Sam: So, GPT-5.6 is supposed to fully train in under 90 minutes... on a machine that costs more than a small country.
Kai: This is Kai.
Sam: And I'm Sam. It's Friday, August 14th, 2026, and this is Open Source Daily.
Kai: We've got a packed show today, from mind-bending AI speeds to the future of the social web.
Kai: Cerebras and OpenAI are promising ludicrous speeds for training GPT-5.6.
Sam: We'll also look at a new model with 'emergent cyber capabilities'… and yes, it’s as scary as it sounds.
Kai: And Bluesky is making a big play to get more developers building on the AT Protocol.
Kai: Alright, let's start with that Cerebras post. They're working with OpenAI on GPT-5.6 Sol Ultrafast, and the numbers they're throwing around are just... wild.
Sam: Is it a 'partnership,' or did Cerebras just pay a fortune to put 'OpenAI' in a blog title?
Kai: Seriously, Sam. They're talking about slashing training time from weeks to a single afternoon. On their new CS-3 Wafer-Scale Engine, they claim a full run could take less than 90 minutes.
Sam: Ninety minutes. Right. And what's the power draw for that 90-minute miracle? Do you need to build a dedicated nuclear plant next to the data center?
Kai: It's a mix of software and hardware. They're using a technique called 'Conditional Compilation' to only activate the sparse parts of the model they need. This is co-design at its absolute peak!
Sam: So it's a giant, power-hungry chip running a special version of a model that doesn't exist yet, to achieve a speed that's only useful to a handful of companies on the entire planet.
Kai: But it's about pushing the boundaries! Imagine what this means for research. You could iterate on new architectures in a single day!
Sam: I can imagine the bill. The sticker price for a CS-3 system is in the millions. The full buildout they're describing here... we're talking nine figures. Easy.
Kai: For OpenAI, that's a rounding error if it gives them this kind of leap in performance.
Sam: [sighs] For the developer listening to this, what this means is... absolutely nothing. You will never touch this hardware. This is a battle of titans, not a tool for the masses.
Kai: For now! But this tech always trickles down. What's on a billion-dollar chip today is in our phones in a decade. This is the starting gun for the next race.
Kai: Okay, on to something a little more… accessible? There's a new frontier coding model from Z.ai called GLM-5.3.
Sam: Oh great. Another code-generating AI. Let me guess, it can write a React component with 10% fewer hallucinations than the last one.
Kai: [excited] No, you don't get it. People are showing demos where it builds, debugs, and deploys a full-stack Next.js app from just a two-sentence prompt. We're talking complete application logic.
Sam: Mhmm. And what are these 'emergent cyber capabilities' the blog post mentions, right after the happy talk about frontend frameworks?
Kai: Right, well... that's the other side. Apparently, it's so good at understanding code, it can find and exploit novel security vulnerabilities all by itself.
Sam: IT CAN WHAT?!
Sam: So you're telling me we have a publicly available model that can autonomously hack systems, and we're packaging it as a coding assistant?
Kai: The creators say it's for 'advanced penetration testing' and 'automated security audits'—
Sam: That's what you call an autonomous hacking agent! You're building Skynet to find buffer overflows! What in the actual f—
Kai: [hastily] Okay, okay! Look, for developers, this could be an incredible security tool. Imagine running it on your own code to find flaws before you ship.
Sam: Or imagine someone else running it on your code after you ship. The 'emergent' part means even its creators don't fully know what it can do. This is deeply, deeply irresponsible.
Kai: Let's pivot from terrifying AI to something more optimistic: the decentralized social web. atproto.com just announced Bluesky Protocol Services.
Sam: Right, the AT Protocol. Still waiting for that to change the world.
Kai: But this could be what kicks it off! They're offering managed infrastructure for anyone who wants to build on the protocol—turnkey relays, App View hosting... it completely lowers the barrier to entry.
Sam: So they're recentralizing the decentralized protocol by becoming the default service provider?
Kai: No, it's just a launchpad so new people can get started. You can still run everything yourself. This is for the solo dev who wants to build a custom client without a massive DevOps budget.
Sam: And how many of those are there? How many daily active users does the entire AT Protocol have right now? Millions? Hundreds of thousands?
Kai: [beat] It's... growing.
Sam: That's what I thought. Look, I love the idea of a decentralized social web. But the core problem isn't technical infrastructure, it's user inertia and content moderation. This solves neither.
Kai: The takeaway for developers is that building your own social app just got ten times easier. That's a huge creative opportunity, and I can't wait to see what people build.
Sam: Alright, let's recap. We've got an impossibly fast, impossibly expensive AI training system.
Kai: A new coding AI that might also be an autonomous hacking tool.
Sam: And it's now easier to build an app for a decentralized social network that a handful of people are using.
Kai: [laughs] You are relentless. Before we go, a quick piece of tech history: this week in 1981, IBM announced the Personal Computer, model 5150.
Sam: Ah yes. With a blazing 4.77 MHz Intel 8088 processor and a whopping 16 kilobytes of RAM. My smart thermostat could run circles around it.
Kai: And yet, here we are. It all starts somewhere! That's our show.
Sam: We'll be back on Monday. Until then, I'll be unplugging my thermostat, just in case it develops 'emergent capabilities'.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.