{
  "version": "1.2.0",
  "chapters": [
    {
      "title": "Cold open",
      "startTime": 0,
      "endTime": 62.79
    },
    {
      "title": "Who’s Neel Nanda?",
      "startTime": 62.79,
      "endTime": 119.87
    },
    {
      "title": "How would mechanistic interpretability help with AGI",
      "startTime": 119.87,
      "endTime": 309.49
    },
    {
      "title": "What's mech interp?",
      "startTime": 309.49,
      "endTime": 587.34
    },
    {
      "title": "How Neel changed his take on mech interp",
      "startTime": 587.34,
      "endTime": 953.4
    },
    {
      "title": "Top successes in interpretability",
      "startTime": 953.4,
      "endTime": 1206.3
    },
    {
      "title": "Probes can cheaply detect harmful intentions in AIs",
      "startTime": 1206.3,
      "endTime": 1609.81
    },
    {
      "title": "In some ways we understand AIs better than human minds",
      "startTime": 1609.81,
      "endTime": 1761.88
    },
    {
      "title": "Mech interp won't solve all our AI alignment problems",
      "startTime": 1761.88,
      "endTime": 2287.46
    },
    {
      "title": "Why mech interp is the 'biology' of neural networks",
      "startTime": 2287.46,
      "endTime": 2428.36
    },
    {
      "title": "Interpretability can't reliably find deceptive AI – nothing can",
      "startTime": 2428.36,
      "endTime": 2979.35
    },
    {
      "title": "'Black box' interpretability — reading the chain of thought",
      "startTime": 2979.35,
      "endTime": 3186
    },
    {
      "title": "'Self-preservation' isn't always what it seems",
      "startTime": 3186,
      "endTime": 3729.24
    },
    {
      "title": "For how long can we trust the chain of thought",
      "startTime": 3729.24,
      "endTime": 4299.82
    },
    {
      "title": "We could accidentally destroy chain of thought's usefulness",
      "startTime": 4299.82,
      "endTime": 4616.46
    },
    {
      "title": "Models can tell when they’re being tested and act differently",
      "startTime": 4616.46,
      "endTime": 5030.96
    },
    {
      "title": "Top complaints about mech interp",
      "startTime": 5030.96,
      "endTime": 5872.04
    },
    {
      "title": "Why everyone's excited about sparse autoencoders (SAEs)",
      "startTime": 5872.04,
      "endTime": 6436.19
    },
    {
      "title": "Limitations of SAEs",
      "startTime": 6436.19,
      "endTime": 6889.92
    },
    {
      "title": "SAEs performance on real-world tasks",
      "startTime": 6889.92,
      "endTime": 7690.89
    },
    {
      "title": "Best arguments in favour of mech interp",
      "startTime": 7690.89,
      "endTime": 7923.68
    },
    {
      "title": "Lessons from the hype around mech interp",
      "startTime": 7923.68,
      "endTime": 8270.05
    },
    {
      "title": "Where mech interp will shine in coming years",
      "startTime": 8270.05,
      "endTime": 8462.95
    },
    {
      "title": "Why focus on understanding over control",
      "startTime": 8462.95,
      "endTime": 8649.12
    },
    {
      "title": "If AI models are conscious, will mech interp help us figure it out",
      "startTime": 8649.12,
      "endTime": 8779.79
    },
    {
      "title": "Neel’s new research philosophy",
      "startTime": 8779.79,
      "endTime": 9511.35
    },
    {
      "title": "Who should join the mech interp field",
      "startTime": 9511.35,
      "endTime": 10015.08
    },
    {
      "title": "Advice for getting started in mech interp",
      "startTime": 10015.08,
      "endTime": 10481.59
    },
    {
      "title": "Keeping up to date with mech interp results",
      "startTime": 10481.59,
      "endTime": 10664
    },
    {
      "title": "Who’s hiring and where to work?",
      "startTime": 10664,
      "endTime": 10870.57
    }
  ]
}