Episode publish date
January 7, 2026 6:30 AM (UTC)
Last edit date
Jan 24, 2026 5:03 PM
Last snip date
January 24, 2026 5:02 PM (UTC)
Last sync date
January 24, 2026 5:03 PM (UTC)
Show
Last Week in AI
Show notes link
Snips
21
Warning
⚠️ Any content within the episode information, snip blocks might be updated or overwritten by Snipd in a future sync. Add your edits or additional notes outside these blocks to keep them safe.
‣
Your snips
‣
[02:01] 2025's Core AI Trends
‣
[03:39] Interpretability's Evolving Role
‣
[04:51] Hardware And ROI Became Central
‣
[07:10] The 'Acquire-But-Not-Acquire' Playbook
‣
[26:46] Plan For Inference-First Hardware
‣
[29:41] Inference Rising vs Training Scale
‣
[39:15] Micron's Quiet HBM Surge
‣
[42:17] China's Retrofit Strategy For Lithography
‣
[47:59] Open-Source Models Closing The Gap
‣
[50:15] Harder Benchmarks For Scientific Research
‣
[54:42] From LLM Outputs To Causal Graphs
‣
[57:34] Convergence Of Scientific Representations
‣
[01:02:17] Meta-RL Boosts Exploration And Adaptation
‣
[01:07:24] Long-Horizon Tasks Increase Inference Costs
‣
[01:11:17] Meter Plots Are Noisy And Threshold-Sensitive
‣
[01:13:38] Agentic Workloads and Success Rates
‣
[01:20:50] Monitor Internals With Activation Oracles
‣
[01:27:13] Prioritize Chain-Of-Thought For Monitoring
‣
[01:29:40] Monitorability Tax
‣
[01:32:10] Centralizing Preparedness At Frontier Labs
‣