Hello Jim,

> Followed by, I predict, distorted advice based on product placement within the model itself. Instead of good advice interspersed with annoying ads, you get bad advice. There isn't any obvious way to "block" that, other than switching to a different model,

The remote AI thinks it's talking to a human. The local AI isn't a human. It also isn't even an adblocker, in the regular sense. A classic ad blocker might pick out suspicious domain names and HTML/JS structures associated with ads. Rather crude, but good enough. A local AI can systematically fact check every claim being made against external evidence, entertain counter-arguments, look for signs of bias, use the answer as a starting point for further research, and so on. It can study pictures for signs of bias, product placement, prices, etc.

Still, the issue of using pay-for AIs *at all* for *anything* is increasingly open to question, because ....


> presumably one that costs a lot more money. Including a locally run one.

Presumably?

Did you know that the capability/cost of local AI has improved by about a factor of say, 5x, in each of the last two years?

Please read that sentence once more and think about what that rate of improvement means, and how willing you'd be to bet real money against it.

2 years ago, for $3000-4000 you could get a system that was basically a toy, to see how 'real' AI works.

1 year ago, for $2500 you could get a system that was starting to be a bit useful in some ways.

This year, a tired-looking gaming PC built in 2024 for $2500 will get you to about 85-90% of frontier-level AI (or, frontier minus 3 months). Though, running more slowly.

Google have destroyed every moat they had, in AI and in search/selling ads.

Don't take my word for any of this - check for yourself, here's a suitable prompt you could try:

"Suppose someone spent $2000-2500 on a gaming PC in 2024-2025. Let's say 192GB DDR5, 9600X or 245k, motherboard, 2TB SSD, a PSU, cheap full ATX case 8-slots, and 3 x 4060/5060 tis at sale prices. A $200 ATX motherboard with several PCIE slots e.g. 16x/4x/4x. Maybe a cheap riser to let them space the cards out.

Q1. Does that price total / spec seem realistic for the pricing at that time?

Q2. Adding up the VRAM and RAM, and assuming currently best available local models using distillation / e.g. 3-4 bits/weight, how close could they get to 'state of the art' frontier AI? "

Response:

"The proposed configuration is generally realistic, but it sits right at the upper boundary ($2,400–$2,600) of the target price range."

"For $2,500, this rig delivers ~85–90% of current frontier AI capabilities for standard generation, coding, and dynamic reasoning tasks by utilizing 70B distilled models running fully in VRAM."

... which I know for a fact is the correct answer.

TRS

PS. Local AI doesn't actually need to be local; for 30 cents/hour you can rent a GPU system and run a 'local' AI fully under your own control in the cloud. And if your best investing ideas are worth more than 30c/hour to you, perhaps that would be the right thing to do, rather than feed them into OpenAI/Google's training sets for everyone else to make use of.