AI News #18

Mixture of Depths

  • Dynamically allocate computation to specific parts of input text, for 50% faster LLM inference.
  • Paper


  • A new finetuning method that is 10x–50x more parameter-efficient than LoRA. Adds “intervention” functions to specific layers and token positions.
  • Paper

Language Models as Compilers

  • To solve a problem, ask the LLM to generate and simulate execution of pseudocode.
  • Paper


  • An initial attempt at getting LLMs to show what they’re thinking, to help arrive at the correct answer for complex spatial reasoning tasks.
  • Paper

Command R+

  • New open-weights multilingual model, with 128k-token context, specialized for retrieval augmented generation (RAG).
  • Announcement

Many-shot Jailbreaking

  • Use many-shot prompting of desired jailbreaking behavior, to bypass LLM guardrails.
  • Blog post

Claude can call functions now

  • Anthropic announced that you can ask Claude to use external tools via its Messages API.
  • Docs

Alternate image flipping

  • Instead of random horizontal flipping of images, “flip half the training images on even epochs and the other half on odd epochs” for faster training convergence (on CIFAR-10).
  • Paper

Stay up to date

Interested in future weekly updates? Stay up to date by joining our Slack Community!