Mark Watson

@mark_watson
533 Followers
366 Following
350 Posts
Author 20+ books, mostly on AI, LLMs, deep learning, semantic web, Lisp. 50+ patents. My books are free to read (or buy DRM free copies) online https://leanpub.com/u/markwatson
Artificial IntelligenceAI
Deep LearningSymbolic AI
LispHaskell
Consultant
Latest tiny models are fantastic for many tool use and embedded applications. I really like the tiny nemotron-3-nano:4b model that is the smallest of the nemotron-3 family of models that Nvidia just released. When thinking, it thinks quickly not getting caught is loops. I also just ran 6 tool tests: all worked fine. So good.

@veer66 If I am accessing Ollama form a different computer via Tailscale I set both:

OLLAMA_HOST=0.0.0.0 OLLAMA_CONTEXT_LENGTH=32768 ollama serve

If I am running Ollama on the same computer and just want to increase context length I set:

OLLAMA_CONTEXT_LENGTH=32768 ollama serve

Good catch! Coding agents need larger context length.

My current local dev setup: running qwen3.5:35b on Ollama. OpenCode config: choose "remote Ollama" provider (important!! don't use OpenAI compatible provider). Then select qwen3.5:35b from your list of available remote Ollama models. (Optional: I run Ollama on a 32G Mac mini that is stored away running headless out of sight in a closet, and I use my new MacBook Air to work on via TailScale from any location.)

I never dreamed I would have such a good setup on my own inexpensive hardware.

RE: https://masto.canadiancivil.com/@paige/116180225338582120

This is pretty shocking, but logical: the news media and special interests in the USA spend a lot of energy splitting the country into democrats vs. republicans who hate each other. Segmenting the population is by design, to more easily keep people from realizing that the real enemies are elites and special interests. Diabolically clever!

The only logical thing for US citizens to do is to register as Independents - make it harder for special interests to control the population:

I don't think the UK media have made nearly enough noise about the US Navy sinking an *unarmed* Iranian warship and then, in clear violation of the Geneva Conventions, leaving them to drown.

This is a heinous war crime which the media would be rightly up in arms about if any country that wasn't nominally an "ally" had done it.

https://ihl-databases.icrc.org/en/ihl-treaties/gcii-1949/article-18/commentary/2017

#IRISDena #IranWar #WarCrimes #GenevaConventions

AI is central to this illegal war.

New from me -
https://buttondown.com/creativegood/archive/ai-and-the-illegal-war/

AI and the illegal war

“No stupid rules of engagement.” That was Pete Hegseth, the American war secretary, boasting during a press conference about the approach with which he and...

Creative Good
I just posted a note for readers of my books https://open.substack.com/pub/marklwatson/p/i-removed-my-openai-account-notes explaining how I deleted by OpenAI account and where older material on using OpenAI APIs will be stored for their access if they want it.
I removed my OpenAI Account - notes for readers of my books

I was outraged by the pressure the Department of Defense and the Trump administration put on Anthropic last week to accept changes in their contract.

Mark Watson’s Artificial Intelligence Books and Blog
@spritely Wow, this is a busy time for me so I will try to not carve out a half day to experiment with Hoot, right now. Dev experience is everything for me. Hoot repl looks good in the video, and I would like to try the geiser-hoot Emacs support.

OK, I am an old man so take my ramblings as you will.

The suppression of the Epstein files is one of the worse things I have seen in my lifetime - not as bad as some lies to justify wars for profit, but still really bad. I hope that other countries prosecuting Epstein network child rapists will force the release of more information in the US. I have friends who for political reasons don't want this info released. I accept that everyone gets their opinion but it is my right to be disappointed.

Fast, portable, and private: I added a short chapter "RAG Using zvec Vector Datastore and Local Model" to my Ollama In Action book. You can read the new chapter online using the link https://leanpub.com/read/ollama/rag-using-zvec-vector-datastore-and-local-model

The new vector store library zvec was recently released by Alibaba and it is full of awesomeness.

Leanpub: Publish Early, Publish Often

Leanpub is a platform for authors to write, publish and sell in-progress and completed ebooks and online courses.

Leanpub