Dines: Encoder-decoder models beat decoder-only models on semi-structured documents
Daniel Dines · From Tiny Romanian Startup to Global AI Automation Leader | Daniel Dines, CEO of UIPath · May 31, 2024 · at 48:04
Daniel Dines, co-founder and CEO of UiPath, explains why UiPath builds document extraction models on top of T5 rather than relying solely on decoder-only LLM architectures.
0:00 / 0:08exact quote · 8.8s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“We believe that encoder-decoder models are better for when you deal with semi-structured documents than decoder-only models.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Daniel Dines
Scaling compute and GPUs alone won't yield autonomous enterprise AI
“I'm not an expert in the field, but I believe that this is gonna we need Another giant leap of innovation to get there. I'm not sure just throwing compute and increasing the number of GPUs without some fundamental discovery, it's gonna create this.”
UiPath's specialized models will outperform GPT-4 on enterprise document tasks
“But in the end, we will beat, you know, any performance of GPT-IV or whatever, and they would be extremely reliable.”
Future AI systems will rely on small, dedicated models for routine tasks
“I don't think so, because I think nature is pretty smart. And if you look at how our Brain operates. We have a very good cognitive engine, but we are not using that cognitive engine to do trivial tasks. It's way too expensive. If you want to learn to swim or t…”
LLMs cannot be used autonomously in enterprise production today
“You cannot use an LLM in an enterprise production in an autonomous fashion today. It's simply impossible because of the hallucination aspect and the lack of reliability.”