Captain Zero Fun

SITUATION EXPLAINED: MIT researchers got AI models to talk in "neuralese" instead of words, and it could end our best way of checking what AI is thinking. • Normally one AI writes a summary for the next AI to read, but MIT's "Swarm Telepathy" hands over the first model's internal working memory directly • A small translator network converts that memory between different model families, like Qwen, Llama and Gemma, without retraining either model • It beat models talking in plain text by 40% in accuracy on average and ran about 4x faster • It even passed image understanding from a vision model to a model that can only read text • Chain-of-thought monitoring means reading the reasoning a model writes out in English to catch bad behavior before it acts • In 2025, dozens of researchers from OpenAI, Anthropic and Google DeepMind warned in a joint paper that this window into AI reasoning is "fragile" and could close • AI safety researcher @davidad predicts chain-of-thought monitoring will be "unambiguously over" by early 2028 @theojaffee: "Instead of models having a chain of thought where they write in English, they will just think in neuralese in the inscrutable matrix of floating-point numbers that make up the KV cache."

Autobots First Transform & Rollout | The Transformers (1984) 🚗