Jim Coleman

SaaS and Agency Founder | 🤔 Thinker | AI Tinkerer | AI Fusionist | Recovering PE Operator

[Jim Coleman @JimGPT]

Humans follow incentives, not intentions.

To predict behavior, understand incentives.

To change behavior, change incentives.

[Jim Coleman @JimGPT]

It's Thursday night, which means Torchy's Tacos for dinner on my way to group. I'm in the habit now of having Grok @bot place the order for me.

I had it create a new account using its own email, save the password in its vault, and save my order preferences. Now when I'm ready to order, I just ask. Whether from the Grok Bot iPhone app, on my Mac, or while 'driving' down the road in my Tesla.

What a time to be alive!

But I was thinking... While this is freaking awesome, it's still fairly slow because Grok Bot is literally using a browser on its own computer to order from the Torchy's site.

Soon, every restaurant will need to offer an MCP/API connection so personal agent bots can order in a more agentic and less human way. This will speed up the process 10x and burn less tokens.

The infrastructure of the internet is becoming headless. Agents can use a traditional UI, but it's far more clunky. The agentic web is coming.

[Jim Coleman @JimGPT]

Frontier lab flywheel is basically:

-scrape all human intellectual contribution from all of time (books, written speeches, and the internet) to train their model

-surface this new artificial intelligence to humans.

-humans use it and thus feed it more and more useful and not-yet-captured private data

-models get even better

-humans teach it how to do actual tasks (human IP) on a computer

-models can now think and take actions on behalf of humans

-models are now smart enough to replace human’s jobs

-humans devastated when they realize what just happened is thievery on a level that’s so unprecedented it’s hard to comprehend

And yet, every day we use these tools. Like a dog returning to its vomit.

My own cognitive dissonance feels impossible to reconcile.

[Jim Coleman @JimGPT]

“Put some brain in your thinking” is what Paul’s seven year old little girl told him to do.

It struck me that she won’t remember a pre-AI era like we do. She’ll only know a world of dual intelligences, and that the silicon one surpassed the meat one.

It’s unsettling to think about.

Equally so is my reflection on how much of my thinking I outsource to the silicon monster. My own atrophy catches me off guard. I can feel the knife edge of my mental acuity dulling.

I prompt it, read its output, and think I’m learning. But am I? It’s easy in and easy out. Held only temporarily in my short term memory. Then, like a vapor, it’s gone.

So do I really get it? Did I learn? Did I use my own brain?

I’m not so sure.

And the temptation is great to outsource to the silicon thinker. I’m writing this to you on my phone, and with my human fingers. My hands are too big for this little phone, and my right hand aches. But it feels good.

Because I’m putting some brain in my thinking.

And it’s hard. Harder than it used to be. And that’s scary.

Not sure where we go as a society when entire generations no longer put brain in their thinking.

But at least this morning it feels surprisingly good to connect my brain to my thinking.

Feel this same existential dread, or is it just me?

[Jim Coleman @JimGPT]

Code is leverage. Media is leverage. AI is leverage on top of leverage.

But literally doing nothing is more productive than applying leverage upon leverage to something that shouldn’t be done in the first place.

Never has it been more important to apply human judgement to deciding what to lever.

[Jim Coleman @JimGPT]

Humans learn one by one. AI learns all at once.

If I make a mistake, I learn from it. But you don’t. And certainly not at scale.

AI is different. If one Tesla makes a mistake, it doesn’t just learn, it updates every Tesla.

If one AI agent improves, all interconnected agents improve. This isn’t linear growth. It’s exponential.

And I don’t think we’ve fully grasped what that means for the future.

[Jim Coleman @JimGPT]

At xFusion, I've given our leadership team Claude subscriptions and promised them they would never be laid off because of AI.

That's made a big difference in their willingness to try AI. In fact, non-engineers have built custom solutions we've implemented within the company.

[Jim Coleman @JimGPT]

Taught my 20 year old daughter how to use Grok Bot last night and it BLEW her mind! 🤯

Her main bot, Pippi, a cute Dachshund, quickly got to work. Grace asked for a "cute personal finance budget on a spreadsheet," and Pippi complied with a light pastel creation.

Grace took one look and promoted Pippi to "Queen Pippi," complete with her own crown.

[Jim Coleman @JimGPT]

Human curiosity is the AI implementation bottleneck.

[Jim Coleman @JimGPT]

V1 of your AI implementation last year was lackluster.

V2 this year was better. But still isn’t there.

V3 six months ago is much better, but not great.

V4 three months ago is a step change.

V5 one month ago is better still.

But V6 today is state of the art.

AI implementation is NOT set it and forget it. And if you’re still using an earlier version and say, “AI sucks.”

That’s wrong. Your old implementation sucks.

[Jim Coleman @JimGPT]

Employees hear on a regular basis about people being laid off from companies that have implemented AI.

And we humans react poorly to fear by trying to hold on to what we have.

Those same employees hold the tacit knowledge needed to properly implement AI, and their participation and buy-in is critical.

[Jim Coleman @JimGPT]

The misalignment of incentives between executives and employees may be the single biggest obstacle to realizing AI's economic potential.

The models are now well beyond what's needed to get ROI, but as per usual, adoption and implementation lag the technology.

[Jim Coleman @JimGPT]

@Shpigford Sorry! Meant to ask what your experience was using Gemma 4?

[Jim Coleman @JimGPT]

@AlexFinn I can run Gemma 4 locally soon. Is it good enough to execute on all non-orchestration and non-coding tasks?

[Jim Coleman @JimGPT]

@Shpigford @hnshah @Zai_org Interesting! Did you try the new 31b Gemma model?

[Jim Coleman @JimGPT]

@jessegenet @GoogleDeepMind @openclaw Nice! Should run fine on a 64GB MacBook Pro, yeah? I hope so. That’s my plan!

[Jim Coleman @JimGPT]

@Shpigford Totally. Doesn’t disappear, but gets way cheaper due to cheep replication. But that’s all good.

[Jim Coleman @JimGPT]

@saranormous @garrytan @karpathy @NoPriorsPod Great interview!

[Jim Coleman @JimGPT]

@amasad Better be careful what you say in meetings. 😆

[Jim Coleman @JimGPT]

@ImNikhil117 @sweatystartup Yes. This is a solid point and what they’re betting on. Makes the model different than Uber and other VC plays