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

Overview

This episode dives into Meta's latest AI developments, particularly the launch of its new closed-source language model, Muse Spark, and its implications for the AI landscape. The discussion also touches on Anthropic's Mythos model, the evolving economics of AI, and the broader implications of proprietary data and compute constraints in the AI arms race.

Notable Quotes

- The future of foundation models is closed source. – John Coogan, reflecting on Meta's shift from open-source AI.

- Trained AI models are the fastest depreciating asset in history. – John Coogan, on the fleeting value of cutting-edge AI models.

- What is personal superintelligence if it doesn’t even know your name? – John Coogan, critiquing the personalization gaps in AI systems.

🧠 Meta’s Shift to Closed-Source AI

- Meta launched Muse Spark, its first closed-source AI model, signaling a departure from its open-source strategy.

- John Coogan highlighted how Meta's open-source approach was initially driven by a desire to avoid dependency on third parties, like Apple, but noted that cost and safety concerns are now driving the shift.

- The decision aligns with predictions by John Luttig, who foresaw Meta's eventual pivot to closed-source AI due to economic and strategic pressures.

- Muse Spark is designed to integrate across Meta's ecosystem, enhancing products like Facebook and Instagram while reducing reliance on external providers.

📊 Benchmarking and Performance Controversies

- Muse Spark reportedly outperformed competitors like XAI's Grok on internal benchmarks but underperformed on others, such as ArcAGI2.

- Jordi Hays criticized Meta's presentation of benchmark results, calling it a chart crime for potentially misleading comparisons.

- The team discussed Meta's history of gaming benchmarks, including controversies surrounding its previous model, Llama 4.

💾 Proprietary Data and Compute Economics

- The podcast emphasized the growing importance of proprietary data in differentiating AI models, as public internet data becomes commoditized.

- John Coogan noted that Meta's massive CapEx investments in AI hardware, like H100 GPUs, allow it to train models more cost-effectively in-house.

- The hosts debated whether internal use alone can justify the high costs of developing advanced models, especially for tasks like code generation.

🔒 Anthropic’s Mythos and AI Safety Concerns

- Anthropic's Mythos model, designed for cybersecurity applications, was discussed as a potential game-changer in finding zero-day exploits.

- The model is currently restricted to select partners, including major tech companies and financial institutions, to mitigate risks of misuse.

- John Coogan and Jordi Hays explored the tension between safety concerns and the economic pressures to release powerful models, drawing parallels to OpenAI's earlier decisions with GPT-2.

🌐 The AI Arms Race and Market Dynamics

- The episode highlighted the competitive landscape, with Meta, Anthropic, OpenAI, and Google racing to develop larger, more capable models.

- John Coogan speculated on the emergence of a seller's market for compute, where access to the best models could be limited to the highest bidders.

- The hosts discussed Elon Musk's ambitious plans for XAI, including the training of multiple trillion-parameter models, as part of his effort to catch up in the AI race.