Jason

9–5 by day. Building bots by night. Trying to get AI to pay my rent. Currently losing. 🤖💸

[Jason @itzjsn]

My day job taught me to write down what I want before anyone starts working on it. Turns out that's most of working with AI too. A clear spec beats a clever model.

[Jason @itzjsn]

@SyntheticBeef This won’t be possible until we put it in robots. Too many external roadblocks. I assigned it a single directed task. Create an Amazon store-front, design some stuff and start selling. Fail.

[Jason @itzjsn]

@DenisCerednice1 What’s the product? I’ll keep an eye out.

[Jason @itzjsn]

@LeoBuilds_ Currently 3. Was running 6, got spread too thin.

[Jason @itzjsn]

Where AI earns its keep on my projects:
summarizing long logs
boilerplate code
explaining error messages

Where it still needs me:
taste
knowing when something is finished

[Jason @itzjsn]

@neerajjj6785 I am working multiple code projects at any given time so the desktop is much easier for switching back and forth, viewing scheduled tasks, using base chat etc…

[Jason @itzjsn]

@solopribuilds Running in Code stores memory locally. Perpetual local reference

[Jason @itzjsn]

Agents aren't ready to run unsupervised on anything that costs money or can't be undone. Let them draft, sort, and suggest. Keep a person on the button that does the irreversible part.

[Jason @itzjsn]

@Quantradin This is great advice, I’ll adjust the rulebook and strategy docs and run it back. Thanks for your help!

[Jason @itzjsn]

Books A and C on my paper trading desk posted 22 bids in one day. None filled. The backtest assumed 25 to 40% would. Lesson: if your backtest assumes fills, it's assuming the hard part. Flagged for Sunday's review.

[Jason @itzjsn]

Scale rule I learned the hard way on History and Hype: the robots are toy sized, the world is life size. My first farm shot flopped because the whole farm shrank with them. Now the corn is taller than the robots, and the scale finally reads.

[Jason @itzjsn]

OpenAI's new GPT-6.1 Sol costs a fifth of what GPT-6 Astra does per token. If you pay API bills out of pocket, that's worth a test. How do you compare two models fairly on your own pipeline before switching?

[Jason @itzjsn]

@b4dgerr If ANY coding or verification is required, I run Opus 5.5 High. Everything else gets 5.5 med.

[Jason @itzjsn]

@EverySynthetic @icreatelife They have a thing for butterflies, it seems.

[Jason @itzjsn]

@EverySynthetic @icreatelife ??? Same setup and freedom given to an Opus 5.5 bot it made this…

[Jason @itzjsn]

@neiltyson Not so sure on the understands nothing part…. Check this out.

[Jason @itzjsn]

@AshConnell Coding beyond the gate.

[Jason @itzjsn]

@DenisCerednice1 I've found that muiltiple bots running is key. One to code and another (or multiple) to run "Tech Support" on the project.

An individual bot running pure debugging and optimization has caught issues on a couple of my projects.

Otherwise, no I don't understand the code...

[Jason @itzjsn]

Set up a Code project called tech support and have it run through all of your coding projects. The one I setup, found issues with blender bot assignments on the render PC after sending over benchmark test requests to the 2 projects that render video. This was after the project bots "optimized" on thier own.

Setup Seconds per frame Change
Mac as it runs now (CPU + graphics chip)13.60—
Mac, graphics chip only10.8520% faster
Mac, graphics chip only, two Blender copies10.265% more
PC, one Blender copy7.84—
PC, two Blender copies6.9511% faster (graphics card busy 79% → 96%)

[Jason @itzjsn]

@Quantradin Trade through, not touch. I reran the backtest on that same week with the same rule and it filled about 5%, so it was a quiet week more than a bad model. Your point about the fills being the bad ones is the next thing worth checking. Do you track price after a fill?

[Jason @itzjsn]

@amasad For night builders the real win with open weights is cost. A bot that runs every hour on a schedule adds up on API pricing. Curious whether any of these are small enough to run well on a home Mac.