2026-03-30 10:00:41
Although generative language models have found little widespread, profitable adoption outside of putting artists out of work and giving tech companies an easy scapegoat for cutting staff, their their underlying technology remains a fascinating area of study. Stepping back to the more innocent time of the late 2010s, before the cultural backlash, we could examine these models in their early stages. Or, we could see how even older technology processes these types of machine learning algorithms in order to understand more about their fundamentals. [Damien] has put a 60s-era IBM as well as a PDP-11 to work training a transformer algorithm in order to take a closer look at it.
For such old hardware, the task [Damien] is training his transformer to do is to reverse a list of digits. This is a trivial problem for something like a Python program but much more difficult for a transformer. The model relies solely on self-attention and a residual connection. To fit within the 32KB memory limit of the PDP-11, it employs fixed-point arithmetic and lookup tables to replace computationally expensive functions. Training is optimized with hand-tuned learning rates and stochastic gradient descent, achieving 100% accuracy in 350 steps. In the real world, this means that he was able to get the training time down from hours or days to around five minutes.
Not only does a project like this help understand these tools, but it also goes a long way towards demonstrating that not every task needs a gigawatt datacenter to be useful. In fact, we’ve seen plenty of large language models and other generative AI running on computers no more powerful than an ESP32 or, if you need slightly more computing power, on consumer-grade PCs with or without GPUs.
2026-03-30 07:00:36
Whether it’s a new couch or a rare piece of hardware picked up on eBay, we all know what it feels like to eagerly await a delivery truck. But the CERN researchers involved in a delivery earlier this week weren’t transporting anyone’s Amazon Prime packages, they were hauling antimatter.
Moving antimatter, specifically antiprotons, via trucks might seem a bit ridiculous. But ultimately CERN wants to transfer samples between various European laboratories, and that means they need a practical and reliable way of getting the temperamental stuff from point A to B. To demonstrate this capability, the researchers loaded a truck with 92 antiprotons and drove it around for 30 minutes. Of course, you can’t just put antiprotons in a cardboard box, the experiment utilized a cryogenically cooled magnetic containment unit that they hope will eventually be able to keep antimatter from rudely annihilating itself on trips lasting as long as 8 hours.
Speaking of deliveries, anyone building a new computer should be careful when ordering components. Shady companies are looking to capitalize on the currently sky high prices of solid-state drives by counterfeiting popular models, and according to the Japanese site AKIBA PC Hotline, there are some examples in the wild that would fool all but the most advanced users. They examine a bootleg drive that’s a nearly identical replica of the Samsung 990 PRO — the unit and its packaging are basically a mirror image of the real deal, the stated capacity appears valid, and it even exhibits similar performance when put through a basic benchmark test.
But while the drive’s sequential read and write speeds are within striking distance of the official numbers from Samsung, things start to fall apart when doing random speed tests or performing real-world operations. It took the fake drive over 25 minutes to write a 370 GB file, while the authentic one ripped through the same file in less than 4: giving a true write speed of 261 MB/s and 1,861 MB/s, respectively.
Luckily you don’t have to time how long it takes to dump 100+ GB of data on the drive just to see if it’s legitimate, Samsung offers a tool that can communicate with the drive and determine if it’s an original or not. If they don’t already, we imagine other manufacturers will roll out similar capabilities in an effort to combat these sophisticated clones.
Of course, computers aren’t the only things in our modern world that are impacted by the rising prices of memory and flash storage. On Friday, Sony announced that they would be implementing higher prices across their PlayStation line starting this week to compensate for what they call “pressures in the global economic landscape.”
Starting April 2nd (presumably they didn’t want consumers to think this was a joke), the base model PS5 will be bumped up to $649.99 in the US and €649.99 in Europe, while the PS5 Pro will be set at an eye-watering 899.99 in both currencies. Admittedly we’ve done absolutely no research to support this, but surely that must make the latter system the most expensive home game console in history by a considerable margin. In comparison, Microsoft’s top of the line Xbox Series X is currently priced at $799, though the model with the smaller 1 TB drive is still available for $649.
One might think that the skyrocketing cost of memory would force developers to take a lesson from the early days of computing, and usher in a new era of highly optimized code that manages to do more with less. That would be nice. Instead, we have now have DOOM rendered in the browser using CSS.
As Niels Leenheer explains in the write-up, the original goal was to have the entire game running in CSS. But he quickly ran into issues trying to implement the game logic. So he settled for letting Claude port the open source C code for the base game over to JavaScript, which freed him up to work on doing the graphics in CSS.
If you’re interested in web development it’s a fascinating look at how far the modern browser can be pushed, and even if you don’t, it’s a surprisingly smooth way to play the classic shooter without having to install anything.
Lastly, the public is finally getting some information about the health scare aboard the International Space Station that triggered the first-ever medical evacuation from the orbiting laboratory back in January. As we predicted in our previous coverage, NASA was unwilling to put personal information about one of their astronauts on the public record, and have remained tight-lipped about the situation. So it was Crew-11 Pilot Mike Fincke himself that decided to not only come forward as the individual who experienced the issue, but to detail what he went through in an interview with the Associated Press.
So what happened? Well, nobody is quite sure yet. Fincke says he was eating dinner the night before he was scheduled to go on a spacewalk outside the Station, and suddenly realized he couldn’t speak. His crewmates realized he was in distress, and contacted medical personnel at Mission Control on his behalf. Testing performed both on the Station and back on Earth has yet to provide any explanation for the episode. It lasted approximately 20 minutes, and he’s experienced no issues since. Space is kinda crazy like that sometimes.
See something interesting that you think would be a good fit for our weekly Links column? Drop us a line, we’d love to hear about it.
2026-03-30 04:00:22
Although GNSS systems like GPS have made pin-pointing locations on Earth’s sphere-approximating surface significantly easier and more precise, it’s always possible to go a bit further. The latest innovation involves strapping laser retroreflector arrays (LRAs) to newly launched GPS satellites, enabling ground-based lasers to accurately determine the distance to these satellites.
Similar to the retroreflector array that was left on the Moon during the Apollo missions, these LRAs will be most helpful with scientific pursuits, such as geodesy. This is the science of studying Earth’s shape, gravity and rotation over time, which is information that is also incredibly useful for Earth-observing satellites.
Laser ranging is also essential for determining the geocentric orbit of a satellite, which enables precise calibration of altimeters and increasing the accuracy of long-term measurements. Now that the newly launched GPS III SV-09 satellite is operational this means more information for NASA’s geodesy project, and increased accuracy for GPS measurements as more of its still to be launched satellites are equipped with LRAs.
2026-03-30 01:00:20
Unless you’re into circuit sculptures, generally speaking, a working circuit isn’t the end-point of a lot of electronics projects. To protect your new creation from grabby hands, curious paws, and the ravages of nature, you’ll probably want some kind of enclosure. These days a lot of us would probably run it off on the 3D printer, but some people would rather stay electronics hobbiests without getting into the 3D printing hobby. For those people, [mircemk] shares how he creates professonal-looking enclosures with handtools.
The name [mircemk] will seem familiar to longtime readers– we’ve featured many of his projects, and they’ve always stood out for the simple but elegant enclosures he uses. The secret, it turns out, is thin PVC sheeting from a sign shop. At thicknesses upto and including 5 mm, the material can be bent by hand and cut with hobby knives. It’s obviously also amenable to drilling and cutting with woodworking tools as well. Drilling is especially useful to make holes for indicator LEDs. [mircemk] recommends cyanoacrylate ‘crazy’ glue to hold pieces together. For holding down the PCB, the suggestion of double-sided tape will work for components that won’t get too hot.
Rather than paint, the bold contrasting colours we’ve become used to are applied using peel-and-stick wallpaper, which is a great idea. It’s quick, zero mess, and the colour is guaranteed to be evenly applied. It might even help hold the PVC enclosure together ever so slightly. You can watch him do it in the video embedded below.
We hate to say it, but for a one-off project, this technique probably does beat a 3D printed box for professional looks, assuming you have [mircemk]’s motorskills. If you don’t have said motor skills, check out this parametric project box generator. If you’d rather avoid PVC while making a square box to hold a PCB, have you considered using PCBs?
Thanks to [mircemk] for the tip! If you have a tip or technique you want to share, please box it up and send it to the tipsline
2026-03-29 22:00:58
Ionizing radiation damage from electrons, protons and gamma rays will over time damage a CMOS circuit, through e.g. degrading the oxide layer and damaging the lattice structure. For a space-based camera that’s inside a probe orbiting a planet like Jupiter it’s thus a bit of a bummer if this will massively shorted useful observation time before the sensor has been fully degraded. A potential workaround here is by using thermal energy to anneal the damaged part of a CMOS imager.
The first step is to detect damaged pixels by performing a read-out while the sensor is not exposed to light. If a pixel still carries significant current it’s marked as damaged and a high current is passed through it to significantly raise its temperature. For the digital logic part of the circuit a similar approach is used, where the detection of logic errors is cause for a high voltage pulse that should also result in annealing of any damage.
During testing the chip was exposed to the same level of radiation to what it would experience during thirty days in orbit around Jupiter, which rendered the sensor basically unusable with a massive increase in leakage current. After four rounds of annealing the image was almost restored to full health, showing that it is a viable approach.
Naturally, this self-healing method is only intended as another line of defense against ionizing radiation, with radiation shielding and radiation-resistant semiconductor technologies serving as the primary defenses.
2026-03-29 19:00:36
Usually, when we see non-planar 3D printers, they’re rather rudimentary prototypes, intended more as development frames than as workhorse machines. [multipoleguy]’s Archer five-axis printer, on the other hand, breaks this trend with automatic four-hotend toolchanging, a CoreXY motion system, and print results as good-looking as any Voron’s.
The print bed rests on three ball joints, two on one side and one in the center of the opposite side. Each joint can be raised and lowered on an independent rail, which allows the bed to be tilted on two axes. The dimensions of the extruders’ motion system limit how much the bed can be angled when the extruder is close to the bed, but it can reach sharp angles further out.
The biggest difficulty with non-planar printing is developing a slicer; [multipoleguy] is working on a slicer (MaxiSlicer), but it’s still in development. It looks as though it’s already working rather well, to the point that [multipoleguy] has been optimizing purge settings for tool changes. It seems that when a toolhead is docked, the temperature inside the melt chamber rises above the normal temperature in use, which causes stringing. To compensate for this, the firmware runs a more extensive purge when a hotend’s been sitting for a longer time. The results speak for themselves: a full three-color double helix, involving 830 tool changes, could be printed with as little as six grams of purge waste.
As three-axis 3D printers become consumer products, hackers have kept looking for further improvements to make, which perhaps explains the number of non-planar printing projects appearing recently, including a few five-axis machines. Alternatively, some have experimented with non-planar print ironing.
