Open source versus closed source, memory chips eat AI profits, Comcast restructures

This episode revisits the open-source versus closed-source AI debate through the lens of Z.ai’s GLM 5.2. The model is discussed as an open-weight release strong enough to reopen questions about cybersecurity, geopolitics, and the durability of the U.S. frontier-model lead.

What the episode highlights

• GLM 5.2 can be downloaded, run, and modified outside a proprietary API, which makes it attractive and harder to supervise.
• Reported bug-finding performance puts open-weight models back into national-security and cyber-risk conversations.
• Benchmarks are only part of the story: cost per task, token hunger, and possible distillation matter just as much.
• The market appears to be splitting between expensive frontier models for critical work and small, cheap models for repeatable workflows.

Why it matters

If open-weight models remain close enough to frontier systems, access-control strategies around closed models become harder to enforce. A company waiting for approval to use a restricted frontier model may eventually find that a newer open-weight alternative is good enough.

The episode also points to infrastructure as the real bottleneck. Google reportedly capped Meta’s access to Gemini capacity, while memory makers such as Micron, Samsung, and SK Hynix are benefiting from soaring demand for HBM and DRAM. AI economics are therefore not only about model labs; cloud capacity and memory suppliers may capture a large share of the value.

Key takeaways

• Cost per completed task is becoming more important than price per token.
• Chinese open-weight models remain close enough to shape U.S. policy and business strategy.
• Compute and memory constraints are turning AI infrastructure into a strategic choke point.
• Comcast’s restructuring underscores how media, connectivity, streaming, and physical experiences are increasingly different businesses.

Source

• Chaîne: TBPN
• Vidéo source: https://www.youtube.com/watch?v=jgJYskq-1Yo