Tomasz Tunguz blog · 2023-10-07
· 971d
Chaining Large Language Models for Email: Capabilities and Limitations
Tomasz Tunguz explores his experiments combining Whisper (speech-to-text) and LLaMA 2 models to automate email drafting through dictation. While promising, the approach faces significant challenges around tone calibration, contextual understanding, and multi-party email thread management. The article highlights the difficulty of achieving the final 15-20% of performance quality in machine learning systems.
Metrics in this report
80/20 ML Quality Rule
80%
baseline performance
Typical machine learning system performance
Remaining Effort for Quality
15%
marginal improvement
Additional performance needed to achieve production-quality ML
Subscriber Base
150000readers
current
Tunguz's newsletter audience of founders and operators