Claude AI consumes tokens at a high rate, similar to how a sports car burns fuel, and since tokens incur costs, frequent users constantly seek strategies to reduce expenses. The newest approach involves a bold and divisive technique: prompting the AI to respond in a primitive, caveman-like manner.
This technique operates through a custom initial instruction or dedicated extension, directing Claude to eliminate unnecessary elements including introductions, compliments, side comments, intensifiers such as 'very' or 'really', repetitions, and definite or indefinite articles like 'a' or 'the'.
Among various 'caveman' instructions circulating on Reddit, one popular example states: 'Respond in basic form. Employ nouns and verbs only. Avoid structural words (the, is, are, of). Use brief terms. Reduce tokens. Stay direct.'
Outcomes differ based on the precise instruction applied. In an experiment with a modified version that specified articles to omit and requested removal of courtesies and extras, the responses were highly concise, lacking casual conversation or tangents, though not fully primitive without further refinement.
For instance, querying this modified Claude about fundamental neural network concepts yielded: 'Neural networks consist of linked node layers. Input layer receives data. Hidden layers detect patterns. Output layer delivers outcome. Nodes transmit signals, adjusted by significance weights. Training modifies weights via backpropagation to reduce errors, iterating millions of times. Main variants: CNNs for visuals, RNNs for series, Transformers for text. Current large language models rely on expansive Transformers.' This 50-word reply equates to about 80 tokens, as models tokenize words, punctuation, and other elements into sub-units for processing, each counting separately. In contrast, a standard query without the modification produced a 330-word answer using 460 tokens.
However, this is not a straightforward solution for everyone. Since its emergence this month, the primitive prompt has ignited intense discussions across Reddit and other platforms. Critics claim it results in oversimplified outputs, while others point out that efficiencies apply solely to generated tokens, not input ones. Moreover, a detailed primitive instruction adds to the cost each time a fresh query is submitted.
The primitive approach may not suit all scenarios. It excels in areas like programming, where straightforward answers suffice, but falls short for in-depth explanations needing subtlety, as seen in the neural network query.
Nevertheless, this method highlights user ingenuity in optimizing AI resource use, a growing concern for everyday interactors as advanced autonomous systems, such as Claude Cowork, become widespread.
Developers of prominent primitive extensions are innovating to extend token limits while preserving response standards.
Ben has covered consumer tech for over two decades, currently emphasizing AI's impact on daily life. His reporting examines cutting-edge language models and their applications in professional and personal settings to equip people for the upcoming AI shifts. 'AI will transform existence faster than anticipated,' Ben observes. 'Daily engagement offers the optimal adaptation path.' A PCWorld contributor since 2014, he has reported on devices from computers to surveillance tools prior to spearheading the site's AI section. His work also features in PC Magazine, TIME, Wired, CNET, Men's Fitness, Mobile Magazine, and others. Ben earned a master's in English literature.