The voice interface of the Star Trek holodeck allowed users to create items by saying things like “create a table” and “now turn it into a steel table,” etc. , all with quick feedback. This type of interface would have arguably been a natural fantasy at launch, however, with the advent of AI and LLMs (large language models), this type of natural language interface is almost setting itself up.
[Dominic Pajak’s] demo task called VoxelAstra is a fun demonstration of this. This is a WebXR demo that works on the Meta Quest 3 VR headset (just skip to the demo page in the headset’s internet browser) as well as on desktop.
The challenge is that since the program uses OpenAI APIs on the backend, you want to provide a running OpenAI API key. Otherwise, the demo may not be able to do anything. Providing your API key to someone’s webpage is rarely a very smart security practice, but there’s also the option to run the demo locally.
Either way, once the demo is up and running, the user simply tells the formula what to create. Keep it simple. It’s first and foremost a fun and educational demonstration, and you’ll try to do your homework with primitive shapes like spheres, cubes, and cylinders. “Build a snowman” is recommended as a starting point.
Are you intrigued by what you see and have your own concepts?WebXR can be a wonderful way to bring those concepts to life and seeing how someone else did something similar is a smart way to start. Check out [Dominic’s] other WebXR projects: a simulation from BBC Micro, in virtual reality.
There are many claims circulating about the functions of AI formulas, as the generation continues to climb the dizzying heights of the advertising cycle. Some of them are true, some expand the definitions a bit, while others cross the line and are permanently wrong. . [J] has one that is subsidized through genuine code, a text compression formula that an AI employs, and while its technique might have limitations, it demonstrates an attractive feature of giant language models.
Compression works by assuming that for a large enough style, chances are that there are many source texts somewhere in education. Using llama. cpp, it is possible to extract tokenization data from a piece of text contained in your educational knowledge and store it. as a compressed output. The decompressor can then use this tokenization knowledge as a series of keys to reassemble the original from its formation. We’re not experts on AI, but we believe that source text that has little is not unusual with any educational text. It would do badly, and we expect the same style to be used for compression and decompression. However, it’s still an attractive strategy, and possibly because it comprises AI fairy dust, there’s a blind venture capitalist out there who would pay millions for it. How global we live!
Oddly enough, this is the first time we’ve looked at AI text compression.
[Andrej Karpathy] recently published llm. c, a task focused on natural C LLM training, showing once again that running with that equipment doesn’t have to rely on extensive progression environments. GPT-2 may possibly be older, but it’s perfectly relevant, being the granddaddy of modern LLMs (large language models) with an apparent legacy of more modern offerings.
LLMs are incredibly smart at communicating even if they don’t know what they’re saying, and their regular education is based on the PyTorch deep learning library, written in Python. llm. c uses a simpler technique by directly implementing education in neural networks. ruleset for GPT-2. The result is very specific and oddly short: about a thousand lines of C in a single file. This is a very sublime procedure that does the same thing as the larger, clumsier methods. It can run entirely on a CPU or take advantage of GPU acceleration if available.
It’s not the first time [Andrej Karpathy] has used his abundant skills to summarize those kinds of concepts into undeniable implementations. We’ve already covered one of their projects, which is the “hello world” of GPT, a little style that predicts the next part of a given series and gives low-level insights into how GPT (pre-trained generative transformer) styles work.
What if there was a magical device that could scan all your LEGOs and tell you what you can do with them?It’s the dream of your formative years come true, right?Well, this device is in your pocket. Simply toss your LEGO stash on the mat, roll it out so that only one layer remains, scan it with your phone, and after a short wait, you’ll get a list of all the fun things you can do. With structure instructions. And yes, it shows you where each of the bricks in the pile is.
We’re talking about the BrickIt app, available for Android and Apple. Find out in the short demo after the break. After personally looking at the app, we can say that it does what it claims to do and is pretty good.
While it might hurt to have to pick up all those bricks when you’re done, it works best on a neutral background like a light-colored rug. In an attempt to keep the bricks in circles, we tried a wooden tray, and it didn’t look in the paintings as well as it probably could have as well – it didn’t have a lot of bricks and they can’t extend that far either.
And the only genuine problem is that the effects are limited because there is a paid version. And the app constantly reminds you of what you’re missing. But it’s still great, so check it out.
We don’t want to tell you how flexible LEGO is, but have you noticed that keyboard stand or PCB vise?
Continue “LEGO AI = A Brickton of Ideas” →
One of the scariest facets of AI as we know it today has been the proliferation of the deepfake generation that allows you to take nude pictures of anyone you want. What if you got rid of abstraction and put the trickster and the subject in the same space?That’s what the NUCA camera was designed for. [via 404 Media]
[Mathias Vef] and [Benedikt Groß] designed the NUCA camera “with the aim of critiquing the existing trajectory of AI symbol generation. “The camera itself is modest, a virtual camera published in 3D (19. 5 × 6 × 1. 5 cm) with a 37mm lens. When the camera’s shutter button is pressed, a simple theme symbol is generated.
The final symbol is generated from a combination of the photo taken of the subject, pose data, and facial landmarks. The photo goes through a classifier that identifies characteristics such as age, gender, frame type, etc. , and then uses them to generate a spark of text for a solid delivery. The subject’s original face is then sewn onto the nude symbol and aligned with the estimated pose. Many pattern symbols on the project’s online page show the bias in favor of good-looking safe ideals from AI datasets.
Looking for more tactics for using AI with cameras?How about this one that uses GPS to visualize a scene?Would you rather keep AI out of your efforts to invade your non-public space?How about building your own ASD picture scanner?
Some LLMs (large language models) can serve as useful programming assistants when provided with a project’s source code, but experimenting with this can be a bit tricky if the chatbot has no way to download it from the internet. In such cases, the code will need to be provided by pasting it into the Spark Off or by uploading a record manually. This is appropriate for undeniable things, but for more complex projects, it temporarily becomes inconvenient.
To facilitate this, [Eric Hartford] created github2record, a Python script that generates a single text record containing the combined source code of a specific repository. This text log can be downloaded (or pasted into the message), which makes it much more useful. Percentage code with chatbots.
Continue reading “Flushing a Code Repository as a Text Record to Share with Chatbots” →
When a new generation comes along, there are a large number of other people who need to let you know. Some need to teach you how to use their tools, some need you to pay for education, and some will use flexible education to trap you into acquiring more education. Since AI is the new buzzword, there are plenty of free courses from trusted sources. The wonderful thing about a free course is that if you find that it doesn’t do it for you, there’s no penalty for dropping out. .
We’ve found that NVIDIA, one of the corporations that has benefited the most from the AI boom, offers some courses (but not all of them are free). We were struck by the explanation of generative AI and the construction of its LLM using augmented retrieval generation. There’s also How to Build a Brain in 10 Minutes and Introduction to Physics-Based Machine Learning with Module. However, this is all brief.
Continue reading “A Crowd of AI Catch-Ups” →