@omarsar0: If you aren't looking at the data, you aren't doing AI at all. It was true before and continues to be the case today.
Summary
Omar Saroo emphasizes that data inspection is crucial in AI development, illustrated by an example where ignoring data led to model degradation due to incorrect chat template usage.
View Cached Full Text
Cached at: 08/19/26, 02:53 PM
If you aren’t looking at the data, you aren’t doing AI at all.
It was true before and continues to be the case today.
anton (@abacaj): I tried using Fable to train a model (LFM 2.6B) because I didn’t want to spend time on the data. Turns out neither did Fable and it ended up making the model worse on every attempt until I decided to look at the data. It was using the wrong chat template on like 1/3 of the data
Similar Articles
@addyosmani: https://x.com/addyosmani/status/2056078124346228860
A cautionary note from Addy Osmani that relying too heavily on AI to write code can hinder learning and degrade one's mental model of software development.
AI is getting better at analysis. The problem is still the data.
The author argues that AI analysis quality is limited more by data access and reliability than by reasoning, and that structured datasets dramatically improve outputs.
Why AI Cannot Save an Enterprise That Doesn't Understand Its Data
The article argues that AI cannot compensate for an enterprise's lack of data understanding, emphasizing the critical role of human judgment and context in transforming data into informed decisions.
Everyone is learning how to use AI. I'm starting to think the scarcer skill is knowing what the AI is actually doing.
The author argues that understanding how AI works is more valuable than just learning to use it, illustrated by a clinic's failed AI implementation that caused claim denials, and emphasizes the need for human oversight in AI systems.
The failure that worries me most in AI data analysts is the confident, made up explanation
A workshop on Oct 31 led by Bruno Gonçalves that starts with a naive AI data analyst failing on quiet, trust-eroding cases, then builds a robust harness with an MCP tool layer, context management, layered guardrails, and a run supervisor with timeouts and traces.