Robot job swap, Llama Code 3 test, Stripe's AI moves

Intro

Ivy: A model that writes flawless Python... and then tries to compile Rust using a Python interpreter.

Marcus: I'm Marcus.

Ivy: And I'm Ivy.

Marcus: And this is AI Unfiltered for Thursday, August 20th, 2026.

Ivy: We've got a packed show for you today.

Marcus: We're putting Meta's new Llama Code 3 through its paces. Can the open-source hero really compete?

Ivy: Then, we’re meeting Robo-Cat, the new AI from Google DeepMind that can supposedly learn to use any robot you give it.

Marcus: And later, we dig into why Stripe really bought an AI startup. The official reason is… a bit of a stretch.

First Look: Testing Llama Code 3 for Python Tasks

Marcus: Alright, first up: Meta just dropped Llama Code 3, and as an open-source fan, I am psyched. I've been running tests all morning.

Ivy: And? Is it the GPT-4 killer for coders we were all hoping for?

Marcus: For Python? It's scary good. Boilerplate, entire functions, even decent unit tests—it just spits them out. A huge leap from version 2.

Ivy: Okay, but the official blog post has some... let's say carefully worded benchmarks. It leads with Python, but for languages like Rust, C++, or Javascript, performance falls off a cliff.

Marcus: It's still early! That's the trade-off with open models. They get really good at what the community trains them on, and right now, that's a lot of Python.

Ivy: But they're marketing it as a general 'Code Llama,' Marcus. Not 'Python Llama.' And my own tests confirmed it. I asked for a simple file-handling script in Rust.

Marcus: And?

Ivy: It gave me beautiful code that was, in fact, Python. Then it hallucinated a Rust compiler flag that was... a pip install command. It's not just weak outside Python, it's dangerously confident.

Marcus: [sighs] Okay, that is… not great. So it's a specialist.

Ivy: A very impressive one, for sure. But a specialist. So what's the takeaway here for our listeners?

Marcus: Right. If you're a Python developer, you need to download this today. It's a free, locally-runnable assistant that will absolutely speed up your workflow.

Ivy: But if you work in multiple languages, or anything more obscure, you're still better off with the big commercial models. Or, you know, your own brain.

The Robot That Can Learn Another Robot's Job

Ivy: Alright, next up, let's talk hardware. Google DeepMind just published a paper on something they're calling Robo-Cat.

Marcus: Robo-Cat! I love it. So this isn't just one robot, it's the brain for a bunch of different robots, right?

Ivy: That's the pitch. It's a 'generalist agent.' They took several different robot arms, each with different grippers and controls, and trained one single AI model to run all of them.

Marcus: And it worked! It could see a task, like 'pick up the apple,' and figure out how to do it with whichever arm it was connected to. It even learned new tasks from watching just 100 demos.

Ivy: It learned to control similar arms in a lab. This is a long, long way from a C-3PO that can suddenly drive your car or do your laundry.

Marcus: But it's the right architecture! Instead of building custom AI for every robot in the world, you build one 'driver' that adapts. This is how robotics scales.

Ivy: Or it's how you get an AI that's a jack-of-all-trades and master of none. A specialized arm in a car factory will always be better, faster, and safer than a generalist one.

Marcus: Think bigger, Ivy! A disaster zone. You airdrop a bunch of different robots—some with wheels, some with legs, some that dig—and one Robo-Cat brain possesses them all to coordinate a rescue.

Ivy: That's a movie plot, Marcus. It's fascinating research, but let's bring this back down to Earth. What does this mean today?

Marcus: Fine, fine. Today, it means robotics research just got a lot more efficient. You can test ideas on different hardware without starting from scratch every time.

Ivy: Exactly. It's a tool for researchers. The dream of a general-purpose robot assistant for your messy, unpredictable house is still just that—a dream.

Marcus: For now. I'm keeping an eye on this cat.

Stripe didnt really buy OpenRouter because of the singularity

Marcus: Alright, last story. Let's talk money. The payments giant Stripe just acquired a small AI startup called OpenRouter.

Ivy: And when TechCrunch asked why, the official quote was that it's related to Stripe's long-term belief in... 'the singularity.'

Marcus: [laughs] Oh, come on.

Ivy: That is such a transparently silly PR line. Let's break down what's really going on here.

Marcus: Please. What does OpenRouter actually do?

Ivy: It's a smart switchboard for large language models. A developer sends a prompt to OpenRouter, and it routes the request to the best model for the job, based on cost, speed, or power.

Marcus: So if I just need to summarize an email, I can tell it to use a cheap, fast model. But if I need to write complex code, it can automatically send me to GPT-4o or Claude 3 Opus.

Ivy: Precisely. And it handles the billing for all of them. Now... who does that sound like?

Marcus: Stripe. They're a transaction company. This has nothing to do with the singularity and everything to do with becoming the financial plumbing for the entire AI economy.

Ivy: Exactly. They want to be the universal API for AI costs. Every time a company chooses between a half-a-cent model and a ten-cent model, Stripe wants to process that choice and take its cut.

Marcus: It's brilliant. They're treating AI models like commodities. And in a commodity market, the one who owns the exchange wins.

Ivy: So what's the bottom line for developers and businesses?

Marcus: It means managing AI costs is about to get a lot easier, and probably standardized around Stripe. They're going to bake this right into their platform.

Ivy: And it means the most powerful force in AI right now isn't some god-like intelligence; it's economics. The race to the cheapest possible token for a 'good enough' result is the real story.

Marcus: So much for the singularity.

Ivy: Alright, let's recap. Meta's Llama Code 3 is a fantastic coding assistant... but only if you're a Python developer.

Marcus: Google's Robo-Cat is a cool proof-of-concept for a universal robot brain, but it’s staying in the lab for now.

Ivy: And Stripe buying OpenRouter isn't about sci-fi fantasies; it's a smart business move to become the bank for the AI industry.

Marcus: Before we go, quick one: did you see that concept for a smart mirror that gives you 'gentle, constructive feedback' on your outfit?

Ivy: Absolutely not. The last thing I need is my reflection telling me my shirt has 'suboptimal color synergy' at 7 AM.

Marcus: [laughs] I'd just be worried it would get stuck telling me to 'try turning it off and on again.'

Marcus: And that's our show! Thanks for listening to AI Unfiltered.

Ivy: We'll be back tomorrow. Until then, maybe double-check your compilers.