mark_l_watson

That is a big spend. Hopefully (my last job before retiring was working for a South Korean AI company, so I am rooting for them) they will heavily reuse the research by DeepSeek, etc.

As much as possible countries should have sovereign technology, and as human "sovereign individuals" we should mix local models and public APIs for open models as much and makes sense - case by case.

I agree that pi is very good, don’t agree that it is the last harness (I have written 4 harnesses myself this year, fantastic experience.)

I mostly use local models. I use pi with the strongest models my 64g mac can handle, and for fast light work I use ‘pool’ with Laguna 2.1 XS 6 bit quant. I am happy :-)

Nice story! As an aside: my wife and I only buy pasta imported from Italy because it is simply so much better tasting than pasta made with commercial US-grown wheat. I don’t know why.

The food industry in the US is pretty much messed up. We try to only buy fresh produce and fresh fish, staying away from processed food - but it is difficult. I read something about the composition of the modern Butter Finger candy bar (fake chocolate, lots is chemically processed seed oils) compared to the almost wholesome ingredients used 30 or 40 years ago.

Indeed, recommendation algorithms have some positive benefits and a lot of downsides also. As a deep learning practitioner I worked on recommendation systems for two companies, but as a human being I have my own way of dealing with them: I like to once or twice a month dedicate some time to finding good content to consume: my wife likes streaming media so we have all the subscriptions, so I look around for about three movies that look good and download them so when I want to watch something I choose from my downloads. I spend time on Kobo (eBooks) and Libre.fm (audio books) to make sure I have purchased good reads that are all queued up for my enjoyment.

Finally, Apple has a great tool called ScreenTime that I can set limits on HN, YouTube, etc. I set a 15 minutes a day limit for each, but I can always add 15 minute increments if I find a good rabbit hole to explore.

Nice! I have been using herdr for a few months, love it and for me it is a step up from years using tmux. The new native Herdr app is yet another step up. I use multiple Macs with Tailscale connections and the Herdr app makes jumping around to different projects and different machines easier.

I think I bought the Emacs Manual from Richard Stallman in 1979, give or take a year. I still live in Emacs, even writing my own AI coding harness inside Emacs.

I suspect that I am slightly more productive in VSCode for large code bases, but it just makes me feel happier and content to use Emacs.

This is just a hunch since I have not used Argon very much, and I have not used it for coding (deepseek 4.1 flash 1017 with dsh, and qwen 3.8 max are so good for coding tasks that local models have difficulty with).

My hunch is that Google will end up being best for non coding tasks. Google can concede coding to other vendors who pay Google for infra.

While I agree with you in general, I think the real value of AI (for me, I wouldn’t want to try to change anyone else’s mind on this) lays in three areas: 1. all the little wins saving time at little tasks like analyzing a Haskell, Java, etc. error stacks; and all sorts of little time wasting problems. All those saved minutes add up. 2. Brainstorming before coding, and even before design. I like to write up ideas in a Markdown file and then ask one or two strong AIs to push back, to try to fill in what I might have missed. 3. Non-tech thinking: without any formal training, I enjoy reading philosophy. I frequently ask for second-takes from Gemini on what I am reading; once again, I am looking to just improve my own understanding.

AI should make us better, not just churn out trillions of lines of code.

The very idea of massive vibe coding efforts, where devs go to sleep with a hoard of agents churning away, makes me uncomfortable.

Even the Alibaba $6/month plan supplies a reasonable amount of API use, and Qwen 3.8 Max is a killer model for long time frame agentic use.

Literally everything I do is Open Source software and Open Content writing so I am fine with the lack of privacy using direct Chinese APIs. And, for a little more spend, just use US companies like Fireworks.ai or Ollama Cloud for more privacy, or run on-premise locally and get the best grade of privacy.

As of this morning, that does not work for me.

Another new annoyance: the Settings app has 2 red notification icons because Apple is trying to sell me extended AppleCare. Apparently no way to turn off these notifications.

Have you used Nemotron-3.5-lightening? I don’t use it as much as Poolside’s (excellent!!) Laguna XS 2.1 6bit, but the new Nemotron model is good.

I think NVIDIA does want small open models running on-prem to explode as a market! Lots of smaller GPU installations for companies who wisely want on-prem inference.

Of course NVIDIA will also keep making a ton of money selling to hyper scalers, but not forever: Chinese chips are getting better, Google, Microsoft, Amazon, etc. designing their own inference chips.

NVIDIA is handling this brilliantly.

Maybe? I am using a 3 bit quant for Qwen Next Flash in a 64G Mac using Sushi and the results for coding, design, refactoring, etc. are very good.

That said, after being satisfied with Poolside.ai’s Laguna XS 2.1 4 bit quant, when I upgraded to a 64G Mac, I now only run the 6 bit quant and the results are better.

Unfortunately, I think that benchmarks are not useful for me, and I need to take time seeing what works for me: this is one reason that when I find a combination of model, harness, etc. that works for me, then I tend to use it for a month or two before looking around for new stuff.

Great product idea. I have not paid OpenAI for a subscription for a couple of years, I’m tempted now. I do about 75% of my work using local models and most of the rest using the deepseek 4.1 flash API. That said $20 to experiment with this new product for a month sounds like a pretty fun idea.

For routine coding I run Laguna XS 2.1 6 bit quant with either the poolside.ai harness ‘pool’ or pi-dev. That is fast and almost as good.

re: “”In Python, can you fix the buggy function, re-compile it, and resume execution from the stack you want?””

Yes. I code Python in Emacs with the ancient Python support for running a REPL, loading buffers, editing a function and just hot reloading the modified function, etc.

I am an old man, and I like old man tools :-)

Qwen 3.8 Flash Next is amazing. I only have a 64G Mac so I have to run Sushi project’s 3 bit quant. Amazing results with pi-dev. More for fun than anything else, but I am trying to do as much as possible with local models, now rarely falling back to a paid deepseek-4.1-flash API.

Progress on running local models has been amazing.

I use Muse and only give it access to my Apple Reminders - just to play with it.

I also have two daily tasks for local events and things to do and also a tech summary of new tech news.

Looks interesting. I just went to download from HF, but they only have fp16 which won't fit on my Mac.

Good to see Europe adding toe what Mistral is doing. +100

That would be interesting. I hate to say it, but he might have given us something like Muse by Meta. I made a comment on HN a year or so ago about wanting a strong personal AI agent from a company I more or less trust (e.g., Apple or American Express).

This looks excellent but I only have a 64G M5-Pro Mac so I probably need to stick with Sushi running Qwen3.8-Flash-Next.

It was inevitable that the providers would stop giving away free cookies and milkshakes.

I pay Google $5/month for Chat and my wife gets to share that: a fantastic deal. I even get a short weekly (one each) opus and gemini vibe coding sessions with agy.

I pay for inference: I paid $3300 for a Mac mini 64G M5-Pro which runs Qwen-3.8-Flash-Next (3 bit quant) with pi-dev and the results are so very much excellent enough for about 60% of my work and paying Deepseek for deepseek-4.1-flash with Deepseek Harness handles almost everything else.

I worked at Google in 2013 and I mostly like them as a company. I like their approach to AI in general, but I hate the business models of Anthropic and OpenAI and I fear the US taxpayer is going to pick up the tab on this one.

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.