TechyMagThings

There is a persistent belief in the ‘AI’ community that large language models (LLMs) have the ability to learn and self-improve by tweaking the weights in their vector space. Although there’s scant evidence that tweaking a probability vector space is anything like the learning process in biological brains, we nevertheless get sold the idea that artificial general intelligence (AGI) is just around the corner if we do just enough tweaking.

Instead of emerging super intelligence, the most likely outcome is what is called model collapse, with a recent paper by [Hector Zenil] going over the details on why self-training/learning in LLMs and similar systems is a fool’s errand. For those who just want the brief summary with all the memes, [Metin] wrote a blog post covering the basics.

In the end an LLM as well as a diffusion model (DM) is a statistical model of input data using which a statistically likely output can be generated (inferred) based on an input query. It follows intuitively that by using said output  to adjust the model with, the model will over time converge on a kind of statistical singularity rather than some ‘AI singularity’ event. This is also why these models need to be constantly trained with external, human-generated data in order to prevent such a collapse.

In the paper by [Hector] a mathematical model is created to demonstrate that an LLM, DM or similar statistical model undergoes degenerative dynamics whenever said external input is reduced. Although in the paper a mechanism is suggested to counter the entropy decay within the model, the ultimate point is that a statistical model cannot improve itself without continuous external anchoring.

The idea of LLMs being at all intelligent in any sense has been a contentious one, with the concept of language models being equated with ‘AI’ dating back to the 20th century, including as fun home computer projects. Much of the problem probably lies in humans projecting intelligent behavior onto these statistical models, turning LLMs into ‘counterfeit humans’, not helped by how closely generated text can resemble something written by a human, even if completely confabulated.

Thanks to [deshipu] for the tip.

An often overlooked section in the datasheets for popular humidity sensors like the BME280 and DHT22 is the ‘non-condensing humidity’ bit, which puts an important constraint on which environments you can use this sensor in. This was the painful lesson that [Mellow Labs] recently had to learn when multiple of such sensors had kicked the bucket after being used in a nicely steamed-up bathroom. Fortunately, it introduced him to sensors that are rated for use in condensing humidity environments, such as the SHT40 that’s demonstrated in the video.

This particular sensor is made by Sensirion, and as we can see in the datasheet it features a built-in heater that allows it to keep working even in a condensing environment. This heater has three heating levels which are controlled via the I2C interface, though duration is limited to one second in order to prevent overheating the sensor.

Of note is that you cannot take measurements while the heater is operating, and its use obviously increases power draw significantly. This then mostly leaves when to turn on the heater as an exercise to the engineer, with [Mellow Labs] opting to start the heater when relative humidity hit 70% as a conservative choice.

In the comments to the video other options for suitable sensors were pitched, including the Bosch BME690 which is similarly rated for condensing environments. All of which condenses down to the importance of reading the datasheet for any part that you intend to use in possibly demanding environments.

You may not know what a ADM-3, a TV910, or a H1420 are, but you probably have at least heard of a VT-100. They are all terminals from around the same time, but the DEC VT-100 is the terminal that practically everything today at least somewhat emulates. Even though a real VT-100 is rare, since it defined what have become ANSI escape sequences, most computers you’ve used in the last few decades speak some variation of the VT-100’s language. [Nikhil] wanted to see if you could use a VT-100 for real work today.

While the VT-100 wasn’t a general-purpose computer, it did have an 8080 inside. It only had about 3K of RAM, which was enough to act as a serial terminal. A USB serial port and a terminal with modern Linux, how hard could it be?

As it turns out there were a few issues. MacOS assumes terminals can take data at 9600 baud with no handshaking, apparently. It also means that any application that assumes redrawing the whole terminal is fast will be sorry for that choice.

Of course, there are commands modern VT-100-like terminals accept that the original didn’t. However, as you’ll see in the post, all of these things you can either live with or solve.

It is easy to make your own VT-100 replica. While the VT-100 may seem simple today, it was a marvel compared to even older terminals.

This week Jonathan chats with Andrei, Mahir, and Praneeth, live on location at Texas Instruments! The team at TI has been working hard to provide really good Open Source support for Sitara processors, including upstreaming support to the mainline Linux kernel. We talk about the CI pipeline for these devices, the challenges of doing Open Source at a big company, and more. Check it out!

• https://github.com/texasinstruments
• https://www.beagleboard.org/
• https://www.ti.com/design-development/software-design/open-source.html

Did you know you can watch the live recording of the show right on our YouTube Channel? Have someone you’d like us to interview? Let us know, or have the guest contact us! Take a look at the schedule here.

Direct Download in DRM-free MP3.

If you’d rather read along, here’s the transcript for this week’s episode.

Theme music: “Newer Wave” Kevin MacLeod (incompetech.com)

Licensed under Creative Commons: By Attribution 4.0 License

After previously trying out low-tech compression molding with a toaster oven and 3D printed molds, [future things] is back with a video that seeks to explore some of the questions raised after the first video. Questions such as how well this method works with HDPE and PLA thermoplastics, whether the flashing could be cut off by the mold and the right temperatures and times to heat the plastic before a charge is ready for inserting into the mold.

In this video the same PHA-based mold is used, but in a three-piece configuration to allow for a more complex shape. This way game tokens could be made for use by the son of the author, which also shows one straightforward and very practical use of this method.

A big change here is that no more metal chopsticks are used to handle the charge, as this was found to cool down the heated plastic too much. Instead the hot charge is handled with fingers and wooden chopsticks, with the plastic heated until it has about the consistency of thick honey. For LDPE this takes about 5-7 minutes at 130°C. After compressing the charge into the mold, about 30 seconds are all it takes for the plastic to cool down enough.

There was a question about the use of mold release spray, but this didn’t seem to cause any issues, so can probably be used safely. As for other plastic types, HDPE works fine too when you heat it up at a slightly higher temperature and don’t mind it being tougher to handle.

Easiest is probably PLA, which would seem unsurprising. Using some chopped-up PLA printing waste it was easy enough to make a few more game tokens, demonstrating that this method is very viable for converting scrap FDM print waste into such items. As noted in the comments by [edmundchao] this method works great too for PETG, using PETG molds, while using a ratcheting clamp for extra pressure instead of just pressing by hand.

It seems fair to say that pinball machines are among the most universally loved gaming systems known today, yet the full-sized ones are both very expensive and very large, while even the good quality table-sized ones tend to be on the expensive side. That raises the question of whether a fully 3D printed pinball machine could at all be fun and not just feel like a cheapo toy? A recent video by [Steven] from [3D Printer Academy] on YouTube makes here a compelling argument that it might actually be worth something to consider.

In addition to being fully modular and customizable the most compelling element is probably that the design supports two- and four-player multiplayer. This sees the metal balls leaving at the rear and from there entering the playing field of another player’s machine, which can probably get pretty chaotic.

Unfortunately this is part of a Kickstarter campaign, so you’ll have to either shell out some cash to get access to the print files or DIY your own version. We’d also be remiss to not address the durability concerns of a 100% plastic pinball machine like this, plus the lack of serious heft to compensate for more enthusiastic playing styles.

If you are more into traditional DIY pinball machines, we have covered these as well, along with small screen-based machines, and their miniature brethren for when space is really at a premium.

Australia’s payphones are an iconic part of the national landscape, even if they’re not as important as they once used to be. However, they’re having a resurgence of late, in part thanks to a new national pastime—the sport of Payphone Tag!

Created by [Alex Allchin], the game is simple. To play, you first sign up on the website and get your emoji and 5-digit PIN. You then go out and find a payphone, dial the Payphone Tag number, and enter your PIN when prompted. This lets you “capture” the phone, raising your score in the game. If a phone is already captured, no matter—just head out there, dial the number, and key in your own PIN to steal it. You can also push your score even higher by capturing three payphones in a triangle on the map to get bonus points.

It’s a fun geospatial game that’s also free to play, because Telstra made payphone calls free back in 2022. It might cost you a bit to get out to some phones, but there are plenty you can reach with the aid of free public transport at the moment, anyway. Protip—at the time of writing, there are a ton of easy captures to be had on Kangaroo Island. It might just cost you a pretty penny to get out there. Have at it!

We’d love to see some stats from Telstra as to whether this is making a dent in overall payphone usage rates. In any case, there were 800 players in the last 7 days and a full 36,640 captures so far, so a lot is happening out there. We fully expect to see this concept spread to other nations in turn, though it might be less attractive in places where you still need to dig out a coin to make a call.

We’ve featured a few payphone hacks over the years. If you’re doing something rad with these telecommunication devices of yesteryear, we’d love to hear about it on the tipsline.