Daigo

Just my thought dumpster 🚮 | Bootstrapped @OmniCreatorClub & @handoffcloud | 🇯🇵 @DaigoTanaka_jp

[Daigo @DaigoTanaka]

I live in Mountain View and we rent.

The other day, someone asked me on LinkedIn what was the point not spending while earning high.

We don't live in a mansion but our kids are happy.

Instead of eating out expensive food, our kids love what my wife and I cook every day.

We see a neighbor kid sent to school by a hired driver while we walk and talk to the elementary school every morning as a whole family.

Instead of stressing out at a big tech, I run a one-person tech business on my term.

My wife and I saved and invested enough and now our annual investment return alone is much bigger than the money we spend. So whether we work is a choice. We are in our healthy 40s and still love working.

We aren't too thrifty but never upgraded our lifestyle like others do around here in Silicon Valley. That was the key for the growing assets.

We don't look like rich people, but we have time freedom, choice, and peace of mind thanks to that.

(Wow, that was a long reply.)

[Daigo @DaigoTanaka]

My favorite local LLM setup (updated):

- DELL Pro Max with GB10 (125GB unified memory)
- NemoClaw with Hermes
- LLM: nvidia/Qwen3.6-35B-A3B-NVFP4 (Optimized model for Nvidia GPU)
- vLLM (max-model-len 196608, max-num-batched-tokens 8192, "temperature": 0.2, "top_p": 0.9, "top_k": 20)

It's pretty fast for general conversion. Powerful enough for research, writing, and light coding.

[Daigo @DaigoTanaka]

I was sick on Father's Day 😿
So my family celebrated belatedly for me 😊
I sent this picture to my home AI, and
as you would expect, it interpreted the image impressively.

What you might not expect is that this is not Claude or ChatGPT.

This image interpretation is done by an open model, running on my home computer.

It means more than you think.

First of all, I'm not paying Anthropic or OpenAI for tokens.

Second, I'm not sending my personal data to those companies.

Third, the free, open models are becoming smarter and more useful.

Not every task requires Fable or GPT 5.6.

The future is locally running AI.

Would you stop paying money for tokens if you could?

[Patrick Collison @patrickc]

New from Stripe Economics: The Age of the Solopreneur

https://t.co/GTS8jEcK5z

[Daigo @DaigoTanaka]

For fun, I uploaded a picture of my latte to ChatGPT.
And this is what it made for me to improve ❤️

I was amazed by how AI understood the level of my latte technique and precisely suggested what I need to work on.

It even generated an illustrated piece of advice tailored to my situation!

Can this be achieved merely by correlating the training images and their associated text?

Or does it require "real" understanding of physics and art? (Define real for me...)

Or is it just making things up and convincing me that it's valid advice with polished writing and design?

If AI does all of the amazing things without a "real" understanding and reasoning, we are dancing with a "ghost."

And we may need to admit that a spiritless hollow is just what we need to keep us entertained.

What were you impressed or spooked by AI recently?

[Daigo @DaigoTanaka]

Now my day-to-day local LLM operation is back again with newer NemoClaw (v0.0.66)...except I switched to Hermes + vLLM (from OpenClaw + Ollama).

Qwen3.6-35B-A3B is doing a great job with coding and research writing. I had to modify my system prompts for Hermes, but it was worth.

Not that I hate OpenClaw, but the tool-calling breakdown was a deal breaker for now.

[Daigo @DaigoTanaka]

After upgrading NemoClaw/OpenClaw, tool calling through Qwen 3.6 broke. As an experiment, I switched to Hermes, and the issue disappeared. Now I need to rebuild my system prompts for the Hermes environment. It's been especially struggling with the OpenShell gateway-related components and keeps encountering issues.

[Daigo @DaigoTanaka]

Here is the project page:

https://t.co/p92Hgg95l1

If you (or your kids) happen to own a Dash robot, give it a try. Or drop the link above to your AI agent (Claude Code, Codex, OpenClaw, ...) to get started.

[Daigo @DaigoTanaka]

What if your AI had a body?
So I grabbed my kid's robot and let Claude Code control it.

(The project page is in the post directly after this one.)

I'm not a robotics guy.
I had no clue what it took.

But Opus 4.8 was smart enough to:

- Learn how to send signals to the robot from other projects.
- Design the tests to calibrate the wheels.
- Develop navigation algorithms and track the robot live on a web dashboard.

Claude wrote over 10,000 lines of code, just by us working together a couple of hours a day for a week.

I gave the AI research goals and thought through ways to improve together.

I learned a lot about today's AI capabilities and the roles we humans need to play when AIs are mostly executing tasks.

Have you ever used AI to solve something you had no idea how to tackle? What did you build or learn?

[Daigo @DaigoTanaka]

The complete death of SaaS may ultimately come from the next phase of AI:

Personal AI computing.

Today, AI is largely a cloud computing business.

But eventually, more AIs will run directly on our own computers.

When that happens, the entire house of cards built around “SaaS businesses serving other SaaS businesses” could collapse surprisingly fast.

Thoughts?

[Daigo @DaigoTanaka]

Those who just judge are nobody.
Nobody is judging you.
Go be yourself.

[Daigo @DaigoTanaka]

@TaraBull Dr Nakamatsu’s jumping shoes invented in the 80s:

[Daigo @DaigoTanaka]

Any sale items I should consider for this week's meal plans based on my purchase history? I asked my AI agent.

...It told me St. Louis-style ribs are $12.99 each. I'm going to smoke ribs this week.

How hard was it to achieve this "user experience"?

The easy part: Asking AI.
The hard part: Preparing the data.

The grocery web app is where you'd hope to have public API access available to retrieve the data automatically.

But we have an AI agent now.

After I reverse-engineered the site's API with an AI coder, now I can fetch:

- Item-level purchase history
- Total spending per visit
- Savings for sale items
- Member savings
- Coupons

Before the AI, we've been saving an average of more than 11% per visit through the membership discount and sales.

We also saved an extra 10% from digital coupons when my wife claimed them.

Total of $1,288 saving in the past year.

Now with the AI agent, I just need to ask AI to clip the relevant coupon for us, so my wife doesn't have to do it manually. (They are automatically applied when we shop; no need to do anything to get the discount.)

Before AI, I did not bother to check the weekly deals before deciding on the week's meal plan.

And it's often hard to decide what to cook.

Now AI suggests the sale items, and inspires me to decide what to cook.

[Daigo @DaigoTanaka]

For AI agents' recurring tasks, I recommend stopping the use of web drivers. It's token-wasteful and slow. Reverse-engineering the API for most sites is easy. Start by dumping HAR file from Chrome. Keep it safe, though.

[Daigo @DaigoTanaka]

Market with emotion and sell the information.

[Daigo @DaigoTanaka]

Write prompts for every task you do today.
Let agents handle them tomorrow.
That's the spirit, at least.

[Daigo @DaigoTanaka]

For agents, prompts are algorithms written in human language. Craft accordingly.

[Daigo @DaigoTanaka]

Writing demands an algorithm.

Good papers share a common blueprint; bad ones fail in unique ways — I learned this while I wrote my PhD thesis.

Social media content may look very different from academic papers, but it also requires a clear method. It miserably fails otherwise.

I believe AI, with a good algorithm or guideline, can write better than humans without one.

And most of us are bad writers.

[Daigo @DaigoTanaka]

At AI Agent Conference in NYC, I just heard @simonchannet said AI agents are better at doing deterministic tasks.

True, but it’s always the frustration their outputs aren’t deterministic enough.

Once one engineer it, one will know the pain of building reliable systems.

I’ve been automating part of my one person company by AI agents.

After weeks of intensive work, I finally starter see something reliable enough that I can let it do autonomously while I sleep.

But it required surprising amount of engineering knowledge.

You still need engineers.

Just the engineers who are eager to learn the new agentic paradigm.

[Daigo @DaigoTanaka]

Tech employees are to be either laid off or subjected to incessant cognitive overload while babysitting fast-typing AI agents. Such dystopia. Own a business run by AI instead of working for AI.