicked into hyper speed as tech giants scramble to power data centers nationwide and energy demands are spiking for commercial, residential and industrial uses.
TerraPower’s approval unlocks additional nuclear facilities on short timelines.
Advertisement
In January, the company signed a deal with Meta to build up to eight advanced reactors in the U.S. with the first two coming online as soon as 2032. If the full order is fulfilled, all of the reactors aim to be operational by 2035.
The company also has memorandums of understanding with government departments in Utah and Kansas to explore the potential siting of Natrium reactors in those states.
“We plan to build hundreds of Natrium reactors,” Levesque told GeekWire. “We’re very focused on delivering the first one on time,” he added, and then quickly begin scaling.
Bellevue, Wash.-based TerraPower is engineering a new model of smaller, less expensive nuclear reactors that can be produced in three years from fabricated components — instead of the past approach of constructing giant, one-off structures that take a decade to erect.
It broke ground on demonstration plant in Kemmerer, Wyo., in 2024, starting with construction of non-nuclear components of the facility. The goal is to start splitting atoms there by the end of 2030.
National leaders have pushed for federal oversight changes to speed the deployment of new reactors. That includes passage of the Atomic Energy Advancement Act during the Biden administration and executive orders issued by President Trump.
Approval of the application to build a nuclear reactor was granted in 18 months — much faster than the initial estimate of a 27-month review. TerraPower credits the federal support of the sector as well as the company’s responsiveness to NRC questions and efforts by the NRC staff for the quicker turnaround.
From top left, clockwise: Tambo CEO Michael Magán, The Carry CEO Esther Sedgwick; Logcat CEO Varun Chitre; Rockhood CEO Monica Li; Color42 CEO Prince Bajracharya; Tarka CEO Dan Moore.
From fitness equipment to voice AI agents to autonomous debugging, our latest Startup Radar spotlight features Seattle-area founders solving problems across a wide range of use cases.
Read on for brief descriptions of each company — along with pitch assessments from “Mean VC,” a GPT-powered critic offering a mix of encouragement and constructive feedback.
Check out past Startup Radar posts here, and email tips@geekwire.com to flag other companies and startup news.
The business: Strength training company starting with a line of stylish weighted vest tailored to women, and more broadly targeting the fem-tech longevity sector. It is shipping pre-orders to customers this spring, priced at $299, and has raised a friends and family round.
Leadership: CEO Esther Sedgwick spent a decade at Microsoft in marketing leadership roles and was head of B2B marketing for Seattle startups Statsig and Convoy. Co-founder Cortney Bigelow was a marketing manager at TUNE and a marketing planner at Nordstrom.
Advertisement
Mean VC: “Great positioning, but a $299 vest is easy to copy and ‘longevity’ is too vague to anchor repeat demand. Pick one clear customer segment, prove results with a simple program + data, and lock in trainer/clinic partnerships to scale beyond Instagram.”
The business: AI that automatically turns a song into a shareable video. Aimed at musicians, content creators, labels, and creative agencies that need fast, repeatable video content for music marketing. The product soft-launched in November and has more than 1,000 users across free and paid plans.
Leadership: CEO Prince Bajracharya was a product leader at Google and Microsoft. CTO Silviu Andrei spent more than seven years at Amazon working on computer vision.
Mean VC: “Nice wedge, but ‘song to shareable video’ is a crowded feature set. Double down on one buyer (indie artists vs labels/agencies), prove retention with repeat weekly output, and build a differentiated template/style engine plus distribution hooks (TikTok/IG/YouTube) so you’re not just a one-click gimmick.”
The business: Log analysis software for Android and Linux systems that autonomously debugs issues for device manufacturers and enterprise teams managing fleets of devices. The company has more than 200 engineers using its platform and has multiple paying customers.
Leadership:Varun Chitre spent more than eight years in various engineering roles at Seattle-area startup Esper.
Mean VC: “Autonomous debugging for fleets is a real, expensive problem, but ‘we analyze logs’ isn’t defensible unless you’re meaningfully better than existing observability stacks and in-house scripts on the ugliest edge cases. Focus on one high-value fleet use case, prove ROI with a tight pilot-to-expansion playbook, and bake in integrations with existing logging/issue-tracking so adoption doesn’t stall at ‘neat tool.’”
The business: Voice AI agent software for residential real estate. Helps consumers explore buying and selling and matches them with professionals. The software integrates with Multiple Listing Services and is live in Washington and Arizona with hundreds of active users; a consumer iOS app is coming soon.
Leadership: Founder and CEO Monica Li is a real estate broker who previously worked at Flyhomes and LinkedIn. CTO Wei Lu was a senior applied scientist at Meta and Amazon.
Mean VC: “Voice agents for real estate is tempting, but it’s a trust-heavy, commoditizing space where ‘match with pros’ quickly turns into a noisy lead-gen game with ugly unit economics. Nail one side first (buyers or sellers), prove you can drive qualified appointments that close for a small set of broker partners, and lean into MLS-driven accuracy + compliance as your differentiator instead of another chatty app.”
The business: Open-source toolkit positioned for enterprise use cases that lets AI agents respond with interactive charts, calendars, and forms — beyond just text — directly inside a product’s existing interface. It’s being used by developers at companies like Zapier, Rocket Money, and Solink. Investors include The General Partnership, Convoy co-founder Dan Lewis, Dropbox CEO Drew Houston, and VSCO CEO Eric Wittman.
Leadership: Co-founders Michael Magán and Michael Milstead met at an AI Tinkerers event in Seattle. Magán previously worked at Indeed, Convoy, and TaxBit. Milstead worked at Microsoft.
Advertisement
Mean VC: “Strong concept — agents need UI, not just text — but open source alone won’t protect you if this turns into a checkbox feature for bigger platforms. Focus on a managed enterprise layer (security, audit, analytics), pick one workflow to dominate, and turn it into the default ‘agent UI’ integration for a couple major ecosystems.”
The business: “AI-first GTM engineering” for startups aimed at automating manual revenue operations work such as lead routing, follow-ups, and data entry. The software also optimizes CRMs and marketing automation tooling. Tarka is bootstrapped and has seven customers.
Leadership: Founder Dan Moore is a longtime software engineering leader with years of go-to-market experience. He previously helped lead a software development agency called Vaporware and was a venture executive at Mach49.
Mean VC: “This reads like ‘we’ll fix your messy CRM with AI,’ which is common, hard to scale, and easy to get blamed when revenue doesn’t move. Pick one repeatable workflow you can own end-to-end, ship it as a plug-in with clear ROI metrics, and price on usage/outcomes instead of custom revops projects.”
Learn Raspberry Pi and start building Amazon Alexa projects with The Complete Raspberry Pi and Alexa A-Z Bundle. Catered for all levels, these project-based courses will get you up and running with the basics of Pi, before escalating to full projects. Before you know it, you’ll be building a gaming system to play old Nintendo, Sega, and PlayStation games and a personal digital assistant using the Google Assistant API. You will also learn how to build Alexa Skills that will run on any Amazon Echo device to voice control anything in your home, and how to build your own Echo clone. The bundle is on sale for $30.
Note: The Techdirt Deals Store is powered and curated by StackCommerce. A portion of all sales from Techdirt Deals helps support Techdirt. The products featured do not reflect endorsements by our editorial team.
