Ben and Ryan chat about how the human body can act as a resonance chamber for remote car keys, the potential for GenAI to revolutionize game development by creating an infinite array of dynamic game worlds, and advancements in brain-computer interfaces.
Now you know: The human body can serve as a resonance chamber for remote car keys, effectively extending their range.
A hackathon team used GenAI can create a fully playable D&D-style game in just one day.
Skybox AI from Blockade Labs allows users to generate 360° skybox experiences from text prompts.
A significant advancement in the brain-computer interfaces (BCI) space: a novel framework called DeWave integrates “discrete encoding sequences into open-vocabulary EEG-to-text translation tasks” without the need for “eye-tracking fixations or event markers to segment brain dynamics into word-level features.”
Shoutout to Stack Overflow user Vineeth Chitteti, who earned a Favorite Question badge with Is it possible to hit multiple pods with a single request in Kubernetes cluster?.
[intro music plays]
Ben Popper Hello, and welcome back to another episode of the Stack Overflow Podcast, a place to talk all things software and technology. I'm your host, Ben Popper, Director of Content here at Stack Overflow, joined as I often am by the illustrious editor of our blog and newsletter, Ryan Donovan. Hey, Ryan.
Ryan Donovan Hey, Ben. How’re you doing today?
BP I'm pretty good. So you sent me an interesting link for us to chat about today. This came up on the physics Stack Exchange. Everyone knows physics is one of the hardest things to wrap your head around. It acts spooky and weird at a certain level.
RD It's got some weird stuff.
BP It's got some weird stuff. So this made it to the top of Hacker News, I guess a pretty interesting question. You want to run this through for me?
RD Yeah. This got posted on Hacker News today. The question is from nine years ago, where somebody found that they were able to use the remote car key out of range by pressing it on their head and their chest. The idea is that the person got a little extra range by using their head and/or chest as a resonance chamber. And then later on in the answers, somebody actually does an experiment pressing it to all parts of their bodies, and there's some great pictures of this person pressing the key to their head.
BP Oh my God. Amazing. The experiment.
RD That's science.
BP Science at its finest. “This is a really interesting question,” says the top answer. It turns out your body is reasonably conductive, think saltwater– more on the answer to that in this question, you could go down a real rabbit hole on this one– in that it can couple to RF radio frequency sources capacitively. The effect is probably caused by one or more of the cavities in your body –your head or chest– acting as a resonance chamber for the RF signal. For example, think of how a resonance chamber can amplify waves– the hollow body of a guitar and the strings that are attached. So that's you. You're the guitar playing the remote. That's pretty cool. I like that one a lot.
RD I mean the conductiveness of the human body is one of the things that's sort of exploited by touchscreens. I think they use light conductivity to get there.
BP Right. Have you ever known or played on a theremin?
RD Oh, yeah. I've gone to science museums and played on a theremin.
BP You are acting as the L circuit. You don't have to touch it, you just have to move your hand around it. Very cool.
RD Just recreating all the old spooky movies from the 60’s and 70’s.
BP Yeah, exactly. The Wolfman Jack Radio Hour. All right, I had a neat one. This came from a Substack or something, but it read like a blog. Somebody went to the AI Tech Week Virtual Worlds Hackathon and they created a fully playable D&D role playing-style game in the span of one day. And the title of the article: “Semantic Programming and Software 2.0.” So basically, the conceit here is that they used Anthropic's AI model, Claude, which can work with a big 100k tokens of text, and then something called the Blockade Labs Skybox Generator, which will create dynamic 3D scenes based on the description. And so you can get a few random attributes at the beginning, and then from there, the rest of the user interface, what the user is seeing, is being generated in roughly real time, or maybe not roughly real time, but it's being generated by these Gen AI models. Now, it's not all newfangled stuff. It's a natural language to XML subsystem, so that's the way they exchange information between traditional coding and semantic inputs.
RD It's definitely an interesting use case of AI. I do think AI is going to become integrated into games a lot more, and I think there's already a sort of generative AI on bushes and trees. Nobody's modeling all the trees.
BP Yeah, there are these open source games like No Man's Sky that are procedurally generated and they just continue to expand and move in new directions, and Gen AI might soup that up and sort of follow you wherever you decide to go. There's no more side quests. Whatever you're interested in, that's what the game is about.
RD Right. You want to open the tree? Sure, there's a door in the tree. You can climb in, you can explore the tree universe.
BP Right, you've fallen into a new dimension
RD In the second part of this series, they talk about comparing this to old text mods– the old multi-user dungeons were you describe the room and now they can build a whole visual sequence off of that and use the room description as part of a vector database to determine how close people are.
BP That's cool.
RD I think one of the things that's most interesting to come out of this is the sort of rapid prototyping. You can have it generating stuff, and then if you like it, you can then export it as a sort of static model or this code be static or make the script static, but you can sort of let the computer surprise you.
BP You definitely should go and check out the Blockade Labs website. You basically give it a simple prompt, you dream up the world, and then it's infinite. You can just go from there and it'll try to cook up something and you can remix it when you choose to. A friend of mine was at Union Square Ventures, which is a well known VC shop in New York, and they have a little demo set up there where it's basically like a projector, a big screen and you type in whatever you want to say and they try in real time to create that environment for you and you can lead it on a bunch of different paths. My friend said it was a fun and immersive experience that sometimes worked really well and you felt like you were just manifesting this dream, and sometimes it didn't work at all or lagged out on you.
RD Right. Right now we're at the sort of “Oh, cool toy” stage of things where it's like, “Oh, what a fun little tech demo they put together,” but I don't know, what are we– five years away from the first AAA game to sort of generate side quests and infinite dungeons?
BP What does the AAA game bring at that point? Maybe they bring the architecture underneath that does it really well, but if you're not writing the story and you're not deciding what the maps look like, what do I need you for? But they'll take a good franchise and they'll spin it that way.
RD You just subscribe to it because you're going to be costing them money on each of your steps forward. You're going to be calling to their generative models.
BP Ooh, I don't like that. The deeper you go, the more you pay. I don't like the pay to explore, I don't like that.
RD The SaaS model is going to take over.
BP That's a grim prediction, my friend. Pay by the minute. Uh-oh, this sounds very Blade Runner. All right, I had one other one that I wanted to share with you about grounding LLMs to prevent hallucinations. So this team created something they call ‘Wikichat’ and they took three different LLM models– and it is model agnostic. You could use other ones, but they took GPT-3.5, GPT-4, and LLaMa, and they set it up so that it only drew on the context of Wikipedia and it had a bit of a multi-step process. You'd get the query: Have you heard about Oppenheimer? Oh, what do you think about the casting? So go read Wikipedia, summarize and filter the content, generate your answer, refine what you generated, fact check it, draft it, refine again, and then answer. And they found that this could increase the accuracy of these conversations to the point where, let's see, “Our best system achieved 97.3 factual accuracy in a simulated conversation, and that is 55% better than GPT-4.” So quite a jump.
RD That's pretty good. So the multi-prompt thing is just chain of thought, verifying what it's saying using vector databases underneath?
BP I'll walk you through the TL;DR here. Stage one, you generate a search query that captures the user's interest with the prompt. Stage two, Wikichat extracts the relevant sections of the retrieved passages from the ground truth, which is Wikipedia. You prompt the LLM to generate a response to the history of the conversation. It contains interesting and relatable knowledge, but is inherently unreliable and prone to hallucination. Then that response is broken down to multiple claims and then you go through a verification prompt, chain of thought prompting to assign each claim to one of three classes: retrieved evidence supports the claim, refuses the claim, or there is not enough information in the evidence to make this decision. Only claims that are supported by evidence are kept. So this deals with two fundamental issues of large language models, one being that they're trained on this enormous corpus of text and if you ask them about current events or things that are less well known, they tend to hallucinate, and two, they don't give you a confidence score. They just respond and it's up to you to then go check the facts. So this has a new sort of stage within it where they try to check the facts, and then if the facts are verified, they can respond. If they're not, they'll cut it out of the answer, and if nothing is found, they'll just say we couldn't find anything about this.
RD I think that's one of the end goals for a lot of people– making sure it's verified and not making stuff up. There's still folks who are wary about the accuracy of Wikipedia just because it's open to editing from everyone, but I think Wikipedia is pretty reliable at this point. It's pretty well monitored by mods.
BP Yeah. All right, one last thing before we go. My favorite subject, the brain-computer interface. I love this stuff. So here's a new paper. It doesn't say when this was released, but it was just shared with me today so I think it's pretty new. This was the first one to work directly from brain waves without putting any markers in there, meaning that you don't look to see what the person's brain waves look like when they're reading, or you don't have them repeat a certain letter a bunch of times, you don't have them say a few words. You just look at the brainwaves, and then you predict on that. So Bob attended the University of Texas at Austin where he graduated Phi Beta Kappa, and it predicted the University of California at Austin where he studied in Beta Kappa, a degree in Latin American Studies, a degree in History and American Studies. 1973, they said 1975, taking only two years to complete his work. Taking a half a year to complete the degree, and generally obtaining excellent grades and was excellent grades. So it gets the gist of it. It gets some of it perfect. Of course, the fact that it's coming straight from your brain without any sort of setup or fine-tuning to get it to know your brain is the wild part.
RD Yeah, and it's doing this without any invasive tech. It's just scanning the electrical signals from your head.
BP Oh, good. Excellent point. Yes, exactly. So it sort of has the possibility to be much more widely applied.
RD Man.
BP All right, bringing you a few notes from the future here.
RD A few notes from the future. Instead of turning off your microphone around your phone, you'll have to take off your wearable hat.
BP Right. Well the tinfoil hat will finally come back, the thing that blocks the EEG or whatever. It'll finally make sense.
RD They're Faraday cages, Ben.
BP Exactly, your portable Faraday cage.
[music plays]
BP All right, everybody. It is that time of the show. Let's shout out a user who came on Stack Overflow and helped to spread a little knowledge. A Favorite Question Badge was awarded to Vineeth Chitteti: “Is it possible to hit multiple pods with a single request in a Kubernetes cluster?” If you want to hit multiple pods, that is the question and there is an answer for you, and helped over 32,000 people, so we appreciate the curiosity and the knowledge. As always, I am Ben Popper, Director of Content here at Stack Overflow. You can find me on X @BenPopper. Hit us up, podcast@stackoverflow.com if you have questions or suggestions for the program. And if you liked it, leave us a rating and a review.
RD I'm Ryan Donovan. I edit the blog here at Stack Overflow. You can find it at stackoverflow.blog. And you can reach out to me on X @RThorDonovan.
BP Thanks for listening, and we'll talk to you soon.
[outro music plays]