mark_l_watson

Sorry to be off topic, but I have a question: has anyone installed the latest beta iOS and macOS, and if so what is the current status of Gemini integration?

We agree more than you may believe. I have worked in the field of AI since the early 80s, symbolic AI, simpler neural networks. I only believe in using any tech if it serves human needs, is privacy preserving, etc.

I agree that many AI businesses will go bust and they deserve it, but the tech is good.

I can recommend my own layered approach, using the lowest capability models that get stuff done:

1. I maximally use local models like gemma4:26b-a4b-it-qat for everything that works with this free option.

2. I like paying for inexpensive APIs for mid-tier models like deepseek v4 flash, gcp-5-mini, gemini-2-flash for things that option 1. fails at. This option is almost free.

3. Pay for more expensive APIs like deepseek v4 pro, gemini 3.5 flash, etc. This option is not too expensive.

4. If all else fails on a class of tasks, then pay for awesomeness of Claude Opus. $$ expensive, I try not to use unless absolutely necessary.

I think developers and companies that just cram everything into Claude Opus are unprofessional.

If you pay for Gemini, then it is good. I recently used Gemini Ultra for a month and the gemini models are very good (and of course, you get a lot of Claude Opus tokens to use through the same plan).

I also pay for Proton's Lumo+ private chat and for what it is it is also good.

The free plans from all the providers are bad, which is fare enough.

I use Apple devices and I expect to be paying for Gemini tokens after the integration.

I have been using their on define AFM models for a year - for small models they are good. Their Secure Enclave server bases AFM model is good, but not in the same class as gemini 3.5 flash or deep seek v4 flash.

I don't think so. They will be running on their servers, or running in the future on Google's servers with privacy guarantees.

Do you think Google doesn't protect privacy for large paying customers?

For years I have enjoyed using Google products that I pay for, and they are clear about privacy guarantees.

I have loved using AI technology for 45 years (symbolic AI, old fashioned NNs, … to the present). I am also skeptical about the apparently desperation-driven ‘bet the farm’ approach we are taking here in the USA.

Slow is Fast.

I had odd Gemma 4 12B results: it was ‘almost excellent’ for writing code in a variety of languages if I was using a detailed one-shot prompt describing new code to write.

I had horrible luck with Gemma 4 12B with a variety of coding harnesses - but as usual Qwen 3.5 9B did OK.

EDIT: CORRECTION: I pulled a fresh copy of Gemma 4 12B and inference code and the tool use problems in my test harnesses are fixed. Gemma 4 12B is slow on my 16B MacBook Air, put produces OK results.

I tend to move back and forth between Common Lisp and Racket - so many good things from both communities and tech.

I turn off running scripts on installation. So far, no inconveniences.

I love the zero dependency implementation. I do this style of breathing during specific time periods of practicing Qi Gong. I will try your script when I get to my laptop. Thanks.

Interesting ideas for generalizing goals to reduce human labor in human <—> agent interactions. That said, maybe it is better to set up customized skills and infrastructure for large projects? At our early stage of trying to capture value of agentic systems, the good ideas in this article might be premature optimization.

Until Anthropic, OpenAI, and Tesla have IPOs and are then bound by some laws to be truthful, I don’t want to bother about their possible valuations.

I do care about: how useful their products are vs. cost and how secure are their businesses. Actually I only care about the first thing since these services are hot swap-able with some effort.

That is a very cool web app with lots of interesting examples and a good place to read code examples and be able to run them. I tried using the editor Mine written in Coalton. Coalton is a nice language project but I can’t get into it because I have been using Common Lisp since 1982, old habbits die hard.

Sometimes I see things that make me reevaluate assumptions I make. I am all in on flexibility using LLMs and agentic coding harnesses: I hate feeling locked down to one platform.

Then I saw this in the article:

>> I've discovered that Claude Desktop supports MCPB. The MCPB provides a Node runtime. Therefore, our application would only contain our code. This means the size of the application would be ~1MB.

I don't know why this is so appealing to me but it is. I currently use Claude Code with a DeepSeek v4 Pro API backend. I will check out if Claude Code itself has MCPB support.

Well, good for them that they are charging enterprises API rates. Why in the world not do something similar for consumers? Use for free a few times a day, have a $5 dollar plan for light use, and perhaps $10 to $15 for heavier use. If 90% of consumers pay nothing then ‘drop them’ except for letting them have an account and a few queries a day.

It is easy for me to change providers. Right now I use the open source Claud Code harness with two paid API venders for DeepSeek v4 (flash and Pro). I like seeing how much each session costs.

I feel sorry for people who have to use strong AI agentic agents all day long for their jobs. I just came off of a 30 day experiment using Gemini Ultra (all the Antigravity+Claude Opus I could use) and while it was great to re-work a few dozen of my open source projects and to check my Open Content books for inconsistencies and make improvements, the awful thing was it felt dehumanizing. I am now just using DeepSeek v4 for less than 1 hour a day and that feels better: a good mix of getting help when really needed and doing my own thing by myself.

Thanks for the references.

BTW, Google is my pick for the winner in the USA tech giants AI race. I worked at Google about 12 years ago and was impressed by their use of renewable energy, etc.

Great article that reinforces my own opinion but adding the cleverness of adding low cost human labor into the equation. Nice.

I spent a month comparing Gemini Ultra plan to using much lower cost DeepSeek v4 with open source coding harnesses and, spoiler alert: I was happier using the much cheaper and more environmentally friendly open models: https://marklwatson.substack.com/p/my-evaluation-of-ai-agent...

I started doing this at work in the late 1970s: if I had to talk with someone at work about new code, design, etc., I would always suggest we walk outside for a while and talk+walk. Big win in creativity, making good group decisions, and making the work day better.

Pardon the almost 50 year old nostalgia, but I got so much out of manually typing in code from old BYTE Magazine and other articles.

Twenty years ago my neighbor, a retired surgeon, made me a PVC didgeridoo and did the wax buildup thing - I still mostly play that didgeridoo. Years later my wife bought me a traditional heavy didgeridoo from Australia, but it doesn’t play as well; still, when I played at a friend’s wedding I used the Australian one because it looks better :-)

My didgeridoo teacher had the class practice at home continuously blowing air through a straw - it still took me almost half a year to reliably be able to do circular breathing.

I have read a few references that humming or ‘ohming’ help sinus health and breathing so I guess it makes sense playing the didgeridoo would help also. Blowing bubbles through a straw won’t cause vibration, so probably in itself won’t help.

I tried it and the text input area was black with a dark font. I checked the documentation, and asked DeepSeek v4, Claude, and Gemini for help with the fonts/style and nothing works except to run in a terminal with a dark theme. Crazy. None of the devs on the project use a light theme?

While you are correct that something like Antigravity 2 + Opus 4.6 can handle large scale software engineering tasks, I would argue that it is usually (but not always) better "coding agent hygiene" to work on smaller code modules and as the human in the loop be a partner, not someone who prompts and then disengages.

Breaking code up into composable chunks has worked well for me over 50+ years as a professional software developer, and I can't get away from the idea that it is still usually the way to go using agentic coding tools.

> "which use cases earn the inference cost they burn?"

That is the question. I love using OpenCode with paid inference providers and seeing the cost of every little thing I do. On the other hand, right now I am flipping between Antigravity CLI and the two Antigravity apps burning Claude Opus tokens like crazy, knocking off a ton of work. Google must be losing money on me.

The app over-write thing is not good. It took me 90 minutes to get the new chat Antigravity, the new Antigravity IDE, and Antigravity CLI installed and one task done on each.

My 1 month subscription of Gemini Ultra is finished in three days and I revert to their $20/month plan. Assuming that daily and weekly quotas are OK for casual use, I will probably use AntiGravity CLI most of the time.

Off topic, but maybe interesting: during my one month test of Gemini Ultra I did several tasks in parallel to compare (old) Antigravity+Claude Opus vs. OpenCode with a fast provider for deepseek-v4-pro, kimi-k2p6, and minimax-m2p7. In almost all cases I could get stuff done in about 60% to 70% of the time using Antivravity+Claude Opus -- but!, OpenCode with the open models is so much cheaper. I get that in a work environment when someone else is paying for tokens, why not burn someone else's money. Paradoxically, I felt more relaxed after the tests with OpenCode with the open models even though I was actively doing more work myself.

EDIT: two months ago I wrote my own small coding agent in Emacs Lisp that I enjoy using. I am researching redoing my Emacs project using the new Antigravity SDK.

I agree on search, really like DDG. I only rely on Google products that I pay for (Gemini APIs, YouTube Premium, occasionally pay for Colab, purchased Books on Play).

You are using Q6 6 bit quantization; on my 32G MacMini I use Q4 and it is faster but when I use it with OpenCode, I set up a task and go outside to walk for ten minutes. Smart, capable, and slow. Still, I love using local models.

EDIT: I run with context wired at 64K

Strong agree with the premises of the article: I like the framing that the AI hype-masters are successful because they instill a fear of missing out in corporate leaders.

I have worked with old fashioned neural networks, deep learning, and now LLM-specific deep learning: wonderful technology, but over hyped, and advice to go a little slowly, with firm use cases that are financially viable is great advice!