← All episodes

Meta’s AI Comeback Moment, Claude Mythos | Diet TBPN

| 9 products mentioned
TBPN TBPN host
Watch on YouTube artificial intelligence ai models meta platforms frontier models anthropic open source vs proprietary ai safety

Hosts discuss Meta's AI comeback with the launch of Muse Spark, a closed-source frontier-level model that marks Meta's strategic shift away from open-source AI amid rising computational costs and competitive pressure from OpenAI, Anthropic, and Google. The episode contextualizes Meta's move within broader AI industry dynamics, including Anthropic's Mythos model release and its careful rollout strategy to prevent exploit weaponization, while examining what frontier AI access means for competitive advantage and market consolidation. Key insight: as AI models become more capable and expensive to train, the "best models may well not be in public"—shifting power to companies that can afford compute access and raising questions about who controls the future of AI infrastructure.

Key takeaways
  • Meta is shifting from open-source to proprietary AI because frontier models at $10B+ training costs demand clear shareholder ROI, and open-sourcing expensive models no longer makes competitive sense when internal use cases (feed algorithms, recommendations, image generation) provide sufficient justification.
  • Benchmark rankings are becoming unreliable signals of actual model quality; labs that previously gamed benchmarks should be viewed skeptically, and product performance ("the vibe") matters more than narrow test scores—meaning builders should test models directly rather than rely on highlighted charts.
  • Anthropic's gated release of Mythos to 50 critical infrastructure companies (Apple, Google, Microsoft, JP Morgan Chase, etc.) reveals a compute scarcity strategy: restrict access to manage demand, justify premium pricing tied to zero-day vulnerability prevention, and prevent Chinese model makers from distilling the technology.
  • Trained AI models depreciate rapidly (GPT-4 costs dropped 99%+ in two years); companies have incentive to gate-keep and monetize frontier models through high-margin API access or exclusive partnerships rather than broad public release, creating an economy where compute becomes a seller's market.
  • Cyber offense is frontier AI's first killer app—finding zero-day exploits offers clear, fast feedback loops and measurable rewards (system compromised or not), making it ideal for reinforcement learning agents; beneficial security applications (hardening infrastructure) require deliberate staging to prevent widespread weaponization.
  • Elon Musk's XAI is training seven models simultaneously (variants at 1T, 1.5T, and 10T parameters), signaling that frontier labs will continue aggressive model proliferation despite current market saturation, keeping pressure on pricing and access.

Recommendations (1)

Meta AI
Meta AI uses

"I have I of course we started asking it jokes right away. So I'll pull up the joke. I said, 'Tell me a joke.'"

TBPN · ▶ 4:57

Mentioned (8)

Muse Spark "The roll out of the model called Muse Spark is a critical moment for Meta which is up seven and a..." ▶ 0:18
Claude
Claude "what Okay. What does it mean if if the entire company has been like maxing their their Claude tok..." ▶ 8:46
Llama
Llama "And Llama has been great developer marketing for Facebook" ▶ 1:37
Mythos "anthropic new model mythos which sort of was announced loosely and and the model card dropped yes..." ▶ 3:54
GPT-2 "Open AI says its text generating algorithm GPT2 is too dangerous" ▶ 17:12
ChatGPT
ChatGPT "This was the story of GPT3 through 2, the story of of Chat GBT" ▶ 16:59
Grok
Grok "It significantly outscored XAI's Grock on most tests" ▶ 7:24
H100
H100 "He bought hundreds of thousands of H100s for improving social feed algorithms across products" ▶ 1:29