If I can tell ChatGPT to order me what I need to replace the bottom element on my water heater, and here’s a picture of the ridiculously long model number, that saves me quite a bit of time.
I used that as an example because I did that last week, apart from me just going to the store to get the links that it brought up.
Of course anecdotal but 2 months ago I tried to do that for a missing bolt in a popular scooter (roughly 7 years old, bought 2nd hand so didn't know exact specs). I fed it various images that I took + internet images. It found a bunch of local shops but always found the wrong part (but very similar looking). I double-checked everything multiple times, via multiple contexts, even different pre-prompts from various sources, and asking plenty of questions. We chatted in thinking mode for what felt like ages, and according to paragraphs explaining why it HAD to be the right one, multiple times, with evidence that it sometimes gladly generated in image form (often completely garbage imagery e.g. with half-drawn bolts extruding from multiple surfaces). Eventually it found something very plausible which I ordered. It was the wrong part.
I had to get someone on the phone to help me find and order the right part (which was on their website, for many years according to waybackmachien).
I love LLMs but it's still totally hit/miss what you get. I'd rather not give it write-access to my bank account just yet.
Conversely, I did something similar, but I took a picture of the model number, copy/pasted the text from the image into a search engine, and that was that. Either way you're taking a picture and performing a manual step (query an LLM, query a search engine)
I would never trust an LLM to accurately identify and purchase something for me based on a picture, a prompt, and a prayer.
I mean, it was easy to confirm they were the right parts. Sure, I could have done it in a few minutes of Googling, but I was already asking it to help me troubleshoot (which it did) and instead it took me seconds.
Today, ads are based on user information you can reasonably collect from the users historical actions on your website, and then whatever search term they enter.
But soon, ads can be based on your current chat context + (derived interests of yours from your entire chat history across all chats. Shhhh.) passed in full to the e-commerce website that will use it to choose ads, generates creatives on the fly, all that crap, hyper-specific to you.
I'm so excited. Aren't you?
Now, as a side effect, searching through these can become better experience wise as well. They can use all that context and genuinely surface fewer, better results. But that's not the motivation of the e-commerce player anyways. If the ads work they'll be happy.
ChatGPT doesn’t know what the best-converting dog food in Scranton, PA is. Amazon does.
Anything that starts with chat history presumes the theoretical limit for ad effectiveness is higher than it is now, and chat is a better way of getting there than actual purchase history. I have a feeling it’s not.
Imagine a person who shops on Amazon for basically everything. So theoretically Amazon should know a ton about them, more than enough to put together a profile on that person.
To say OpenAI could do a better job of selling products is to say they can do better than Amazon already does if you scroll through their personalized product recommendations. There is some better feed out there, or some better way of presenting the feed that Amazon hasn’t thought of.
I don’t doubt it can be marginally better.
I do doubt whether it can be double or triple digits better that can justify trillion dollar valuations. And I do doubt whether a model trained on Internet text rather than user interactions can do better.
Ok fair point, and the general principle of “more data=higher accuracy” probably applies.
I think the issue is the tradeoff between accuracy and cost (on the seller’s side). If you get more accurate and convert more but your costs go up too much, you actually lose money.
Current systems are basically in a sweet spot of speed, cost, and accuracy.
And I will go back to my previous point that I believe there is simply a limit to how much people will buy, and it might already be saturated. I could be wrong though.
Yeah, this is true even today. Many advertisers are competing to be the thing that people buy. This is why many ad systems have naturally evolved into a bidding based system.
All of this just increases the base bid. If you don't pay the ad platform whatever they ask, you are simply invisible to whoever uses chatgpt to search for things they want to buy.
This is true for all improvements in targeting ads. Because the platform is in a position where they can rent-seek.
The marketing costs of running any consumer company simply "inflates". For example with LLMs, people can control the number of results they see, which is a stark difference from previous ad serving systems. Meaning the bids get even more fierce since if you're not in the top three, you might as well not exist.
The question is whether or not increasingly better targeted ads provide increasing value to the consumer. If it does, then great, the ever "inflating" ad economy is justified. But my opinion is that they don't. This is why I think major ad agencies have to bolt it on to an existing useful service. Amazon, instagram, google's services, etc,. In that sense, it is unaccounted inflation. A "correct" economy would involve paying directly for the service they are using based on what they value that service at. Instead of a random disconnected marketing economy inflating without account to cover their costs. If someone is actually keeping track of this.. and increase in cost or reduction in quality of every product we buy (due to the company passing on marketing costs to us...) actually is <= the cost we would otherwise pay instagram, google, etc,. then great. I think it would be wildly off the mark though!
Ssshhh there’s an “VP of AI Transformation” getting paid $600,000 to do this plus the budget. They need their “AI transformation journey” to show to the board.
Already your favourite e commerce site has all your data. You can switch on the "buy this automatically" feature.