In the past week, Gemini, which had successfully built PicoDB and PicoRR V1 over the course of about 8 weeks, has become unworkable for me.

First it spent several days botching a compound IF statement which resolved into 4 conditions. I noticed at one point that it had dropped me from a PRO to a FAST model--which meant it would guess rather than look at the actual code. Then, for the last 3 days I have gotten hours-worth of "Error: can you re-prompt?" or "I'm having a problem, can I help you with something else?". As a result, issues I had hoped to resolve a week ago remain unresolved.

I had seen some of this coming (though I hadn't expected it to be so bad). I bought a MINISFORUM UM870 PC: 32GB DDR5, 256GB SSD, Ubuntu 24.04.4 LTS. Yesterday I set up Claude Code Pro on the device, loaded my files, and asked it to do its initialization. Quite impressive--128 lines of assessment, spot on. Then I asked it to analyze the PicoDB library.bas file. Again impressive, with 8 issue areas located. 6 of them were easily fixed; the others are deeper, and I'll have to ask it to resolve them.

We'll see. It was a good 7-week run with Gemini. It showed me what the capabilities were. Claude Code goes to the Anthropic servers for intelligence, so it's also going to be susceptible to this tsunami of token usage that is coming.

I'm hoping, if need be, I'll be able to go entirely local with my setup. But first I want to see what Claude Code can do at the present time.

~
Edited 2026-03-23 00:34 by lizby

More on AI coding on the MINISFORUM UM870 PC. I did a brief test of ollama & qwen2.5-coder running entirely locally--a few hours. It seemed somewhat primitive, but probably workable in the case of an internet outage or a funding shortfall.

So I went back to Claude Code API (separate from the monthly subscription). With this, your files are stored locally, but Claude understands the structure. All the AI is done on Anthropic's servers, not locally. This was around 5,000 lines of code in PicoDB and PicoRR.

I was very impressed by the way it worked through the PicoDB problem Gemini had cycled around for several days. Fixing that bit of code revealed some other dependent problems which I systematically worked through with Claude. Then I set it on some difficulties in PicoRR, and quickly got those sorted. I liked the way it presented its suggested changes as a diff which you accepted or rejected. Several times I accepted code I knew to have an error, then, before testing, suggested a change which it incorporated.

This was very smooth, but expensive (by a retired person's standards)--I paid $25 over the course of 5 days. I thought it was well worth it, both as a learning experience and for fixing problems I had been stuck on with Gemini.

I can't in the long run spend $5 a day for hobby coding--but if it was for a job, it would be a trivial expense.

So now I'm back to Claude Code in the browser with the $20 monthly subscription. Plus, for some reason, it has given me $50 in extra credits to be used if subscription limits are hit. That will probably expire before I use much (or any) of it. The "Page not responding" issue I had some weeks ago seems to be resolved.

Gemini, Claude, Chat are all trying to respond to the same basic dilemma: they want to throw all the compute they can at the effort to reach AGI before anyone else does (the holy grail), but they also need to address exploding demand from customers who just want current AI to solve their immediate problems.

Here is a take on one aspect of the dilemma from a major doubter of the finances of all AI-related companies ( "A massive bubble" ):

  Quote  

"As a result, there really isn’t any way that these services make sense at a monthly rate, and every single AI company loses incredible amounts of money, all while failing to make that much revenue in the first place."

~
Edited 2026-04-02 00:36 by lizby