You can imagine a world where advertising is designed to trick the concierge of course, but at a certain point there's just objective data, e.g. "don't pay more for [x] because it actually comes from the same factor." If something like that could be achieved, what does it do for the ad business generally, and how do companies react?

@ZachWeinersmith Advertisers would probably pivot to bribing bot operators to give preference to their product, lying as needed.
Basically, it just means that instead of Ad Exchange operators grabbing rent on purchases, it'll be bot operators.

@ZachWeinersmith Adversarial Noise Injection to the concierge AI. https://arxiv.org/abs/1412.6572

It still takes multiple orders of magnitude less energy to create spam than to prevent it in this paradigm. The general large scale system thermodynamics are the same. Perhaps even worse due to Turing Siren effect (linguistic competence bias)

@ZachWeinersmith Flash sales would probably beat this system. Even AIs need time to update their search results, and flash sales would compel more shoppers to buy out of FOMO without doing more research first.

@ZachWeinersmith

Not AI, but this was one of the goals of General Magic and its Telescript language. You’d write an agent in Telescript and it would go run itself on other systems to find what you were looking for, and potentially buy it or book it or whatever.

It was ahead of its time.

@ZachWeinersmith
In principle we don’t need ai to do this now. But there is very little incentive for sellers to make the prices easily accessible, and there is no reason to believe that gets better when the data queries are coming from a few central AI inference farms.

@ZachWeinersmith brands will fight to have their ai plugin be the thing that decides what to buy

@ZachWeinersmith huh. You found a use case where it's not going to be more unreliable than the existing method

@ZachWeinersmith would be interesting to know what the computational cost would be to have an updated layer somewhat close in time? 🤔

The basic model could probably be trained once, but to be up to date one would harvest the relevant sources regularly and embed the changes into a new layer or so.
Plain index diffs like with classical search engines are probably much cheaper - tbh I am not sure, if attracting more users/advertisers would be enough?

@ZachWeinersmith
Assuming the AI works for you looks like a logical flaw.

@ZachWeinersmith unless you have an open source AI not funded by the same people who sell you ads then it won't work correctly or do exactly what Google does and recommend paid stuff first

@ZachWeinersmith I don't think AI fundamentally changes the reasons why we don't have that now. We do have semi-trusted reviewers, but they're not the go to for most people. Why, for instance, doesn't everyone just use e.g NYT Wirecutter's reviews to guide their purchases?

If AI gains even a small foothold in that space the "facts" they receive will be manipulated, competitors who take money from big sellers to warp the advice will get enough funding to seem ubiquitous to the avg. consumer. Etc

@ZachWeinersmith This is why my answer is still going to be the well-stocked spice shop down the street, a 10-min walk away.

@ZachWeinersmith

Which do you think most often happens first....

A desire to buy something, or
An advertisement of that something.

@ZachWeinersmith I mean... I already have a spreadsheet that does that, no AI required

@ZachWeinersmith I think brands double down on branding (i.e. target ads to persuade you to ask your AI concierge for them by name) and use other AIs to poison the well of your concierge’s training (e.g., flooding the internet with favourable reviews and comparison lists).

@ZachWeinersmith you would just have ai companies start selling spices for 4.28 when normally you could get them for 2.00