{cas} a journal by Cas Stantonius

Can I make a reliable data agent that can identify how to integrate (synthetic) JSON into SQL.

Can I use DSPy to optimize converting JSON into existing SQL TL;DR Objective: Generate synthetic data with known data manipulations (transformations and corruptions); optimize prompts so an agent can spot said manipulations to ultimately suggest how the json could be incorporated into the SQL Process: Created a mock database, Python tools for querying the database, defined manipulations, generated synthetic data based on existing schema, built agent to predict these known data anomalies, attempted to score these predictions Outcome: The agent did a good job in identifying the issues in general but we don't know how well yet, since scoring proved to be a challenge.

o3 summary of application of principles to a SaaS solution from this excellent post Principle in the essay What it means for your platform Malleability over rigidity – software should be “clay” users reshape, not sealed appliances.

Related to an AI-coordinated HRIS: For an AI layer mediating front-end requests and SQL data, design the agent to issue small, incremental SQL queries and adjust each next step based on returned results (“doing more”) instead of crafting one large query plan upfront (“thinking more”).

If software becomes easier to build and faster to ship, who sustains the share - market and mind? Established players will initially but may relinquish share to fast moving disruptors.

General Be explicit & literal: Spell everything out exactly as you want it Induce planning/CoT: +4% gain.

I've been thinking about the "meaning economy" for a while, given the possibility that cheap access to AI knowledge fundamentally changes what is valuable in our economy.

Paul Graham describes the following principles: take care of people and the world, and make good new things where PG describes emphasizes make (as opposed to critique) and new as being critical requirements of this philosophy.

Source: https://paulrobertlloyd.

One of my favourite bloggers, Simon Willison, has said the following several times I believe that the price you have to pay for taking on a project is writing about it afterwards.

Source: https://www.

I am petrified of doing things in public.

Source: https://open.

so much of AI Engineering isn’t an intelligence play; it’s an architecture play.

Source: https://www.

Source: https://www.

Source: https://simonwillison.

I am annoyed that I have to write this, because I knew this risk.

Source: https://kupajo.

Source: https://stripe.

I love everything about this post - because it describes me so much.

Superglue - LLM-generated API connectors .

I think I needed these types of reminders right now: Every line of code you write is a potential bug.

Found this doc from Google Cloud.

I keep hearing about the value of a mentor (Tyler Cowen talks about this a lot), or at the very least a strong community - this is not new information, but maybe I'm just actually listening now? In trying to be an optimist, here is a take of how AI will positively help folks like me achieve the benefits of a mentor/community without actually having one.

A machine need not be perfect, or even particularly good at something, to be educationally useful.

A very useful summary for how to use Cursor properly from bhl I use it everyday.

This is a live document that will evolve over time.

Source: https://www.

https://ghuntley.