Paperclip 001: The Allocation Economy; Prompt Engineering Techniques; and AI Transparency
The Changing Nature of Work
Some argue that the knowledge economy is giving way to an “allocation economy” in which knowledge workers will need to become AI managers to stay employed.
Prompt engineering is a weird new field that’s all about tricking GPTs into producing better answers. Some techniques that actually work (from Michael Taylor at Every): USING ALL CAPS, adding emotional statements to the prompt, telling the AI to take a deep breath, and threatening to kill a real person if the AI doesn’t do exactly as you instructed.
High ROI
Poolside.ai raised $400 million on a $2 billion valuation – with no product and no revenue – making it a potential pure play.
AI Transparency
In “Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet]" (a search paper TheSequence calls “one of the most important papers of 2024”) Anthropic researchers document a significant leap forward in the interpretability of LLMs. Improving interpretability surfaces the inner working of LLMs and has implications for training and safety.
Researchers used a sparse autoencoder technique to map model activations in the Claude 3 Sonnet model to a create a dictionary of interpretable features and demonstrated that the technique can be scaled up to larger model.
The interpretable features are abstract, meaning they are “multilingual, multimodal, and generalizing between concrete and abstract references.” Examples include sycophancy, empathy, sarcasm, the Golden Gate Bridge, brain sciences, and transit infrastructure. The features can be used for model steering, which has implications for AI training and for AI safety.