mark_l_watson

I agree.

I might be wrong about this, but obviously Google would like to provide inferencing at the lowest cost to themselves, so perhaps their slow ‘pro’ releases and rapid ‘flash’ releases is an attempt to guide people to use more profitable models?

I usually use OpenCode for all open weight models but for Gemini I use Google’s agy coding harness (or my own).

Venders coupling coding harnesses with their own models is usually a good thing. Poolside.ai has a combined harness with their own models that works well locally, and the DeepSeek harness with their models is very interesting.

That is my understanding, which is a nice thing. I use Exa and Brave search APIs separately but the Google bundling is convenient.

Ollama Cloud offers the same thing: they supply a web search tool bundled with cloud API inference services.

The Gemini APIs have options for using Google Search grounding, or in simpler terms: to process a prompt by first using search to augment the context.

good summary, thanks. I used to use Google Cloud for consulting and projects (I worked at Google for a while, and there is some nostalgia) so I have Gemini API via Google Cloud, but I am retired now. Your post reminded me that I need to shut that all down and switch to getting an API key using AI Studio.

I have spent two months experimenting with a wide range of US and Chinese models, and I had a lot of fun doing that, but I am in the process of switching to just using local models, using Gemini on an API if I need it, and once or twice a month when I really need help on something difficult, I use something top-tier like Kimi K3.

I use Kimi K3 and GLM 5.3 to find old bugs and improve old code, and would try Fable 5.1 if I could afford $50 for 1M output tokens (I am retired and all my research is self funded now)

I like to start by prompting with “examine this code base for problems and improvements and write to IMPROVEMENTS.MD” and then carefully look over the suggestions, and either fix myself or let the model+coding harness try.

and $0.15 1M input, $0.50 1M output

I am a huge enthusiast of running local models, but when multiple quality USA vendors provide models like GLM 5.3-flash, I run locally just for the fun of it.

For the purposes of comparing to Fable 5.1, I would mention GLM 5.3 that is about 1/12 the cost.

I have worked at three companies (Capital One, Google, and SAIC) where for high value work the cost of compute was no real concern. I understand the economics of spending big for huge payoffs, so this is a serious question:

Does Fable 5.1 really provide much benefit over models like Kimi K3 that are 1/3 the cost? Or GLM-3 that are 1/12 the cost?

If you can talk about your work, what kind of tasks do you work on where the higher cost is very much worth it?

but the article cherry picks. the article ignores Zitron’s economic arguments about data center costs vs. future profite.

This article cherry picks:

Ed Zitron mostly covers the costs of data centers, circular spending, and predictions of large the market for AI has to be to justify the data center expenditures.

I was disappointed this article didn’t really cover Zitron’s main arguments.

Maybe a paid article placement? I don’t know, but I was dissapointed: I read Zitron’s material and I wanted to see good counter arguments to his rants about costs of data centers, circular spending, and predictions of large the market for AI has to be to justify the data center expenditures arguments.

Nice improvement. For me, uv is the ‘Quicklisp for Python.’ uv just let me enjoy using Python like Quicklisp just makes Common Lisp nicer to use.

I have always been a Lisp devotee, but a few years ago when I started using uv, I then started seeing Python as a language I could really enjoy using so I put effort into making my Python dev setup nearly frictionless.

I liked Prometheus so much that I purchased the movie.

that would be incredible, Forever War was a great book

re: Ridley Scott making another Alien franchise movie: if he is doing it because he really wants to do the movie then that is great; if he is doing it for business reasons to make money, that is not so great

I tried Ornith-1.5-35B-A3B-MLX-4bit on a 32G M2-Pro mac mini with pi and OpenCode. I didn’t get very good results coding in Python, Racket, and TypeScript. I have seen several positive comments like yours so I was probably doing something wrong. I amgetting the new 64G mac mini in 4 weeks, and I made a note to try Ornith-1.5-35B-A3B-MLX-4bit again.

That is saying a lot if Ox Alpha is also small and relatively cheap computationally. I hope so; I love deepseek-v4-flash-0731 and use it frequently. Fast inference is good and fits with my dev style: I like to be in the loop, not let an agent code on its own for long periods of time.

I have seen studies from MIT and Stanford that the majority or US startups are using much less expensive open weight models so consumers of their products are open model users whether they know it or not. These are often Chinese models.

Not to go off topic but I am pleased to see open model support from US companies like Poolside.ai, NVIDIA, IBM, Google, etc.

Yes, GPUs are much better for dense models. On Macs, MOE models run better, so I agree Macs are more limited and expensive. I have a Linux laptop with a 10GB 1080 GPU, dated, but I should add even more system RAM and try that.

I loved the book The Infinite Game. Really changed how I looked at work, personal research, and being an author. Recommended!

I run a lot of local models (I am always experimenting) on my 32G M2-Pro MacMini - I would love to upgrade.

The financial aspects don’t work however: I can learn and experiment with what I have for local models, and I pay as I go on FireWorks.ai for open model inferencing and no matter how much I use this service my monthly bill is between $10 and $40 and much faster than any reasonable home rig.

Hybrid ‘small local’ and buying inference is the way I choose.

I have lower standards than you: I pay for deepseek-v4-flash-0731 tokens from a fast and reliable US vendor and I feel like working on one task at a time is fast and gets almost everything done I need.

Yes! My main use of very strong models is in writing my own coding harnesses for small local models, tailored for my needs. I also use very strong models to get much smaller skill files and also writing tools for my harnesses.

re: data centers: pump and dump. Wealthy investors will have made their money and walked away, and the corrupt democrat and republican politicians in Washington will, as usual, protect the interests of the ultra wealthy and leave the general public to pay for poor decisions. There will be a government bailout.

Anyway, on a positive note, I am all in for small local models that are augmented by strong hosted models for specific tasks. Use technology to help people, not make billionaires even more money.

I wrote my own AI coding harness in Racket, it works fine, meets my own needs. I did earlier versions in Common Lisp and Python but the Racket version is the most full featured.

What you say is in agreement with someone I follow on X. My family has been traveling for two weeks and I have been looking forward to trying qwen3.8-37b when I get home.

Now I might try Ornith-1.5-35B first even though I don’t see an official MLX version.

You make a reasonable argument. I will take my chances and more people I talk with now seem more concerned about privacy.

That is a really good question. My personal intuition is that LLMs do well with any language as long as you set up for fast syntax checking tools, access to either a REPL or very fast to run tests, etc.

I usually use small local models, and the work to set up very concise skills and efficient tooling is a big part of the fun. I have also adopted the practice of writing my own custom coding harnesses (these can be less than 2000 lines of code, not the huge project you might expect.)

Since 1982 I have spent much of the time being paid to work in Lisp languages.

A weird thing: whichever Lisp language I am currently using for work (or a side project) is my favorite.

I love Racket, and have a ton of Racket code I have written over the years. When I get home from a vacation I will have an AI translate a small sample of my Racket code to Rhombus: easiest way for me to play with the language.

Good list! Removing as much as possible all AI and surveillance features, then carefully adding back the very few I want is the way I roll.

On our car I can hook up my iPhone via Bluetooth to listen to audio books and music, bypassing carplay.

There are a near infinite number of Qi Gong and Tai Chi videos on YouTube. I prefer to following new videos, and where my wife and I live they have Tai Chi and Yoga classes. A great use of time, and the feeling good benefits are manefest.