I love it! I wish the underlying data was more accurate. I asked @fintoolx to identify data discrepancies and there are a lot:
- $NOW number of employees is 26,293 not 22,000
- $NOW FY24 revenue is $10.9B not $8.5B
- $NOW Subscription rev Q4 2024 is $2.8B not $2.96B
- etc etc
Most of the numbers like net income, gross margin etc are wrong. That’s the issue with general agents, the underlying web data is trash and not verifiable
I tried several AI assistants to book an international flight:
- GPT-5 thought for 5 minutes and then hallucinated a Google Flights URL
- @ChatGPTapp agent took 6 minutes but did a better job and found the route
- @PerplexityComet found the same route in half the time
My article on the decline of RAG is at the top of Hacker News: The RAG Obituary: Killed by Agents, Buried by Context Windows.
We built vector databases, embeddings, rerankers to help LLMs survive tiny context windows. GPT-3.5 could only read ~6 pages, GPT-4 about 12, while a
