🤖 AI/LLMs - Speculative Decoding Observation
This article describes an observation regarding speculative decoding setups, specifically the EAGLE-3 style. It notes a performance degradation when training for k=4 results in lower acceptance rates for k>4.
Key Points:
• Acceptance rates for k>4 decrease significantly after training for k=4.
• The observation relates to EAGLE-3 style speculative decoding setups.
• Research indicates specific behaviors in speculative decoding performance.
🔗 Resources:
• Discussion Thread ↗ - Research notes on speculative decoding acceptance rates
🤖 Blockchain Security - Exploit Report
This report details a blockchain exploit involving funds from summerfinance_. An exploiter converted DAI to ETH and subsequently transferred a significant portion to Tornado Cash.
Key Points:
• An exploiter swapped 740.1K DAI for 421 ETH.
• 400 ETH was transferred to Tornado Cash.
• A total of 600 ETH has been sent to Tornado Cash.
• The exploit originated from summerfinance_.
🔗 Resources:
• Exploit Alert ↗ - Details of a recent cryptocurrency exploit
• summerfinance_ ↗ - Affected platform in the security incident
🤖 Machine Learning - Equivariant Transformers
This article introduces Platonic Transformers, a model for scalable equivariant Transformers. It highlights their ability to maintain speed and simplicity compared to standard Transformers.
Key Points:
• Platonic Transformers offer scalable equivariant properties.
• They retain the speed and simplicity of standard Transformers.
• The research was presented at ICML.
🔗 Resources:
• Platonic Transformers Overview ↗ - Information on scalable equivariant Transformers
Image
✨ AI Models - Muse Image and Muse Video
This article introduces Muse Image and Muse Video, media generation models developed by Meta Superintelligence Labs. Muse Image is highlighted for its advanced capabilities in image generation.
Key Points:
• Muse Image and Muse Video are new media generation models from Meta.
• Muse Image is an agentic model.
• It plans, writes code, and uses search tools.
• The model refines its outputs using a chain-of-thought process.
• Image performance improves with increased test-time compute and reasoning.
🔗 Resources:
• Meta AI ↗ - Experience Muse Image and Muse Video
• Official Announcement ↗ - Introduction of Meta's media generation models
• Muse Image Launch Details ↗ - Blog post about Muse Image's agentic capabilities
Image
Image
⭐️ Support
If you liked reading this report, please star ⭐️ this repository and follow me on Github ↗, 𝕏 (previously known as Twitter) ↗ to help others discover these resources and regular updates.
