mark_l_watson

Because of the drastic cached input costs, they used the published DeepSeek KV caching techniques, right? I didn’t see any public acknowledgement.

The article resonates with me. I am easing back off of vibe coding. Now I will use dsh with a deepseek-4.1-flash commercial API occasionally, but now run Laguna-xs-2.1 running on oMLX locally for code reviews, and routine code changes. I am “devolving” back to doing most work myself.

I have lived in a tech-rich life since I bought my first single board home system in around 1975.

I think it is as easy as setting boundaries: practice meditation. Learn to enjoy sitting and doing nothing, both in the living room and sitting outside.

Then when the time is right, work hard for several hours to get important stuff done.

Food tastes best when we are very hungry, so we must go long periods of time doing mini-fasting a few times a day in order to enjoy food. Similarly, earning a living or experimenting with tech should be something that you sit down and do for a while, then do something else.

I have a different take on this: I bought a Quest 2 several years ago, spent some money and bought about 10 experiences/games/etc. for it, and it is good enough that I continue using it with no real desire to buy a newer model. My wife was watching me use it last week and she commented “isn’t there an upgrade for that?”

Wonderful toy, even though I only use it about 90 minutes a month it is worthwhile for me to own.

Love this article, right on.

re: “”Start with the harness, because it is what makes everything else usable.””

I have written three harnesses this year (in Common Lisp, Python, and Racket Scheme) and it has been a fantastic learning experience. I have been using neural network tech since I was on a DARPA neural network advisory panel in the 198os, and I am even more excited (by an order of magnitude) than I was back then.

I think Tim has it right, and the hyperscalers serve a function of bootstrapping smaller systems, but they are far from the whole story.

Since I am retired, my uses of agentic coding harnesses include:

1. update my old open source projects by searching for and fixing defects, adding tests and documentation

2. working on my own agentic coding harnesses, using the coding harness I am modifying to update itself. I am tightly in the loop

Sure, not highly practical use of AI, but I am retired!

Wow, this is a pessimistic (but accurate!) view.

I was the second featured Creative Commoner decades ago and I have published all my books using Creative Commons share and share alike, no commercial reuse for the last few decades.

I haven’t been considering AI training use to violate my use of this license, but I suppose it does. When I have some free time I think I will switch the license on my books to the CC license that permits commercial reuse.

I think there are (at least) 2 important things here:

AI training web scrapers should respect robots.txt, ai.txt, etc. files associated with content.

Privacy issues are a severe problem, and privacy and security issues should be ‘top of mind.’

This is why only synthetic and highly tailored training data should be used.

As someone else here said: the Deepseek team makes training runs in tightly controlled sandboxes, and any hacking behavior is scored as a failure.

The problem we have in the USA is that financial (and political influence) are misaligned from what is good for society.

This paper is the most reasonable one I have read on AI safety. We need to fundamentally change the training pipelines by figuring out better ways to ‘reward’ behavior. Yoshua didn’t explicitly mention training data, but we probably need to only use synthetic data that contains no text that could motivate bad behavior via imitation.

I feel like a heretic for saying this, but I will say it anyway: AI agents are great for activities like `writing that bash script, proof reading our writing and interactively brainstorming when designing and writing code but I feel like all of this can be done with any similar model to a super-inexpensive deepseek-4.1-flash API and sometimes even qwen3.8:27b running locally. When is good enough, good enough?

Concentrating on commercial exploitation of small, efficient (fewer new data centers!) models and agentic harnesses crafted for more practical things than just software development would allow AI investors (who have too much political influence) to make money short term while we figure out how to do AI correctly.

I like the API, and since this is open source I will copy the design with attribution: I have been organically hacking little bits of LLM client code for most providers in Common Lisp for about 3 1/2 years and it is time to clean up all my old code. I probably need to do the same sort of refactoring for my search API wrappers.

Thanks for the cool project.

I am 75, and have a ton of computer gear. I can imagine myself as an old man later in life, not wanting to do any programming, and just wanting a single digital device like a future version of this product. Let’s see what Apple has available in 10 years.

I am shocked. After playing with Muse by Meta for an hour in just chat mode, I was impressed enough to:

I never use Apple Reminders so I gave Muse read access to Reminders. I asked Muse to schedule repeating tasks that only involve web search e.g., every other morning find sci-fi movies playing in my city, 3pm EST stock market summary, etc.

I also permanently set my food preferences and just naming any restaurant in town gets me an accurate list of recommendations with costs and descriptions.

Your three examples in the README.md file help. If you have the time, please add more examples and more example use cases. Sorry to sound lazy, I started reading the code in yayster.el, then stopped, wanting more docs. I wrote an Emacs based coding harness earlier this year, so your ‘opposite direction’ idea is interesting.

I am starting to wonder if I am doing something wrong: I ignore evals and instead I just try new models or new harnesses (or tweak my own harnesses) by solving problems I want to solve in any case; I just use new tools and form my own subjective opinions of them.

I need to manage skill files across 2 Macs and 1 VPS, and also across fast inferencing APIs vs. very slow local models.

The first dimension is easy: I simply keep copies of debugged skill files in iCloud and copy them where I need them.

The second dimension is where I spend my time: I use short skill files for fast inference APIs and tiny skill files when I am running slow local models, and I simply spend a lot of time writing and tuning tiny skills files.

Of course, with increasingly better models, skill files become less relevant, but not totally irrelevant.

I watched a 90 second clip from this film earlier today: Holmes unblinking asking ‘Why would I deceive you…’ Really strange stuff. I have to admit that I watched the clip a second time.

My older books have real photographs on them (e.g., https://leanpub.com/lovinglisp) and newer ones have AI generated covers and some have AI generated figures. Look for books with photos I have taken of nature - those are the old books (exception is Java AI book that is 90% very old material, but has a new AI generated cover).

In the last year I have started vibe coding a lot, but I still take personal ownership of my latest code examples because I set requirements and provide a design, and ask AI coding harnesses to let me run and test the code so I tweak it. Anyway, it still feels like mostly my code and I prefer the new AI + Mark code over what I used to produce.

This has been a strange ride! I learned to program in 1965 and have always loved coding, but it is a weird new world we live in, my friend.

And NC doesn’t always work.

I use no commercial reuse (NC) yet years ago a polite young man from China emailed me letting me know that he was translating my stuff to resell under his name for money. To be honest, I didn’t mind and it made a good story to tell friends and family.

I have been releasing books under Creative Commons for 20+ years. Share alike, no modifications, no commercial reuse. I sell my books via Leanpub (my books https://leanpub.com/u/markwatson) and they can all be read free online (recommended to see if you like them), and if purchased the price can be reset to free, zero cost.

I like CC because I like my books to be read and I still make money. When people choose to pay I use that as a signal for which topics to put future effort into.

My wife has been my editor for ever, so this is a family business.

EDIT: over 20 years ago, I was what I think was Larry Lessig’s second ‘Featured Commoner’ and that was fun because they interviewed me and posted the interview in the CC web site.

I like the AGPL but don’t often use it. (I did recently license two coding harnesses I wrote as AGPL.)

A long time ago, I didn’t mind huge tech companies getting rich partially via open source, now I do. In the USA the collusion between government and big tech is starting to smell a little bit like fascism. Sorry, I have only complaints, not any good answers.

I should have said first commercial Go program because I actively marketed it. There were several previous Go programs.

I have used Lockdown mode also. Not that inconvenient but I didn’t mind the slight no JIT slowdown. I did find not seeing some images annoying.

thanks!

I love this.

I watched every game years ago between Lee Sedol, even though it was late at night. I love AI and I love Go.

Off topic, but I wrote the first commercial Go program for the Apple II in the late 1970s.

Please, I have a question: in September 2026, what is the price differential between buying GPT inference tokens vs. a $20/month subscription?

I have moved to just using Google Gemini occasionally paying for tokens, no subscription, and it is OK, given my low level of use.

I would like to add gpt-6 astra to my toolbox, but I probably would use it infrequently.

Meta is getting most of my casual vibe coding business as long as they keep giving about 90% discount in the ‘we train on your prompt interactions data.’

I have had sos-so results with their local muse 30b model, but the hosted API is very fast and I have been getting good results.

When deepseek-v4-flash-0731 was released I used it constantly for everything, and I loved the very low cost and speed. I think Google has the same game plan with their flash models.

I am not surprised. Google is a real company that wants to make money on selling services.

Anthropic and OpenAI are in a different game of spending investor money to buy market share.

It is trivial to set up a spending cap on Google Cloud and also on Google AI Studio.

That said, it is probably easier to use APIs from FireWorks.ai, Ollama Cloud, Open Router, etc.

I agree, I use my own harness (link to the most stable version, from my Racket book: https://github.com/mark-watson/Racket-AI-book/tree/main/sour...) and except for handling user interrupts correctly, writing harnesses that are customized to your workflow is fairly easy.