@a_karvonen@jeremyphoward +1. I was using GPT-4-vision to build a diagramming app. It had no reliable ability to produce basic geometric operations despite prompting & a robust set of functions to call. Same for Claude. It makes sense - they don’t interact physically during training. Is RL a way forward?
@EMostaque I recall you talking a while back about Apple silicon chips (I think). And recently saw you mentioning edge LLMs. Also Render (I think) will mostly be useful for models with low VRAM reqs?
For text models - what do you see as the future here? Fine-tuned quantized OSS models?
@blader The “move slower” people invented and created the underlying technology. The “move faster” people can move faster with applications and improvements to delivery of the current tech, sure.
@GaryMarcus@satyanadella And Excel. And Word. And PPT. And VSCode. And honestly many aspects of Windows as an OS. I hope they take it srsly and implement well so I can do less busywork.
@GaryMarcus@satyanadella No opinion on the Bing bot/their plans for autoregressive LM. But of all big techs, MSFT has an unbelievable opportunity in front of them for incorporating BERT-like NLU capabilities on top of their Outlook & Teams user datasets.
@Ethan_smith_20 Awesome, tyvm. Going to give GPT another shot then with that approach.
Someone needs to create a model that generates prompts that prompt GPT to generate prompts, dammit.