Monday, 2025-08-04 to Sunday, 2025-08-10
Ever since I started working at MISO, I had stopped making pour-over coffee at home, and today (Saturday, 2025-08-09) I tried to make some. I found myself having to think about what I had to do. It was no longer a muscle memory!
Went to play Pickleball at Meadowlark Park, played for an hour and left. Many things about Pickleball were elusive too after almost three months of hiatus.
Indiana State Fair 2025
Went to Indiana State Fair on Saturday, Aug 9, 2025. We have probably gone there 4 or 5 times since 2008.
I haven’t read David Foster Wallace’s “Ticket to the Fair”, about his visit to the Illinois State Fair. I guess we are like that only.
Original sign of “Indiana Sheep Association” at a shop window on the fair grounds
I would wear this t-shirt ( recreated using the above image on chatgpt)
Zenias have so many varieties
yama pāśa has a new meaning. (Yama’s hair braiding, somewhere near Indiana state fair .. 37th street).
Paying for Online Services: Paying for Gmail, and Youtube has definitely err.. paid off. No one should watch the crappy youtube ads (or anywhere else for that matter), especially children, unmonitored by you. So, please buy your family Youtube Premium, if you it daily.
Gmail is kinda subtle.. I resented paying for it initially when I went over the 100GB limit on my primary gmail account, but the familiarity of @gmail.com is hard to beat especially since my email id at this domain (which I consider my primary email), is kinda hard to spell out on phone, plus the gmail login makes it easier and easier to use my @gmail.com address. And the benefits have added ever since I started using it more. So, Paying $3/mo is totally worth it IMO.
If I can only figure how not to pay $35/mo to Apple, that would be great.
Reading
I read these books this week:
Observations
(Killingsworth & Gilbert, 2010) — “In conclusion, a human mind is a wandering mind, and a wandering mind is an unhappy mind. The ability to think about what is not happening is a cognitive achievement that comes at an emotional cost.”
This is also at the heart of performance coaches telling you to “shut out all the noise, and focus on your work[, workout, goals].”
—
What I need to practice:
- identifying what sells
- building things that attract positive attention
- using others interest as a way to guide exploration
- developing my first interest into something that can be grown beyond me
Programming
I created Natty.CSS as minimal, yet not afraid of css stylesheet for minimal styling of content heavy pages.
James Bennett, a long time Django developer makes a a good case for using Litestar python web framework over currently hyped fastapi. James captures my discomfort with Flask and frameworks inspired by it (like fastapi).
AI Engineering
Search is the natural abstraction of augmenting AI with evolving reality — Jo Bergum
Is a fantastic way of framing what’s needed for AI applications to evolve into even more useful things in the real world, because Searching is how one expands their external horizons.
—
Someone (JL) from TensorZero emailed me to take their framework for a spin. TensorZero is a stack for building LLM applications, that provides the following: (GOOEE)
- Gateway to access different LLMs through an unified API
- Observability to store inferences and feedback in a database, making it available via code and UI
- Optimization - collect metrics and human feedback to optmize prompts, models, and inference strategies
- Evaluations benchmark inferences and end-to-end workflows using heuristics, LLM judges etc
- Experimentation using A/B testing, routing, fallbacks, retries etc.,
This does look promising because it tries to encapsulate the common elements needed for building a “production” application using LLMs.
todo Next I’ll take a look at the code, configuration, and library choices, in the context of building an actual service. How does TZ compare to bifrost in MCP space?
bifrost does some portions of what TensorZero does — provides gateway for connecting to multiple LLM backends. todo understand what “MCP support for seamless tool integration” means.
—
gpt-oss from openai seems to be a good deal overall in the sense that it is quite capable, and is available under Apache license. Even if it does not compare to hosted GPT models. You can fine tune it and built products on top of it and make money. Developing tooling to work with capable Offline AI models is a good thing, and we should expect more things around it.
From GPT-2 to gpt-oss: Analyzing the Architectural Advances by Sebastian Raschka (author of “Build a Large Language Model”) is a great overview. mustread