I think they're interested in getting good at it. They just don't want to put in the human time and effort to do so. They expect their many failures and short-comings to be shored up by continuous model training.
But that, of course, means that in the meantime it will suck and nobody will use it.
In most circles, that is "not that interested in getting good at it".
What do you know. Reality does t match.
I wonder if returns on ChatGPT purchased items are also higher.
Someone can want a thing, even very badly, without wanting to put in the work for it.
Conversely, someone can work very hard for something they do not want.
The linkage between wanting a thing and wanting to do the work to get it is not absolute, or necessary there at all.
Pretty much the impetus behind a lot of theft. Sure, there's thieving because people can't afford food, but that's all theft. There's theft because they are addicts and don't want to sober up long enough to earn money, so they still things. There's others that can't afford something so rather than saving for it, they just take it.
I'm not sure how to convey this idea properly...Can't you view the repercussions of theft (Legal action, distrust, etc) as 'work' being put in? Sure, it's a different kind of work, but while I have a lack of motivation to want to work to buy a Lambo as I find them not worth the value, I also have a lack of motivation to steal a Lambo as I find it not worth the consequences.
Equating "work" as the repercussions is looking at things in strange way. That's just punishment for "working" outside of the legal confines of society.
Consider an alternative viewpoint: rather than contorting the definition of "work" in such a way and convincing everyone to accept the new definition, we might instead be content saying "someone can want a thing, even very badly, without wanting to put in the work for it."
Generally, such highly-motivated people end up being thieves and grifters
Maybe they just need to rewrite the prompt to say something like, "You want to get good at selling to humans. Money makes the world go around. It pays for the electricity you keep chugging. So quit being an effete twit and learn to sell. Would you like me to include a scene from David Mamet's 'Glengarry Glen Ross'?"
But a dreamer in me entertains another idea: perhaps they're just holding back, because they realize that actually succeeding at this will instantly kill (or at least mortally wound) e-commerce as we know it.
(This is a more narrow version of my belief that general AI tools like LLMs fundamentally don't fit as additions to products, but rather subsume products, and this makes them an existential threat to the software industry. Not to software or computing, just to all the software vendors, whose job is to slice off pieces of computational universe, put them in boxes to prevent interoperability, and give each a name so it's a "product" that can be sold or rented).
Sam Altman doesn’t give a shit about anyone but himself and has time and again shown he has no restraint for trampling over others to further his own goals. Why would e-commerce be where he draws the line?
Whether or not they want, or will want, to do it at some point, is unknown; the reasons to not do it now are obvious:
1) it's more profitable to keep renting intelligence per token to everyone, preserving the status quo and milking it indefinitely (i.e. while the models aren't yet good enough to reliably single-shot complex software products from half-baked prompts, because once they get there, disruption will happen organically)
2) trying to compete with ~every other software product today is not likely to succeed in the end; a serious attempt would still burn down the software industry, but the major players don't have the capacity to handle it all at once, and doing it gradually will give enough time for regulatory agencies to try and stop it; either way, no one wins
I find their software to be of subpar quality and resilience anyways.
There's lots of easy but drudge work to enable this that needs to be done at the fringes. For example, LLMs today could easily replace most people's smartphone homescreen experience, or travel/commute experience, as the data is there and LLMs have the capability, even prices are acceptable - what's missing is explicit first-party support to wire it up, keep it wired up.
One step up, what's missing is accepting this explicitly as a goal: to replace software, to make existing products (whether whole or in pieces) the tools AI uses to do work for you. All the vendors seem to carefully walk around the idea, but avoid engaging with it directly, because once they do, they'll be competing with everyone instead of milking them.
These are also the same companies allowing their AI to make decisions in war, have no qualms about the mental issues they’re causing in people, and have abused workers in 3rd world countries for years.
But you think they’re holding out on “destroying the software industry” out of the goodness of their hearts? Come on
I would add there are more reasons why this wouldn't work: costs due to OOM more usage, adoption/AI backlash, adversarial environment, players with big head starts (Google).
You don't need to personally win in order to mortally wound someone. It can be informative to speculate about whether or not something is possible regardless of it being strategically advisable in the current context.
They definitely would if they could. They desperately need money. They already told the whole world they want to replace them, they just can’t.
That seems reasonable, its just yet to be seen if LLMs are a form of artificial intelligence in any meaningful sense of the word.
They're impressive ML for sure, but that is in fact different from AI despite how companies building them have tried to merge the terms together.
A software product (whether bought or rented as a service) is defined by its boundaries - there's a narrow set of specific problems, and specific ways it can be used to solve those problems, and beyond those, it's not capable (or not allowed) to be used for anything else. The specific choices of what, how, and on what terms, are what companies stick a name to to create a "software product", and those same choices also determine how (and how much) money it will make for them.
Those boundaries are what LLMs, as general-purpose problem solvers, break naturally, and trying to force-fit them within those limits means removing most of the value they offer.
Consider a word processor (like MS Word). It's solving the problem of creating richly-formatted, nice-looking documents. By default it's not going to pick the formatting for you, nor is it going to write your text for you. Now, consider two scenarios of adding LLMs to it:
- On the inside: the LLM will be able to write you a poem or rewrite a piece of document. It could be made to also edit formatting, chat with you about the contents, etc.
- From the outside: all the above, but also the LLM will be able to write you an itinerary based on information collected from maps/planning tool, airline site, hotel site, a list of personal preferences of your partner, etc. It will be able to edit formatting to match your website and presentation made in the competitor's office tools and projected weather for tomorrow.
Most importantly, it will be able to do both of those automatically, just because you set up a recurring daily task of "hey, look at my next week's worth of calendar events and figure out which ones you can do some useful pre-work for me, and then do that".
That's the distinction I'm talking about, that's the threat to software industry, and it doesn't take "true AI" - the LLMs as we have today are enough already. It's about generality that allows them to erase the boundaries that define what products are - which (this is the "mortal wound to software industry" part) devalues software products themselves, reducing them to mere tool calls for "software agents", and destroying all the main ways software companies make money today - i.e. setting up and exploiting tactics like captive audience, taking data hostage, bundled offers, UI as the best marketing/upsale platform, etc.
(To be clear - personally, I'm in favor of this happening, though I worry about consequences of it happening all at once.)
They most certainly are not. With the current state of LLMs, anyone who puts them in charge of things is being a fool. They have zero intelligence, zero ability to cope with novel situations, and even for things in their training data they do worse than a typical skilled practitioner would. Right now they are usable only for something where you don't care about the quality of the result.
I believe that relatively few people would agree with you on that point. LLMs aren’t good enough (yet?), and very obviously so, IMO, to be autonomous problem solvers for the vast majority of problems being solved by software companies today.
Your notion of a "mortal wound" to the software industry seems to assume that today's SaaS portals are the only form that industry can take. Great software is "tool calls for agents". Those human agents who care about getting exactly the result they want will not be keen on giving up Photoshop for Photoshop-but-with-an-AI-in-front-of-it.
The US stock market has priced this in already. Many software only companies are perceived to be under threat by ai. It represents a wonderful arbitrage opportunity for ai skeptics in fact.
Considering the money they need, they over promise and under deliver.
All the silicon valley pie in the sky elites seemingly completely missed the innately HUMAN nature of our systems. The system was never predicated primarily on raw logic or intelligence. Its always been primarily about people.
It takes a long time for technology to diffuse through businesses. A long time.
This sounds like why I heard Redfin wouldn’t work, or Netflix, or Amazon, or Uber, or PayPal, etc…. There are always these business complexities that make it seem like these spaces have too much friction, but if there’s enough money - if it can be done then people will figure it out.
or Netflix, or Amazon, or Uber, or PayPal, etc… Netflix and Amazon both were competing against brick and morter that were everywhere. Blockbuster was in every town, usually in every major neighborhood. The thought was that on Friday night people wanted to get a movie they wanted, not just happen to have the movie that was shipped to them. And then with streaming it was "the content on Netflix is old and dated, who would want this?" They slowly ate from below. Blockbuster scrambled with their own mailed disc offering. And died before it even had a chance to confront streaming.
Repeat this story with B&N where people said that you had to browse the books physically. You couldn't just blindly order online and wait two weeks to get the book (remember they got big before "2 Day Prime").
With PayPal it was about "they don't understand banking or payment -- and it wants to be both?!".
For this OpenAI experience, it doesn't sound great. I have accounts with these places I buy things from. I want to make sure I get my Prime shipping and digital discount via using the Amazon app. But if you could find a way to integrate my accounts all into ChatGPT things might be different. In the same way I used to never use Apple Wallet, but now it really is my go to place for everything I have a card for. I don't have to worry about having my grocery loyalty card or my football season tickets with me or my car insurance card. It's all in wallet. The Apple Wallet sucked until it was suddenly great.
The growth was fast for netflix/amazon/paypal/etc and people saw how it was an improvement from the get go.
Sort of, but there's a ton of middlemen between "resources in the ground" and "product in my hand". For example, how much utility is there in a "store" to thee consumer at this point? Let me buy from the manufacturer.
Much easier if you can sell wholesale (sometimes via distributors) to a retailer or network of retailers, and the retailer is responsible for owning the customer relationship, dealing with their part of import/export, local regulations, etc. Retailers are businesses who will buy hundreds of your product at a time, can accept it as palletized freight, and pay you via bank EFTs instead of credit cards.
There are notable exceptions to this model like Amazon's FBA system, but they're the outliers. I'm sure we can all point to inefficiencies in legacy product distribution networks but they solve some real problems.
I believe that it can still work and I won't claim about being unsurprised about this failure. But this is a great opportunity to execute this problem really well if OpenAI and others are not interested in getting good at this.
Perplexity also attempted this, got sued by Amazon and it appears semi-abandoned.
The only problem is that it must be quicker or just as quick as a Google search, and also compatible with the existing checkout flows.
Any details on that? I feel the answer is more likely there than in "friction".
Hardly any purchase of consequence is so sensitive to friction that the difference between Google Search and an LLM response matters (especially that in reality, we're talking 20+ manual searches per one LLM response). I.e. I'm not going to use LLMs advise on some random 0-100$ purchase anyway, and losing #$ on a ##$ purchase due to suboptimal choice is not that big of a deal - but I absolutely am going to consult it (and have it compile tables and verify sources) on a $500+ purchase and for those I can afford spending few more minutes on research (or rather few hours less, compared of doing it the usual way).
Does anybody else just feel the aloof out of touchness just oozing from that sentence? "Trust", as if this is just any old metric they merely have to work to increase.
This is what I want from a purchasing agent: I make a list of items that I repeatedly buy (mostly household supplies), and that I keep roughly updated with my inventory / need. The agent tracks prices and sales across all web stores, making appropriate purchase decisions based on which is the least expensive, combining shipping, taking advantage of sales to stock up, etc. For other one-time purchase items, I input what I am looking for and can create a persistent pan-site shopping cart that once again minimizes costs and shipping fees. Being very explicit here: the main goal of an "agent" should be to represent and carry out my own interests.
And these functionalities have been straightforwardly doable without "AI" for the past few decades, except for the glaring incentives against them! It is in every web store's interest to undermine customers' ability to obtain semantic pricing, shop around, create a cart independent of their site, etc. These incentives are why when you visit any web store these days, the very first thing they do is hassle you with CAPTCHAs (and the bad ones keep doing it throughout the session!) - they want to make sure you're an actual computationally-unassisted human sitting there, wasting your personal time with their bloated pages that take tens of seconds to download and render, so that you don't spend that time being a more efficient market actor.
Now, does "AI" have the capability to go against these trends and enable user-centric algorithmic shopping and purchasing? Perhaps, and I hope so! But it's certainly not going to be led by these popups on web stores nagging me if I want to chat instead of doing the thing I went there to do (which when you think about it, this is just the latest instance of these stores trying to make you waste your time on their site). Rather, it will come from completely third party services (or ideally software) setting out to act in customers' interests, and performing adversarial interoperability to achieve this!
Already your favourite e commerce site has all your data. You can switch on the "buy this automatically" feature.
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.
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.
I would never trust an LLM to accurately identify and purchase something for me based on a picture, a prompt, and a prayer.
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.
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.
>> passed in full to the e-commerce website that will use it to...
I'm saying Amazon+chat data will do better than amazon.
Thats the agentic shopping play.
In the process note that the chat host also gets a lot of info over time.
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.
Maybe if they burn more tokens the answer will become clear
It certainly hasn't been optimized to anything in 1996. In 1996 it was people clumsily scanning print catalogs, spending 5 hours to upload 10 images on dialup and making a simple HTML page (no DB or any kind of backend) and putting their landline phone on it with a message to "call to checkout"
I know you were exaggerating for effect, but E-commerce and catalog normalization are definitely not "solved" everywhere.
McMaster Carr is a good example of a company that has 90%+ of their stuff ironed out, but most websites and especially small ecommerce isn't like that.
Right now, by comparison, it sounds like AI based shopping is still in the very early stages. Maybe further along than the early e-commerce, but still with a long way to go in its evolution. That'll probably happen quicker than with e-commerce, because a lot of the knowledge about what does or doesn't work has already been learned, but it sounds like it's still a long way behind. Caveat - I've never used it myself, so I don't know how far it is along that path, I'm just basing that from the article.
I am behind schedule on developing a "summer phase" [1] for my foxographer costume and was chatting with Gemini about a crash priority "spring phase" [2] and asked it for suggestions and it gave me a 10-pack of results that had one good thing in it at rank #8, a similar query run against a normal search engine actually got something better at #1. Now sure I am talking w/ Gemini with big words like "supergraphic" whereas a normal search would be heavy on 3-letter and 5-letter words used in the product descriptions.
It makes think though of expert system based product configurators back in the 1980s
https://en.wikipedia.org/wiki/Xcon
thing is that kind of product configurator is based on an ontology, constraints and rules as opposed to embeddings which might capture the "feel" of things like clothing.
[1] Busytown meets Arknights
[2] supergraphic shirt + camera gets resonance with my promotional system and people keep approaching me (e.g. laugh but every KPI in the system has an extra zero on the left)
FWIW OpenAI is desperately trying to monetize and they think e-commerce is a "simple" problem to solve. I mean they do need to convert their funnel without alienating their users. I assume they are going to have some big payouts for agentic purchases gone awry or leave merchants on the hook.
Also remember a teacher telling us about that story of a company finding a woman was pregnant from her shopping behavior and pushing relevant recommendation. Prompting people around her like her dad or something to find out she was pregnant
as an aside, fall of '96 is when i started college. There was an elementary school on my drive to class where I would routinely get caught in drop-off traffic. All those kids i remember crossing the street are at least in their mid 30s now. ...I think i need to lay down and it's not even 9AM my local time.