5 ms·
Fundamentals are outdated - it's all based on old GPT-2/3 models which needed a lot of hand-holding and the whole point of chains was that those models were too
by treprinum 3y ago
Fundamentals are outdated - it's all based on old GPT-2/3 models which needed a lot of hand-holding and the whole point of chains was that those models were too dumb to run multi-task prompts, not to mention that by default some tasks are executed sequentially while they can be run in parallel, slowing everything down (see how they did NER).
- wokwokwok 3y ago> it's all based on old GPT-2/3 models Are you sure? There are examples of using mistral eg. https://github.com/langchain-ai/langchain/blob/master/templates/sql-llamacpp/sql_llamacpp/chain.py https://github.com/langchain-ai/langchain/blob/master/templa... This is exactly what I’m talking about. How can you say that when there is evidence that blatently contradicts it? This reeks of “…or so I’ve heard, but I never actually looked into it myself…”
- treprinum 3y agoThey keep adding new models but it's a bolt-on on an underlying architecture based on old assumptions that no longer hold for LLMs with emergent abilities like GPT-3.5/4.