Leaked Internal Google Document Claims Open Source AI Will Outcompete Google And OpenAI

May 5, 2023 by Maya Posch 37 Comments

In the world of large language models (LLM), the focus has for the longest time been on proprietary technologies from companies such as OpenAI (GPT-3 & 4, ChatGPT, etc.) as well as increasingly everyone from Google to Meta and Microsoft. What’s remained underexposed in this whole discussion about which LLM will do more things better are the efforts by hobbyists, unaffiliated researchers and everyone else you may find in Open Source LLM projects. According to a leaked document from a researcher at Google (anonymous, but apparently verified), Google is very worried that Open Source LLMs will wipe the floor with both Google’s and OpenAI’s efforts.

According to the document, after the open source community got their hands on the leaked LLaMA foundation model, motivated and highly knowledgeable individuals set to work to take a fairly basic model to new levels where it could begin to compete with the offerings by OpenAI and Google. Major innovations are the scaling issues, allowing these LLMs to work on far less powerful systems (like a laptop or even smartphone).

An important factor here is Low-Rank adaptation (LoRa), which massively cuts down the effort and resources required to train a model. Ultimately, as this document phrases it, Google and in extension OpenAI do not have a ‘secret sauce’ that makes their approaches better than anything the wider community can come up with. Noted is also that essentially Meta has won out here by having their LLM leak, as it has meant that the OSS community has been improving on the Meta foundations, allowing Meta to benefit from those improvements in their products.

The dire prediction is thus that in the end the proprietary LLMs by Google, OpenAI and others will cease to be relevant, as the open source community will have steamrolled them into fine, digital dust. Whether this will indeed work out this way remains to be seen, but things are not looking up for proprietary LLMs.

(Thanks to [Mike Szczys] for the tip)

Holograms Display Time With ESP32

May 5, 2023 by Bryan Cockfield 30 Comments

Holograms and holographic imagery are typically viewed within the frame of science fiction, with perhaps the most iconic examples being Princess Leia’s message to Obi-Wan in Star Wars, or the holodecks from Star Trek. In reality, holograms have been around for a surprising amount of time, with early holographic images being produced in the late 1940s. There are plenty of uses outside of imagery for modern holographic systems as well, and it’s a common enough technology that it’s possible to construct one using an ESP32 as well.

In this build, [Fiberpunk] demonstrates the construction and operation of a holographic clock. The image is three-dimensional and somewhat transparent and is driven by an ESP32 microcontroller. The display is based around a beamsplitter prism which, when viewed from the front, is almost completely invisible to the viewer. The ESP32 is housed in a casing beneath this prism, and [Fiberpunk] has two firmware versions available for the device. The first is the clock which displays an image as well as the time, and the second is more of a demonstration which can show more in-depth 3D videos using gcode models and also has motion sensing controls.

For anyone interested in holography, a platform like this is might make an excellent entry point to explore, and with the source for this build available becomes even easier. It’s almost certainly less expensive than these 3D printers that can turn out custom holographic images, and has the added benefit of being customizable and programmable as well.

Continue reading “Holograms Display Time With ESP32” →

Tactical Build Makes Machining Splined Shaft A Snap

May 5, 2023 by Dan Maloney 9 Comments

Quick, what’s 360 divided by 23? It’s easy enough to get the answer, of course, but if you need to machine a feature every 15.652 degrees around a shaft, how exactly would you accomplish that? There are a number of ways, but they all involve some degree of machining wizardry. Or, you can just make the problem go away with a little automation.

The story behind [Tony Goacher]’s Rotary Table Buddy begins with some ATV tracks he got off AliExpress. His idea is to build a specialty electric vehicle for next year’s EMF Camp. The tracks require a splined shaft to drive them, which would need to be custom-made on a milling machine. A rotary table with a dividing plate — not as fancy as this one, of course –is usually the answer, but [Tony] was a little worried about getting everything set up correctly, so he embarked on a tactical automation solution to the problem.

An RP2040 provided the brains of the project, while a NEMA 23 stepper provides the brawn. [Tony] whipped up a quick PCB and 3D printed a case for the microcontroller, a stepper driver, an LCD display, and a few buttons. He 3D printed an adapter and a shaft coupler to mount the stepper motor to a rotary table. From there it was just a matter of coming up with a bit of code to run everything.

There’s a brief video in [Tony]’s blog post that shows Rotary Table Buddy in action, indexing to the next position after cutting one of the 23 splines. He says it took about ten minutes to cut each spline using this setup, which probably makes to total cutting time far less than the amount of time invested in the tool. But that’s hardly the point, and besides, now he’s set up for all kinds of machining operations in the future.

And we sure hope we hear about the EMF Camp build, too.

Op-Amp Challenge: Light Up Breadboard Shows Us The Signals

May 5, 2023 by Jenny List 5 Comments

Most Hackaday readers will no doubt at some point used a solderless breadboard for prototyping. They do the job, but sometimes their layout can be inflexible and keeping track of signals can be a pain. There’s a neat idea from [rasmusviil0] which might go some way to making the humble breadboard easier to use, it’s a breadboard in which each line is coupled via an op-amp buffer to an LED. In this way it can be seen at a glance some indication of the DC voltage present.

It’s an idea reminiscent of those simple logic probes which were popular years ago, but its implementation is not entirely easy. Each circuit is simple enough, but to replicate it across all the lines in a breadboard makes for a huge amount of quad op-amp chips stuffed onto one piece of stripboard as well as a veritable forest of wires beneath the board.

The effect is of a breadboard crossed with a set of blinkenlights, and we could see that for simple digital circuits it could have some utility if not so much for higher frequency or analogue signals. Certainly it’s an experiment worth doing, and indeed it’s not the first tricked out breadboard we’ve seen.

Linux Fu: Supercharge Bash History

May 5, 2023 by Al Williams 18 Comments

Having a history of shell commands is a great idea. It is, of course, enormously handy when you have to run something repetitively or you make a simple mistake that needs correction. However, as I’ve mentioned in the past, bash history isn’t without its problems. For one thing, by default, you don’t get history in one window from typing in another window. If you use a terminal multiplexer or a GUI, you are very likely to have many shells open. You can make them share history, but that comes with its own baggage. If you think about it, we have super fast computers with tons of storage compared to the “old days,” yet shell history is pretty much the same as it has been for decades. But [Rcaloras] did think about it and created Bashhub, a history database for bash, zsh, and probably some other shells, too.

You might think you don’t need anything more than what you have, and, of course, you don’t. However, Bashhub offers privately stored and encrypted history across machines. It also provides context about commands you’ve executed in the past. In other words, you can see the directory you were in, the exact time and date, the system you were on, and the last return code of the command.

Continue reading “Linux Fu: Supercharge Bash History” →

Hackaday Podcast 217: The Unintentional Space And 3D Printing Episode

May 5, 2023 by Tom Nardi No comments

Hackaday Editors Elliot Williams and Tom Nardi definitely didn’t plan on devoting most of this episode to 3D printing and space stories, but let’s be honest, it was bound to happen sooner or later. After an update on the Hackaday Prize, the discussion moves on to a pair of troubled spacecraft and the challenges of exploring the final frontier. From there you’ll hear about a chocolate 3D printer we’ve had our eyes on for years, the tools you should have next to your own (non-chocolate) 3D printer, and a bit of contemplation of what it really means to design for 3D printing versus traditional manufacturing methods. But it’s not all plastic fantastic — by the end of the episode you’ll also hear about some particularly bold high-altitude aviators and the surprisingly short time we have left with the humble barcode.

Check out the links below if you want to follow along, and as always, tell us what you think about this episode in the comments!

Going to be stranded on a desert island? Download the podcast now!

Continue reading “Hackaday Podcast 217: The Unintentional Space And 3D Printing Episode” →

A History Of NASA Supercomputers, Among Others

May 5, 2023 by Dave Rowntree 8 Comments

The History Guy on YouTube has posted an interesting video on the history of the supercomputer, with a specific focus on their use by NASA for the implementation of computational fluid dynamics (CFD) models of aeronautical assemblies.

The aero designers of the day were quickly finding out the limitations of the wind tunnel testing approach, especially for so-called transonic flow conditions. This occurs when an object moving through a fluid (like air can be modeled) produces regions of supersonic flow mixed in with subsonic flow and makes for additional drag scenarios. This severely impacts aircraft performance. Not accounting for these effects is not an option, hence the great industry interest in CFD modeling. But the equations for which (usually based around the Navier-Stokes system) are non-linear, and extremely computationally intensive.

Obviously, a certain Mr. Cray is a prominent player in this story, who, as the story goes, exhausted the financial tolerance of his employer, CDC, and subsequently formed Cray Research Inc, and the rest is (an interesting) history. Many Cray machines were instrumental in the development of the space program, and now adorn computing museums the world over. You simply haven’t lived until you’ve sipped your weak lemon drink whilst sitting on the ‘bench’ around an early Cray machine.

You see, supercomputers are a different beast from those machines mere mortals have access to, or at least the earlier ones were. The focus is on pure performance, ideally for floating-point computation, with cost far less of a concern, than getting to the next computational milestone. The Cray-1 for example, is a 64-bit machine capable of 80 MIPS scalar performance (whilst eating over 100 kW of juice), and some very limited parallel processing ability.

While this was immensely faster than anything else available at the time, the modern approach to supercomputing is less about fancy processor design and more about the massive use of parallelism of existing chips with lots of local fast storage mixed in. Every hacker out there should experience these old machines if they can, because the tricks they used and the lengths the designers went to get squeeze out every ounce of processing grunt, can be a real eye-opener.

Want to see what happens when you really push out the boat and use the whole wafer for parallel computation? Checkout the Cerberus. If your needs are somewhat less, but dabbling in parallel computing gets you all pumped, you could build a small array out of Pine64s. Finally, the story wouldn’t be complete without talking about the life and sad early demise of Seymour Cray.
Continue reading “A History Of NASA Supercomputers, Among Others” →