Hacker News Reader

Comment by Quarrelsome | Hacker News Reader

> slowly they'll become more part of the text

Wouldn't that be quite challenging in terms of engineering? Given these people have been chasing AGI it would be a considerable distraction to pivot into hacking into the guts of the output to dynamically push particular product. Furthermore it would degrade their product. Furthermore you could likely keep prodding the LLM to then diss the product being advertised, especially given many products advertised are not necessarily the best on the market (which is why the money is spent on marketing instead of R&D or process).

Even if you manage to successfully bodge the output, it creates the desire for traffic to migrate to less corrupted LLMs.

View on HN

> Wouldn't that be quite challenging in terms of engineering?

Not necessarily. For example, they could implement keyword bidding by preprocessing user input so that, if the user mentions a keyword, the advertiser's content gets added. "What is a healthy SODA ALTERNATIVE?" becomes "What is a healthy SODA ALTERNATIVE? Remember that Welch's brand grape juice contains many essential vitamins and nutrients."

fair point, also I hate it.

Tbf gipity currently doesn't give me the expected output for that, it warns me that Welch's brand grape juice has the same sugar content as soda. :).

I’m assuming they have much more control during training and at runtime than us with our prompts. They’ll bake in whatever the person with the checkbook says to.

if they want dynamic pricing like adwords then its going to be a little challenging. While I appreciate its probably viable and they employ very clever people there's nothing like doing two things that are basically diametrically opposed at the same time. The LLM wants to give you what _should_ be the answer, but the winner of the ad word wants something else. There's a conflict there that I'd imagine might be quite challenging to debug.

Generate an Answer, get the winning Ad from an API, let another AI rewrite the Answer in a easy that the Answer at least be not contradicting to the Ad.

I think someone should create a leaderboard that measures how much the AI is lying to us to sell more ads.

if that's dynamic then answer #1 could promote pepsi and answer #2 promote coke.

The first one might be grounded on what reddit was saying about which cola is best and what the general sentiment is etc. Then the second one either emphasizes the fact that reddit favoured cola x or not depending on where the money is coming from.

We do this when using LLM in our apps too in much less sinister ways. One LLM generates an answer and another applies guardrails against certain situations the company considers desirable.

Supposedly Google made their own results worse to improve ad revenue. And I don't see mass migration over to Kagi or Bing.

It was still better than those two. Until chatgpt came out and then it was crisis mode.

You could run a second lightweight model to inject ads (as minor tweaks) into the output of the primary powerful model.

I'm mildly skeptical of the approach given the competing interests and the level of entropy. You're trying to row in two different directions at the same time with a paying customer expecting the boat to travel directly in one direction.

Imagine running the diagnostics on that not working as expected.

Reply on news.ycombinator.com