Interesting Tweets I Read Recently

• A friend’s list of high signal-to-noise pages
• Reproducibility crisis as fraud
• Married, conservative, college graduates who trust the system and others are happy
• Mesaoptimisation in transformers
• Reducing PPO memory in RLHF
• Visualising with TorchLens
• Why LLMs work so well
• Prediction is insufficient for making money
• A 7 point recipe for single managers
• Transformer model inference optimisation
• Training large models on many GPUs
• Training stability of transformers
• Drone racing with deep RL
• Speculative execution for LLMs
• Transformers learning to find GCDs
• Inverting embeddings
• Transformers as SVMs
• Transformers being bottlenecked by memory bandwidth, not quadratic complexity
• The real Jensen’s law