videos

]]

• Need DM to explain why the tempature fluctations where 10x smaller than expected. There already had to be structure there at the time of last scattering…
• Di-pole anamoly… basicaly the velocity inferred from the CMB doesnt match the velocity inferred from other measures

• great video
  • [>>] Bone up on TMUX
  • [>>] fzf
  • [>>] SSH Port Forwarding Into Remote Containers
  • [ ] pass: Unix Password Management With GPG Encryption
• [>>] Process

• 50% unemployment = longer childhood + more education + shorter work hours + bigger retirement
• better lives

• Mix of optimism and realism
• Jobs in turning the world into data.
• Energy / compute

• Good discussion of impact of AI on production function

• Total vs Active Model parameters:

  In a dense transformer (GPT-3, Llama 3, etc.), every parameter participates in every forward pass, so active = total. In an MoE model, the FFN block at each layer is replaced by N separate "expert" FFNs plus a small router network. For each token, the router picks the top-k experts (commonly k=2 out of 8, or 8 out of 256 in DeepSeek-style models) and only those experts' weights are used in the matmuls for that token. The other experts sit idle for that token but still need to be in memory because a different token might route to them.

For long-context inference, KV cache often dominates memory usage over the weights themselves.

• https://videos.cern.ch/record/3025356