@baldur @ljrk LLMs alone yes but # the idea of combining and constraining them with knowledge graphs is quite exciting .

@NoraReed so true. Gpt4 can do it a bit with the right prompt but you’re always fighting the RLHF

@jasongorman I guess it depends how broad your definition is . Many of the patterns I build are generalisable to multiple domains and often co developed with smes in those domains

@hobs totally .

@hobs https://www.vice.com/en/article/pkadgm/man-dies-by-suicide-after-talking-with-ai-chatbot-widow-says

@baldur @hobs Every AI company Inc EU ones have this very same problem, most LLMs are trained on the pile . Openai developer api has much better terms than chat GPT (30 day retention and no retraining on your prompts, same as Microsoft Azure OpenAI service ). My personal view , especially with GPT-4 and it’s extensive RLHF , and superior reasoning capability , is considerably less dangerous than open source models , eg GPT-J powered chatbot that was linked to a suicide.

@baldur @hobs not in to selection inference frameworks or toolformer ? LLMs are much better when augmented with other tools , and by making them behave more agentically . Better performance and more ‘explainable’ in a human sense , but more worrying from an alignment pov .

@drewharwell fully expect my job as prompt engineer in its current form to be obsolete in 3 years . Hell, the amount you have to prompt engineer Chat-GPT compared to Text-Davinci-002 eg is already much less . But by then there will be hundreds of other jobs in this space that don’t even have names yet .

@simon @dahukanna @timrburnham my guess would be QNRs. Graph + embedding . https://www.fhi.ox.ac.uk/qnrs/

@emilymbender @danmcquillan agree with most of the broader societal points . But you can make LLMs cite their sources by creating embeddings and running similarity search as the context within the LLM prompt , then matching the context back to the original text. You can see where the data came from , and the inference the model made . relying on the base knowledge in the model is not necessary with a ground truth corpus stored as vectors .