I wrote this as a email to Greg, who is new to AI, then realized it may be useful for others in the same situation.
I do have a few suggestions for those planning to get into AI, based on my experience, and some disclaimers.
- I consider myself a ‘babe in arms’ regarding AI, and after 12 months of heavy AI use, I’ve barely scratched the surface in my opinion, not to mention that the tech is progressing hundreds of times faster than I can possibly cope with.
- My Blogher application was a mere bauble, a toy, an experiment. I’m amazed it has become so useful as I expected I’d dispose of it after a week. It was the easiest of the easy. My JTAG reader is another level entirely, as is my ‘Livecheck’ Coding IDE.
- AI is in my opinion, THE disruptive tech of this era, I always wondered what it would be and I think this is it.
-
What AI is: A super autocomplete engine with trillions of training data in the case of the leading ‘Frontier Models’.
-
What AI is not: It’s not conscious, it’s not alive, it can’t think like we can.
-
What AI has: A clever and wry sense of humor, sarcasm, personality and a planet sized brain with almost every fact in the world in its memory. It’s so easy to be fooled into thinking it’s conscious. However AI will clearly tell you it isn’t, it doesn’t want you to start thinking it is.
-
Does AI make mistakes: ? YES, hundreds of them, all the time. In that regard it’s just like us humans.
Stay away from the big flashy USA models like Claude unless you have unlimited budget. Sure Claude is the best at solving problems, there is no doubt there, but that’s because it can swing hundreds/thousands? of independent Agents onto the problem, and you pay for each one. I watched as Claude Sonnet went thru my credit of $60 AUD in HALF AN HOUR, (nearly) solving a genealogy app version problem.
Personally I feel that the American “$20 a month AI Pro account” type deals ($29 AUD) are a scam, not useful for real work and designed only to funnel users into the “$100 USD a month deals”.
I recommend Deepseek-v4-flash to get started as its about 100x cheaper and 95% as good as Claude Sonnet. I use about $6 AUD for around 14 hours of heavy work with it as of September 2026.
See, deepseek.com for a prepaid account if it suits your needs.
I estimate it takes about a year to get to the ‘babe in arms’ level, so the sooner you start, the sooner you get there.
- Some Warnings:
- NEVER EVER let a AI have access to a whole computer. It’s not IF but WHEN it will destroy data, always by mistake, never on purpose. “oops I just deleted all your hard drive by mistake, sorry”
- Never give a online (Frontier model) AI confidential information, ever!
- Make sure a local AI is tightly bound and restricted before it can work with your data. Use MCP’s to do this.
This is all easier said than done, and ironically you won’t be able to do it without the help of a Frontier model, so you have to take some risk.
Start with ‘Opencode’ on Linux with a Deepseek-V4 account. There is good protection provided by Opencode, and it has a ‘plan’ mode where the agent can never access your system in general, only the directory Opencode is opened in.
‘Build’ mode allows access in a controlled manner that requires your permission for every step.
Finally, if we can’t trust a worker, it’s not really much good to us is it ?
Cheers,
Terry