Jason Thorsness Blog

I compensate for my poor intuition of how wooden pieces will hang by coding an interactive ornament builder and physics simulation.

We condemn media created in massive volume with low effort by AI as slop, but it's the industrial software machine distributing the slop that's the problem.

The book Soul of a New Machine follows a group of computer engineers developing a novel product.

If someone writes an mp3 player in a forest (presumably on a laptop), does it make a sound?

As I showed in my last article, LLMs can generate web sites in real-time, complete with rich CSS and images.

In a previous article I joked about browsers generating pages with an LLM when the real server is slow.

I produced a 580-byte monospace font that renders all code points in Unicode planes 0, 1, and 15 as empty glyphs.

I've played a lot of Tower Defense over the years, starting with Desktop Tower Defense in 2007.

I downloaded the entirety of Hacker News and fed it to DuckDB, ultimately uncovering the secret to building a successful startup (or at least to drawing attention to a successful article).

You might have used an LLM to translate between high-level languages like Python to TypeScript, or Go to C#.

I set out to write an article on how big O notation doesn't help to compare algorithms at small input sizes.

One universal technique for getting software systems to perform faster is to make them do less work.

SingleStore started as MemSQL (YC11) with a memory-optimized skiplist-based table structure.

You've maintained an uneasy contract with the web for years -- you know web sites scrape every last byte of your information, but you tolerate it because it's invisible.

Another article only two days later?

Ever wonder how to build a scalable high-performance real-time application?

Images here are powered by Tractor Loader, a glue library for NextJS and Sharp that enables inline image transformation in JSX or MDX, like !

I designed and built some gift boxes out of maple, walnut and padauk.

With only ChatGPT, a web browser, and a random topic, can I generate and load millions of rows of highly-tailored data into SingleStore and highlight key features, all in the space of 5 minutes?

The JSON Schema project has fought the fight against schema-free JSON for many years.

SingleStore now offers a novel scheduled jobs service, enabling any Python or SQL task to be run on a schedule.

SQL makes writing UDFs for complex string, JSON, and blob manipulation awkward.

The Streamlit project is an easy way to visualize data.

You've probably seen the computer language benchmark game \-- but how about a real-time comparison in your browser?

Faster AND Cheaper?

I hate waiting for data loads.

The magic of vector search comes from the model, but even today's most sophisticated models might still be at the "first year at Hogwarts" level.

The location search on weather.bingo is fast.

Did it rain yesterday?

I recently wrote an article on the SingleStore Engineering Blog on our use of the BSON Data Type in SingleStore's latest release.

Hundreds of readers wrote to complain about the sluggish performance of the slider on the previous article.

NativeAOT-LLVM does not yet support SIMD through System.Numerics.Vector.

DotNet WebAssembly has been tied up with the Blazor project which is an awkward fit for use cases that don't involve a Blazor UI.

Most of the fun of having a personal website is in recreating it every few years...?