> Agents made bad inferences because they had no context on the business

We've been working on this since before the chatgpt launch.

We started with a semantic layer since there were already good open source options and LLMs at the time were good at writing the JSON (remember function calling?) to run a semantic query.

But as LLMs have gotten smarter and people wanted to do more data work in agents, we found we needed something more flexible, so we built an "Ontology" that lets you store all the terms you use in your company and connect them to the data points (e.g. tables, columns, metrics) that matter.

https://www.definite.app/blog/ontology-ai-analytics

I index all my local Claude Code sessions in DuckDB. I have 202,381 messages in the last 30 days.

There's been a steady increase since Opus 4.6 in the model saying "honest".

It probably shouldn't, but this bugs me.

Should I assume most of the time you're lying and you're being honest in this one message?

I was pumped in the first few hours of Fable where this had seemingly been "fixed". 100+ messages and no "honest" to be seen. But it didn't last.

Within a few hours, Fable proved itself to be the most honest model to date.

Here is the rate at which visible assistant text contained the string "honest" (case-insensitive), split by model:

  claude-fable-5:             25 / 1,397   = 1.7895%
  claude-opus-4-8:            83 / 5,818   = 1.4266%
  claude-opus-4-7:           163 / 16,432  = 0.9920%
  claude-opus-4-6:            18 / 5,877   = 0.3063%
  claude-haiku-4-5-20251001:   0 / 71      = 0.0000%
  claude-sonnet-4-6:           0 / 4       = 0.0000%

Comments URL: https://news.ycombinator.com/item?id=48503904

Points: 3

# Comments: 2

Comments URL: https://news.ycombinator.com/item?id=48403422

Points: 2

# Comments: 1

If someone from Anthropic sees this, would love to know if I can use my max plan here.

> This unexpected shift completely broke my preferred workflow

it might not have been so unexpected if you knew you were one of ~15 people that start their day with Antigravity

does anyone know how much they're thinking for Flipper One?

> Not yet, but we are working on it!

Seems like a niche use case, but it's the one I'm most interested in.

Our lakehouse uses ducklake with postgres as the catalog. Seems like a DuckDB / Quack catalog would be an excellent alternative.

For example, user says "build a report with revenue and orders by month and show 100 most recent orders". The agent would write a spec that would get rendered by our frontend.

This runs fast, but we were drowning in feature requests for what the framework could render (e.g. "I don't want labels here", "I DO want labels there", "can this chart be a heatmap", etc.)

A few months ago, we let the agent just write HTML instead. It takes longer to generate, but you get unlimited customization.

There are a host of issues with the new approach (non-technical users debugging a monstrous app they created), but net-net our customers like it much better.

I completely understand a "people who give a shit stick around" mentality if you work there, but you can't expect users who run a business on it to stick around if it's broken.

Show HN: I've built a C# IDE, Runtime, and AppStore inside Excel

670 points | 179 comments

One of the main use cases was to analyze Excel data with SQL. I'm the kind of nerd that loves stuff like that, but stuff like that seems completely obsolete now.

[0] https://news.ycombinator.com/item?id=34516366

we have a custom .yaml spec for data pipelines in our product and the agent follows it as well as anything in the training data.

while I agree you don't need to build a new thing "for agents", you can get them to understand new things, that are not in the training data, very easily.

yep, that's what Definite is for: https://www.definite.app/

All the data infra (datalake + ELT/ETL + dashboards) you need in 5 minutes.

we run datalakes using DuckLake and this sounds really useful. GCP should follow suit quickly.