Okay, I’ll admit it I just needed an excuse to finally buy a Mac Mini (been wanting one for way too long).

The serious answer though: some workflows really do benefit from a device like this especially when you’re trying to avoid website rate limits or need to integrate with actual macOS apps. It’s not quite the same comparison as a Raspberry Pi for that reason.

Bonus: I’m running local inference for the tool directly on the Mac Mini using LM Studio. For compute per dollar, that’s honestly some of the best value out there.

Prompt injections are real and there is no real solution for them... yet. Forget about how irresponsible your Clawd/MoltBot usage is if you don't 1)use a password managers for unique passwords per account 2) don't enable 2FA on everything (whether authenticator code apps or passkey vs SMS based 2FA) you can consider yourself lucky that your accounts have not yet been compromised.

For the paranoid a hardware security key where applicable is even better. (I'm looking at you U.S Banks with dismal security that still don't support hardware keys for customers and rely on SMS).

Expecting people to "not use" technology because it's unsafe is a losing proposition for consumer or enterprise apps. It's our responsibility as people building in technology that products are secure. Passing the blame to the user is 1) not an excuse 2) a sure way to not succeed as a business.

P.S: The Clawd/MoltBot product has in-depth documentation and tooling to avoid security issues. I'm sure as an awesome open source project it will evolve overtime to be more secure by default. Even if it doesn't it's definitely bringing the right conversations around proper credential management practices that shouldn't exist in 2026.

AI Assisted White Collar Work: Optimistically Pessimistic

A while back, we heard about tools limiting usage because of abuse in coding tasks. At the time, I couldn’t really grasp how someone could hit those limits so quickly.

Today, I accidentally burned through 89% of my “MAX” plan in a single workflow. Is my workflow inefficient? Very possible. Am I not optimized for token usage? Also yes.

But that’s kind of the point. I’m growing more optimistic about the quality of what these models produce, yet less optimistic about our ability to scale inference. That matters even more for text-heavy work, not just coding.

A lot of white-collar work is really just planning, researching, and curating. So what happens if we actually find the high-value workflows that justify the cost? I’m not sure. But it feels like we’re still far from the power we’ll need, assuming today’s generation of models is even the right path forward.

Probabilistic output has enabled generative tools to become both more accessible, as they do not require explicit programming, and highly unpredictable.

As generative tools mature, they tend to become increasingly constrained, whether through the system prompt, sub-agent context, or memory.

This evolution is not inherently negative; rather, it highlights an interesting shift away from purely probabilistic behavior toward more programmed and deterministic outputs. This dynamic warrants careful consideration when developing new products or frameworks built around such systems.