Meta Drops New Model, Mythos, RoboLamp
TL;DR
Meta finally went closed with Muse Spark — After years of Llama-style open releases, Meta’s new model is closed-source, which the hosts frame as the predictable endgame John Luttig called back in 2024 as CAPEX climbs into the $10B+ range and shareholders demand ROI.
Muse Spark looks competitive but not dominant — and the personalization story feels half-baked — On Meta’s own charts it beats rivals on some benchmarks but trails badly on others like ARC-AGI-2, while the hosts’ live testing found it weirdly suggesting “Malibu appropriate surf puns” yet denying it knew anything personal, which undercuts the whole “personal superintelligence” pitch.
Anthropic’s Mythos is being treated like a cyber weapon, not a normal model launch — Access is limited to roughly 50 critical-infrastructure partners including Apple, Google, Microsoft, Amazon, Nvidia, JP Morgan Chase, Cisco, CrowdStrike, and Palo Alto because the model is reportedly strong at finding zero-days and exploits.
The next AI race is about scale, compute, and gated access — The hosts connect Mythos, Muse Spark, and xAI’s rumored seven-model training lineup (including 1T, 1.5T, 6T, and 10T variants) to a world where frontier models may increasingly stay private, expensive, and available first to the highest-value enterprise buyers.
RoboLamp was the day’s most charming hardware demo — Aaron Tan’s Loom lamp folds shirts, remakes beds, drops laundry, and even plays a record, and the hosts’ reaction was basically that this “Pixar lamp” form factor might work precisely because people already want lamps and only need one or two useful chores, à la Roomba, for the product to click.
YC’s Luther Lowe says Apple is the real bottleneck for vibe coding on mobile — His argument is that AI has made app creation radically easier, but the App Store remains “the worst DMV in the world,” with review friction, self-preferencing, and hardware-level lock-in keeping new AI-native apps and assistants from reaching users freely.
The Breakdown
Meta’s new model lands, and the open-source era may be over
The show opens with the big AI headline: Meta launched Muse Spark, its first major model in over a year and, more importantly, its first big closed model after years of open-source posturing. The hosts connect that move to John Luttig’s 2024 argument that Meta would stay open only “as long as it’s helping us,” especially once training costs hit the $10 billion zone and Wall Street starts asking harder ROI questions.
Muse Spark benchmark drama and the weirdly personal joke test
They pull up the model card and immediately get into “chart crime” territory: Muse Spark is highlighted in blue and looks like the winner at first glance, but once you read the rows carefully, it beats some frontier models and clearly loses on others, including ARC-AGI-2. Then the live demo gets funnier — asked for a joke, Meta offers “Malibu appropriate surf puns,” which feels oddly specific, then repeatedly denies knowing anything personal, creating a very Meta moment where the product hints at personalization while refusing to admit it.
Why Meta might really be building this: economics, internal coding, and future video
The hosts frame the bigger strategic question as simple economics: if Meta engineers have been burning huge amounts of Claude tokens internally, owning the model stack could turn expensive opex into cheaper capex and inference on Meta’s own hardware. They note reports of an internal token dashboard showing more than 60 trillion tokens over 30 days before it was taken offline, and they sound especially curious about what comes next — image/video models like the rumored “Mango,” where Meta’s massive proprietary data advantage might matter more.
Mythos enters like a restricted cyber model, not a consumer chatbot
From there they move to Anthropic’s Mythos, which got maybe the most dramatic framing of the episode: a model preview available only to around 50 critical-infrastructure organizations because it’s unusually capable at bug-finding, exploit generation, and cybersecurity work. They run through the partner list — Apple, Google, Microsoft, Amazon, Nvidia, JP Morgan Chase, Broadcom, Linux Foundation, Cisco, CrowdStrike, Palo Alto — and talk through why cyber is the perfect reinforcement-learning domain: binary rewards, no real-world lab delay, just endless virtual machines and “did you break in or not?”
The backlash to Anthropic’s safety theater — and why the hosts still take the capability jump seriously
They also give airtime to the backlash: George Hotz, Mike from Also Capital, and others mocking the “too dangerous to release” framing as familiar AI marketing. But even with the eye-rolling, the hosts’ real takeaway is that something meaningful has changed — frontier labs may now have models whose weights really matter geopolitically, whose release cadence is gated by compute and safety, and whose best versions may increasingly be invisible to the public.
Elon’s seven-model rumor mill, Nvidia mania, and the next compute wave
The conversation widens into a broader scaling-laws bull case: Martin Casado says pretraining isn’t saturated, RL works, and Blackwell-era training is just getting started. They mention Elon Musk saying xAI has seven models in training, with rumored sizes ranging from 1 trillion to 10 trillion parameters, and pair that with wild Nvidia upside chatter — including a $22 trillion valuation model — to capture the “buckle your chin straps” energy around the next compute cycle.
RoboLamp steals the vibe with a Pixar-friendly home robot demo
Then the mood lightens with Aaron Tan’s Loom lamp, a domestic robot disguised as a warm, friendly lamp that can drop laundry, fold shirts, remake beds, and even place a record on a turntable. The hosts genuinely love the demo’s tone — “doesn’t feel dystopian, feels delightful” — and land on a practical insight: if the form factor is already something people want in a bedroom, the robot only needs to do a few useful chores reliably for the Roomba-style adoption flywheel to start.
YC’s Luther Lowe goes after Apple, calling the App Store “the worst DMV in the world”
In the first interview block, Lowe argues that AI has democratized app creation the way browser-based HTML editors democratized websites 25 years ago — but Apple and Google now sit between makers and users. His sharpest line is that once a computer fits in your pocket, your freedom disappears; he argues side-loading, alternative app stores, and anti-self-preferencing policy like California’s proposed BASTA Act are necessary if AI assistants and vibe-coded apps are actually going to reach consumers at scale.