What's ahead in Q4 from LPs and VCs (Origins Podcast) - more shutdowns?, hard choices for seed funds?, net new company funding 🔥, default 1st 10 in office, TVPI vs. DPI
The two Honeycomb posts about building with LLM's are some of the more honest takes on that experience that I've seen. I'm currently a founding member at a generative AI startup, and I've been working with ML since 2018, when I started leading an ML solutions team at AWS. My co-founders (who are experts in NLP and have been working with state of the art neural network architectures like transformers for many years) have long said that the future will be smaller LLM's trained for specific purposes. This is exactly what my company is doing, and it greatly reduces the complexity of prompt engineering and retrieval augmented generation because you can show the model many specific examples during pre-training, instruction training and fine-tuning. This means you don't have to put examples into the prompt itself, which greatly reduces the impact of context window limits. Of course, training LLM's (even ones with far fewer parameters than the models from OpenAI, Google, or Anthropic) is hard, so many companies are going to have to build on the 3rd party foundational models, at least in the near term.
The two Honeycomb posts about building with LLM's are some of the more honest takes on that experience that I've seen. I'm currently a founding member at a generative AI startup, and I've been working with ML since 2018, when I started leading an ML solutions team at AWS. My co-founders (who are experts in NLP and have been working with state of the art neural network architectures like transformers for many years) have long said that the future will be smaller LLM's trained for specific purposes. This is exactly what my company is doing, and it greatly reduces the complexity of prompt engineering and retrieval augmented generation because you can show the model many specific examples during pre-training, instruction training and fine-tuning. This means you don't have to put examples into the prompt itself, which greatly reduces the impact of context window limits. Of course, training LLM's (even ones with far fewer parameters than the models from OpenAI, Google, or Anthropic) is hard, so many companies are going to have to build on the 3rd party foundational models, at least in the near term.
Thanks for including the time to ARR chart.
Downloadable slides with it + more are here:
https://www.saasletter.com/p/saas-benchmarks-historical
Thank you - great data!