Examining The Vulnerability Of Large Language Models To Data-Poisoning

February 3, 2025 by Maya Posch 23 Comments

Large language models (LLMs) are wholly dependent on the quality of the input data with which these models are trained. While suggestions that people eat rocks are funny to you and me, in the case of LLMs intended to help out medical professionals, any false claims or statements dripping out of such an LLM can have dire consequences, ranging from incorrect diagnoses to much worse. In a recent study published in Nature Medicine by [Daniel Alexander Alber] et al. the ease with which this data poisoning can occur is demonstrated.

According to their findings, only 0.001% of training tokens have to be replaced with medical misinformation to order to create models that are likely to produce medically erroneous statement. Most concerning is that such a corrupted model isn’t readily discovered using standard medical LLM benchmarks. There are filters for erroneous content, but these tend to be limited in scope due to the overhead. Post-training adjustments can be made, as can the addition of RAG, but none of this helps with the confident bull excrement due to corruption.

The mitigation approach that the researchers developed cross-references LLM output against biomedical knowledge graphs, to reduce the LLM mostly for generating natural language. In this approach LLM outputs are matched against the graphs and if LLM ‘facts’ cannot be verified, it’s marked as potential misinformation. In a test with 1,000 random passages detected issues with a claimed effectiveness of 91.9%.

Naturally, this does not guarantee that misinformation does not make it past these knowledge graphs, and largely leaves the original problem with LLMs in place, namely that their outputs can never be fully trusted. This study also makes it abundantly clear how easy it is to corrupt an LLM via the input training data, as well as underlining the broader problem that AI is making mistakes that we don’t expect.

Stepper Motor Operating Principle And Microstepping Explained

February 3, 2025 by Maya Posch 14 Comments

The [Denki Otaku] YouTube channel took a look recently at some stepper motors, or ‘stepping motors’ as they’re called in Japanese. Using a 2-phase stepper motor as an example, the stepper motor is taken apart and its components explained. Next a primer on the types and the ways of driving stepper motors is given, providing a decent overview of the basics at the hand of practical examples.

As great as theoretical explanations are, there’s a lot of value in watching the internals of a stepper motor move when its coils are activated in order. Also demonstrated are PWM-controlled stepper motor drivers before diving into the peculiarities of microstepping, whereby the driving of the coils is done such that the stator moves in the smallest possible increments, often through flux levels in these coils. This allows for significantly finer positioning of the output shaft than with wave stepping and similar methods that are highly dependent on the number of phases and coils.

As demonstrated in the video, another major benefit of microstepping is that it creates much smoother movement while moving, but also noted is that servo motors are often what you want instead. This is a topic which we addressed in our recent article on the workings of stepper motors, with particular focus on the 4-phase 28BYJ-48 stepper motor and the disadvantages of steppers versus servos.

Continue reading “Stepper Motor Operating Principle And Microstepping Explained” →

How The Main Frame Became The Mainframe: An Etymological Dissertation

February 2, 2025 by Maya Posch 12 Comments

In his most recent article, [Ken Shirriff] takes a break from putting ASICs under a microscope, and instead does the same in a proverbial manner with the word ‘mainframe’. Although these days the word ‘mainframe’ brings to mind a lumbering behemoth of a system that probably handles things like finances and other business things, but originally the ‘main frame’ was just one of many ‘frames’. Which brings us to the early computer systems.

We have all seen the photos of early computer systems, which not only filled rooms, but which also tended to consist of multiple units. This was something which the designers of the IBM 701 computer seem to have come up with, to make it possible to transport and install computer systems without cranes and the breaking out of walls. Within the IBM 701 system’s internal documentation, the unit containing the core logic was referred to as the ‘main frame’, alongside the ‘power frame’, the ‘core frame’, etc.

From this [Ken] then traces how the word ‘main frame’ got reused over the years, eventually making it outside of the IBM world, with a 1978 Radio Electronics magazine defining the ‘mainframe’ as the enclosure for the computer, separating it seemingly from peripherals. This definition seems to have stuck, with BYTE and other magazines using this definition.

By the 1960s the two words ‘main frame’ had already seen itself hyphenated and smushed together into a singular word before the 1980s redefined it as ‘a large computer’. Naturally marketing at IBM and elsewhere leaned into the word ‘mainframe’ as a token of power and reliability, as well as a way to distinguish it from the dinky little computers that people had at home or on their office desk.

Truly, after three-quarters of a century, the word ‘mainframe’ has become a reflection of computing history itself.

Is Fire Conductive Enough To Power A Lamp?

February 2, 2025 by Maya Posch 17 Comments

Is fire conductive? As ridiculous that may sound at first glance, from a physics perspective the rapid oxidation process we call ‘fire’ produces a lot of substances that can reduce the electrical insulating (dielectric) properties of air. Is this change enough to allow for significant current to pass? To test this, [The Action Lab] on YouTube ran some experiments after being called out on this apparent fact in the comments to an earlier video.

Ultimately what you need to make ‘fire’ conductive is to have an appreciable amount of plasma to reduce the dielectric constant, which means that you cannot just use any rapid oxidation process. In the demonstration with lights and what appears to be a (relatively clean-burning) butane torch, the current conducted is not enough to light up an incandescent or LED light bulb, but can light up a 5 mm LED. When using his arm as a de-facto sensor, it does not conduct enough current to be noticeable.

The more interesting experiment here demonstrates the difference in dielectric breakdown of air at different temperatures. As the dielectric constant for hot air is much lower than for room temperature air, even a clean burning torch is enough to register on a multimeter. Ultimately this seems to be the biggest hazard with fire around exposed (HV) electrical systems, as the ionic density of most types of fire just isn’t high enough.

To reliably strike a conductive plasma arc, you’d need something like explosive (copper) wire and a few thousand joules to pump through it.

Continue reading “Is Fire Conductive Enough To Power A Lamp?” →

Inside A Vintage Oven Controlled Crystal Oscillator

February 1, 2025 by Maya Posch 8 Comments

Crystal oscillators are incredibly useful components, but they come with one little snag: their oscillation is temperature-dependent. For many applications the relatively small deviation is not a problem, but especially for precision instruments this is a deal breaker. Enter the oven controlled crystal oscillator, or OCXO. These do basically what it says on the tin, but what’s inside them? [Kerry Wong] took apart a vintage Toyocom TCO-627VC 10 MHz OCXO, revealing a lot more complexity than one might assume.

Inside the insulated enclosure there is of course the crystal oscillator itself, which has a heating coil wrapped around it. Of note is that other OCXOs that [Kerry] took apart had more insulation, as well as other ways of providing the thermal energy. In this particular unit a thermistor is attached to the crystal’s metal case to measure its temperature and provide feedback to the heating circuit. The ICs on the PCB are hard to identify due to the conformal coating, but at least one appears to be a 74LS00, alongside a 78L05 voltage regulator which reduces the 12V input voltage.

As an older OCXO it probably is a lot chunkier than newer units, but the basic principle remains the same, with a heating loop that ensures that the crystal inside the unit remains at the same temperature.

Continue reading “Inside A Vintage Oven Controlled Crystal Oscillator” →

Using Microwave Heating To Locally Anneal CNT-Coated FDM Prints

February 1, 2025 by Maya Posch 19 Comments

The CNT coating between the layers is heated with microwaves to locally anneal. (Credit: Sweeney et al., Science Adv., 2017)

Layer adhesion is one of the weak points with FDM 3D printing, with annealing often recommended as a post-processing step. An interestingly creative method for this was published in Science Advances back in 2017, featuring the work of researchers at Texas A&M University and citing previous work by other teams. In the paper by [Charles B. Sweeney] et al, they describe how they coated PLA filament with carbon nanotubes (CNTs), resulting in this CNT being distributed primarily between the individual layers of polymer.

This is useful because CNTs are quite sensitive to microwave radiation, resulting in the conversion to thermal energy, i.e. heat. Compared to traditional annealing where the entire part is placed into an oven or similar, this microwave-based heating – or locally induced RF (LIRF) as they call this method – localizes the heat to the interface between two layers.

The advantages of this approach are that it doesn’t change the dimensions of the part noticeably, it’s faster and more efficient, and the annealing between layers approaches the strength of traditional manufacturing. Unfortunately not too much seems to have happened with this approach since then, but considering that both CNTs (single & double-walled) and microwaves are readily available, there’s not much standing in the way of replicating these results.

Taking A $15 Casio F91W 5,000 Meters Underwater

February 1, 2025 by Maya Posch 42 Comments

When considering our favorite spy movies and kin that involve deep-sea diving, we’d generally expect to see some high-end watch that costs thousands of dollars and is specially engineered to withstand the immense pressures kilometers below the ocean’s surface. Yet what about a humble Casio F91W that can be bought for about $15 if it’s the genuine article and not one of the millions of fakes? Over at the Watches of Espionage site they figured that they’d dress up one of these famous watches to give it the best possible shot at surviving the crushing pressures at a depth of 5 km.

The actual modification to the F91W was pretty mild, involving nothing but a ‘hydro-mod’ whereby oil is used to replace the air inside the watch case. Since oil is incompressible, nothing bad should happen to the watch. Theoretically at least. The Watch-Under-Test (WUT) was strapped to a US Navy’s CURV 21 remotely operated vehicle and dunked into the ocean before starting its descend into the inky darkness of the deep sea.

Although only hitting a measly 4,950 m, the watch survived just fine, showing that even if you’re a secret US operative on a deep-dive espionage mission, all you really need is one of these Casio watches.