Jakob Serlier's RSS feed

After trying a set of LLM-based and 'claw' personal doc management approaches, I've settled on a pragmatic and practical setup for extending Karpathy's llm-wiki into something I actually use day-to-day.

A while back a friend and I played a variant of Texas hold'em.

"I have only made this letter longer because I have not had the time to make it shorter."

I've had Spotify Premium continuously for 10 years.

I want an always-on agent that can read a subset of my notes and manage tasks.

Running coding agents with free reign is very powerful for a certain class of tasks, especially ones that require little human supervision, or where you want to close (or disconnect) your laptop, walk away, and come back to results.

In short: Atlas performs individual browsing steps correctly, but it breaks down when asked to repeat the same well-scoped action across multiple iterations.

I've been working on telemetry redaction using Microsoft Presidio, and recently contributed a sample to the Presidio repository.

I’ve recently been experimenting with PII masking in observability pipelines using Presidio.

Coding with AI agents creates this peculiar problem: you get these short 10-30 second waits while the model thinks, and it's just long enough to lose focus.

It can be challenging to pinpoint the 'concrete' impact AI assistants have on daily development work.

When building a SaaS offering publishing platform that helps customers publish azure marketplace offers, we faced an interesting observability challenge: how do you maintain separate telemetry streams for both your platform and each customer's application?

Learnings from ingesting millions of technical pages for RAG on Azure.

Showcase: Azure AI Hybrid Search unexpected results gotcha This document describes a gotcha in Azure AI Search hybrid queries where unexpected results are returned.

The recent publication on LLM "sleeper agents" prompted me to re-explore influencing LLMs to pursue alternative objectives.

What does a statement like “AI will take my job” look like in practice?

A short one: I needed a GPT-4 CLI interface (that's a RAS Syndrome), but most options seemed quite cluttered, and I like code-golf.

The why Recording my daily thoughts and notes is a challenge I return to every few months, and so far, I hadn't figured out a system which works well for me.

In short: The RAG pattern for LLMs can be evaluated using QA pairs.

Dear colleagues, I have a confession to make: I have been delegating some of my unit tests to my Jr. engineer, Gary-Pete Truman.

Addendum after hitting the HN frontpage: MNIST is a straightforward dataset, and higher accuracies are possible with various methods.

Disclaimer: I am an employee at Microsoft ISE.

Note: I will not discuss the alignment issue or responsible AI / ethics.

TL;DR: I had been avoiding properly setting up remote networking to my homeserver.